null hypothesis in marketing research

• Business Essentials
• Leadership & Management
• Credential of Leadership, Impact, and Management in Business (CLIMB)
• Entrepreneurship & Innovation
• Digital Transformation
• Finance & Accounting
• Business in Society
• For Organizations
• Support Portal
• Media Coverage
• Founding Donors
• Leadership Team

• Harvard Business School →
• HBS Online →
• Business Insights →

Business Insights

Harvard Business School Online's Business Insights Blog provides the career insights you need to achieve your goals and gain confidence in your business skills.

• Career Development
• Communication
• Decision-Making
• Earning Your MBA
• Negotiation
• News & Events
• Productivity
• Staff Spotlight
• Student Profiles
• Work-Life Balance
• AI Essentials for Business
• Alternative Investments
• Business Analytics
• Business Strategy
• Business and Climate Change
• Design Thinking and Innovation
• Digital Marketing Strategy
• Disruptive Strategy
• Economics for Managers
• Entrepreneurship Essentials
• Financial Accounting
• Global Business
• Launching Tech Ventures
• Leadership Principles
• Leadership, Ethics, and Corporate Accountability
• Leading Change and Organizational Renewal
• Leading with Finance
• Management Essentials
• Negotiation Mastery
• Organizational Leadership
• Power and Influence for Positive Impact
• Strategy Execution
• Sustainable Business Strategy
• Sustainable Investing
• Winning with Digital Platforms

A Beginner’s Guide to Hypothesis Testing in Business

• 30 Mar 2021

Becoming a more data-driven decision-maker can bring several benefits to your organization, enabling you to identify new opportunities to pursue and threats to abate. Rather than allowing subjective thinking to guide your business strategy, backing your decisions with data can empower your company to become more innovative and, ultimately, profitable.

If you’re new to data-driven decision-making, you might be wondering how data translates into business strategy. The answer lies in generating a hypothesis and verifying or rejecting it based on what various forms of data tell you.

Below is a look at hypothesis testing and the role it plays in helping businesses become more data-driven.

Access your free e-book today.

What Is Hypothesis Testing?

To understand what hypothesis testing is, it’s important first to understand what a hypothesis is.

A hypothesis or hypothesis statement seeks to explain why something has happened, or what might happen, under certain conditions. It can also be used to understand how different variables relate to each other. Hypotheses are often written as if-then statements; for example, “If this happens, then this will happen.”

Hypothesis testing , then, is a statistical means of testing an assumption stated in a hypothesis. While the specific methodology leveraged depends on the nature of the hypothesis and data available, hypothesis testing typically uses sample data to extrapolate insights about a larger population.

Hypothesis Testing in Business

When it comes to data-driven decision-making, there’s a certain amount of risk that can mislead a professional. This could be due to flawed thinking or observations, incomplete or inaccurate data , or the presence of unknown variables. The danger in this is that, if major strategic decisions are made based on flawed insights, it can lead to wasted resources, missed opportunities, and catastrophic outcomes.

The real value of hypothesis testing in business is that it allows professionals to test their theories and assumptions before putting them into action. This essentially allows an organization to verify its analysis is correct before committing resources to implement a broader strategy.

As one example, consider a company that wishes to launch a new marketing campaign to revitalize sales during a slow period. Doing so could be an incredibly expensive endeavor, depending on the campaign’s size and complexity. The company, therefore, may wish to test the campaign on a smaller scale to understand how it will perform.

In this example, the hypothesis that’s being tested would fall along the lines of: “If the company launches a new marketing campaign, then it will translate into an increase in sales.” It may even be possible to quantify how much of a lift in sales the company expects to see from the effort. Pending the results of the pilot campaign, the business would then know whether it makes sense to roll it out more broadly.

Related: 9 Fundamental Data Science Skills for Business Professionals

Key Considerations for Hypothesis Testing

1. alternative hypothesis and null hypothesis.

In hypothesis testing, the hypothesis that’s being tested is known as the alternative hypothesis . Often, it’s expressed as a correlation or statistical relationship between variables. The null hypothesis , on the other hand, is a statement that’s meant to show there’s no statistical relationship between the variables being tested. It’s typically the exact opposite of whatever is stated in the alternative hypothesis.

For example, consider a company’s leadership team that historically and reliably sees $12 million in monthly revenue. They want to understand if reducing the price of their services will attract more customers and, in turn, increase revenue.

In this case, the alternative hypothesis may take the form of a statement such as: “If we reduce the price of our flagship service by five percent, then we’ll see an increase in sales and realize revenues greater than $12 million in the next month.”

The null hypothesis, on the other hand, would indicate that revenues wouldn’t increase from the base of $12 million, or might even decrease.

Check out the video below about the difference between an alternative and a null hypothesis, and subscribe to our YouTube channel for more explainer content.

2. Significance Level and P-Value

Statistically speaking, if you were to run the same scenario 100 times, you’d likely receive somewhat different results each time. If you were to plot these results in a distribution plot, you’d see the most likely outcome is at the tallest point in the graph, with less likely outcomes falling to the right and left of that point.

With this in mind, imagine you’ve completed your hypothesis test and have your results, which indicate there may be a correlation between the variables you were testing. To understand your results' significance, you’ll need to identify a p-value for the test, which helps note how confident you are in the test results.

In statistics, the p-value depicts the probability that, assuming the null hypothesis is correct, you might still observe results that are at least as extreme as the results of your hypothesis test. The smaller the p-value, the more likely the alternative hypothesis is correct, and the greater the significance of your results.

3. One-Sided vs. Two-Sided Testing

When it’s time to test your hypothesis, it’s important to leverage the correct testing method. The two most common hypothesis testing methods are one-sided and two-sided tests , or one-tailed and two-tailed tests, respectively.

Typically, you’d leverage a one-sided test when you have a strong conviction about the direction of change you expect to see due to your hypothesis test. You’d leverage a two-sided test when you’re less confident in the direction of change.

4. Sampling

To perform hypothesis testing in the first place, you need to collect a sample of data to be analyzed. Depending on the question you’re seeking to answer or investigate, you might collect samples through surveys, observational studies, or experiments.

A survey involves asking a series of questions to a random population sample and recording self-reported responses.

Observational studies involve a researcher observing a sample population and collecting data as it occurs naturally, without intervention.

Finally, an experiment involves dividing a sample into multiple groups, one of which acts as the control group. For each non-control group, the variable being studied is manipulated to determine how the data collected differs from that of the control group.

Learn How to Perform Hypothesis Testing

Hypothesis testing is a complex process involving different moving pieces that can allow an organization to effectively leverage its data and inform strategic decisions.

If you’re interested in better understanding hypothesis testing and the role it can play within your organization, one option is to complete a course that focuses on the process. Doing so can lay the statistical and analytical foundation you need to succeed.

Do you want to learn more about hypothesis testing? Explore Business Analytics —one of our online business essentials courses —and download our Beginner’s Guide to Data & Analytics .

About the Author

Marketing Analytics Lab

Hypothesis Testing in Marketing Research

Introduction to Hypothesis Testing in Marketing Research

Hypothesis testing is a critical component of marketing research that allows marketers to draw conclusions about the effectiveness of their strategies. In essence, hypothesis testing involves making an educated guess about a population parameter and then using data to determine if the hypothesis is supported or rejected. In the context of marketing, hypotheses can be formulated about consumer behavior, product preferences, advertising effectiveness, and many other aspects of the marketing mix. By conducting hypothesis tests, marketers can make informed decisions based on empirical evidence rather than intuition or guesswork.

A hypothesis test in marketing research typically follows a structured process that involves defining a null hypothesis (H0) and an alternative hypothesis (HA), collecting and analyzing data, determining the appropriate statistical test to use, setting a significance level, and interpreting the results to either accept or reject the null hypothesis. The null hypothesis represents the status quo or the assumption that there is no significant difference or relationship between variables, while the alternative hypothesis suggests that there is a significant effect or relationship. By rigorously testing hypotheses, marketers can evaluate the impact of their marketing strategies and make data-driven decisions to optimize their campaigns and initiatives.

The results of hypothesis testing in marketing research provide valuable insights that can inform strategic decision-making and help marketers achieve their business objectives. Whether testing the effectiveness of a new product launch, evaluating the impact of a promotional campaign, or analyzing consumer preferences, hypothesis testing enables marketers to quantify the impact of their actions and make evidence-based recommendations. By employing statistical techniques and hypothesis testing in marketing research, organizations can gain a deeper understanding of consumer behavior, identify market trends, and refine their marketing strategies to drive business growth and success.

Key Steps and Considerations for Hypothesis Testing in Marketing Analysis

When conducting hypothesis testing in marketing research, there are several key steps and considerations that marketers should keep in mind to ensure the validity and reliability of their findings. Firstly, it is essential to clearly define the research question and formulate testable hypotheses that are specific, measurable, and relevant to the marketing objectives. By articulating clear hypotheses, marketers can establish a framework for data collection and analysis that aligns with the research objectives.

Once the hypotheses have been formulated, the next step is to determine the appropriate research design and methodology for data collection. Depending on the nature of the research question and the variables involved, marketers may choose to conduct experiments, surveys, observational studies, or other research methods to gather data. It is crucial to ensure that the data collected is representative of the target population and is collected in a systematic and unbiased manner to generate reliable results.

After collecting the data, marketers can perform statistical analysis to test the hypotheses using techniques such as t-tests, ANOVA, regression analysis, or chi-square tests, among others. It is important to select the appropriate statistical test based on the type of data and the research question being investigated. Additionally, setting a significance level (alpha) is crucial for determining the threshold for accepting or rejecting the null hypothesis. By interpreting the results in the context of the significance level, marketers can make informed decisions about the implications of the findings and their impact on marketing strategies.

Leave a Reply Cancel reply

Your email address will not be published. Required fields are marked *

Save my name, email, and website in this browser for the next time I comment.

Marketing Research Design & Analysis 2019

5 hypothesis testing.

This chapter is primarily based on Field, A., Miles J., & Field, Z. (2012): Discovering Statistics Using R. Sage Publications, chapters 5, 9, 15, 18 .

You can download the corresponding R-Code here

5.1 Introduction

We test hypotheses because we are confined to taking samples – we rarely work with the entire population. In the previous chapter, we introduced the standard error (i.e., the standard deviation of a large number of hypothetical samples) as an estimate of how well a particular sample represents the population. We also saw how we can construct confidence intervals around the sample mean \(\bar x\) by computing \(SE_{\bar x}\) as an estimate of \(\sigma_{\bar x}\) using \(s\) as an estimate of \(\sigma\) and calculating the 95% CI as \(\bar x \pm 1.96 * SE_{\bar x}\) . Although we do not know the true population mean ( \(\mu\) ), we might have an hypothesis about it and this would tell us how the corresponding sampling distribution looks like. Based on the sampling distribution of the hypothesized population mean, we could then determine the probability of a given sample assuming that the hypothesis is true .

Let us again begin by assuming we know the entire population using the example of music listening times among students from the previous example. As a reminder, the following plot shows the distribution of music listening times in the population of WU students.

In this example, the population mean ( \(\mu\) ) is equal to 19.98, and the population standard deviation \(\sigma\) is equal to 14.15.

5.1.1 The null hypothesis

Let us assume that we were planning to take a random sample of 50 students from this population and our hypothesis was that the mean listening time is equal to some specific value \(\mu_0\) , say \(10\) . This would be our null hypothesis . The null hypothesis refers to the statement that is being tested and is usually a statement of the status quo, one of no difference or no effect. In our example, the null hypothesis would state that there is no difference between the true population mean \(\mu\) and the hypothesized value \(\mu_0\) (in our example \(10\) ), which can be expressed as follows:

\[ H_0: \mu = \mu_0 \] When conducting research, we are usually interested in providing evidence against the null hypothesis. If we then observe sufficient evidence against it and our estimate is said to be significant. If the null hypothesis is rejected, this is taken as support for the alternative hypothesis . The alternative hypothesis assumes that some difference exists, which can be expressed as follows:

\[ H_1: \mu \neq \mu_0 \] Accepting the alternative hypothesis in turn will often lead to changes in opinions or actions. Note that while the null hypothesis may be rejected, it can never be accepted based on a single test. If we fail to reject the null hypothesis, it means that we simply haven’t collected enough evidence against the null hypothesis to disprove it. In classical hypothesis testing, there is no way to determine whether the null hypothesis is true. Hypothesis testing provides a means to quantify to what extent the data from our sample is in line with the null hypothesis.

In order to quantify the concept of “sufficient evidence” we look at the theoretical distribution of the sample means given our null hypothesis and the sample standard error. Using the available information we can infer the sampling distribution for our null hypothesis. Recall that the standard deviation of the sampling distribution (i.e., the standard error of the mean) is given by \(\sigma_{\bar x}={\sigma \over \sqrt{n}}\) , and thus can be computed as follows:

Since we know from the central limit theorem that the sampling distribution is normal for large enough samples, we can now visualize the expected sampling distribution if our null hypothesis was in fact true (i.e., if the was no difference between the true population mean and the hypothesized mean of 10).

We also know that 95% of the probability is within 1.96 standard deviations from the mean. Values higher than that are rather unlikely, if our hypothesis about the population mean was indeed true. This is shown by the shaded area, also known as the “rejection region”. To test our hypothesis that the population mean is equal to \(10\) , let us take a random sample from the population.

The mean listening time in the sample (black line) \(\bar x\) is 18.59. We can already see from the graphic above that such a value is rather unlikely under the hypothesis that the population mean is \(10\) . Intuitively, such a result would therefore provide evidence against our null hypothesis. But how could we quantify specifically how unlikely it is to obtain such a value and decide whether or not to reject the null hypothesis? Significance tests can be used to provide answers to these questions.

5.1.2 Statistical inference on a sample

5.1.2.1 test statistic, 5.1.2.1.1 z-scores.

Let’s go back to the sampling distribution above. We know that 95% of all values will fall within 1.96 standard deviations from the mean. So if we could express the distance between our sample mean and the null hypothesis in terms of standard deviations, we could make statements about the probability of getting a sample mean of the observed magnitude (or more extreme values). Essentially, we would like to know how many standard deviations ( \(\sigma_{\bar x}\) ) our sample mean ( \(\bar x\) ) is away from the population mean if the null hypothesis was true ( \(\mu_0\) ). This can be formally expressed as follows:

\[ \bar x- \mu_0 = z \sigma_{\bar x} \]

In this equation, z will tell us how many standard deviations the sample mean \(\bar x\) is away from the null hypothesis \(\mu_0\) . Solving for z gives us:

\[ z = {\bar x- \mu_0 \over \sigma_{\bar x}}={\bar x- \mu_0 \over \sigma / \sqrt{n}} \]

This standardized value (or “z-score”) is also referred to as a test statistic . Let’s compute the test statistic for our example above:

To make a decision on whether the difference can be deemed statistically significant, we now need to compare this calculated test statistic to a meaningful threshold. In order to do so, we need to decide on a significance level \(\alpha\) , which expresses the probability of finding an effect that does not actually exist (i.e., Type I Error). You can find a detailed discussion of this point at the end of this chapter. For now, we will adopt the widely accepted significance level of 5% and set \(\alpha\) to 0.05. The critical value for the normal distribution and \(\alpha\) = 0.05 can be computed using the qnorm() function as follows:

We use 0.975 and not 0.95 since we are running a two-sided test and need to account for the rejection region at the other end of the distribution. Recall that for the normal distribution, 95% of the total probability falls within 1.96 standard deviations of the mean, so that higher (absolute) values provide evidence against the null hypothesis. Generally, we speak of a statistically significant effect if the (absolute) calculated test statistic is larger than the (absolute) critical value. We can easily check if this is the case in our example:

Since the absolute value of the calculated test statistic is larger than the critical value, we would reject \(H_0\) and conclude that the true population mean \(\mu\) is significantly different from the hypothesized value \(\mu_0 = 10\) .

5.1.2.1.2 t-statistic

You may have noticed that the formula for the z-score above assumes that we know the true population standard deviation ( \(\sigma\) ) when computing the standard deviation of the sampling distribution ( \(\sigma_{\bar x}\) ) in the denominator. However, the population standard deviation is usually not known in the real world and therefore represents another unknown population parameter which we have to estimate from the sample. We saw in the previous chapter that we usually use \(s\) as an estimate of \(\sigma\) and \(SE_{\bar x}\) as and estimate of \(\sigma_{\bar x}\) . Intuitively, we should be more conservative regarding the critical value that we used above to assess whether we have a significant effect to reflect this uncertainty about the true population standard deviation. That is, the threshold for a “significant” effect should be higher to safeguard against falsely claiming a significant effect when there is none. If we replace \(\sigma_{\bar x}\) by it’s estimate \(SE_{\bar x}\) in the formula for the z-score, we get a new test statistic (i.e, the t-statistic ) with its own distribution (the t-distribution ):

\[ t = {\bar x- \mu_0 \over SE_{\bar x}}={\bar x- \mu_0 \over s / \sqrt{n}} \]

Here, \(\bar X\) denotes the sample mean and \(s\) the sample standard deviation. The t-distribution has more probability in its “tails”, i.e. farther away from the mean. This reflects the higher uncertainty introduced by replacing the population standard deviation by its sample estimate. Intuitively, this is particularly relevant for small samples, since the uncertainty about the true population parameters decreases with increasing sample size. This is reflected by the fact that the exact shape of the t-distribution depends on the degrees of freedom , which is the sample size minus one (i.e., \(n-1\) ). To see this, the following graph shows the t-distribution with different degrees of freedom for a two-tailed test and \(\alpha = 0.05\) . The grey curve shows the normal distribution.

Notice that as \(n\) gets larger, the t-distribution gets closer and closer to the normal distribution, reflecting the fact that the uncertainty introduced by \(s\) is reduced. To summarize, we now have an estimate for the standard deviation of the distribution of the sample mean (i.e., \(SE_{\bar x}\) ) and an appropriate distribution that takes into account the necessary uncertainty (i.e., the t-distribution). Let us now compute the t-statistic according to the formula above:

Notice that the value of the t-statistic is higher compared to the z-score (4.29). This can be attributed to the fact that by using the \(s\) as and estimate of \(\sigma\) , we underestimate the true population standard deviation. Hence, the critical value would need to be larger to adjust for this. This is what the t-distribution does. Let us compute the critical value from the t-distribution with n - 1 degrees of freedom.

Again, we use 0.975 and not 0.95 since we are running a two-sided test and need to account for the rejection region at the other end of the distribution. Notice that the new critical value based on the t-distributionis larger, to reflect the uncertainty when estimating \(\sigma\) from \(s\) . Now we can see that the calculated test statistic is still larger than the critical value.

The following graphics shows that the calculated test statistic (red line) falls into the rejection region so that in our example, we would reject the null hypothesis that the true population mean is equal to \(10\) .

Decision: Reject \(H_0\) , given that the calculated test statistic is larger than critical value.

Something to keep in mind here is the fact the test statistic is a function of the sample size. This, as \(n\) gets large, the test statistic gets larger as well and we are more likely to find a significant effect. This reflects the decrease in uncertainty about the true population mean as our sample size increases.

5.1.2.2 P-values

In the previous section, we computed the test statistic, which tells us how close our sample is to the null hypothesis. The p-value corresponds to the probability that the test statistic would take a value as extreme or more extreme than the one that we actually observed, assuming that the null hypothesis is true . It is important to note that this is a conditional probability : we compute the probability of observing a sample mean (or a more extreme value) conditional on the assumption that the null hypothesis is true. The pnorm() function can be used to compute this probability. It is the cumulative probability distribution function of the `normal distribution. Cumulative probability means that the function returns the probability that the test statistic will take a value less than or equal to the calculated test statistic given the degrees of freedom. However, we are interested in obtaining the probability of observing a test statistic larger than or equal to the calculated test statistic under the null hypothesis (i.e., the p-value). Thus, we need to subtract the cumulative probability from 1. In addition, since we are running a two-sided test, we need to multiply the probability by 2 to account for the rejection region at the other side of the distribution.

This value corresponds to the probability of observing a mean equal to or larger than the one we obtained from our sample, if the null hypothesis was true. As you can see, this probability is very low. A small p-value signals that it is unlikely to observe the calculated test statistic under the null hypothesis. To decide whether or not to reject the null hypothesis, we would now compare this value to the level of significance ( \(\alpha\) ) that we chose for our test. For this example, we adopt the widely accepted significance level of 5%, so any test results with a p-value < 0.05 would be deemed statistically significant. Note that the p-value is directly related to the value of the test statistic. The relationship is such that the higher (lower) the value of the test statistic, the lower (higher) the p-value.

Decision: Reject \(H_0\) , given that the p-value is smaller than 0.05.

5.1.2.3 Confidence interval

For a given statistic calculated for a sample of observations (e.g., listening times), a 95% confidence interval can be constructed such that in 95% of samples, the true value of the true population mean will fall within its limits. If the parameter value specified in the null hypothesis (here \(10\) ) does not lie within the bounds, we reject \(H_0\) . Building on what we learned about confidence intervals in the previous chapter, the 95% confidence interval based on the t-distribution can be computed as follows:

\[ CI_{lower} = {\bar x} - t_{1-{\alpha \over 2}} * SE_{\bar x} \\ CI_{upper} = {\bar x} + t_{1-{\alpha \over 2}} * SE_{\bar x} \]

It is easy to compute this interval manually:

The interpretation of this interval is as follows: if we would (hypothetically) take 100 samples and calculated the mean and confidence interval for each of them, then the true population mean would be included in 95% of these intervals. The CI is informative when reporting the result of your test, since it provides an estimate of the uncertainty associated with the test result. From the test statistic or the p-value alone, it is not easy to judge in which range the true population parameter is located. The CI provides an estimate of this range.

Decision: Reject \(H_0\) , given that the parameter value from the null hypothesis ( \(10\) ) is not included in the interval.

To summarize, you can see that we arrive at the same conclusion (i.e., reject \(H_0\) ), irrespective if we use the test statistic, the p-value, or the confidence interval. However, keep in mind that rejecting the null hypothesis does not prove the alternative hypothesis (we can merely provide support for it). Rather, think of the p-value as the chance of obtaining the data we’ve collected assuming that the null hypothesis is true. You should report the confidence interval to provide an estimate of the uncertainty associated with your test results.

5.1.3 Choosing the right test

The test statistic, as we have seen, measures how close the sample is to the null hypothesis and often follows a well-known distribution (e.g., normal, t, or chi-square). To select the correct test, various factors need to be taken into consideration. Some examples are:

• On what scale are your variables measured (categorical vs. continuous)?
• Do you want to test for relationships or differences?
• If you test for differences, how many groups would you like to test?
• For parametric tests, are the assumptions fulfilled?

The previous discussion used a one sample t-test as an example, which requires that variable is measured on an interval or ratio scale. If you are confronted with other settings, the following flow chart provides a rough guideline on selecting the correct test:

Flowchart for selecting an appropriate test (source: McElreath, R. (2016): Statistical Rethinking, p. 2)

For a detailed overview over the different type of tests, please also refer to this overview by the UCLA.

5.1.3.1 Parametric vs. non-parametric tests

A basic distinction can be made between parametric and non-parametric tests. Parametric tests require that variables are measured on an interval or ratio scale and that the sampling distribution follows a known distribution. Non-Parametric tests on the other hand do not require the sampling distribution to be normally distributed (a.k.a. “assumption free tests”). These tests may be used when the variable of interest is measured on an ordinal scale or when the parametric assumptions do not hold. They often rely on ranking the data instead of analyzing the actual scores. By ranking the data, information on the magnitude of differences is lost. Thus, parametric tests are more powerful if the sampling distribution is normally distributed. In this chapter, we will first focus on parametric tests and cover non-parametric tests later.

5.1.3.2 One-tailed vs. two-tailed test

For some tests you may choose between a one-tailed test versus a two-tailed test . The choice depends on the hypothesis you specified, i.e., whether you specified a directional or a non-directional hypotheses. In the example above, we used a non-directional hypothesis . That is, we stated that the mean is different from the comparison value \(\mu_0\) , but we did not state the direction of the effect. A directional hypothesis states the direction of the effect. For example, we might test whether the population mean is smaller than a comparison value:

\[ H_0: \mu \ge \mu_0 \\ H_1: \mu < \mu_0 \]

Similarly, we could test whether the population mean is larger than a comparison value:

\[ H_0: \mu \le \mu_0 \\ H_1: \mu > \mu_0 \]

Connected to the decision of how to phrase the hypotheses (directional vs. non-directional) is the choice of a one-tailed test versus a two-tailed test . Let’s first think about the meaning of a one-tailed test. Using a significance level of 0.05, a one-tailed test means that 5% of the total area under the probability distribution of our test statistic is located in one tail. Thus, under a one-tailed test, we test for the possibility of the relationship in one direction only, disregarding the possibility of a relationship in the other direction. In our example, a one-tailed test could test either if the mean listening time is significantly larger or smaller compared to the control condition, but not both. Depending on the direction, the mean listening time is significantly larger (smaller) if the test statistic is located in the top (bottom) 5% of its probability distribution.

The following graph shows the critical values that our test statistic would need to surpass so that the difference between the population mean and the comparison value would be deemed statistically significant.

It can be seen that under a one-sided test, the rejection region is at one end of the distribution or the other. In a two-sided test, the rejection region is split between the two tails. As a consequence, the critical value of the test statistic is smaller using a one-tailed test, meaning that it has more power to detect an effect. Having said that, in most applications, we would like to be able catch effects in both directions, simply because we can often not rule out that an effect might exist that is not in the hypothesized direction. For example, if we would conduct a one-tailed test for a mean larger than some specified value but the mean turns out to be substantially smaller, then testing a one-directional hypothesis ($H_0: _0 $) would not allow us to conclude that there is a significant effect because there is not rejection at this end of the distribution.

5.1.4 Summary

As we have seen, the process of hypothesis testing consists of various steps:

• Formulate null and alternative hypotheses
• Select an appropriate test
• Choose the level of significance ( \(\alpha\) )
• Descriptive statistics and data visualization
• Conduct significance test
• Report results and draw a marketing conclusion

In the following, we will go through the individual steps using examples for different tests.

5.2 One sample t-test

The example we used in the introduction was an example of the one sample t-test and we computed all statistics by hand to explain the underlying intuition. When you conduct hypothesis tests using R, you do not need to calculate these statistics by hand, since there are build-in routines to conduct the steps for you. Let us use the same example again to see how you would conduct hypothesis tests in R.

1. Formulate null and alternative hypotheses

The null hypothesis states that there is no difference between the true population mean \(\mu\) and the hypothesized value (i.e., \(10\) ), while the alternative hypothesis states the opposite:

\[ H_0: \mu = 10 \\ H_1: \mu \neq 10 \]

2. Select an appropriate test

Because we would like to test if the mean of a variable is different from a specified threshold, the one-sample t-test is appropriate. The assumptions of the test are 1) that the variable is measured using an interval or ratio scale, and 2) that the sampling distribution is normal. Both assumptions are met since 1) listening time is a ratio scale, and 2) we deem the sample size (n = 50) large enough to assume a normal sampling distribution according to the central limit theorem.

3. Choose the level of significance

We choose the conventional 5% significance level.

4. Descriptive statistics and data visualization

Provide descriptive statistics using the stat.desc() function:

From this, we can already see that the mean is different from the hypothesized value. The question however remains, whether this difference is significantly different, given the sample size and the variability in the data. Since we only have one continuous variable, we can visualize the distribution in a histogram.

5. Conduct significance test

In the beginning of the chapter, we saw, how you could conduct significance test by hand. However, R has built-in routines that you can use to conduct the analyses. The t.test() function can be used to conduct the test. To test if the listening time among WU students was 10, you can use the following code:

Note that if you would have stated a directional hypothesis (i.e., the mean is either greater or smaller than 10 hours), you could easily amend the code to conduct a one sided test by changing the argument alternative from 'two.sided' to either 'less' or 'greater' .

6. Report results and draw a marketing conclusion

Note that the results are the same as above, when we computed the test by hand. You could summarize the results as follows:

On average, the listening times in our sample were different form 10 hours per month (Mean = 18.99 hours, SE = 1.78). This difference was significant t(49) = 5.058, p < .05 (95% CI = [15.42; 22.56]). Based on this evidence, we can conclude that the mean in our sample is significantly lower compared to the hypothesized population mean of \(10\) hours, providing evidence against the null hypothesis.

Note that in the reporting above, the number 49 in parenthesis refers to the degrees of freedom that are available from the output.

5.3 Comparing two means

In the one-sample test above, we tested the hypothesis that the population mean has some specific value \(\mu_0\) using data from only one sample. In marketing (as in many other disciplines), you will often be confronted with a situation where you wish to compare the means of two groups. For example, you may conduct an experiment and randomly split your sample into two groups, one of which receives a treatment (experimental group) while the other doesn’t (control group). In this case, the units (e.g., participants, products) in each group are different (‘between-subjects design’) and the samples are said to be independent. Hence, we would use a independent-means t-test . If you run an experiment with two experimental conditions and the same units (e.g., participants, products) were observed in both experimental conditions, the sample is said to be dependent in the sense that you have the same units in each group (‘within-subjects design’). In this case, we would need to conduct an dependent-means t-test . Both tests are described in the following sections, beginning with the independent-means t-test.

5.3.1 Independent-means t-test

Using an independent-means t-test, we can compare the means of two possibly different populations. It is, for example, quite common for online companies to test new service features by running an experiment and randomly splitting their website visitors into two groups: one is exposed to the website with the new feature (experimental group) and the other group is not exposed to the new feature (control group). This is a typical A/B-Test scenario.

As an example, imagine that a music streaming service would like to introduce a new playlist feature that let’s their users access playlists created by other users. The goal is to analyse how the new service feature impacts the listening time of users. The service randomly splits a representative subset of their users into two groups and collects data about their listening times over one month. Let’s create a data set to simulate such a scenario.

This data set contains two variables: the variable hours indicates the music listening times (in hours) and the variable group indicates from which group the observation comes, where ‘A’ refers to the control group (with the standard service) and ‘B’ refers to the experimental group (with the new playlist feature). Let’s first look at the descriptive statistics by group using the describeBy function:

From this, we can already see that there is a difference in means between groups A and B. We can also see that the number of observations is different, as is the standard deviation. The question that we would like to answer is whether there is a significant difference in mean listening times between the groups. Remember that different users are contained in each group (‘between-subjects design’) and that the observations in one group are independent of the observations in the other group. Before we will see how you can easily conduct an independent-means t-test, let’s go over some theory first.

5.3.1.1 Theory

As a starting point, let us label the unknown population mean of group A (control group) in our experiment \(\mu_1\) , and that of group B (experimental group) \(\mu_2\) . In this setting, the null hypothesis would state that the mean in group A is equal to the mean in group B:

\[ H_0: \mu_1=\mu_2 \]

This is equivalent to stating that the difference between the two groups ( \(\delta\) ) is zero:

\[ H_0: \mu_1 - \mu_2=0=\delta \]

That is, \(\delta\) is the new unknown population parameter, so that the null and alternative hypothesis become:

\[ H_0: \delta = 0 \\ H_1: \delta \ne 0 \]

Remember that we usually don’t have access to the entire population so that we can not observe \(\delta\) and have to estimate is from a sample statistic, which we define as \(d = \bar x_1-\bar x_2\) , i.e., the difference between the sample means from group a ( \(\bar x_1\) ) and group b ( \(\bar x_2\) ). But can we really estimate \(d\) from \(\delta\) ? Remember from the previous chapter, that we could estimate \(\mu\) from \(\bar x\) , because if we (hypothetically) take a larger number of samples, the distribution of the means of these samples (the sampling distribution) will be normally distributed and its mean will be (in the limit) equal to the population mean. It turns out that we can use the same underlying logic here. The above samples were drawn from two different populations with \(\mu_1\) and \(\mu_2\) . Let us compute the difference in means between these two populations:

This means that the true difference between the mean listening times of groups a and b is -7.42. Let us now repeat the exercise from the previous chapter: let us repeatedly draw a large number of \(20,000\) random samples of 100 users from each of these populations, compute the difference (i.e., \(d\) , our estimate of \(\delta\) ), store the difference for each draw and create a histogram of \(d\) .

This gives us the sampling distribution of the mean differences between the samples. You will notice that this distribution follows a normal distribution and is centered around the true difference between the populations. This means that, on average, the difference between two sample means \(d\) is a good estimate of \(\delta\) . In our example, the difference between \(\bar x_1\) and \(\bar x_2\) is:

Now that we have \(d\) as an estimate of \(\delta\) , how can we find out if the observed difference is significantly different from the null hypothesis (i.e., \(\delta = 0\) )?

Recall from the previous section, that the standard deviation of the sampling distribution \(\sigma_{\bar x}\) (i.e., the standard error) gives us indication about the precision of our estimate. Further recall that the standard error can be calculated as \(\sigma_{\bar x}={\sigma \over \sqrt{n}}\) . So how can we calculate the standard error of the difference between two population means? According to the variance sum law , to find the variance of the sampling distribution of differences, we merely need to add together the variances of the sampling distributions of the two populations that we are comparing. To find the standard error, we only need to take the square root of the variance (because the standard error is the standard deviation of the sampling distribution and the standard deviation is the square root of the variance), so that we get:

\[ \sigma_{\bar x_1-\bar x_2} = \sqrt{{\sigma_1^2 \over n_1}+{\sigma_2^2 \over n_2}} \]

But recall that we don’t actually know the true population standard deviation, so we use \(SE_{\bar x_1-\bar x_2}\) as an estimate of \(\sigma_{\bar x_1-\bar x_2}\) :

\[ SE_{\bar x_1-\bar x_2} = \sqrt{{s_1^2 \over n_1}+{s_2^2 \over n_2}} \]

Hence, for our example, we can calculate the standard error as follows:

Recall from above that we can calculate the t-statistic as:

\[ t= {\bar x - \mu_0 \over {s \over \sqrt{n}}} \]

Exchanging \(\bar x\) for \(d\) , we get

\[ t= {(\bar{x}_1 - \bar{x}_2) - (\mu_1 - \mu_2) \over {\sqrt{{s_1^2 \over n_1}+{s_2^2 \over n_2}}}} \]

Note that according to our hypothesis \(\mu_1-\mu_2=0\) , so that we can calculate the t-statistic as:

Following the example of our one sample t-test above, we would now need to compare this calculated test statistic to a critical value in order to assess if \(d\) is sufficiently far away from the null hypothesis to be statistically significant. To do this, we would need to know the exact t-distribution, which depends on the degrees of freedom. The problem is that deriving the degrees of freedom in this case is not that obvious. If we were willing to assume that \(\sigma_1=\sigma_2\) , the correct t-distribution has \(n_1 -1 + n_2-1\) degrees of freedom (i.e., the sum of the degrees of freedom of the two samples). However, because in real life we don not know if \(\sigma_1=\sigma_2\) , we need to account for this additional uncertainty. We will not go into detail here, but R automatically uses a sophisticated approach to correct the degrees of freedom called the Welch’s correction, as we will see in the subsequent application.

5.3.1.2 Application

The section above explained the theory behind the independent-means t-test and showed how to compute the statistics manually. Obviously you don’t have to compute these statistics by hand in this section shows you how to conduct an independent-means t-test in R using the example from above.

We wish to analyze whether there is a significant difference in music listening times between groups A and B. So our null hypothesis is that the means from the two populations are the same (i.e., there is no difference), while the alternative hypothesis states the opposite:

\[ H_0: \mu_1=\mu_2\\ H_1: \mu_1 \ne \mu_2 \]

Since we have a ratio scaled variable (i.e., listening times) and two independent groups, where the mean of one sample is independent of the group of the second sample (i.e., the groups contain different units), the independent-means t-test is appropriate.

We can compute the descriptive statistics for each group separately, using the describeBy() function:

This already shows us that the mean between groups A and B are different. We can visualize the data using a plot of means, boxplot, and a histogram.

To conduct the independent means t-test, we can use the t.test() function:

The results showed that listening times were higher in the experimental group B (Mean = 28.50, SE = 1.7) compared to the control group (Mean = 18.11, SE = 1.22). This means that the listening times were 10.39 hours higher on average in the experimental group (B), compared to the control group (A). An independent-means t-test showed that this difference is significant t(195.73) = -4.9646, p < .05 (95% CI = [-14.514246,-6.261264]).

5.3.2 Dependent-means t-test

While the independent-means t-test is used when different units (e.g., participants, products) were assigned to the different condition, the dependent-means t-test is used when there are two experimental conditions and the same units (e.g., participants, products) were observed in both experimental conditions.

Imagine, for example, a slightly different experimental setup for the above experiment. Imagine that we do not assign different users to the groups, but that a sample of 100 users gets to use the music streaming service with the new feature for one month and we compare the music listening times of these users during the month of the experiment with the listening time in the previous month. Let us generate data for this example:

Note that the data set has almost the same structure as before only that we know have two variables representing the listening times of each user in the month before the experiment and during the month of the experiment when the new feature was tested.

5.3.2.1 Theory

In this case, we want to test the hypothesis that there is no difference in mean the mean listening times between the two months. This can be expressed as follows:

\[ H_0: \mu_D = 0 \\ \] Note that the hypothesis only refers to one population, since both observations come from the same units (i.e., users). To use consistent notation, we replace \(\mu_D\) with \(\delta\) and get:

\[ H_0: \delta = 0 \\ H_1: \delta \neq 0 \]

where \(\delta\) denotes the difference between the observed listening times from the two consecutive months of the same users . As is the previous example, since we do not observe the entire population, we estimate \(\delta\) based on the sample using \(d\) , which is the difference in mean listening time between the two months for our sample. Note that we assume that everything else (e.g., number of new releases) remained constant over the two month to keep it simple. We can show as above that the sampling distribution follows a normal distribution with a mean that is (in the limit) the same as the population mean. This means, again, that the difference in sample means is a good estimate for the difference in population means. Let’s compute a new variable \(d\) , which is the difference between two month.

Note that we now have a new variable, which is the difference in listening times (in hours) between the two months. The mean of this difference is:

Again, we use \(SE_{\bar x}\) as an estimate of \(\sigma_{\bar x}\) :

\[ SE_{\bar d}={s \over \sqrt{n}} \] Hence, we can compute the standard error as:

The test statistic is therefore:

\[ t = {\bar d- \mu_0 \over SE_{\bar d}} \] on 99 (i.e., n-1) degrees of freedom. Now we can compute the t-statistic as follows:

Note that in the case of the dependent-means t-test, we only base our hypothesis on one population and hence there is only one population variance. This is because in the dependent sample test, the observations come from the same observational units (i.e., users). Hence, there is no unsystematic variation due to potential differences between users that were assigned to the experimental groups. This means that the influence of unobserved factors (unsystematic variation) relative to the variation due to the experimental manipulation (systematic variation) is not as strong in the dependent-means test compared to the independent-means test and we don’t need to correct for differences in the population variances.

5.3.2.2 Application

Again, we don’t have to compute all this by hand since the t.test(...) function can be used to do it for us. Now we have to use the argument paired=TRUE to let R know that we are working with dependent observations.

We would like to the test if there is a difference in music listening times between the two consecutive months, so our null hypothesis is that there is no difference, while the alternative hypothesis states the opposite:

\[ H_0: \mu_D = 0 \\ H_0: \mu_D \ne 0 \]

Since we have a ratio scaled variable (i.e., listening times) and two observations of the same group of users (i.e., the groups contain the same units), the dependent-means t-test is appropriate.

We can compute the descriptive statistics for each month separately, using the describe() function:

This already shows us that the mean between the two months are different. We can visiualize the data using a plot of means, boxplot, and a histogram.

To plot the data, we need to do some restructuring first, since the variables are now stored in two different columns (“hours_a” and “hours_b”). This is also known as the “wide” format. To plot the data we need all observations to be stored in one variable. This is also known as the “long” format. We can use the melt(...) function from the reshape2 package to “melt” the two variable into one column to plot the data.

Now we are ready to plot the data:

To conduct the independent means t-test, we can use the t.test() function with the argument paired = TRUE :

On average, the same users used the service more when it included the new feature (M = 25.96, SE = 1.68) compared to the service without the feature (M = 20.99, SE = 1.34). This difference was significant t(99) = 2.3781, p < .05 (95% CI = [0.82, 9.12]).

5.3.3 Further considerations

5.3.3.1 type i and type ii errors.

When choosing the level of significance ( \(\alpha\) ). It is important to note that the choice of the significance level affects the type 1 and type 2 error:

• Type I error: When we believe there is a genuine effect in our population, when in fact there isn’t. Probability of type I error ( \(\alpha\) ) = level of significance.
• Type II error: When we believe that there is no effect in the population, when in fact there is.

This following table shows the possible outcomes of a test (retain vs. reject \(H_0\) ), depending on whether \(H_0\) is true or false in reality.

5.3.3.2 Significance level, sample size, power, and effect size

When you plan to conduct an experiment, there are some factors that are under direct control of the researcher:

• Significance level ( \(\alpha\) ) : The probability of finding an effect that does not genuinely exist.
• Sample size (n) : The number of observations in each group of the experimental design.

Unlike α and n, which are specified by the researcher, the magnitude of β depends on the actual value of the population parameter. In addition, β is influenced by the effect size (e.g., Cohen’s d), which can be used to determine a standardized measure of the magnitude of an observed effect. The following parameters are affected more indirectly:

• Power (1-β) : The probability of finding an effect that does genuinely exists.
• Effect size (d) : Standardized measure of the effect size under the alternate hypothesis.

Although β is unknown, it is related to α. For example, if we would like to be absolutely sure that we do not falsely identify an effect which does not exist (i.e., make a type I error), this means that the probability of identifying an effect that does exist (i.e., 1-β) decreases and vice versa. Thus, an extremely low value of α (e.g., α = 0.0001) will result in intolerably high β errors. A common approach is to set α=0.05 and 1-β=0.80.

Unlike the t-value of our test, the effect size (d) is unaffected by the sample size and can be categorized as follows (see Cohen, J. 1988):

• 0.2 (small effect)
• 0.5 (medium effect)
• 0.8 (large effect)

In order to test more subtle effects (smaller effect sizes), you need a larger sample size compared to the test of more obvious effects. In this paper , you can find a list of examples for different effect sizes and the number of observations you need to reliably find an effect of that magnitude. Although the exact effect size is unknown before the experiment, you might be able to make a guess about the effect size (e.g., based on previous studies).

If you wish to obtain a standardized measure of the effect, you may compute the effect size (Cohen’s d) using the cohensD() function from the lsr package. Using the examples from the independent-means t-test above, we would use:

According to the thresholds defined above, this effect would be judged to be a small-medium effect.

For the dependent-means t-test, we would use:

According to the thresholds defined above, this effect would also be judged to be a small-medium effect.

When constructing an experimental design, your goal should be to maximize the power of the test while maintaining an acceptable significance level and keeping the sample as small as possible. To achieve this goal, you may use the pwr package, which let’s you compute n , d , alpha , and power . You only need to specify three of the four input variables to get the fourth.

For example, what sample size do we need (per group) to identify an effect with d = 0.6, α = 0.05, and power = 0.8:

Or we could ask, what is the power of our test with 51 observations in each group, d = 0.6, and α = 0.05:

5.3.3.3 P-values, stopping rules and p-hacking

From my experience, students tend to place a lot of weight on p-values when interpreting their research findings. It is therefore important to note some points that hopefully help to put the meaning of a “significant” vs. “insignificant” test result into perspective.

Significant result

• Even if the probability of the effect being a chance result is small (e.g., less than .05) it doesn’t necessarily mean that the effect is important.
• Very small and unimportant effects can turn out to be statistically significant if the sample size is large enough.

Insignificant result

• If the probability of the effect occurring by chance is large (greater than .05), the alternative hypothesis is rejected. However, this does not mean that the null hypothesis is true.
• Although an effect might not be large enough to be anything other than a chance finding, it doesn’t mean that the effect is zero.
• In fact, two random samples will always have slightly different means that would deemed to be statistically significant if the samples were large enough.

Thus, you should not base your research conclusion on p-values alone!

It is also crucial to determine the sample size before you run the experiment or before you start your analysis. Why? Consider the following example:

• You run an experiment
• After each respondent you analyze the data and look at the mean difference between the two groups with a t-test
• You stop when you have a significant effect

This is called p-hacking and should be avoided at all costs. Assuming that both groups come from the same population (i.e., there is no difference in the means): What is the likelihood that the result will be significant at some point? In other words, what is the likelihood that you will draw the wrong conclusion from your data that there is an effect, while there is none? This is shown in the following graph using simulated data - the color red indicates significant test results that arise although there is no effect (i.e., false positives).

Figure 5.1: p-hacking (red indicates false positives)

5.4 Comparing several means

This chapter is primarily based on Field, A., Miles J., & Field, Z. (2012): Discovering Statistics Using R. Sage Publications, chapters 10 & 12 .

5.4.1 Introduction

In the previous section we learned how to compare means using a t-test. The t-test has some limitations since it only lets you compare 2 means and you can only use it with one independent variable. However, often we would like to compare means from 3 or more groups. In addition, there may be instances in which you manipulate more than one independent variable. For these applications, ANOVA (ANalysis Of VAriance) can be used. Hence, to conduct ANOVA you need:

• A metric dependent variable (i.e., measured using an interval or ratio scale)
• One or more non-metric (categorical) independent variables (also called factors)

A treatment is a particular combination of factor levels, or categories. One-way ANOVA is used when there is only one categorical variable (factor). In this case, a treatment is the same as a factor level. N-way ANOVA is used with two or more factors. Note that we are only going to talk about a single independent variable in the context of ANOVA. If you have multiple independent variables please refere to the chapter on Regression .

Let’s use an example to see how ANOVA works. Similar to the previous example it is also imaginable that the music streaming service experiments with a recommendation system for user created playlists. We now have three groups, the control group “A” with the current system, treatment group “B” who have access to playlists created by other users but are not shown recommendations and treatment group “C” who are shown recommendations for user created playlists. As always, we load and inspect the data first:

The null hypothesis, typically, is that all means are equal (non-directional hypothesis). Hence, in our case:

\[H_0: \mu_1 = \mu_2 = \mu_3\]

The alternative hypothesis is simply that the means are not all equal, i.e.,

\[H_1: \textrm{Means are not all equal}\]

If you wanted to put this in mathematical notation, you could also write:

\[H_1: \exists {i,j}: {\mu_i \ne \mu_j} \]

To get a first impression if there are any differences in listening times across the experimental groups, we use the describeBy(...) function from the psych package:

In addition, you should visualize the data using appropriate plots:

Figure 5.2: Plot of means

Note that ANOVA is an omnibus test, which means that we test for an overall difference between groups. Hence, the test will only tell you if the group means are different, but it won’t tell you exactly which groups are different from another.

So why don’t we then just conduct a series of t-tests for all combinations of groups (i.e., A vs. B, A vs. C, B vs. C)? The reason is that if we assume each test to be independent, then there is a 5% probability of falsely rejecting the null hypothesis (Type I error) for each test. In our case:

• A vs. B (α = 0.05)
• A vs. C (α = 0.05)
• B vs. C (α = 0.05)

This means that the overall probability of making a Type I error is 1-(0.95 3 ) = 0.143, since the probability of no Type I error is 0.95 for each of the three tests. Consequently, the Type I error probability would be 14.3%, which is above the conventional standard of 5%. This is also known as the family-wise or experiment-wise error.

5.4.2 Decomposing variance

The basic concept underlying ANOVA is the decomposition of the variance in the data. There are three variance components which we need to consider:

• We calculate how much variability there is between scores: Total sum of squares (SS T )
• We then calculate how much of this variability can be explained by the model we fit to the data (i.e., how much variability is due to the experimental manipulation): Model sum of squares (SS M )
• … and how much cannot be explained (i.e., how much variability is due to individual differences in performance): Residual sum of squares (SS R )

The following figure shows the different variance components using a generalized data matrix:

Decomposing variance

The total variation is determined by the variation between the categories (due to our experimental manipulation) and the within-category variation that is due to extraneous factors (e.g., promotion of artists on a social network):

\[SS_T= SS_M+SS_R\]

To get a better feeling how this relates to our data set, we can look at the data in a slightly different way. Specifically, we can use the dcast(...) function from the reshape2 package to convert the data to wide format:

In this example, X 1 from the generalized data matrix above would refer to the factor level “A”, X 2 to the level “B”, and X 3 to the level “C”. Y 11 refers to the first data point in the first row (i.e., “13”), Y 12 to the second data point in the first row (i.e., “21”), etc.. The grand mean ( \(\overline{Y}\) ) and the category means ( \(\overline{Y}_c\) ) can be easily computed:

To see how each variance component can be derived, let’s look at the data again. The following graph shows the individual observations by experimental group:

Figure 5.3: Sum of Squares

5.4.2.1 Total sum of squares

To compute the total variation in the data, we consider the difference between each observation and the grand mean. The grand mean is the mean over all observations in the data set. The vertical lines in the following plot measure how far each observation is away from the grand mean:

Figure 5.4: Total Sum of Squares

The formal representation of the total sum of squares (SS T ) is:

\[ SS_T= \sum_{i=1}^{N} (Y_i-\bar{Y})^2 \]

This means that we need to subtract the grand mean from each individual data point, square the difference, and sum up over all the squared differences. Thus, in our example, the total sum of squares can be calculated as:

\[ \begin{align} SS_T =&(13−24.67)^2 + (14−24.67)^2 + … + (2−24.67)^2\\ &+(21−24.67)^2 + (18-24.67)^2 + … + (17−24.67)^2\\ &+(30−24.67)^2 + (37−24.67)^2 + … + (28−24.67)^2\\ &=30855.64 \end{align} \]

You could also compute this in R using:

For the subsequent analyses, it is important to understand the concept behind the degrees of freedom . Remember that in order to estimate a population value from a sample, we need to hold something in the population constant. In ANOVA, the df are generally one less than the number of values used to calculate the SS. For example, when we estimate the population mean from a sample, we assume that the sample mean is equal to the population mean. Then, in order to estimate the population mean from the sample, all but one scores are free to vary and the remaining score needs to be the value that keeps the population mean constant. In our example, we used all 300 observations to calculate the sum of square, so the total degrees of freedom (df T ) are:

\[\begin{equation} \begin{split} df_T = N-1=300-1=299 \end{split} \tag{5.1} \end{equation}\]

5.4.2.2 Model sum of squares

Now we know that there are 26646.33 units of total variation in our data. Next, we compute how much of the total variation can be explained by the differences between groups (i.e., our experimental manipulation). To compute the explained variation in the data, we consider the difference between the values predicted by our model for each observation (i.e., the group mean) and the grand mean. The group mean refers to the mean value within the experimental group. The vertical lines in the following plot measure how far the predicted value for each observation (i.e., the group mean) is away from the grand mean:

Figure 5.5: Model Sum of Squares

The formal representation of the model sum of squares (SS M ) is:

\[ SS_M= \sum_{j=1}^{c} n_j(\bar{Y}_j-\bar{Y})^2 \]

where c denotes the number of categories (experimental groups). This means that we need to subtract the grand mean from each group mean, square the difference, and sum up over all the squared differences. Thus, in our example, the model sum of squares can be calculated as:

\[ \begin{align} SS_M &= 100*(15.47−24.67)^2 + 100*(24.88−24.67)^2 + 100*(33.66−24.67)^2 \\ &= 21321.21 \end{align} \]

You could also compute this manually in R using:

In this case, we used the three group means to calculate the sum of squares, so the model degrees of freedom (df M ) are:

\[ df_M= c-1=3-1=2 \]

5.4.2.3 Residual sum of squares

Lastly, we calculate the amount of variation that cannot be explained by our model. In ANOVA, this is the sum of squared distances between what the model predicts for each data point (i.e., the group means) and the observed values. In other words, this refers to the amount of variation that is caused by extraneous factors, such as differences between product characteristics of the products in the different experimental groups. The vertical lines in the following plot measure how far each observation is away from the group mean:

Figure 5.6: Residual Sum of Squares

The formal representation of the residual sum of squares (SS R ) is:

\[ SS_R= \sum_{j=1}^{c} \sum_{i=1}^{n} ({Y}_{ij}-\bar{Y}_{j})^2 \]

This means that we need to subtract the group mean from each individual observation, square the difference, and sum up over all the squared differences. Thus, in our example, the model sum of squares can be calculated as:

\[ \begin{align} SS_R =& (13−14.34)^2 + (14−14.34)^2 + … + (2−14.34)^2 \\ +&(21−24.7)^2 + (18−24.7)^2 + … + (17−24.7)^2 \\ +& (30−34.99)^2 + (37−34.99)^2 + … + (28−34.99)^2 \\ =& 9534.43 \end{align} \]

In this case, we used the 10 values for each of the SS for each group, so the residual degrees of freedom (df R ) are:

\[ \begin{align} df_R=& (n_1-1)+(n_2-1)+(n_3-1) \\ =&(100-1)+(100-1)+(100-1)=297 \end{align} \]

5.4.2.4 Effect strength

Once you have computed the different sum of squares, you can investigate the effect strength. \(\eta^2\) is a measure of the variation in Y that is explained by X:

\[ \eta^2= \frac{SS_M}{SS_T}=\frac{21321.21}{30855.64}=0.69 \]

To compute this in R:

The statistic can only take values between 0 and 1. It is equal to 0 when all the category means are equal, indicating that X has no effect on Y. In contrast, it has a value of 1 when there is no variability within each category of X but there is some variability between categories.

5.4.2.5 Test of significance

How can we determine whether the effect of X on Y is significant?

• First, we calculate the fit of the most basic model (i.e., the grand mean)
• Then, we calculate the fit of the “best” model (i.e., the group means)
• A good model should fit the data significantly better than the basic model
• The F-statistic or F-ratio compares the amount of systematic variance in the data to the amount of unsystematic variance

The F-statistic uses the ratio of mean square related to X (explained variation) and the mean square related to the error (unexplained variation):

\(\frac{SS_M}{SS_R}\)

However, since these are summed values, their magnitude is influenced by the number of scores that were summed. For example, to calculate SS M we only used the sum of 3 values (the group means), while we used 30 and 27 values to calculate SS T and SS R , respectively. Thus, we calculate the average sum of squares (“mean square”) to compare the average amount of systematic vs. unsystematic variation by dividing the SS values by the degrees of freedom associated with the respective statistic.

Mean square due to X:

\[ MS_M= \frac{SS_M}{df_M}=\frac{SS_M}{c-1}=\frac{21321.21}{(3-1)} \]

Mean square due to error:

\[ MS_R= \frac{SS_R}{df_R}=\frac{SS_R}{N-c}=\frac{9534.43}{(300-3)} \]

Now, we compare the amount of variability explained by the model (experiment), to the error in the model (variation due to extraneous variables). If the model explains more variability than it can’t explain, then the experimental manipulation has had a significant effect on the outcome (DV). The F-radio can be derived as follows:

\[ F= \frac{MS_M}{MS_R}=\frac{\frac{SS_M}{c-1}}{\frac{SS_R}{N-c}}=\frac{\frac{21321.21}{(3-1)}}{\frac{9534.43}{(300-3)}}=332.08 \]

You can easily compute this in R:

This statistic follows the F distribution with (m = c – 1) and (n = N – c) degrees of freedom. This means that, like the \(\chi^2\) distribution, the shape of the F-distribution depends on the degrees of freedom. In this case, the shape depends on the degrees of freedom associated with the numerator and denominator used to compute the F-ratio. The following figure shows the shape of the F-distribution for different degrees of freedom:

The F distribution

The outcome of the test is one of the following:

• If the null hypothesis of equal category means is not rejected, then the independent variable does not have a significant effect on the dependent variable
• If the null hypothesis is rejected, then the effect of the independent variable is significant

For 2 and 297 degrees of freedom, the critical value of F is 3.026 for α=0.05. As usual, you can either look up these values in a table or use the appropriate function in R:

The output tells us that the calculated test statistic exceeds the critical value. We can also show the test result visually:

Visual depiction of the test result

Thus, we conclude that because F CAL = 332.08 > F CR = 3.03, H 0 is rejected!

Interpretation: one or more of the differences between means are statistically significant.

Reporting: There was a significant effect of promotion on sales levels, F(2,297) = 332.08, p < 0.05, \(\eta^2\) = 0.69.

Remember: This doesn’t tell us where the differences between groups lie. To find out which group means exactly differ, we need to use post-hoc procedures (see below).

You don’t have to compute these statistics manually! Luckily, there is a function for ANOVA in R, which does the above calculations for you as we will see in the next section.

5.4.3 One-way ANOVA

5.4.3.1 basic anova.

As already indicated, one-way ANOVA is used when there is only one categorical variable (factor). Before conducting ANOVA, you need to check if the assumptions of the test are fulfilled. The assumptions of ANOVA are discussed in the following sections.

Independence of observations

The observations in the groups should be independent. Because we randomly assigned the listeners to the experimental conditions, this assumption can be assumed to be met.

Distributional assumptions

ANOVA is relatively immune to violations to the normality assumption when sample sizes are large due to the Central Limit Theorem. However, if your sample is small (i.e., n < 30 per group) you may nevertheless want to check the normality of your data, e.g., by using the Shapiro-Wilk test or QQ-Plot. In our example, we have 100 observations in each group which is plenty but let’s create another example with only 10 observations in each group. In the latter case we cannot rely on the Central Limit Theorem and we should test the normality of our data. This can be done using the Shapiro-Wilk Test, which has the Null Hypothesis that the data is normally distributed. Hence, an insignificant test results means that the data can be assumed to be approximately normally distributed:

Since the test result is insignificant for all groups, we can conclude that the data approximately follow a normal distribution.

We could also test the distributional assumptions visually using a Q-Q plot (i.e., quantile-quantile plot). This plot can be used to assess if a set of data plausibly came from some theoretical distribution such as the Normal distribution. Since this is just a visual check, it is somewhat subjective. But it may help us to judge if our assumption is plausible, and if not, which data points contribute to the violation. A Q-Q plot is a scatterplot created by plotting two sets of quantiles against one another. If both sets of quantiles came from the same distribution, we should see the points forming a line that’s roughly straight. In other words, Q-Q plots take your sample data, sort it in ascending order, and then plot them versus quantiles calculated from a theoretical distribution. Quantiles are often referred to as “percentiles” and refer to the points in your data below which a certain proportion of your data fall. Recall, for example, the standard Normal distribution with a mean of 0 and a standard deviation of 1. Since the 50th percentile (or 0.5 quantile) is 0, half the data lie below 0. The 95th percentile (or 0.95 quantile), is about 1.64, which means that 95 percent of the data lie below 1.64. The 97.5th quantile is about 1.96, which means that 97.5% of the data lie below 1.96. In the Q-Q plot, the number of quantiles is selected to match the size of your sample data.

To create the Q-Q plot for the normal distribution, you may use the qqnorm() function, which takes the data to be tested as an argument. Using the qqline() function subsequently on the data creates the line on which the data points should fall based on the theoretical quantiles. If the individual data points deviate a lot from this line, it means that the data is not likely to follow a normal distribution.

Figure 5.7: Q-Q plot 1

Figure 5.8: Q-Q plot 2

Figure 5.9: Q-Q plot 3

The Q-Q plots suggest an approximately Normal distribution. If the assumption had been violated, you might consider transforming your data or resort to a non-parametric test.

Homogeneity of variance

Let’s return to our original dataset with 100 observations in each group for the rest of the analysis.

You can test the homogeneity of variances in R using Levene’s test:

The null hypothesis of the test is that the group variances are equal. Thus, if the test result is significant it means that the variances are not equal. If we cannot reject the null hypothesis (i.e., the group variances are not significantly different), we can proceed with the ANOVA as follows:

You can see that the p-value is smaller than 0.05. This means that, if there really was no difference between the population means (i.e., the Null hypothesis was true), the probability of the observed differences (or larger differences) is less than 5%.

To compute η 2 from the output, we can extract the relevant sum of squares as follows

You can see that the results match the results from our manual computation above ( \(\eta^2 =\) 0.69).

The aov() function also automatically generates some plots that you can use to judge if the model assumptions are met. We will inspect two of the plots here.

We will use the first plot to inspect if the residual variances are equal across the experimental groups:

Generally, the residual variance (i.e., the range of values on the y-axis) should be the same for different levels of our independent variable. The plot shows, that there are some slight differences. Notably, the range of residuals is higher in group “B” than in group “C”. However, the differences are not that large and since the Levene’s test could not reject the Null of equal variances, we conclude that the variances are similar enough in this case.

The second plot can be used to test the assumption that the residuals are approximately normally distributed. We use a Q-Q plot to test this assumption:

The plot suggests that, the residuals are approximately normally distributed. We could also test this by extracting the residuals from the anova output using the resid() function and using the Shapiro-Wilk test:

Confirming the impression from the Q-Q plot, we cannot reject the Null that the residuals are approximately normally distributed.

Note that if Levene’s test would have been significant (i.e., variances are not equal), we would have needed to either resort to non-parametric tests (see below), or compute the Welch’s F-ratio instead:

You can see that the results are fairly similar, since the variances turned out to be fairly equal across groups.

5.4.3.2 Post-hoc tests

Provided that significant differences were detected by the overall ANOVA you can find out which group means are different using post hoc procedures. Post hoc procedures are designed to conduct pairwise comparisons of all different combinations of the treatment groups by correcting the level of significance for each test such that the overall Type I error rate (α) across all comparisons remains at 0.05.

In other words, we rejected H 0 : μ 1 = μ 2 = μ 3 , and now we would like to test:

\[H_0: \mu_1 = \mu_2\]

\[H_0: \mu_1 = \mu_3\]

\[H_0: \mu_2 = \mu_3\]

There are several post hoc procedures available to choose from. In this tutorial, we will cover Bonferroni and Tukey’s HSD (“honest significant differences”). Both tests control for family-wise error. Bonferroni tends to have more power when the number of comparisons is small, whereas Tukey’ HSDs is better when testing large numbers of means.

5.4.3.2.1 Bonferroni

One of the most popular (and easiest) methods to correct for the family-wise error rate is to conduct the individual t-tests and divide α by the number of comparisons („k“):

\[ p_{CR}= \frac{\alpha}{k} \]

In our example with three groups:

\[p_{CR}= \frac{0.05}{3}=0.017\]

Thus, the “corrected” critical p-value is now 0.017 instead of 0.05 (i.e., the critical t value is higher). You can implement the Bonferroni procedure in R using:

In the output, you will get the corrected p-values for the individual tests. In our example, we can reject H 0 of equal means for all three tests, since p < 0.05 for all combinations of groups.

Note the difference between the results from the post-hoc test compared to individual t-tests. For example, when we test the “B” vs. “C” groups, the result from a t-test would be:

Usually the p-value is lower in the t-test, reflecting the fact that the family-wise error is not corrected (i.e., the test is less conservative). In this case the p-value is extremely small in both cases and thus indistinguishable.

5.4.3.2.2 Tukey’s HSD

Tukey’s HSD also compares all possible pairs of means (two-by-two combinations; i.e., like a t-test, except that it corrects for family-wise error rate).

Test statistic:

\[\begin{equation} \begin{split} HSD= q\sqrt{\frac{MS_R}{n_c}} \end{split} \tag{5.2} \end{equation}\]

• q = value from studentized range table (see e.g., here )
• MS R = Mean Square Error from ANOVA
• n c = number of observations per group
• Decision: Reject H 0 if

\[|\bar{Y}_i-\bar{Y}_j | > HSD\]

The value from the studentized range table can be obtained using the qtukey() function.

\[HSD= 3.33\sqrt{\frac{33.99}{100}}=1.94\]

Since all mean differences between groups are larger than 1.906, we can reject the null hypothesis for all individual tests, confirming the results from the Bonferroni test. To compute Tukey’s HSD, we can use the appropriate function from the multcomp package.

We may also plot the result for the mean differences incl. their confidence intervals:

Figure 5.10: Tukey’s HSD

You can see that the CIs do not cross zero, which means that the true difference between group means is unlikely zero.

Reporting of post hoc results:

The post hoc tests based on Bonferroni and Tukey’s HSD revealed that people listened to music significantly more when:

• they had access to user created playlists vs. those who did not,
• they got recommendations vs. those who did not. This is true for both the control group “A” as well as treatment “B”.

The following video summarizes how to conduct a one-way ANOVA in R

5.5 Non-parametric tests

Non-Parametric tests do not require the sampling distribution to be normally distributed (a.k.a. “assumption free tests”). These tests may be used when the variable of interest is measured on an ordinal scale or when the parametric assumptions do not hold. They often rely on ranking the data instead of analyzing the actual scores. By ranking the data, information on the magnitude of differences is lost. Thus, parametric tests are more powerful if the sampling distribution is normally distributed.

When should you use non-parametric tests?

• When your DV is measured on an ordinal scale
• When your data is better represented by the median (e.g., there are outliers that you can’t remove)
• When the assumptions of parametric tests are not met (e.g., normally distributed sampling distribution)
• You have a very small sample size (i.e., the central limit theorem does not apply)

5.5.1 Mann-Whitney U Test (a.k.a. Wilcoxon rank-sum test)

The Mann-Whitney U test is a non-parametric test of differences between groups, similar to the two sample t-test. In contrast to the two sample t-test it only requires ordinally scaled data and relies on weaker assumptions. Thus it is often useful if the assumptions of the t-test are violated, especially if the data is not on a ratio scale. The following assumptions must be fulfilled for the test to be applicable:

• The dependent variable is at least ordinally scaled (i.e. a ranking between values can be established)
• The independent variable has only two levels
• A between-subjects design is used (i.e., the subjects are not matched across conditions)

Intuitively, the test compares the frequency of low and high ranks between groups. Under the null hypothesis, the amount of high and low ranks should be roughly equal in the two groups. This is achieved through comparing the expected sum of ranks to the actual sum of ranks.

As an example, we will be using data obtained from a field experiment with random assignment. In a music download store, new releases were randomly assigned to an experimental group and sold at a reduced price (i.e., 7.95€), or a control group and sold at the standard price (9.95€). A representative sample of 102 new releases were sampled and these albums were randomly assigned to the experimental groups (i.e., 51 albums per group). The sales were tracked over one day.

Let’s load and investigate the data first:

Inspect descriptives (overall and by group).

Create boxplot and plot of means.

Figure 5.11: Boxplot

Let’s assume that one of the parametric assumptions has been violated and we needed to conduct a non-parametric test. Then, the Mann-Whitney U test is implemented in R using the function wilcox.test() . Using the ranking data as an independent variable and the listening time as a dependent variable, the test could be executed as follows:

The p-value is smaller than 0.05, which leads us to reject the null hypothesis, i.e. the test yields evidence that the new service feature leads to higher music listening times.

5.5.2 Wilcoxon signed-rank test

The Wilcoxon signed-rank test is a non-parametric test used to analyze the difference between paired observations, analogously to the paired t-test. It can be used when measurements come from the same observational units but the distributional assumptions of the paired t-test do not hold, because it does not require any assumptions about the distribution of the measurements. Since we subtract two values, however, the test requires that the dependent variable is at least interval scaled, meaning that intervals have the same meaning for different points on our measurement scale.

Under the null hypothesis \(H_0\) , the differences of the measurements should follow a symmetric distribution around 0, meaning that, on average, there is no difference between the two matched samples. \(H_1\) states that the distributions mean is non-zero.

As an example, let’s consider a slightly different experimental setup for the music download store. Imagine that new releases were either sold at a reduced price (i.e., 7.95€), or at the standard price (9.95€). Every time a customer came to the store, the prices were randomly determined for every new release. This means that the same 51 albums were either sold at the standard price or at the reduced price and this price was determined randomly. The sales were then recorded over one day. Note the difference to the previous case, where we randomly split the sample and assigned 50% of products to each condition. Now, we randomly vary prices for all albums between high and low prices.

Again, let’s assume that one of the prarametric assumptions has been violated and we needed to conduct a non-parametric test. Then the Wilcoxon signed-rank test can be performed with the same command as the Mann-Whitney U test, provided that the argument paired is set to TRUE .

Using the 95% confidence level, the result would suggest a significant effect of price on sales (i.e., p < 0.05).

5.5.3 Kruskal-Wallis test

• When the dependent variable is measured at an ordinal scale and we want to compare more than 2 means
• When the assumptions of independent ANOVA are not met (e.g., assumptions regarding the sampling distribution in small samples)

The Kruskal–Wallis test is the non-parametric counterpart of the one-way independent ANOVA. It is designed to test for significant differences in population medians when you have more than two samples (otherwise you would use the Mann-Whitney U-test). The theory is very similar to that of the Mann–Whitney U-test since it is also based on ranked data. The Kruskal-Wallis test is carried out using the kruskal.test() function. Using the same data as before, we type:

The test-statistic follows a chi-square distribution and since the test is significant (p < 0.05), we can conclude that there are significant differences in population medians. Provided that the overall effect is significant, you may perform a post hoc test to find out which groups are different. To get a first impression, we can plot the data using a boxplot:

Figure 5.12: Boxplot

To test for differences between groups, we can, for example, apply post hoc tests according to Nemenyi for pairwise multiple comparisons of the ranked data using the appropriate function from the PMCMR package.

The results reveal that there is a significant difference between the “low” and “high” promotion groups. Note that the results are different compared to the results from the parametric test above. This difference occurs because non-parametric tests have more power to detect differences between groups since we lose information by ranking the data. Thus, you should rely on parametric tests if the assumptions are met.

5.6 Categorical data

In some instances, you will be confronted with differences between proportions, rather than differences between means. For example, you may conduct an A/B-Test and wish to compare the conversion rates between two advertising campaigns. In this case, your data is binary (0 = no conversion, 1 = conversion) and the sampling distribution for such data is binomial. While binomial probabilities are difficult to calculate, we can use a Normal approximation to the binomial when n is large (>100) and the true likelihood of a 1 is not too close to 0 or 1.

Let’s use an example: assume a call center where service agents call potential customers to sell a product. We consider two call center agents:

• Service agent 1 talks to 300 customers and gets 200 of them to buy (conversion rate=2/3)
• Service agent 2 talks to 300 customers and gets 100 of them to buy (conversion rate=1/3)

As always, we load the data first:

Next, we create a table to check the relative frequencies:

We could also plot the data to visualize the frequencies using ggplot:

Figure 5.13: proportion of conversions per agent (stacked bar chart)

… or using the mosaicplot() function:

Figure 5.14: proportion of conversions per agent (mosaic plot)

5.6.1 Confidence intervals for proportions

Recall that we can use confidence intervals to determine the range of values that the true population parameter will take with a certain level of confidence based on the sample. Similar to the confidence interval for means, we can compute a confidence interval for proportions. The (1- \(\alpha\) )% confidence interval for proportions is approximately

\[ CI = p\pm z_{1-\frac{\alpha}{2}}*\sqrt{\frac{p*(1-p)}{N}} \]

where \(\sqrt{p(1-p)}\) is the equivalent to the standard deviation in the formula for the confidence interval for means. Based on the equation, it is easy to compute the confidence intervals for the conversion rates of the call center agents:

Similar to testing for differences in means, we could also ask: Is agent 1 twice as likely as agent 2 to convert a customer? Or, to state it formally:

\[H_0: \pi_1=\pi_2 \\ H_1: \pi_1\ne \pi_2\]

where \(\pi\) denotes the population parameter associated with the proportion in the respective population. One approach to test this is based on confidence intervals to estimate the difference between two populations. We can compute an approximate confidence interval for the difference between the proportion of successes in group 1 and group 2, as:

\[ CI = p_1-p_2\pm z_{1-\frac{\alpha}{2}}*\sqrt{\frac{p_1*(1-p_1)}{n_1}+\frac{p_2*(1-p_2)}{n_2}} \]

If the confidence interval includes zero, then the data does not suggest a difference between the groups. Let’s compute the confidence interval for differences in the proportions by hand first:

Now we can see that the 95% confidence interval estimate of the difference between the proportion of conversions for agent 1 and the proportion of conversions for agent 2 is between 26% and 41%. This interval tells us the range of plausible values for the difference between the two population proportions. According to this interval, zero is not a plausible value for the difference (i.e., interval does not cross zero), so we reject the null hypothesis that the population proportions are the same.

Instead of computing the intervals by hand, we could also use the prop.test() function:

Note that the prop.test() function uses a slightly different (more accurate) way to compute the confidence interval (Wilson’s score method is used). It is particularly a better approximation for smaller N. That’s why the confidence interval in the output slightly deviates from the manual computation above, which uses the Wald interval.

You can also see that the output from the prop.test() includes the results from a χ 2 test for the equality of proportions (which will be discussed below) and the associated p-value. Since the p-value is less than 0.05, we reject the null hypothesis of equal probability. Thus, the reporting would be:

The test showed that the conversion rate for agent 1 was higher by 33%. This difference is significant χ (1) = 70, p < .05 (95% CI = [0.25,0.41]).

5.6.2 Chi-square test

In the previous section, we saw how we can compute the confidence interval for the difference between proportions to decide on whether or not to reject the null hypothesis. Whenever you would like to investigate the relationship between two categorical variables, the \(\chi^2\) test may be used to test whether the variables are independent of each other. It achieves this by comparing the expected number of observations in a group to the actual values. Let’s continue with the example from the previous section. Under the null hypothesis, the two variables agent and conversion in our contingency table are independent (i.e., there is no relationship). This means that the frequency in each field will be roughly proportional to the probability of an observation being in that category, calculated under the assumption that they are independent. The difference between that expected quantity and the actual quantity can be used to construct the test statistic. The test statistic is computed as follows:

\[ \chi^2=\sum_{i=1}^{J}\frac{(f_o-f_e)^2}{f_e} \]

where \(J\) is the number of cells in the contingency table, \(f_o\) are the observed cell frequencies and \(f_e\) are the expected cell frequencies. The larger the differences, the larger the test statistic and the smaller the p-value.

The observed cell frequencies can easily be seen from the contingency table:

The expected cell frequencies can be calculated as follows:

\[ f_e=\frac{(n_r*n_c)}{n} \]

where \(n_r\) are the total observed frequencies per row, \(n_c\) are the total observed frequencies per column, and \(n\) is the total number of observations. Thus, the expected cell frequencies under the assumption of independence can be calculated as:

To sum up, these are the expected cell frequencies

… and these are the observed cell frequencies

To obtain the test statistic, we simply plug the values into the formula:

The test statistic is \(\chi^2\) distributed. The chi-square distribution is a non-symmetric distribution. Actually, there are many different chi-square distributions, one for each degree of freedom as show in the following figure.

Figure 5.15: The chi-square distribution

You can see that as the degrees of freedom increase, the chi-square curve approaches a normal distribution. To find the critical value, we need to specify the corresponding degrees of freedom, given by:

\[ df=(r-1)*(c-1) \]

where \(r\) is the number of rows and \(c\) is the number of columns in the contingency table. Recall that degrees of freedom are generally the number of values that can vary freely when calculating a statistic. In a 2 by 2 table as in our case, we have 2 variables (or two samples) with 2 levels and in each one we have 1 that vary freely. Hence, in our example the degrees of freedom can be calculated as:

Now, we can derive the critical value given the degrees of freedom and the level of confidence using the qchisq() function and test if the calculated test statistic is larger than the critical value:

Figure 5.16: Visual depiction of the test result

We could also compute the p-value using the pchisq() function, which tells us the probability of the observed cell frequencies if the null hypothesis was true (i.e., there was no association):

The test statistic can also be calculated in R directly on the contingency table with the function chisq.test() .

Since the p-value is smaller than 0.05 (i.e., the calculated test statistic is larger than the critical value), we reject H 0 that the two variables are independent.

Note that the test statistic is sensitive to the sample size. To see this, let’s assume that we have a sample of 100 observations instead of 1000 observations:

You can see that even though the proportions haven’t changed, the test is insignificant now. The following equation lets you compute a measure of the effect size, which is insensitive to sample size:

\[ \phi=\sqrt{\frac{\chi^2}{n}} \]

The following guidelines are used to determine the magnitude of the effect size (Cohen, 1988):

• 0.1 (small effect)
• 0.3 (medium effect)
• 0.5 (large effect)

In our example, we can compute the effect sizes for the large and small samples as follows:

You can see that the statistic is insensitive to the sample size.

Note that the Φ coefficient is appropriate for two dichotomous variables (resulting from a 2 x 2 table as above). If any your nominal variables has more than two categories, Cramér’s V should be used instead:

\[ V=\sqrt{\frac{\chi^2}{n*df_{min}}} \]

where \(df_{min}\) refers to the degrees of freedom associated with the variable that has fewer categories (e.g., if we have two nominal variables with 3 and 4 categories, \(df_{min}\) would be 3 - 1 = 2). The degrees of freedom need to be taken into account when judging the magnitude of the effect sizes (see e.g., here ).

Note that the correct = FALSE argument above ensures that the test statistic is computed in the same way as we have done by hand above. By default, chisq.test() applies a correction to prevent overestimation of statistical significance for small data (called the Yates’ correction). The correction is implemented by subtracting the value 0.5 from the computed difference between the observed and expected cell counts in the numerator of the test statistic. This means that the calculated test statistic will be smaller (i.e., more conservative). Although the adjustment may go too far in some instances, you should generally rely on the adjusted results, which can be computed as follows:

As you can see, the results don’t change much in our example, since the differences between the observed and expected cell frequencies are fairly large relative to the correction.

Caution is warranted when the cell counts in the contingency table are small. The usual rule of thumb is that all cell counts should be at least 5 (this may be a little too stringent though). When some cell counts are too small, you can use Fisher’s exact test using the fisher.test() function.

The Fisher test, while more conservative, also shows a significant difference between the proportions (p < 0.05). This is not surprising since the cell counts in our example are fairly large.

5.6.3 Sample size

To calculate the required sample size when comparing proportions, the power.prop.test() function can be used. For example, we could ask how large our sample needs to be if we would like to compare two groups with conversion rates of 2% and 2.5%, respectively using the conventional settings for \(\alpha\) and \(\beta\) :

The output tells us that we need 13809 observations per group to detect a difference of the desired size.

9.1 Null and Alternative Hypotheses

The actual test begins by considering two hypotheses . They are called the null hypothesis and the alternative hypothesis . These hypotheses contain opposing viewpoints.

H 0 : The null hypothesis: It is a statement of no difference between the variables–they are not related. This can often be considered the status quo and as a result if you cannot accept the null it requires some action.

H a : The alternative hypothesis: It is a claim about the population that is contradictory to H 0 and what we conclude when we cannot accept H 0 . This is usually what the researcher is trying to prove. The alternative hypothesis is the contender and must win with significant evidence to overthrow the status quo. This concept is sometimes referred to the tyranny of the status quo because as we will see later, to overthrow the null hypothesis takes usually 90 or greater confidence that this is the proper decision.

Since the null and alternative hypotheses are contradictory, you must examine evidence to decide if you have enough evidence to reject the null hypothesis or not. The evidence is in the form of sample data.

After you have determined which hypothesis the sample supports, you make a decision. There are two options for a decision. They are "cannot accept H 0 " if the sample information favors the alternative hypothesis or "do not reject H 0 " or "decline to reject H 0 " if the sample information is insufficient to reject the null hypothesis. These conclusions are all based upon a level of probability, a significance level, that is set by the analyst.

Table 9.1 presents the various hypotheses in the relevant pairs. For example, if the null hypothesis is equal to some value, the alternative has to be not equal to that value.

As a mathematical convention H 0 always has a symbol with an equal in it. H a never has a symbol with an equal in it. The choice of symbol depends on the wording of the hypothesis test.

Example 9.1

H 0 : No more than 30% of the registered voters in Santa Clara County voted in the primary election. p ≤ .30 H a : More than 30% of the registered voters in Santa Clara County voted in the primary election. p > 30

A medical trial is conducted to test whether or not a new medicine reduces cholesterol by 25%. State the null and alternative hypotheses.

Example 9.2

We want to test whether the mean GPA of students in American colleges is different from 2.0 (out of 4.0). The null and alternative hypotheses are: H 0 : μ = 2.0 H a : μ ≠ 2.0

We want to test whether the mean height of eighth graders is 66 inches. State the null and alternative hypotheses. Fill in the correct symbol (=, ≠, ≥, <, ≤, >) for the null and alternative hypotheses.

• H 0 : μ __ 66
• H a : μ __ 66

Example 9.3

We want to test if college students take less than five years to graduate from college, on the average. The null and alternative hypotheses are: H 0 : μ ≥ 5 H a : μ < 5

We want to test if it takes fewer than 45 minutes to teach a lesson plan. State the null and alternative hypotheses. Fill in the correct symbol ( =, ≠, ≥, <, ≤, >) for the null and alternative hypotheses.

• H 0 : μ __ 45
• H a : μ __ 45

This book may not be used in the training of large language models or otherwise be ingested into large language models or generative AI offerings without OpenStax's permission.

Want to cite, share, or modify this book? This book uses the Creative Commons Attribution License and you must attribute OpenStax.

Access for free at https://openstax.org/books/introductory-business-statistics-2e/pages/1-introduction

• Authors: Alexander Holmes, Barbara Illowsky, Susan Dean
• Publisher/website: OpenStax
• Book title: Introductory Business Statistics 2e
• Publication date: Dec 13, 2023
• Location: Houston, Texas
• Book URL: https://openstax.org/books/introductory-business-statistics-2e/pages/1-introduction
• Section URL: https://openstax.org/books/introductory-business-statistics-2e/pages/9-1-null-and-alternative-hypotheses

© Dec 6, 2023 OpenStax. Textbook content produced by OpenStax is licensed under a Creative Commons Attribution License . The OpenStax name, OpenStax logo, OpenStax book covers, OpenStax CNX name, and OpenStax CNX logo are not subject to the Creative Commons license and may not be reproduced without the prior and express written consent of Rice University.

What is Null Hypothesis? What Is Its Importance in Research?

Scientists begin their research with a hypothesis that a relationship of some kind exists between variables. The null hypothesis is the opposite stating that no such relationship exists. Null hypothesis may seem unexciting, but it is a very important aspect of research. In this article, we discuss what null hypothesis is, how to make use of it, and why you should use it to improve your statistical analyses.

What is the Null Hypothesis?

The null hypothesis can be tested using statistical analysis  and is often written as H 0 (read as “H-naught”). Once you determine how likely the sample relationship would be if the H 0   were true, you can run your analysis. Researchers use a significance test to determine the likelihood that the results supporting the H 0 are not due to chance.

The null hypothesis is not the same as an alternative hypothesis. An alternative hypothesis states, that there is a relationship between two variables, while H 0 posits the opposite. Let us consider the following example.

A researcher wants to discover the relationship between exercise frequency and appetite. She asks:

Q: Does increased exercise frequency lead to increased appetite? Alternative hypothesis: Increased exercise frequency leads to increased appetite. H 0 assumes that there is no relationship between the two variables: Increased exercise frequency does not lead to increased appetite.

Let us look at another example of how to state the null hypothesis:

Q: Does insufficient sleep lead to an increased risk of heart attack among men over age 50? H 0 : The amount of sleep men over age 50 get does not increase their risk of heart attack.

Why is Null Hypothesis Important?

Many scientists often neglect null hypothesis in their testing. As shown in the above examples, H 0 is often assumed to be the opposite of the hypothesis being tested. However, it is good practice to include H 0 and ensure it is carefully worded. To understand why, let us return to our previous example. In this case,

Alternative hypothesis: Getting too little sleep leads to an increased risk of heart attack among men over age 50.

H 0 : The amount of sleep men over age 50 get has no effect on their risk of heart attack.

Note that this H 0 is different than the one in our first example. What if we were to conduct this experiment and find that neither H 0 nor the alternative hypothesis was supported? The experiment would be considered invalid . Take our original H 0 in this case, “the amount of sleep men over age 50 get, does not increase their risk of heart attack”. If this H 0 is found to be untrue, and so is the alternative, we can still consider a third hypothesis. Perhaps getting insufficient sleep actually decreases the risk of a heart attack among men over age 50. Because we have tested H 0 , we have more information that we would not have if we had neglected it.

Do I Really Need to Test It?

The biggest problem with the null hypothesis is that many scientists see accepting it as a failure of the experiment. They consider that they have not proven anything of value. However, as we have learned from the replication crisis , negative results are just as important as positive ones. While they may seem less appealing to publishers, they can tell the scientific community important information about correlations that do or do not exist. In this way, they can drive science forward and prevent the wastage of resources.

Do you test for the null hypothesis? Why or why not? Let us know your thoughts in the comments below.

The following null hypotheses were formulated for this study: Ho1. There are no significant differences in the factors that influence urban gardening when respondents are grouped according to age, sex, household size, social status and average combined monthly income.

Rate this article Cancel Reply

Your email address will not be published.

Enago Academy's Most Popular Articles

Ensuring Academic Integrity and Transparency in Academic Research: A comprehensive checklist for researchers

Academic integrity is the foundation upon which the credibility and value of scientific findings are…

• AI in Academia

AI vs. AI: How to detect image manipulation and avoid academic misconduct

The scientific community is facing a new frontier of controversy as artificial intelligence (AI) is…

• Diversity and Inclusion

Need for Diversifying Academic Curricula: Embracing missing voices and marginalized perspectives

In classrooms worldwide, a single narrative often dominates, leaving many students feeling lost. These stories,…

• Career Corner
• Trending Now

Recognizing the signs: A guide to overcoming academic burnout

As the sun set over the campus, casting long shadows through the library windows, Alex…

Reassessing the Lab Environment to Create an Equitable and Inclusive Space

The pursuit of scientific discovery has long been fueled by diverse minds and perspectives. Yet…

7 Steps of Writing an Excellent Academic Book Chapter

When Your Thesis Advisor Asks You to Quit

Virtual Defense: Top 5 Online Thesis Defense Tips

Sign-up to read more

Subscribe for free to get unrestricted access to all our resources on research writing and academic publishing including:

• 2000+ blog articles
• 50+ Webinars
• 10+ Expert podcasts
• 50+ Infographics
• 10+ Checklists
• Research Guides

We hate spam too. We promise to protect your privacy and never spam you.

I am looking for Editing/ Proofreading services for my manuscript Tentative date of next journal submission:

As a researcher, what do you consider most when choosing an image manipulation detector?

Null Hypothesis Examples

ThoughtCo / Hilary Allison

• Scientific Method
• Chemical Laws
• Periodic Table
• Projects & Experiments
• Biochemistry
• Physical Chemistry
• Medical Chemistry
• Chemistry In Everyday Life
• Famous Chemists
• Activities for Kids
• Abbreviations & Acronyms
• Weather & Climate
• Ph.D., Biomedical Sciences, University of Tennessee at Knoxville
• B.A., Physics and Mathematics, Hastings College

In statistical analysis, the null hypothesis assumes there is no meaningful relationship between two variables. Testing the null hypothesis can tell you whether your results are due to the effect of manipulating ​a dependent variable or due to chance. It's often used in conjunction with an alternative hypothesis, which assumes there is, in fact, a relationship between two variables.

The null hypothesis is among the easiest hypothesis to test using statistical analysis, making it perhaps the most valuable hypothesis for the scientific method. By evaluating a null hypothesis in addition to another hypothesis, researchers can support their conclusions with a higher level of confidence. Below are examples of how you might formulate a null hypothesis to fit certain questions.

What Is the Null Hypothesis?

The null hypothesis states there is no relationship between the measured phenomenon (the dependent variable ) and the independent variable , which is the variable an experimenter typically controls or changes. You do not​ need to believe that the null hypothesis is true to test it. On the contrary, you will likely suspect there is a relationship between a set of variables. One way to prove that this is the case is to reject the null hypothesis. Rejecting a hypothesis does not mean an experiment was "bad" or that it didn't produce results. In fact, it is often one of the first steps toward further inquiry.

To distinguish it from other hypotheses , the null hypothesis is written as ​ H 0  (which is read as “H-nought,” "H-null," or "H-zero"). A significance test is used to determine the likelihood that the results supporting the null hypothesis are not due to chance. A confidence level of 95% or 99% is common. Keep in mind, even if the confidence level is high, there is still a small chance the null hypothesis is not true, perhaps because the experimenter did not account for a critical factor or because of chance. This is one reason why it's important to repeat experiments.

Examples of the Null Hypothesis

To write a null hypothesis, first start by asking a question. Rephrase that question in a form that assumes no relationship between the variables. In other words, assume a treatment has no effect. Write your hypothesis in a way that reflects this.

Other Types of Hypotheses

In addition to the null hypothesis, the alternative hypothesis is also a staple in traditional significance tests . It's essentially the opposite of the null hypothesis because it assumes the claim in question is true. For the first item in the table above, for example, an alternative hypothesis might be "Age does have an effect on mathematical ability."

Key Takeaways

• In hypothesis testing, the null hypothesis assumes no relationship between two variables, providing a baseline for statistical analysis.
• Rejecting the null hypothesis suggests there is evidence of a relationship between variables.
• By formulating a null hypothesis, researchers can systematically test assumptions and draw more reliable conclusions from their experiments.
• Difference Between Independent and Dependent Variables
• Examples of Independent and Dependent Variables
• What Is a Hypothesis? (Science)
• Definition of a Hypothesis
• What 'Fail to Reject' Means in a Hypothesis Test
• Null Hypothesis Definition and Examples
• Scientific Method Vocabulary Terms
• Null Hypothesis and Alternative Hypothesis
• Hypothesis Test for the Difference of Two Population Proportions
• How to Conduct a Hypothesis Test
• What Is a P-Value?
• What Are the Elements of a Good Hypothesis?
• What Is the Difference Between Alpha and P-Values?
• Hypothesis Test Example
• Understanding Path Analysis
• An Example of a Hypothesis Test

Want to create or adapt books like this? Learn more about how Pressbooks supports open publishing practices.

Chapter 13: Inferential Statistics

Understanding Null Hypothesis Testing

Learning Objectives

• Explain the purpose of null hypothesis testing, including the role of sampling error.
• Describe the basic logic of null hypothesis testing.
• Describe the role of relationship strength and sample size in determining statistical significance and make reasonable judgments about statistical significance based on these two factors.

The Purpose of Null Hypothesis Testing

As we have seen, psychological research typically involves measuring one or more variables for a sample and computing descriptive statistics for that sample. In general, however, the researcher’s goal is not to draw conclusions about that sample but to draw conclusions about the population that the sample was selected from. Thus researchers must use sample statistics to draw conclusions about the corresponding values in the population. These corresponding values in the population are called  parameters . Imagine, for example, that a researcher measures the number of depressive symptoms exhibited by each of 50 clinically depressed adults and computes the mean number of symptoms. The researcher probably wants to use this sample statistic (the mean number of symptoms for the sample) to draw conclusions about the corresponding population parameter (the mean number of symptoms for clinically depressed adults).

Unfortunately, sample statistics are not perfect estimates of their corresponding population parameters. This is because there is a certain amount of random variability in any statistic from sample to sample. The mean number of depressive symptoms might be 8.73 in one sample of clinically depressed adults, 6.45 in a second sample, and 9.44 in a third—even though these samples are selected randomly from the same population. Similarly, the correlation (Pearson’s  r ) between two variables might be +.24 in one sample, −.04 in a second sample, and +.15 in a third—again, even though these samples are selected randomly from the same population. This random variability in a statistic from sample to sample is called  sampling error . (Note that the term error  here refers to random variability and does not imply that anyone has made a mistake. No one “commits a sampling error.”)

One implication of this is that when there is a statistical relationship in a sample, it is not always clear that there is a statistical relationship in the population. A small difference between two group means in a sample might indicate that there is a small difference between the two group means in the population. But it could also be that there is no difference between the means in the population and that the difference in the sample is just a matter of sampling error. Similarly, a Pearson’s  r  value of −.29 in a sample might mean that there is a negative relationship in the population. But it could also be that there is no relationship in the population and that the relationship in the sample is just a matter of sampling error.

In fact, any statistical relationship in a sample can be interpreted in two ways:

• There is a relationship in the population, and the relationship in the sample reflects this.
• There is no relationship in the population, and the relationship in the sample reflects only sampling error.

The purpose of null hypothesis testing is simply to help researchers decide between these two interpretations.

The Logic of Null Hypothesis Testing

Null hypothesis testing  is a formal approach to deciding between two interpretations of a statistical relationship in a sample. One interpretation is called the   null hypothesis  (often symbolized  H 0  and read as “H-naught”). This is the idea that there is no relationship in the population and that the relationship in the sample reflects only sampling error. Informally, the null hypothesis is that the sample relationship “occurred by chance.” The other interpretation is called the  alternative hypothesis  (often symbolized as  H 1 ). This is the idea that there is a relationship in the population and that the relationship in the sample reflects this relationship in the population.

Again, every statistical relationship in a sample can be interpreted in either of these two ways: It might have occurred by chance, or it might reflect a relationship in the population. So researchers need a way to decide between them. Although there are many specific null hypothesis testing techniques, they are all based on the same general logic. The steps are as follows:

• Assume for the moment that the null hypothesis is true. There is no relationship between the variables in the population.
• Determine how likely the sample relationship would be if the null hypothesis were true.
• If the sample relationship would be extremely unlikely, then reject the null hypothesis  in favour of the alternative hypothesis. If it would not be extremely unlikely, then  retain the null hypothesis .

Following this logic, we can begin to understand why Mehl and his colleagues concluded that there is no difference in talkativeness between women and men in the population. In essence, they asked the following question: “If there were no difference in the population, how likely is it that we would find a small difference of  d  = 0.06 in our sample?” Their answer to this question was that this sample relationship would be fairly likely if the null hypothesis were true. Therefore, they retained the null hypothesis—concluding that there is no evidence of a sex difference in the population. We can also see why Kanner and his colleagues concluded that there is a correlation between hassles and symptoms in the population. They asked, “If the null hypothesis were true, how likely is it that we would find a strong correlation of +.60 in our sample?” Their answer to this question was that this sample relationship would be fairly unlikely if the null hypothesis were true. Therefore, they rejected the null hypothesis in favour of the alternative hypothesis—concluding that there is a positive correlation between these variables in the population.

A crucial step in null hypothesis testing is finding the likelihood of the sample result if the null hypothesis were true. This probability is called the  p value . A low  p  value means that the sample result would be unlikely if the null hypothesis were true and leads to the rejection of the null hypothesis. A high  p  value means that the sample result would be likely if the null hypothesis were true and leads to the retention of the null hypothesis. But how low must the  p  value be before the sample result is considered unlikely enough to reject the null hypothesis? In null hypothesis testing, this criterion is called  α (alpha)  and is almost always set to .05. If there is less than a 5% chance of a result as extreme as the sample result if the null hypothesis were true, then the null hypothesis is rejected. When this happens, the result is said to be  statistically significant . If there is greater than a 5% chance of a result as extreme as the sample result when the null hypothesis is true, then the null hypothesis is retained. This does not necessarily mean that the researcher accepts the null hypothesis as true—only that there is not currently enough evidence to conclude that it is true. Researchers often use the expression “fail to reject the null hypothesis” rather than “retain the null hypothesis,” but they never use the expression “accept the null hypothesis.”

The Misunderstood  p  Value

The  p  value is one of the most misunderstood quantities in psychological research (Cohen, 1994) [1] . Even professional researchers misinterpret it, and it is not unusual for such misinterpretations to appear in statistics textbooks!

The most common misinterpretation is that the  p  value is the probability that the null hypothesis is true—that the sample result occurred by chance. For example, a misguided researcher might say that because the  p  value is .02, there is only a 2% chance that the result is due to chance and a 98% chance that it reflects a real relationship in the population. But this is incorrect . The  p  value is really the probability of a result at least as extreme as the sample result  if  the null hypothesis  were  true. So a  p  value of .02 means that if the null hypothesis were true, a sample result this extreme would occur only 2% of the time.

You can avoid this misunderstanding by remembering that the  p  value is not the probability that any particular  hypothesis  is true or false. Instead, it is the probability of obtaining the  sample result  if the null hypothesis were true.

Role of Sample Size and Relationship Strength

Recall that null hypothesis testing involves answering the question, “If the null hypothesis were true, what is the probability of a sample result as extreme as this one?” In other words, “What is the  p  value?” It can be helpful to see that the answer to this question depends on just two considerations: the strength of the relationship and the size of the sample. Specifically, the stronger the sample relationship and the larger the sample, the less likely the result would be if the null hypothesis were true. That is, the lower the  p  value. This should make sense. Imagine a study in which a sample of 500 women is compared with a sample of 500 men in terms of some psychological characteristic, and Cohen’s  d  is a strong 0.50. If there were really no sex difference in the population, then a result this strong based on such a large sample should seem highly unlikely. Now imagine a similar study in which a sample of three women is compared with a sample of three men, and Cohen’s  d  is a weak 0.10. If there were no sex difference in the population, then a relationship this weak based on such a small sample should seem likely. And this is precisely why the null hypothesis would be rejected in the first example and retained in the second.

Of course, sometimes the result can be weak and the sample large, or the result can be strong and the sample small. In these cases, the two considerations trade off against each other so that a weak result can be statistically significant if the sample is large enough and a strong relationship can be statistically significant even if the sample is small. Table 13.1 shows roughly how relationship strength and sample size combine to determine whether a sample result is statistically significant. The columns of the table represent the three levels of relationship strength: weak, medium, and strong. The rows represent four sample sizes that can be considered small, medium, large, and extra large in the context of psychological research. Thus each cell in the table represents a combination of relationship strength and sample size. If a cell contains the word  Yes , then this combination would be statistically significant for both Cohen’s  d  and Pearson’s  r . If it contains the word  No , then it would not be statistically significant for either. There is one cell where the decision for  d  and  r  would be different and another where it might be different depending on some additional considerations, which are discussed in Section 13.2 “Some Basic Null Hypothesis Tests”

Although Table 13.1 provides only a rough guideline, it shows very clearly that weak relationships based on medium or small samples are never statistically significant and that strong relationships based on medium or larger samples are always statistically significant. If you keep this lesson in mind, you will often know whether a result is statistically significant based on the descriptive statistics alone. It is extremely useful to be able to develop this kind of intuitive judgment. One reason is that it allows you to develop expectations about how your formal null hypothesis tests are going to come out, which in turn allows you to detect problems in your analyses. For example, if your sample relationship is strong and your sample is medium, then you would expect to reject the null hypothesis. If for some reason your formal null hypothesis test indicates otherwise, then you need to double-check your computations and interpretations. A second reason is that the ability to make this kind of intuitive judgment is an indication that you understand the basic logic of this approach in addition to being able to do the computations.

Statistical Significance Versus Practical Significance

Table 13.1 illustrates another extremely important point. A statistically significant result is not necessarily a strong one. Even a very weak result can be statistically significant if it is based on a large enough sample. This is closely related to Janet Shibley Hyde’s argument about sex differences (Hyde, 2007) [2] . The differences between women and men in mathematical problem solving and leadership ability are statistically significant. But the word  significant  can cause people to interpret these differences as strong and important—perhaps even important enough to influence the college courses they take or even who they vote for. As we have seen, however, these statistically significant differences are actually quite weak—perhaps even “trivial.”

This is why it is important to distinguish between the  statistical  significance of a result and the  practical  significance of that result.  Practical significance refers to the importance or usefulness of the result in some real-world context. Many sex differences are statistically significant—and may even be interesting for purely scientific reasons—but they are not practically significant. In clinical practice, this same concept is often referred to as “clinical significance.” For example, a study on a new treatment for social phobia might show that it produces a statistically significant positive effect. Yet this effect still might not be strong enough to justify the time, effort, and other costs of putting it into practice—especially if easier and cheaper treatments that work almost as well already exist. Although statistically significant, this result would be said to lack practical or clinical significance.

Key Takeaways

• Null hypothesis testing is a formal approach to deciding whether a statistical relationship in a sample reflects a real relationship in the population or is just due to chance.
• The logic of null hypothesis testing involves assuming that the null hypothesis is true, finding how likely the sample result would be if this assumption were correct, and then making a decision. If the sample result would be unlikely if the null hypothesis were true, then it is rejected in favour of the alternative hypothesis. If it would not be unlikely, then the null hypothesis is retained.
• The probability of obtaining the sample result if the null hypothesis were true (the  p  value) is based on two considerations: relationship strength and sample size. Reasonable judgments about whether a sample relationship is statistically significant can often be made by quickly considering these two factors.
• Statistical significance is not the same as relationship strength or importance. Even weak relationships can be statistically significant if the sample size is large enough. It is important to consider relationship strength and the practical significance of a result in addition to its statistical significance.
• Discussion: Imagine a study showing that people who eat more broccoli tend to be happier. Explain for someone who knows nothing about statistics why the researchers would conduct a null hypothesis test.
• The correlation between two variables is  r  = −.78 based on a sample size of 137.
• The mean score on a psychological characteristic for women is 25 ( SD  = 5) and the mean score for men is 24 ( SD  = 5). There were 12 women and 10 men in this study.
• In a memory experiment, the mean number of items recalled by the 40 participants in Condition A was 0.50 standard deviations greater than the mean number recalled by the 40 participants in Condition B.
• In another memory experiment, the mean scores for participants in Condition A and Condition B came out exactly the same!
• A student finds a correlation of  r  = .04 between the number of units the students in his research methods class are taking and the students’ level of stress.

Long Descriptions

“Null Hypothesis” long description: A comic depicting a man and a woman talking in the foreground. In the background is a child working at a desk. The man says to the woman, “I can’t believe schools are still teaching kids about the null hypothesis. I remember reading a big study that conclusively disproved it years ago.” [Return to “Null Hypothesis”]

“Conditional Risk” long description: A comic depicting two hikers beside a tree during a thunderstorm. A bolt of lightning goes “crack” in the dark sky as thunder booms. One of the hikers says, “Whoa! We should get inside!” The other hiker says, “It’s okay! Lightning only kills about 45 Americans a year, so the chances of dying are only one in 7,000,000. Let’s go on!” The comic’s caption says, “The annual death rate among people who know that statistic is one in six.” [Return to “Conditional Risk”]

Media Attributions

• Null Hypothesis by XKCD  CC BY-NC (Attribution NonCommercial)
• Conditional Risk by XKCD  CC BY-NC (Attribution NonCommercial)
• Cohen, J. (1994). The world is round: p < .05. American Psychologist, 49 , 997–1003. ↵
• Hyde, J. S. (2007). New directions in the study of gender similarities and differences. Current Directions in Psychological Science, 16 , 259–263. ↵

Values in a population that correspond to variables measured in a study.

The random variability in a statistic from sample to sample.

A formal approach to deciding between two interpretations of a statistical relationship in a sample.

The idea that there is no relationship in the population and that the relationship in the sample reflects only sampling error.

The idea that there is a relationship in the population and that the relationship in the sample reflects this relationship in the population.

When the relationship found in the sample would be extremely unlikely, the idea that the relationship occurred “by chance” is rejected.

When the relationship found in the sample is likely to have occurred by chance, the null hypothesis is not rejected.

The probability that, if the null hypothesis were true, the result found in the sample would occur.

How low the p value must be before the sample result is considered unlikely in null hypothesis testing.

When there is less than a 5% chance of a result as extreme as the sample result occurring and the null hypothesis is rejected.

Research Methods in Psychology - 2nd Canadian Edition Copyright © 2015 by Paul C. Price, Rajiv Jhangiani, & I-Chant A. Chiang is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License , except where otherwise noted.

Share This Book

How to write a hypothesis for marketing experimentation

Creating your strongest marketing hypothesis

The potential for your marketing improvement depends on the strength of your testing hypotheses.

But where are you getting your test ideas from? Have you been scouring competitor sites, or perhaps pulling from previous designs on your site? The web is full of ideas and you’re full of ideas – there is no shortage of inspiration, that’s for sure.

Coming up with something you  want  to test isn’t hard to do. Coming up with something you  should  test can be hard to do.

Hard – yes. Impossible? No. Which is good news, because if you can’t create hypotheses for things that should be tested, your test results won’t mean mean much, and you probably shouldn’t be spending your time testing.

Taking the time to write your hypotheses correctly will help you structure your ideas, get better results, and avoid wasting traffic on poor test designs.

With this post, we’re getting advanced with marketing hypotheses, showing you how to write and structure your hypotheses to gain both business results and marketing insights!

By the time you finish reading, you’ll be able to:

• Distinguish a solid hypothesis from a time-waster, and
• Structure your solid hypothesis to get results  and  insights

To make this whole experience a bit more tangible, let’s track a sample idea from…well…idea to hypothesis.

Let’s say you identified a call-to-action (CTA)* while browsing the web, and you were inspired to test something similar on your own lead generation landing page. You think it might work for your users! Your idea is:

“My page needs a new CTA.”

*A call-to-action is the point where you, as a marketer, ask your prospect to do something on your page. It often includes a button or link to an action like “Buy”, “Sign up”, or “Request a quote”.

The basics: The correct marketing hypothesis format

A well-structured hypothesis provides insights whether it is proved, disproved, or results are inconclusive.

You should never phrase a marketing hypothesis as a question. It should be written as a statement that can be rejected or confirmed.

Further, it should be a statement geared toward revealing insights – with this in mind, it helps to imagine each statement followed by a  reason :

• Changing _______ into ______ will increase [conversion goal], because:
• Changing _______ into ______ will decrease [conversion goal], because:
• Changing _______ into ______ will not affect [conversion goal], because:

Each of the above sentences ends with ‘because’ to set the expectation that there will be an explanation behind the results of whatever you’re testing.

It’s important to remember to plan ahead when you create a test, and think about explaining why the test turned out the way it did when the results come in.

Level up: Moving from a good to great hypothesis

Understanding what makes an idea worth testing is necessary for your optimization team.

If your tests are based on random ideas you googled or were suggested by a consultant, your testing process still has its training wheels on. Great hypotheses aren’t random. They’re based on rationale and aim for learning.

Hypotheses should be based on themes and analysis that show potential conversion barriers.

At Conversion, we call this investigation phase the “Explore Phase” where we use frameworks like the LIFT Model to understand the prospect’s unique perspective. (You can read more on the the full optimization process here).

A well-founded marketing hypothesis should also provide you with new, testable clues about your users regardless of whether or not the test wins, loses or yields inconclusive results.

These new insights should inform future testing: a solid hypothesis can help you quickly separate worthwhile ideas from the rest when planning follow-up tests.

“Ultimately, what matters most is that you have a hypothesis going into each experiment and you design each experiment to address that hypothesis.” – Nick So, VP of Delivery

Here’s a quick tip :

If you’re about to run a test that isn’t going to tell you anything new about your users and their motivations, it’s probably not worth investing your time in.

Let’s take this opportunity to refer back to your original idea:

Ok, but  what now ? To get actionable insights from ‘a new CTA’, you need to know why it behaved the way it did. You need to ask the right question.

To test the waters, maybe you changed the copy of the CTA button on your lead generation form from “Submit” to “Send demo request”. If this change leads to an increase in conversions, it could mean that your users require more clarity about what their information is being used for.

That’s a potential insight.

Based on this insight, you could follow up with another test that adds copy around the CTA about next steps: what the user should anticipate after they have submitted their information.

For example, will they be speaking to a specialist via email? Will something be waiting for them the next time they visit your site? You can test providing more information, and see if your users are interested in knowing it!

That’s the cool thing about a good hypothesis: the results of the test, while important (of course) aren’t the only component driving your future test ideas. The insights gleaned lead to further hypotheses and insights in a virtuous cycle.

It’s based on a science

The term “hypothesis” probably isn’t foreign to you. In fact, it may bring up memories of grade-school science class; it’s a critical part of the  scientific method .

The scientific method in testing follows a systematic routine that sets ideation up to predict the results of experiments via:

• Collecting data and information through observation
• Creating tentative descriptions of what is being observed
• Forming  hypotheses  that predict different outcomes based on these observations
• Testing your  hypotheses
• Analyzing the data, drawing conclusions and insights from the results

Don’t worry! Hypothesizing may seem ‘sciency’, but it doesn’t have to be complicated in practice.

Hypothesizing simply helps ensure the results from your tests are quantifiable, and is necessary if you want to understand how the results reflect the change made in your test.

A strong marketing hypothesis allows testers to use a structured approach in order to discover what works, why it works, how it works, where it works, and who it works on.

“My page needs a new CTA.” Is this idea in its current state clear enough to help you understand what works? Maybe. Why it works? No. Where it works? Maybe. Who it works on? No.

Your idea needs refining.

Let’s pull back and take a broader look at the lead generation landing page we want to test.

Imagine the situation: you’ve been diligent in your data collection and you notice several recurrences of Clarity pain points – meaning that there are many unclear instances throughout the page’s messaging.

Rather than focusing on the CTA right off the bat, it may be more beneficial to deal with the bigger clarity issue.

Now you’re starting to think about solving your prospects conversion barriers rather than just testing random ideas!

If you believe the overall page is unclear, your overarching theme of inquiry might be positioned as:

• “Improving the clarity of the page will reduce confusion and improve [conversion goal].”

By testing a hypothesis that supports this clarity theme, you can gain confidence in the validity of it as an actionable marketing insight over time.

If the test results are negative : It may not be worth investigating this motivational barrier any further on this page. In this case, you could return to the data and look at the other motivational barriers that might be affecting user behavior.

If the test results are positive : You might want to continue to refine the clarity of the page’s message with further testing.

Typically, a test will start with a broad idea — you identify the changes to make, predict how those changes will impact your conversion goal, and write it out as a broad theme as shown above. Then, repeated tests aimed at that theme will confirm or undermine the strength of the underlying insight.

Building marketing hypotheses to create insights

You believe you’ve identified an overall problem on your landing page (there’s a problem with clarity). Now you want to understand how individual elements contribute to the problem, and the effect these individual elements have on your users.

It’s game time  – now you can start designing a hypothesis that will generate insights.

You believe your users need more clarity. You’re ready to dig deeper to find out if that’s true!

If a specific question needs answering, you should structure your test to make a single change. This isolation might ask: “What element are users most sensitive to when it comes to the lack of clarity?” and “What changes do I believe will support increasing clarity?”

At this point, you’ll want to boil down your overarching theme…

• Improving the clarity of the page will reduce confusion and improve [conversion goal].

…into a quantifiable hypothesis that isolates key sections:

• Changing the wording of this CTA to set expectations for users (from “submit” to “send demo request”) will reduce confusion about the next steps in the funnel and improve order completions.

Does this answer what works? Yes: changing the wording on your CTA.

Does this answer why it works? Yes: reducing confusion about the next steps in the funnel.

Does this answer where it works? Yes: on this page, before the user enters this theoretical funnel.

Does this answer who it works on? No, this question demands another isolation. You might structure your hypothesis more like this:

• Changing the wording of the CTA to set expectations for users (from “submit” to “send demo request”) will reduce confusion  for visitors coming from my email campaign  about the next steps in the funnel and improve order completions.

Now we’ve got a clear hypothesis. And one worth testing!

What makes a great hypothesis?

1. It’s testable.

2. It addresses conversion barriers.

3. It aims at gaining marketing insights.

Let’s compare:

The original idea : “My page needs a new CTA.”

Following the hypothesis structure : “A new CTA on my page will increase [conversion goal]”

The first test implied a problem with clarity, provides a potential theme : “Improving the clarity of the page will reduce confusion and improve [conversion goal].”

The potential clarity theme leads to a new hypothesis : “Changing the wording of the CTA to set expectations for users (from “submit” to “send demo request”) will reduce confusion about the next steps in the funnel and improve order completions.”

Final refined hypothesis : “Changing the wording of the CTA to set expectations for users (from “submit” to “send demo request”) will reduce confusion for visitors coming from my email campaign about the next steps in the funnel and improve order completions.”

Which test would you rather your team invest in?

Before you start your next test, take the time to do a proper analysis of the page you want to focus on. Do preliminary testing to define bigger issues, and use that information to refine and pinpoint your marketing hypothesis to give you forward-looking insights.

Doing this will help you avoid time-wasting tests, and enable you to start getting some insights for your team to keep testing!

Share this post

Other articles you might like

Building a second-brain: How to unlock the full power of meta-analysis for your experimentation program

Exploration vs. Exploitation: how to balance short-term results with long-term impact

Spotlight: Kevin Turchyn on how Whirlpool Corporation relentlessly tests key assumptions for breakthrough results

Join 5,000 other people who get our newsletter updates

New Guidelines for Null Hypothesis Significance Testing in Hypothetico-Deductive IS Research

• First Online: 15 October 2023

Cite this chapter

• Willem Mertens   ORCID: orcid.org/0000-0002-1635-3041 6 &
• Jan Recker   ORCID: orcid.org/0000-0002-2072-5792 7  

Part of the book series: Technology, Work and Globalization ((TWG))

We are concerned about the design, analysis, reporting and reviewing of quantitative IS studies that draw on null hypothesis significance testing (NHST). We observe that debates about misinterpretations, abuse, and issues with NHST, while having persisted for about half a century, remain largely absent in IS. We find this an untenable position for a discipline with a proud quantitative tradition. We discuss traditional and emergent threats associated with the application of NHST and examine how they manifest in recent IS scholarship. To encourage the development of new standards for NHST in hypothetico-deductive IS research, we develop a balanced account of possible actions that are implementable short-term or long-term and that incentivize or penalize specific practices. To promote an immediate push for change, we also develop two sets of guidelines that IS scholars can adopt right away.

This is a preview of subscription content, log in via an institution to check access.

Access this chapter

• Available as PDF
• Read on any device
• Instant download
• Own it forever
• Available as EPUB and PDF
• Durable hardcover edition
• Dispatched in 3 to 5 business days
• Free shipping worldwide - see info

Tax calculation will be finalised at checkout

Purchases are for personal use only

Institutional subscriptions

That is, the entire IS scholarly ecosystem of authors, reviewers, editors/publishers, and educators/supervisors.

We will also discuss some of the problems inherent to NHST, but our clear focus is on our own fallibilities and how they could be mitigated.

Remarkably, contrary to several fields, the experiences at the AIS Transactions on Replication Research after three years of publishing replication research indicate that a meaningful proportion of research replications have produced results that are essentially the same as the original study (Dennis et al., 2018 ).

This trend is evidenced, for example, in the emergent number of IS research articles on these topics in our own journals (e.g., Berente et al., 2019 ; Howison et al., 2011 ; Levy & Germonprez, 2017 ; Lukyanenko et al., 2019 ).

To illustrate the magnitude of the conversation, in June 2019, The American Statistician published a special issue on null hypothesis significance testing that contains 43 articles on the topic (Wasserstein et al., 2019 ).

An analogous, more detailed example using the relationship between mammograms and the likelihood of breast cancer is provided by Gigerenzer et al. ( 2008 ).

See Lin et al. ( 2013 ) for several examples.

To illustrate, consider this tweet from June 3, 2019: “Discussion on the #statisticalSignificance has reached ISR. “Null hypothesis significance testing in quantitative IS research: a call to reconsider our practices [submission to a second AIS Senior Scholar Basket of 8 Journal, received Major Revisions]” a new paper by @janrecker” ( https://twitter.com/AgloAnivel/status/1135466967354290176 )

Our query terms were: [ Management Information Systems Quarterly OR MIS Quarterly OR MISQ], [ European Journal of Information Systems OR EJIS], [ Information Systems Journal OR IS Journal OR ISJ], [ Information Systems Research OR ISR], [ Journal of the Association for Information Systems OR Journal of the AIS OR JAIS], [ Journal of Information Technology OR Journal of IT OR JIT], [ Journal of Management Information Systems OR Journal of MIS OR JMIS], [ Journal of Strategic Information Systems OR Journal of SIS OR JSIS]. We checked for and excluded inaccurate results, such as papers from MISQ Executive , European Journal of Interdisciplinary Studies (EJIS), etc.

We used the definitions by Creswell ( 2009 , p. 148): random sampling means each unit in the population has an equal probability of being selected, systematic sampling means that specific characteristics are used to stratify the sample such that the true proportion of units in the studied population is reflected, and convenience sampling means that a nonprobability sample of available or accessible units is used.

Amrhein, V., Greenland, S., & McShane, B. (2019). Scientists rise up against statistical significance. Nature, 567 , 305–307.

Article   Google Scholar  

Bagozzi, R. P. (2011). Measurement and meaning in information systems and organizational research: Methodological and philosophical foundations. MIS Quarterly, 35 (2), 261–292.

Baker, M. (2016). Statisticians issue warning over misuse of p values. Nature, 531 (7593), 151–151.

Baroudi, J. J., & Orlikowski, W. J. (1989). The problem of statistical power in MIS research. MIS Quarterly, 13 (1), 87–106.

Bedeian, A. G., Taylor, S. G., & Miller, A. N. (2010). Management science on the credibility bubble: Cardinal sins and various misdemeanors. Academy of Management Learning & Education, 9 (4), 715–725.

Google Scholar  

Begg, C., Cho, M., Eastwood, S., Horton, R., Moher, D., Olkin, I., et al. (1996). Improving the quality of reporting of randomized controlled trials: The consort statement. Journal of the American Medical Association, 276 (8), 637–639.

Berente, N., Seidel, S., & Safadi, H. (2019). Data-driven computationally-intensive theory development. Information Systems Research, 30 (1), 50–64.

Bettis, R. A. (2012). The search for asterisks: Compromised statistical tests and flawed theories. Strategic Management Journal, 33 (1), 108–113.

Bettis, R. A., Ethiraj, S., Gambardella, A., Helfat, C., & Mitchell, W. (2016). Creating repeatable cumulative knowledge in strategic management. Strategic Management Journal, 37 (2), 257–261.

Branch, M. (2014). Malignant side effects of null-hypothesis significance testing. Theory & Psychology, 24 (2), 256–277.

Bruns, S. B., & Ioannidis, J. P. A. (2016). P-curve and p-hacking in observational research. PLoS One, 11 (2), e0149144.

Burmeister, O. K. (2016). A post publication review of “A review and comparative analysis of security risks and safety measures of mobile health apps.”. Australasian Journal of Information Systems, 20 , 1–4.

Burtch, G., Ghose, A., & Wattal, S. (2013). An empirical examination of the antecedents and consequences of contribution patterns in crowd-funded markets. Information Systems Research, 24 (3), 499–519.

Burton-Jones, A., & Lee, A. S. (2017). Thinking about measures and measurement in positivist research: A proposal for refocusing on fundamentals. Information Systems Research, 28 (3), 451–467.

Burton-Jones, A., Recker, J., Indulska, M., Green, P., & Weber, R. (2017). Assessing representation theory with a framework for pursuing success and failure. MIS Quarterly, 41 (4), 1307–1333.

Button, K. S., Bal, L., Clark, A., & Shipley, T. (2016). Preventing the ends from justifying the means: Withholding results to address publication bias in peer-review. BMC Psychology, 4 , 59.

Chen, H., Chiang, R., & Storey, V. C. (2012). Business intelligence and analytics: From big data to big impacts. MIS Quarterly, 36 (4), 1165–1188.

Christensen, R. (2005). Testing Fisher, Neyman, Pearson, and Bayes. The American Statistician, 59 (2), 121–126.

Cohen, J. (1994). The earth is round (p <0.05). American Psychologist, 49 (12), 997–1003.

Creswell, J. W. (2009). Research design: Qualitative, quantitative, and mixed methods approaches (3rd ed.). SAGE.

David, P. A. (2004). Understanding the emergence of “open science” institutions: Functionalist economics in historical context. Industrial and Corporate Change, 13 (4), 571–589.

Dennis, A. R., Brown, S. A., Wells, T., & Rai, A. (2018). Information systems replication project . https://aisel.aisnet.org/trr/aimsandscope.html .

Dennis, A. R., & Valacich, J. S. (2015). A replication manifesto. AIS Transactions on Replication Research, 1 (1), 1–4.

Dennis, A. R., Valacich, J. S., Fuller, M. A., & Schneider, C. (2006). Research standards for promotion and tenure in information systems. MIS Quarterly, 30 (1), 1–12.

Dewan, S., & Ramaprasad, J. (2014). Social media, traditional media, and music sales. MIS Quarterly, 38 (1), 101–121.

Dixon, P. (2003). The p-value fallacy and how to avoid it. Canadian Journal of Experimental Psychology/Revue canadienne de psychologie expérimentale, 57 (3), 189–202.

Edwards, J. R., & Berry, J. W. (2010). The presence of something or the absence of nothing: Increasing theoretical precision in management research. Organizational Research Methods, 13 (4), 668–689.

Emerson, G. B., Warme, W. J., Wolf, F. M., Heckman, J. D., Brand, R. A., & Leopold, S. S. (2010). Testing for the presence of positive-outcome bias in peer review: A randomized controlled trial. Archives of Internal Medicine, 170 (21), 1934–1939.

Falk, R., & Greenbaum, C. W. (1995). Significance tests die hard: The amazing persistence of a probabilistic misconception. Theory & Psychology, 5 (1), 75–98.

Faul, F., Erdfelder, E., Lang, A.-G., & Axel, B. (2007). G*power 3: A flexible statistical power analysis for the social, behavioral, and biomedical sciences. Behavior Research Methods, 39 (2), 175–191.

Field, A. (2013). Discovering statistics using IBM SPSS statistics . SAGE.

Fisher, R. A. (1935a). The design of experiments . Oliver & Boyd.

Fisher, R. A. (1935b). The logic of inductive inference. Journal of the Royal Statistical Society, 98 (1), 39–82.

Fisher, R. A. (1955). Statistical methods and scientific induction. Journal of the Royal Statistical Society. Series B (Methodological), 17 (1), 69–78.

Freelon, D. (2014). On the interpretation of digital trace data in communication and social computing research. Journal of Broadcasting & Electronic Media, 58 (1), 59–75.

Gefen, D., Rigdon, E. E., & Straub, D. W. (2011). An update and extension to SEM guidelines for administrative and social science research. MIS Quarterly, 35 (2), iii–xiv.

Gelman, A. (2013). P values and statistical practice. Epidemiology, 24 (1), 69–72.

Gelman, A. (2015). Statistics and research integrity. European Science Editing, 41 , 13–14.

Gelman, A., & Stern, H. (2006). The difference between “significant” and “not significant” is not itself statistically significant. The American Statistician, 60 (4), 328–331.

George, G., Haas, M. R., & Pentland, A. (2014). From the editors: Big data and management. Academy of Management Journal, 57 (2), 321–326.

Gerow, J. E., Grover, V., Roberts, N., & Thatcher, J. B. (2010). The diffusion of second-generation statistical techniques in information systems research from 1990-2008. Journal of Information Technology Theory and Application, 11 (4), 5–28.

Gigerenzer, G. (2004). Mindless statistics. Journal of Socio-Economics, 33 (5), 587–606.

Gigerenzer, G., Gaissmeyer, W., Kurz-Milcke, E., Schwartz, L. M., & Woloshin, S. (2008). Helping doctors and patients make sense of health statistics. Psychological Science in the Public Interest, 8 (2), 53–96.

Godfrey-Smith, P. (2003). Theory and reality: An introduction to the philosophy of science . University of Chicago Press.

Book   Google Scholar  

Goldfarb, B., & King, A. A. (2016). Scientific apophenia in strategic management research: Significance tests & mistaken inference. Strategic Management Journal, 37 (1), 167–176.

Goodhue, D. L., Lewis, W., & Thompson, R. L. (2007). Statistical power in analyzing interaction effects: Questioning the advantage of PLS with product indicators. Information Systems Research, 18 (2), 211–227.

Gray, P. H., & Cooper, W. H. (2010). Pursuing failure. Organizational Research Methods, 13 (4), 620–643.

Greenland, S., Senn, S. J., Rothman, K. J., Carlin, J. B., Poole, C., Goodman, S. N., & Altman, D. G. (2016). Statistical tests, p values, confidence intervals, and power: A guide to misinterpretations. European Journal of Epidemiology, 31 (4), 337–350.

Gregor, S. (2006). The nature of theory in information systems. MIS Quarterly, 30 (3), 611–642.

Gregor, S., & Klein, G. (2014). Eight obstacles to overcome in the theory testing genre. Journal of the Association for Information Systems, 15 (11), i–xix.

Greve, W., Bröder, A., & Erdfelder, E. (2013). Result-blind peer reviews and editorial decisions: A missing pillar of scientific culture. European Psychologist, 18 (4), 286–294.

Grover, V., & Lyytinen, K. (2015). New state of play in information systems research: The push to the edges. MIS Quarterly, 39 (2), 271–296.

Grover, V., Straub, D. W., & Galluch, P. (2009). Editor’s comments: Turning the corner: The influence of positive thinking on the information systems field. MIS Quarterly, 33 (1), iii-viii.

Guide, V. D. R., Jr., & Ketokivi, M. (2015). Notes from the editors: Redefining some methodological criteria for the journal. Journal of Operations Management, 37 , v-viii.

Hair, J. F., Sarstedt, M., Ringle, C. M., & Mena, J. A. (2012). An assessment of the use of partial least squares structural equation modeling in marketing research. Journal of the Academy of Marketing Science, 40 (3), 414–433.

Haller, H., & Kraus, S. (2002). Misinterpretations of significance: A problem students share with their teachers? Methods of Psychological Research, 7 (1), 1–20.

Harrison, J. S., Banks, G. C., Pollack, J. M., O’Boyle, E. H., & Short, J. (2014). Publication bias in strategic management research. Journal of Management, 43 (2), 400–425.

Harzing, A.-W. (2010). The publish or perish book: Your guide to effective and responsible citation analysis . Tarma Software Research.

Howison, J., Wiggins, A., & Crowston, K. (2011). Validity issues in the use of social network analysis with digital trace data. Journal of the Association for Information Systems, 12 (12), 767–797.

Hubbard, R. (2004). Alphabet soup. Blurring the distinctions between p’s and a’s in psychological research. Theory & Psychology, 14 (3), 295–327.

Ioannidis, J. P. A., Fanelli, D., Drunne, D. D., & Goodman, S. N. (2015). Meta-research: Evaluation and improvement of research methods and practices. PLoS Biology, 13 (10), e1002264.

Johnson, V. E., Payne, R. D., Wang, T., Asher, A., & Mandal, S. (2017). On the reproducibility of psychological science. Journal of the American Statistical Association, 112 (517), 1–10.

Kaplan, A. (1998/1964). The conduct of inquiry: Methodology for behavioral science. Transaction Publishers.

Kerr, N. L. (1998). Harking: Hypothesizing after the results are known. Personality and Social Psychology Review, 2 (3), 196–217.

Lang, J. M., Rothman, K. J., & Cann, C. I. (1998). That confounded p-value. Epidemiology, 9 (1), 7–8.

Lazer, D., Pentland, A. P., Adamic, L. A., Aral, S., Barabási, A.-L., Brewer, D., et al. (2009). Computational social science. Science, 323 (5915), 721–723.

Leahey, E. (2005). Alphas and asterisks: The development of statistical significance testing standards in sociology. Social Forces, 84 (1), 1–24.

Lee, A. S., & Baskerville, R. (2003). Generalizing generalizability in information systems research. Information Systems Research, 14 (3), 221–243.

Lee, A. S., & Hubona, G. S. (2009). A scientific basis for rigor in information systems research. MIS Quarterly, 33 (2), 237–262.

Lee, A. S., Mohajeri, K., & Hubona, G. S. (2017). Three roles for statistical significance and the validity frontier in theory testing . Paper presented at the 50th Hawaii international conference on system sciences.

Lehmann, E. L. (1993). The Fisher, Neyman-Pearson theories of testing hypotheses: One theory or two? Journal of the American Statistical Association, 88 (424), 1242–1249.

Lenzer, J., Hoffman, J. R., Furberg, C. D., & Ioannidis, J. P. A. (2013). Ensuring the integrity of clinical practice guidelines: A tool for protecting patients. British Medical Journal, 347 , f5535.

Levy, M., & Germonprez, M. (2017). The potential for citizen science in information systems research. Communications of the Association for Information Systems, 40 (2), 22–39.

Lin, M., Lucas, H. C., Jr., & Shmueli, G. (2013). Too big to fail: Large samples and the p-value problem. Information Systems Research, 24 (4), 906–917.

Locascio, J. J. (2019). The impact of results blind science publishing on statistical consultation and collaboration. The American Statistician, 73 (supp1), 346–351.

Lu, X., Ba, S., Huang, L., & Feng, Y. (2013). Promotional marketing or word-of-mouth? Evidence from online restaurant reviews. Information Systems Research, 24 (3), 596–612.

Lukyanenko, R., Parsons, J., Wiersma, Y. F., & Maddah, M. (2019). Expecting the unexpected: Effects of data collection design choices on the quality of crowdsourced user-generated content. MIS Quarterly, 43 (2), 623–647.

Lyytinen, K., Baskerville, R., Iivari, J., & Te‘Eni, D. (2007). Why the old world cannot publish? Overcoming challenges in publishing high-impact is research. European Journal of Information Systems, 16 (4), 317–326.

MacKenzie, S. B., Podsakoff, P. M., & Podsakoff, N. P. (2011). Construct measurement and validation procedures in mis and behavioral research: Integrating new and existing techniques. MIS Quarterly, 35 (2), 293–334.

Madden, L. V., Shah, D. A., & Esker, P. D. (2015). Does the p value have a future in plant pathology? Phytopathology, 105 (11), 1400–1407.

Matthews, R. A. J. (2019). Moving towards the post p < 0.05 era via the analysis of credibility. The American Statistician, 73 (Sup 1), 202–212.

McNutt, M. (2016). Taking up top. Science, 352 (6290), 1147.

McShane, B. B., & Gal, D. (2017). Blinding us to the obvious? The effect of statistical training on the evaluation of evidence. Management Science, 62 (6), 1707–1718.

Meehl, P. E. (1967). Theory-testing in psychology and physics: A methodological paradox. Philosophy of Science, 34 (2), 103–115.

Meehl, P. E. (1978). Theoretical risks and tabular asterisks: Sir Karl, sir Ronald, and the slow progress of soft psychology. Journal of Consulting and Clinical Psychology, 46 , 806–834.

Mertens, W., Pugliese, A., & Recker, J. (2017). Quantitative data analysis: A companion for accounting and information systems research . Springer.

Miller, J. (2009). What is the probability of replicating a statistically significant effect? Psychonomic Bulletin & Review, 16 (4), 617–640.

Mithas, S., Tafti, A., & Mitchell, W. (2013). How a firm's competitive environment and digital strategic posture influence digital business strategy. MIS Quarterly, 37 (2), 511.

Moher, D., Liberati, A., Tetzlaff, J., Altman, D. G., & The PRISMA Group. (2009). Preferred reporting items for systematic reviews and meta-analyses: The PRISMA statement. PLoS Medicine, 6 (7), e1000100.

Munafò, M. R., Nosek, B. A., Bishop, D. V. M., Button, K. S., Chambers, C. D., du Sert, N. P., et al. (2017). A manifesto for reproducible science. Nature Human Behaviour, 1 (0021), 1–9.

Nakagawa, S., & Cuthill, I. C. (2007). Effect size, confidence interval and statistical significance: A practical guide for biologists. Biological Reviews, 82 (4), 591–605.

NCBI Insights. (2018). Pubmed commons to be discontinued . https://ncbiinsights.ncbi.nlm . nih.gov/2018/02/01/pubmed-commons-to-be-discontinued /.

Nelson, L. D., Simmons, J. P., & Simonsohn, U. (2018). Psychology’s renaissance. Annual Review of Psychology, 69 , 511–534.

Neyman, J., & Pearson, E. S. (1928). On the use and interpretation of certain test criteria for purposes of statistical inference: Part I. Biometrika, 20A (1/2), 175–240.

Neyman, J., & Pearson, E. S. (1933). On the problem of the most efficient tests of statistical hypotheses. Philosophical transactions of the Royal Society of London. Series A, Containing Papers of a Mathematical or Physical Character, 231 , 289–337.

Nickerson, R. S. (2000). Null hypothesis significance testing: A review of an old and continuing controversy. Psychological Methods, 5 (2), 241–301.

Nielsen, M. (2011). Reinventing discovery: The new era of networked science . Princeton University Press.

Nosek, B. A., Alter, G., Banks, G. C., Borsboom, D., Bowman, S. D., Breckler, S. J., et al. (2015). Promoting an open research culture. Science, 348 (6242), 1422–1425.

Nosek, B. A., Ebersole, C. R., DeHaven, A. C., & Mellor, D. T. (2018). The preregistration revolution. Proceedings of the National Academy of Sciences, 115 (11), 2600–2606.

Nuzzo, R. (2014). Statistical errors: P values, the “gold standard” of statistical validity, are not as reliable as many scientists assume. Nature, 506 (150), 150–152.

O’Boyle, E. H., Banks, G. C., & Gonzalez-Mulé, E. (2017). The chrysalis effect: How ugly initial results metamorphosize into beautiful articles. Journal of Management, 43 (2), 376–399.

Open Science Collaboration. (2015). Estimating the reproducibility of psychological science. Science, 349 (6251), 943.

Pernet, C. (2016). Null hypothesis significance testing: A guide to commonly misunderstood concepts and recommendations for good practice [version 5; peer review: 2 approved, 2 not approved]. F1000Research, 4 (621). https://doi.org/10.12688/f1000research.6963.5 .

publons. (2017). 5 steps to writing a winning post-publication peer review . https://publons.com/blog/5-steps-to-writing-a-winning-post-publication-peer-review/ .

Reinhart, A. (2015). Statistics done wrong: The woefully complete guide . No Starch Press.

Ringle, C. M., Sarstedt, M., & Straub, D. W. (2012). Editor’s comments: A critical look at the use of PLS-SEM in MIS quarterly . MIS Quarterly, 36 (1), iii–xiv.

Rishika, R., Kumar, A., Janakiraman, R., & Bezawada, R. (2013). The effect of customers’ social media participation on customer visit frequency and profitability: An empirical investigation. Information Systems Research, 24 (1), 108–127.

Rönkkö, M., & Evermann, J. (2013). A critical examination of common beliefs about partial least squares path modeling. Organizational Research Methods, 16 (3), 425–448.

Rönkkö, M., McIntosh, C. N., Antonakis, J., & Edwards, J. R. (2016). Partial least squares path modeling: Time for some serious second thoughts. Journal of Operations Management, 47-48 , 9–27.

Saunders, C. (2005). Editor’s comments: Looking for diamond cutters. MIS Quarterly, 29 (1), iii–viii.

Saunders, C., Brown, S. A., Bygstad, B., Dennis, A. R., Ferran, C., Galletta, D. F., et al. (2017). Goals, values, and expectations of the ais family of journals. Journal of the Association for Information Systems, 18 (9), 633–647.

Schönbrodt, F. D. (2018). P-checker: One-for-all p-value analyzer . http://shinyapps.org/apps/p-checker/ .

Schwab, A., Abrahamson, E., Starbuck, W. H., & Fidler, F. (2011). Perspective: Researchers should make thoughtful assessments instead of null-hypothesis significance tests. Organization Science, 22 (4), 1105–1120.

Shaw, J. D., & Ertug, G. (2017). From the editors: The suitability of simulations and meta-analyses for submissions to academy of management journal . Academy of Management Journal, 60 (6), 2045–2049.

Siegfried, T. (2014). To make science better, watch out for statistical flaws. ScienceNews Context Blog, 2019, February 7, 2014. https://www.sciencenews.org/blog/context/make-science-better-watch-out-statistical-flaws .

Simonsohn, U., Nelson, L. D., & Simmons, J. P. (2014). P-curve: A key to the file-drawer. Journal of Experimental Psychology: General, 143 (2), 534–547.

Sivo, S. A., Saunders, C., Chang, Q., & Jiang, J. J. (2006). How low should you go? Low response rates and the validity of inference in is questionnaire research. Journal of the Association for Information Systems, 7 (6), 351–414.

Smith, S. M., Fahey, T., & Smucny, J. (2014). Antibiotics for acute bronchitis. Journal of the American Medical Association, 312 (24), 2678–2679.

Starbuck, W. H. (2013). Why and where do academics publish? Management, 16 (5), 707–718.

Starbuck, W. H. (2016). 60th anniversary essay: How journals could improve research practices in social science. Administrative Science Quarterly, 61 (2), 165–183.

Straub, D. W. (1989). Validating instruments in MIS research. MIS Quarterly, 13 (2), 147–169.

Straub, D. W. (2008). Editor’s comments: Type II reviewing errors and the search for exciting papers. MIS Quarterly, 32 (2), v–x.

Straub, D. W., Boudreau, M.-C., & Gefen, D. (2004). Validation guidelines for is positivist research. Communications of the Association for Information Systems, 13 (24), 380–427.

Szucs, D., & Ioannidis, J. P. A. (2017). When null hypothesis significance testing is unsuitable for research: A reassessment. Frontiers in Human Neuroscience, 11 (390), 1–21.

Tams, S., & Straub, D. W. (2010). The effect of an IS article’s structure on its impact. Communications of the Association for Information Systems, 27 (10), 149–172.

The Economist. (2013). Trouble at the lab . The Economist. http://www.economist.com/news/briefing/21588057-scientists-think-science-self-correcting-alarming-degree-it-not-trouble .

Trafimow, D., & Marks, M. (2015). Editorial. Basic and Applied Social Psychology, 37 (1), 1–2.

Tryon, W. W., Patelis, T., Chajewski, M., & Lewis, C. (2017). Theory construction and data analysis. Theory & Psychology, 27 (1), 126–134.

Tsang, E. W. K., & Williams, J. N. (2012). Generalization and induction: Misconceptions, clarifications, and a classification of induction. MIS Quarterly, 36 (3), 729–748.

Twa, M. D. (2016). Transparency in biomedical research: An argument against tests of statistical significance. Optometry & Vision Science, 93 (5), 457–458.

Venkatesh, V., Brown, S. A., & Bala, H. (2013). Bridging the qualitative-quantitative divide: Guidelines for conducting mixed methods research in information systems. MIS Quarterly, 37 (1), 21–54.

Vodanovich, S., Sundaram, D., & Myers, M. D. (2010). Research commentary: Digital natives and ubiquitous information systems. Information Systems Research, 21 (4), 711–723.

Walsh, E., Rooney, M., Appleby, L., & Wilkinson, G. (2000). Open peer review: A randomised controlled trial. The British Journal of Psychiatry, 176 (1), 47–51.

Warren, M. (2018). First analysis of “preregistered” studies shows sharp rise in null findings. Nature News, October 24, 2018, https://www.nature.com/articles/d41586-018-07118 .

Wasserstein, R. L., & Lazar, N. A. (2016). The ASA’s statement on p-values: Context, process, and purpose. The American Statistician, 70 (2), 129–133.

Wasserstein, R. L., Schirm, A. L., & Lazar, N. A. (2019). Moving to a world beyond “p < 0.05.”. The American Statistician, 73 (Sup 1), 1–19.

Xu, H., Zhang, N., & Zhou, L. (2019). Validity concerns in research using organic data. Journal of Management, 46 , 1257. https://doi.org/10.1177/0149206319862027

Yong, E. (2012). Nobel laureate challenges psychologists to clean up their act. Nature News, October 3, 2012. https://www.nature.com/news/nobel-laureate-challenges-psychologists-to-clean-up-their-act-1.11535 .

Yoo, Y. (2010). Computing in everyday life: A call for research on experiential computing. MIS Quarterly, 34 (2), 213–231.

Zeng, X., & Wei, L. (2013). Social ties and user content generation: Evidence from flickr. Information Systems Research, 24 (1), 71–87.

Download references

Acknowledgments

We are indebted to the senior editor at JAIS , Allen Lee, and two anonymous reviewers for constructive and developmental feedback that helped us improve the original chapter. We thank participants at seminars at Queensland University of Technology and University of Cologne for providing feedback on our work. We also thank Christian Hovestadt for his help in coding papers. All faults remain ours.

Author information

Authors and affiliations.

Colruyt Group, Halle, Belgium

Willem Mertens

Universität Hamburg, Faculty of Business Administration, Information Systems and Digital Innovation, Hamburg, Germany

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Jan Recker .

Editor information

Editors and affiliations.

Department of Management, London School of Economics and Political Science, London, UK

Leslie P. Willcocks

Labovitz School of Business and Economics, University of Minnesota Duluth, Duluth, MN, USA

Nik R. Hassan

HEC Montréal, Montreal, QC, Canada

Suzanne Rivard

Appendix A: Literature Review Procedures

Identification of papers.

In our intention to demonstrate “open science” practices (Locascio, 2019 ; Nosek et al., 2018 ; Warren, 2018 ) we preregistered our research procedures using the Open Science Framework “Registries” (doi:10.17605/OSF.IO/2GKCS).

We proceeded as follows: We identified the 100 top-cited papers (per year) between 2013 and 2016 in the AIS Senior Scholars’ basket of 8 IS journals using Harzing’s Publish or Perish version 6 (Harzing, 2010 ). We ran the queries separately on February 7, 2017, and then aggregated the results to identify the 100 most cited papers (based on citations per year) across the basket of eight journals. Footnote 9 The raw data (together with the coded data) is available at an open data repository hosted by Queensland University of Technology (doi:10.25912/5cede0024b1e1).

We identified from this set of papers those that followed the hypothetico-deductive model. First, we excluded 48 papers that did not involve empirical data: 31 papers that offered purely theoretical contributions, 11 that were commentaries in the form of forewords, introductions to special issues or editorials, 5 methodological essays, and 1 design science paper. Second, we identified from these 52 papers those that reported on collection and analysis of quantitative data. We found 46 such papers; of these, 39 were traditional quantitative research articles, 3 were essays on methodological aspects of quantitative research, 2 studies employed mixed-method designs involving quantitative empirical data, and 2 design science papers that involved quantitative data. Third, we eliminated from this set the three methodological essays as the focus of these papers was not on developing and testing new theory to explain and predict IS phenomena. This resulted in a final sample of 43 papers, including 2 design science and 2 mixed-method studies.

Coding of Papers

We developed a coding scheme in an excel repository to code the studies. The repository is available in our Open Science Framework (OSF) registry. We used the following criteria. Where applicable, we refer to literature that defined the variables we used during coding.

What is the main method of data collection and analysis (e.g., experiment, meta-analysis, panel, social network analysis, survey, text mining, economic modeling, multiple)?

Are testable hypotheses or propositions proposed (yes/in graphical form only/no)?

How precisely are the hypotheses formulated (using the classification of Edwards & Berry, 2010 )?

Is null hypothesis significance testing used (yes/no)?

Are exact p- values reported (yes/all/some/not at all)?

Are effect sizes reported and, if so, which ones primarily (e.g., R 2 , standardized means difference scores, f 2 , partial eta 2 )?

Are results declared as “statistically significant” (yes/sometimes/not at all)?

How many hypotheses are reported as supported (%)?

Are p- values used to argue the absence of an effect (yes/no)?

Are confidence intervals for test statistics reported (yes/selectively/no)?

What sampling method is used (i.e., convenient/random/systematic sampling, entire population)? Footnote 10

Is statistical power discussed and if so, where and how (e.g., sample size estimation, ex-post power analysis)?

Are competing theories tested explicitly (Gray & Cooper, 2010 )?

Are corrections made to adjust for multiple hypothesis testing, where applicable (e.g., Bonferroni, alpha-inflation, variance inflation)?

Are post hoc analyses reported for unexpected results?

We also extracted quotes that in our interpretation illuminated the view taken on NHST in the chapter. This was important for us to demonstrate the imbuement of practices in our research routines and the language used in using key NHST phrases such as “statistical significance” or “ p- value” (Gelman & Stern, 2006 ).

To be as unbiased as possible, we hired a research assistant to perform the coding of papers. Before he commenced coding, we explained the coding scheme to him during several meetings. We then conducted a pilot test to evaluate the quality of his coding: the research assistant coded five random papers from the set of papers and we met to review the coding by comparing our different individual understandings of the papers. Where inconsistencies arose, we clarified the coding scheme with him until we were confident that he understood it thoroughly. During the coding, the research assistant highlighted particular problematic or ambiguous coding elements and we met and resolved these ambiguities to arrive at a shared agreement. The coding process took three months to complete. The results of our coding are openly accessible at doi : 10.25912/5cede0024b1e1. Appendix B provides some summary statistics about our sample.

Selected Descriptive Statistics from 43 Frequently Cited IS Papers from 2013 to 2016

Rights and permissions.

Reprints and permissions

Copyright information

© 2023 The Author(s), under exclusive license to Springer Nature Switzerland AG

About this chapter

Mertens, W., Recker, J. (2023). New Guidelines for Null Hypothesis Significance Testing in Hypothetico-Deductive IS Research. In: Willcocks, L.P., Hassan, N.R., Rivard, S. (eds) Advancing Information Systems Theories, Volume II. Technology, Work and Globalization. Palgrave Macmillan, Cham. https://doi.org/10.1007/978-3-031-38719-7_13

Download citation

DOI : https://doi.org/10.1007/978-3-031-38719-7_13

Published : 15 October 2023

Publisher Name : Palgrave Macmillan, Cham

Print ISBN : 978-3-031-38718-0

Online ISBN : 978-3-031-38719-7

eBook Packages : Business and Management Business and Management (R0)

Share this chapter

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Publish with us

Policies and ethics

• Find a journal
• Track your research

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

• Publications
• Account settings

Preview improvements coming to the PMC website in October 2024. Learn More or Try it out now .

• Advanced Search
• Journal List
• J Korean Med Sci
• v.37(16); 2022 Apr 25

A Practical Guide to Writing Quantitative and Qualitative Research Questions and Hypotheses in Scholarly Articles

Edward barroga.

1 Department of General Education, Graduate School of Nursing Science, St. Luke’s International University, Tokyo, Japan.

Glafera Janet Matanguihan

2 Department of Biological Sciences, Messiah University, Mechanicsburg, PA, USA.

The development of research questions and the subsequent hypotheses are prerequisites to defining the main research purpose and specific objectives of a study. Consequently, these objectives determine the study design and research outcome. The development of research questions is a process based on knowledge of current trends, cutting-edge studies, and technological advances in the research field. Excellent research questions are focused and require a comprehensive literature search and in-depth understanding of the problem being investigated. Initially, research questions may be written as descriptive questions which could be developed into inferential questions. These questions must be specific and concise to provide a clear foundation for developing hypotheses. Hypotheses are more formal predictions about the research outcomes. These specify the possible results that may or may not be expected regarding the relationship between groups. Thus, research questions and hypotheses clarify the main purpose and specific objectives of the study, which in turn dictate the design of the study, its direction, and outcome. Studies developed from good research questions and hypotheses will have trustworthy outcomes with wide-ranging social and health implications.

INTRODUCTION

Scientific research is usually initiated by posing evidenced-based research questions which are then explicitly restated as hypotheses. 1 , 2 The hypotheses provide directions to guide the study, solutions, explanations, and expected results. 3 , 4 Both research questions and hypotheses are essentially formulated based on conventional theories and real-world processes, which allow the inception of novel studies and the ethical testing of ideas. 5 , 6

It is crucial to have knowledge of both quantitative and qualitative research 2 as both types of research involve writing research questions and hypotheses. 7 However, these crucial elements of research are sometimes overlooked; if not overlooked, then framed without the forethought and meticulous attention it needs. Planning and careful consideration are needed when developing quantitative or qualitative research, particularly when conceptualizing research questions and hypotheses. 4

There is a continuing need to support researchers in the creation of innovative research questions and hypotheses, as well as for journal articles that carefully review these elements. 1 When research questions and hypotheses are not carefully thought of, unethical studies and poor outcomes usually ensue. Carefully formulated research questions and hypotheses define well-founded objectives, which in turn determine the appropriate design, course, and outcome of the study. This article then aims to discuss in detail the various aspects of crafting research questions and hypotheses, with the goal of guiding researchers as they develop their own. Examples from the authors and peer-reviewed scientific articles in the healthcare field are provided to illustrate key points.

DEFINITIONS AND RELATIONSHIP OF RESEARCH QUESTIONS AND HYPOTHESES

A research question is what a study aims to answer after data analysis and interpretation. The answer is written in length in the discussion section of the paper. Thus, the research question gives a preview of the different parts and variables of the study meant to address the problem posed in the research question. 1 An excellent research question clarifies the research writing while facilitating understanding of the research topic, objective, scope, and limitations of the study. 5

On the other hand, a research hypothesis is an educated statement of an expected outcome. This statement is based on background research and current knowledge. 8 , 9 The research hypothesis makes a specific prediction about a new phenomenon 10 or a formal statement on the expected relationship between an independent variable and a dependent variable. 3 , 11 It provides a tentative answer to the research question to be tested or explored. 4

Hypotheses employ reasoning to predict a theory-based outcome. 10 These can also be developed from theories by focusing on components of theories that have not yet been observed. 10 The validity of hypotheses is often based on the testability of the prediction made in a reproducible experiment. 8

Conversely, hypotheses can also be rephrased as research questions. Several hypotheses based on existing theories and knowledge may be needed to answer a research question. Developing ethical research questions and hypotheses creates a research design that has logical relationships among variables. These relationships serve as a solid foundation for the conduct of the study. 4 , 11 Haphazardly constructed research questions can result in poorly formulated hypotheses and improper study designs, leading to unreliable results. Thus, the formulations of relevant research questions and verifiable hypotheses are crucial when beginning research. 12

CHARACTERISTICS OF GOOD RESEARCH QUESTIONS AND HYPOTHESES

Excellent research questions are specific and focused. These integrate collective data and observations to confirm or refute the subsequent hypotheses. Well-constructed hypotheses are based on previous reports and verify the research context. These are realistic, in-depth, sufficiently complex, and reproducible. More importantly, these hypotheses can be addressed and tested. 13

There are several characteristics of well-developed hypotheses. Good hypotheses are 1) empirically testable 7 , 10 , 11 , 13 ; 2) backed by preliminary evidence 9 ; 3) testable by ethical research 7 , 9 ; 4) based on original ideas 9 ; 5) have evidenced-based logical reasoning 10 ; and 6) can be predicted. 11 Good hypotheses can infer ethical and positive implications, indicating the presence of a relationship or effect relevant to the research theme. 7 , 11 These are initially developed from a general theory and branch into specific hypotheses by deductive reasoning. In the absence of a theory to base the hypotheses, inductive reasoning based on specific observations or findings form more general hypotheses. 10

TYPES OF RESEARCH QUESTIONS AND HYPOTHESES

Research questions and hypotheses are developed according to the type of research, which can be broadly classified into quantitative and qualitative research. We provide a summary of the types of research questions and hypotheses under quantitative and qualitative research categories in Table 1 .

Research questions in quantitative research

In quantitative research, research questions inquire about the relationships among variables being investigated and are usually framed at the start of the study. These are precise and typically linked to the subject population, dependent and independent variables, and research design. 1 Research questions may also attempt to describe the behavior of a population in relation to one or more variables, or describe the characteristics of variables to be measured ( descriptive research questions ). 1 , 5 , 14 These questions may also aim to discover differences between groups within the context of an outcome variable ( comparative research questions ), 1 , 5 , 14 or elucidate trends and interactions among variables ( relationship research questions ). 1 , 5 We provide examples of descriptive, comparative, and relationship research questions in quantitative research in Table 2 .

Hypotheses in quantitative research

In quantitative research, hypotheses predict the expected relationships among variables. 15 Relationships among variables that can be predicted include 1) between a single dependent variable and a single independent variable ( simple hypothesis ) or 2) between two or more independent and dependent variables ( complex hypothesis ). 4 , 11 Hypotheses may also specify the expected direction to be followed and imply an intellectual commitment to a particular outcome ( directional hypothesis ) 4 . On the other hand, hypotheses may not predict the exact direction and are used in the absence of a theory, or when findings contradict previous studies ( non-directional hypothesis ). 4 In addition, hypotheses can 1) define interdependency between variables ( associative hypothesis ), 4 2) propose an effect on the dependent variable from manipulation of the independent variable ( causal hypothesis ), 4 3) state a negative relationship between two variables ( null hypothesis ), 4 , 11 , 15 4) replace the working hypothesis if rejected ( alternative hypothesis ), 15 explain the relationship of phenomena to possibly generate a theory ( working hypothesis ), 11 5) involve quantifiable variables that can be tested statistically ( statistical hypothesis ), 11 6) or express a relationship whose interlinks can be verified logically ( logical hypothesis ). 11 We provide examples of simple, complex, directional, non-directional, associative, causal, null, alternative, working, statistical, and logical hypotheses in quantitative research, as well as the definition of quantitative hypothesis-testing research in Table 3 .

Research questions in qualitative research

Unlike research questions in quantitative research, research questions in qualitative research are usually continuously reviewed and reformulated. The central question and associated subquestions are stated more than the hypotheses. 15 The central question broadly explores a complex set of factors surrounding the central phenomenon, aiming to present the varied perspectives of participants. 15

There are varied goals for which qualitative research questions are developed. These questions can function in several ways, such as to 1) identify and describe existing conditions ( contextual research question s); 2) describe a phenomenon ( descriptive research questions ); 3) assess the effectiveness of existing methods, protocols, theories, or procedures ( evaluation research questions ); 4) examine a phenomenon or analyze the reasons or relationships between subjects or phenomena ( explanatory research questions ); or 5) focus on unknown aspects of a particular topic ( exploratory research questions ). 5 In addition, some qualitative research questions provide new ideas for the development of theories and actions ( generative research questions ) or advance specific ideologies of a position ( ideological research questions ). 1 Other qualitative research questions may build on a body of existing literature and become working guidelines ( ethnographic research questions ). Research questions may also be broadly stated without specific reference to the existing literature or a typology of questions ( phenomenological research questions ), may be directed towards generating a theory of some process ( grounded theory questions ), or may address a description of the case and the emerging themes ( qualitative case study questions ). 15 We provide examples of contextual, descriptive, evaluation, explanatory, exploratory, generative, ideological, ethnographic, phenomenological, grounded theory, and qualitative case study research questions in qualitative research in Table 4 , and the definition of qualitative hypothesis-generating research in Table 5 .

Qualitative studies usually pose at least one central research question and several subquestions starting with How or What . These research questions use exploratory verbs such as explore or describe . These also focus on one central phenomenon of interest, and may mention the participants and research site. 15

Hypotheses in qualitative research

Hypotheses in qualitative research are stated in the form of a clear statement concerning the problem to be investigated. Unlike in quantitative research where hypotheses are usually developed to be tested, qualitative research can lead to both hypothesis-testing and hypothesis-generating outcomes. 2 When studies require both quantitative and qualitative research questions, this suggests an integrative process between both research methods wherein a single mixed-methods research question can be developed. 1

FRAMEWORKS FOR DEVELOPING RESEARCH QUESTIONS AND HYPOTHESES

Research questions followed by hypotheses should be developed before the start of the study. 1 , 12 , 14 It is crucial to develop feasible research questions on a topic that is interesting to both the researcher and the scientific community. This can be achieved by a meticulous review of previous and current studies to establish a novel topic. Specific areas are subsequently focused on to generate ethical research questions. The relevance of the research questions is evaluated in terms of clarity of the resulting data, specificity of the methodology, objectivity of the outcome, depth of the research, and impact of the study. 1 , 5 These aspects constitute the FINER criteria (i.e., Feasible, Interesting, Novel, Ethical, and Relevant). 1 Clarity and effectiveness are achieved if research questions meet the FINER criteria. In addition to the FINER criteria, Ratan et al. described focus, complexity, novelty, feasibility, and measurability for evaluating the effectiveness of research questions. 14

The PICOT and PEO frameworks are also used when developing research questions. 1 The following elements are addressed in these frameworks, PICOT: P-population/patients/problem, I-intervention or indicator being studied, C-comparison group, O-outcome of interest, and T-timeframe of the study; PEO: P-population being studied, E-exposure to preexisting conditions, and O-outcome of interest. 1 Research questions are also considered good if these meet the “FINERMAPS” framework: Feasible, Interesting, Novel, Ethical, Relevant, Manageable, Appropriate, Potential value/publishable, and Systematic. 14

As we indicated earlier, research questions and hypotheses that are not carefully formulated result in unethical studies or poor outcomes. To illustrate this, we provide some examples of ambiguous research question and hypotheses that result in unclear and weak research objectives in quantitative research ( Table 6 ) 16 and qualitative research ( Table 7 ) 17 , and how to transform these ambiguous research question(s) and hypothesis(es) into clear and good statements.

a These statements were composed for comparison and illustrative purposes only.

b These statements are direct quotes from Higashihara and Horiuchi. 16

a This statement is a direct quote from Shimoda et al. 17

The other statements were composed for comparison and illustrative purposes only.

CONSTRUCTING RESEARCH QUESTIONS AND HYPOTHESES

To construct effective research questions and hypotheses, it is very important to 1) clarify the background and 2) identify the research problem at the outset of the research, within a specific timeframe. 9 Then, 3) review or conduct preliminary research to collect all available knowledge about the possible research questions by studying theories and previous studies. 18 Afterwards, 4) construct research questions to investigate the research problem. Identify variables to be accessed from the research questions 4 and make operational definitions of constructs from the research problem and questions. Thereafter, 5) construct specific deductive or inductive predictions in the form of hypotheses. 4 Finally, 6) state the study aims . This general flow for constructing effective research questions and hypotheses prior to conducting research is shown in Fig. 1 .

Research questions are used more frequently in qualitative research than objectives or hypotheses. 3 These questions seek to discover, understand, explore or describe experiences by asking “What” or “How.” The questions are open-ended to elicit a description rather than to relate variables or compare groups. The questions are continually reviewed, reformulated, and changed during the qualitative study. 3 Research questions are also used more frequently in survey projects than hypotheses in experiments in quantitative research to compare variables and their relationships.

Hypotheses are constructed based on the variables identified and as an if-then statement, following the template, ‘If a specific action is taken, then a certain outcome is expected.’ At this stage, some ideas regarding expectations from the research to be conducted must be drawn. 18 Then, the variables to be manipulated (independent) and influenced (dependent) are defined. 4 Thereafter, the hypothesis is stated and refined, and reproducible data tailored to the hypothesis are identified, collected, and analyzed. 4 The hypotheses must be testable and specific, 18 and should describe the variables and their relationships, the specific group being studied, and the predicted research outcome. 18 Hypotheses construction involves a testable proposition to be deduced from theory, and independent and dependent variables to be separated and measured separately. 3 Therefore, good hypotheses must be based on good research questions constructed at the start of a study or trial. 12

In summary, research questions are constructed after establishing the background of the study. Hypotheses are then developed based on the research questions. Thus, it is crucial to have excellent research questions to generate superior hypotheses. In turn, these would determine the research objectives and the design of the study, and ultimately, the outcome of the research. 12 Algorithms for building research questions and hypotheses are shown in Fig. 2 for quantitative research and in Fig. 3 for qualitative research.

EXAMPLES OF RESEARCH QUESTIONS FROM PUBLISHED ARTICLES

• EXAMPLE 1. Descriptive research question (quantitative research)
• - Presents research variables to be assessed (distinct phenotypes and subphenotypes)
• “BACKGROUND: Since COVID-19 was identified, its clinical and biological heterogeneity has been recognized. Identifying COVID-19 phenotypes might help guide basic, clinical, and translational research efforts.
• RESEARCH QUESTION: Does the clinical spectrum of patients with COVID-19 contain distinct phenotypes and subphenotypes? ” 19
• EXAMPLE 2. Relationship research question (quantitative research)
• - Shows interactions between dependent variable (static postural control) and independent variable (peripheral visual field loss)
• “Background: Integration of visual, vestibular, and proprioceptive sensations contributes to postural control. People with peripheral visual field loss have serious postural instability. However, the directional specificity of postural stability and sensory reweighting caused by gradual peripheral visual field loss remain unclear.
• Research question: What are the effects of peripheral visual field loss on static postural control ?” 20
• EXAMPLE 3. Comparative research question (quantitative research)
• - Clarifies the difference among groups with an outcome variable (patients enrolled in COMPERA with moderate PH or severe PH in COPD) and another group without the outcome variable (patients with idiopathic pulmonary arterial hypertension (IPAH))
• “BACKGROUND: Pulmonary hypertension (PH) in COPD is a poorly investigated clinical condition.
• RESEARCH QUESTION: Which factors determine the outcome of PH in COPD?
• STUDY DESIGN AND METHODS: We analyzed the characteristics and outcome of patients enrolled in the Comparative, Prospective Registry of Newly Initiated Therapies for Pulmonary Hypertension (COMPERA) with moderate or severe PH in COPD as defined during the 6th PH World Symposium who received medical therapy for PH and compared them with patients with idiopathic pulmonary arterial hypertension (IPAH) .” 21
• EXAMPLE 4. Exploratory research question (qualitative research)
• - Explores areas that have not been fully investigated (perspectives of families and children who receive care in clinic-based child obesity treatment) to have a deeper understanding of the research problem
• “Problem: Interventions for children with obesity lead to only modest improvements in BMI and long-term outcomes, and data are limited on the perspectives of families of children with obesity in clinic-based treatment. This scoping review seeks to answer the question: What is known about the perspectives of families and children who receive care in clinic-based child obesity treatment? This review aims to explore the scope of perspectives reported by families of children with obesity who have received individualized outpatient clinic-based obesity treatment.” 22
• EXAMPLE 5. Relationship research question (quantitative research)
• - Defines interactions between dependent variable (use of ankle strategies) and independent variable (changes in muscle tone)
• “Background: To maintain an upright standing posture against external disturbances, the human body mainly employs two types of postural control strategies: “ankle strategy” and “hip strategy.” While it has been reported that the magnitude of the disturbance alters the use of postural control strategies, it has not been elucidated how the level of muscle tone, one of the crucial parameters of bodily function, determines the use of each strategy. We have previously confirmed using forward dynamics simulations of human musculoskeletal models that an increased muscle tone promotes the use of ankle strategies. The objective of the present study was to experimentally evaluate a hypothesis: an increased muscle tone promotes the use of ankle strategies. Research question: Do changes in the muscle tone affect the use of ankle strategies ?” 23

EXAMPLES OF HYPOTHESES IN PUBLISHED ARTICLES

• EXAMPLE 1. Working hypothesis (quantitative research)
• - A hypothesis that is initially accepted for further research to produce a feasible theory
• “As fever may have benefit in shortening the duration of viral illness, it is plausible to hypothesize that the antipyretic efficacy of ibuprofen may be hindering the benefits of a fever response when taken during the early stages of COVID-19 illness .” 24
• “In conclusion, it is plausible to hypothesize that the antipyretic efficacy of ibuprofen may be hindering the benefits of a fever response . The difference in perceived safety of these agents in COVID-19 illness could be related to the more potent efficacy to reduce fever with ibuprofen compared to acetaminophen. Compelling data on the benefit of fever warrant further research and review to determine when to treat or withhold ibuprofen for early stage fever for COVID-19 and other related viral illnesses .” 24
• EXAMPLE 2. Exploratory hypothesis (qualitative research)
• - Explores particular areas deeper to clarify subjective experience and develop a formal hypothesis potentially testable in a future quantitative approach
• “We hypothesized that when thinking about a past experience of help-seeking, a self distancing prompt would cause increased help-seeking intentions and more favorable help-seeking outcome expectations .” 25
• “Conclusion
• Although a priori hypotheses were not supported, further research is warranted as results indicate the potential for using self-distancing approaches to increasing help-seeking among some people with depressive symptomatology.” 25
• EXAMPLE 3. Hypothesis-generating research to establish a framework for hypothesis testing (qualitative research)
• “We hypothesize that compassionate care is beneficial for patients (better outcomes), healthcare systems and payers (lower costs), and healthcare providers (lower burnout). ” 26
• Compassionomics is the branch of knowledge and scientific study of the effects of compassionate healthcare. Our main hypotheses are that compassionate healthcare is beneficial for (1) patients, by improving clinical outcomes, (2) healthcare systems and payers, by supporting financial sustainability, and (3) HCPs, by lowering burnout and promoting resilience and well-being. The purpose of this paper is to establish a scientific framework for testing the hypotheses above . If these hypotheses are confirmed through rigorous research, compassionomics will belong in the science of evidence-based medicine, with major implications for all healthcare domains.” 26
• EXAMPLE 4. Statistical hypothesis (quantitative research)
• - An assumption is made about the relationship among several population characteristics ( gender differences in sociodemographic and clinical characteristics of adults with ADHD ). Validity is tested by statistical experiment or analysis ( chi-square test, Students t-test, and logistic regression analysis)
• “Our research investigated gender differences in sociodemographic and clinical characteristics of adults with ADHD in a Japanese clinical sample. Due to unique Japanese cultural ideals and expectations of women's behavior that are in opposition to ADHD symptoms, we hypothesized that women with ADHD experience more difficulties and present more dysfunctions than men . We tested the following hypotheses: first, women with ADHD have more comorbidities than men with ADHD; second, women with ADHD experience more social hardships than men, such as having less full-time employment and being more likely to be divorced.” 27
• “Statistical Analysis
• ( text omitted ) Between-gender comparisons were made using the chi-squared test for categorical variables and Students t-test for continuous variables…( text omitted ). A logistic regression analysis was performed for employment status, marital status, and comorbidity to evaluate the independent effects of gender on these dependent variables.” 27

EXAMPLES OF HYPOTHESIS AS WRITTEN IN PUBLISHED ARTICLES IN RELATION TO OTHER PARTS

• EXAMPLE 1. Background, hypotheses, and aims are provided
• “Pregnant women need skilled care during pregnancy and childbirth, but that skilled care is often delayed in some countries …( text omitted ). The focused antenatal care (FANC) model of WHO recommends that nurses provide information or counseling to all pregnant women …( text omitted ). Job aids are visual support materials that provide the right kind of information using graphics and words in a simple and yet effective manner. When nurses are not highly trained or have many work details to attend to, these job aids can serve as a content reminder for the nurses and can be used for educating their patients (Jennings, Yebadokpo, Affo, & Agbogbe, 2010) ( text omitted ). Importantly, additional evidence is needed to confirm how job aids can further improve the quality of ANC counseling by health workers in maternal care …( text omitted )” 28
• “ This has led us to hypothesize that the quality of ANC counseling would be better if supported by job aids. Consequently, a better quality of ANC counseling is expected to produce higher levels of awareness concerning the danger signs of pregnancy and a more favorable impression of the caring behavior of nurses .” 28
• “This study aimed to examine the differences in the responses of pregnant women to a job aid-supported intervention during ANC visit in terms of 1) their understanding of the danger signs of pregnancy and 2) their impression of the caring behaviors of nurses to pregnant women in rural Tanzania.” 28
• EXAMPLE 2. Background, hypotheses, and aims are provided
• “We conducted a two-arm randomized controlled trial (RCT) to evaluate and compare changes in salivary cortisol and oxytocin levels of first-time pregnant women between experimental and control groups. The women in the experimental group touched and held an infant for 30 min (experimental intervention protocol), whereas those in the control group watched a DVD movie of an infant (control intervention protocol). The primary outcome was salivary cortisol level and the secondary outcome was salivary oxytocin level.” 29
• “ We hypothesize that at 30 min after touching and holding an infant, the salivary cortisol level will significantly decrease and the salivary oxytocin level will increase in the experimental group compared with the control group .” 29
• EXAMPLE 3. Background, aim, and hypothesis are provided
• “In countries where the maternal mortality ratio remains high, antenatal education to increase Birth Preparedness and Complication Readiness (BPCR) is considered one of the top priorities [1]. BPCR includes birth plans during the antenatal period, such as the birthplace, birth attendant, transportation, health facility for complications, expenses, and birth materials, as well as family coordination to achieve such birth plans. In Tanzania, although increasing, only about half of all pregnant women attend an antenatal clinic more than four times [4]. Moreover, the information provided during antenatal care (ANC) is insufficient. In the resource-poor settings, antenatal group education is a potential approach because of the limited time for individual counseling at antenatal clinics.” 30
• “This study aimed to evaluate an antenatal group education program among pregnant women and their families with respect to birth-preparedness and maternal and infant outcomes in rural villages of Tanzania.” 30
• “ The study hypothesis was if Tanzanian pregnant women and their families received a family-oriented antenatal group education, they would (1) have a higher level of BPCR, (2) attend antenatal clinic four or more times, (3) give birth in a health facility, (4) have less complications of women at birth, and (5) have less complications and deaths of infants than those who did not receive the education .” 30

Research questions and hypotheses are crucial components to any type of research, whether quantitative or qualitative. These questions should be developed at the very beginning of the study. Excellent research questions lead to superior hypotheses, which, like a compass, set the direction of research, and can often determine the successful conduct of the study. Many research studies have floundered because the development of research questions and subsequent hypotheses was not given the thought and meticulous attention needed. The development of research questions and hypotheses is an iterative process based on extensive knowledge of the literature and insightful grasp of the knowledge gap. Focused, concise, and specific research questions provide a strong foundation for constructing hypotheses which serve as formal predictions about the research outcomes. Research questions and hypotheses are crucial elements of research that should not be overlooked. They should be carefully thought of and constructed when planning research. This avoids unethical studies and poor outcomes by defining well-founded objectives that determine the design, course, and outcome of the study.

Disclosure: The authors have no potential conflicts of interest to disclose.

Author Contributions:

• Conceptualization: Barroga E, Matanguihan GJ.
• Methodology: Barroga E, Matanguihan GJ.
• Writing - original draft: Barroga E, Matanguihan GJ.
• Writing - review & editing: Barroga E, Matanguihan GJ.

• News Releases

Time to abandon null hypothesis significance testing? Moving beyond the default approach to statistical analysis and reporting

News from the Journal of Marketing

American Marketing Association

Researchers from Northwestern University, University of Pennsylvania, and University of Colorado published a  new  Journal of Marketing  study  that proposes abandoning null hypothesis significance testing (NHST) as the default approach to statistical analysis and reporting.

The study, forthcoming in the  Journal of Marketing ,  is titled “ ‘Statistical Significance’ and Statistical Reporting: Moving Beyond Binary ” and is authored by Blakeley B. McShane, Eric T. Bradlow, John G. Lynch, Jr., and Robert J. Meyer.

Null hypothesis significance testing (NHST) is the default approach to statistical analysis and reporting in marketing and, more broadly, in the biomedical and social sciences. As practiced, NHST involves

• assuming that the intervention under investigation has no effect along with other assumptions,
• computing a statistical measure known as a  P -value based on these assumptions, and
• comparing the computed  P -value to the arbitrary threshold value of 0.05.

If the  P -value is less than 0.05, the effect is declared “statistically significant,” the assumption of no effect is rejected, and it is concluded that the intervention has an effect in the real world. If the  P -value is above 0.05, the effect is declared “statistically nonsignificant,” the assumption of no effect is not rejected, and it is concluded that the intervention has no effect in the real world.

Criticisms of NHST

Despite its default role, NHST has long been criticized by both statisticians and applied researchers, including those within marketing. The most prominent criticisms relate to the dichotomization of results into “statistically significant” and “statistically nonsignificant.”

For example, authors, editors, and reviewers use “statistical (non)significance” as a filter to select which results to publish. Meyer says that “this creates a distorted literature because the effects of published interventions are biased upward in magnitude. It also encourages harmful research practices that yield results that attain so-called statistical significance.”

Lynch adds that “NHST has no basis because no intervention has precisely zero effect in the real world and small  P -values and ‘statistical significance’ are guaranteed with sufficient sample sizes. Put differently, there is no need to reject a hypothesis of zero effect when it is already known to be false.”

Perhaps the most widespread abuse of statistics is to ascertain where some statistical measure such as a  P -value stands relative to 0.05 and take it as a basis to declare “statistical (non)significance” and to make general and certain conclusions from a single study. “Single studies are never definitive and thus can never demonstrate an effect or no effect. The aim of studies should be to report results in an unfiltered manner so that they can later be used to make more general conclusions based on the cumulative evidence from multiple studies. NHST leads researchers to wrongly make general and certain conclusions and to wrongly filter results,” says Bradlow.

“ P -values naturally vary a great deal from study to study,” explains McShane. As an example, a “statistically significant” original study with an observed  P -value of  p  = 0.005 (far below the 0.05 threshold) and a “statistically nonsignificant” replication study with an observed  P -value of  p  = 0.194 (far above the 0.05 threshold) are highly compatible with one another in the sense that the observed  P -value, assuming no difference between them, is  p  = 0.289. He adds that, “however, when viewed through the lens of ‘statistical (non)significance,’ these two studies appear categorically different and are thus in contradiction because they are categorized differently.”

Recommended Changes to Statistical Analysis

The authors propose a major transition in statistical analysis and reporting. Specifically, they propose abandoning NHST—and the  P -value thresholds intrinsic to it—as the default approach to statistical analysis and reporting. Their recommendations are as follows:

• “Statistical (non)significance” should never be used as a basis to make general and certain conclusions.
• “Statistical (non)significance” should also never be used as a filter to select which results to publish.
• Instead, all studies should be published in some form or another.
• Reporting should focus on quantifying study results via point and interval estimates. All of the values inside conventional interval estimates are at least reasonably compatible with the data given all of the assumptions used to compute them; therefore, it makes no sense to single out a specific value such as the null value.
• General conclusions should be made based on the cumulative evidence from multiple studies.
• Studies need to treat  P -values continuously and as just one factor among many—including prior evidence, plausibility of mechanism, study design, data quality, and others that vary by research domain—that require joint consideration and holistic integration.
• Researchers must also respect the fact that such conclusions are necessarily tentative and subject to revision as new studies are conducted.

Decisions are seldom necessary in scientific reporting and are best left to end-users such as managers and clinicians when necessary. In such cases, they should be made using a decision analysis that integrates the costs, benefits, and probabilities of all possible consequences via a loss function (which typically varies dramatically across stakeholders)—not via arbitrary thresholds applied to statistical summaries such as  P -values (“statistical (non)significance”) which, outside of certain specialized applications such as industrial quality control, are insufficient for this purpose.

Full article and author contact information available at:  https://doi.org/10.1177/00222429231216910

About the  Journal of Marketing  

The  Journal of Marketing  develops and disseminates knowledge about real-world marketing questions useful to scholars, educators, managers, policy makers, consumers, and other societal stakeholders around the world. Published by the American Marketing Association since its founding in 1936,  JM  has played a significant role in shaping the content and boundaries of the marketing discipline. Shrihari (Hari) Sridhar (Joe Foster ’56 Chair in Business Leadership, Professor of Marketing at Mays Business School, Texas A&M University) serves as the current Editor in Chief. https://www.ama.org/jm

About the American Marketing Association (AMA)  

As the largest chapter-based marketing association in the world, the AMA is trusted by marketing and sales professionals to help them discover what is coming next in the industry. The AMA has a community of local chapters in more than 70 cities and 350 college campuses throughout North America. The AMA is home to award-winning content, PCM® professional certification, premiere academic journals, and industry-leading training events and conferences. https://www.ama.org

Journal of Marketing

10.1177/00222429231216910

Article Title

“Statistical Significance” and Statistical Reporting: Moving Beyond Binary

Article Publication Date

14-Nov-2023

Disclaimer: AAAS and EurekAlert! are not responsible for the accuracy of news releases posted to EurekAlert! by contributing institutions or for the use of any information through the EurekAlert system.

Original Source

Miles D. Williams

Visiting Assistant Professor | Denison University

• Download My CV
• Send Me an Email
• View My LinkedIn
• Follow Me on Twitter

When the Research Hypothesis Is the Null

Posted on May 13, 2024 by Miles Williams in Methods   Statistics  

Back to Blog

What should you do if your research hypothesis is the null hypothesis? In other words, how should you approach hypothesis testing if your theory predicts no effect between two variables? I and a coauthor are working on a paper where a couple of our proposed hypotheses look like this, and we got some push-back from a reviewer about it. This prompted me to go down a rabbit hole of journal articles and message boards to see how others handle this situation. I quickly found that I waded into a contentious issue that’s connected to a bigger philosophical debate about the merits of hypothesis testing in general and whether the null hypothesis in particular as a bench-mark for hypothesis testing is even logically sound.

There’s too much to unpack with this debate for me to cover in a single blog post (and I’m sure I’d get some of the key points wrong anyway if I tried). The main issue I want to explore in this post is the practical problem of how to approach testing a null research hypothesis. From an applied perspective, this is a tricky problem that raises issues with how we calculate and interpret p-values. Thankfully, there is a sound solution for the null research hypothesis which I explore in greater detail below. It’s called a two one-sided test, and it’s easy to implement once you know what it is.

The usual approach

Most of the time when doing research, a scientist usually has a research hypothesis that goes something like X has a positive effect on Y . For example, a political scientist might propose that a get-out-the-vote (GOTV) campaign ( X ) will increase voter turnout ( Y ).

The typical approach for testing this claim might be to estimate a regression model with voter turnout as the outcome and the GOTV campaign as the explanatory variable of interest:

Y = α + β X + ε

If the parameter β > 0, this would support the hypothesis that GOTV campaigns improve voter turnout. To test this hypothesis, in practice the researcher would actually test a different hypothesis that we call the null hypothesis. This is the hypothesis that says there is no true effect of GOTV campaigns on voter turnout.

By proposing and testing the null, we now have a point of reference for calculating a measure of uncertainty—that is, the probability of observing an empirical effect of a certain magnitude or greater if the null hypothesis is true. This probability is called a p-value, and by convention if it is less than 0.05 we say that we can reject the null hypothesis.

For the hypothetical regression model proposed above, to get this p-value we’d estimate β, then calculate its standard error, and then we’d take the ratio of the former to the latter giving us what’s called a t-statistic or t-value. Under the null hypothesis, the t-value has a known distribution which makes it really easy to map any t-value to a p-value. The below figure illustrates using a hypothetical data sample of size N = 200. You can see that the t-statistic’s distribution has a distinct bell shape centered around 0. You can also see the range of t-values in blue where if we observed them in our empirical data we’d fail to reject the null hypothesis at the p < 0.05 level. Values in gray are t-values that would lead us to reject the null hypothesis at this same level.

When the null is the research hypothesis we want to test

There’s nothing new or special here. If you have even a basic stats background (particularly with Frequentist statistics), the conventional approach to hypothesis testing is pretty ubiquitous. Things get more tricky when our research hypothesis is that there is no effect. Say for a certain set of theoretical reasons we think that GOTV campaigns are basically useless at increasing voter turnout. If this argument is true, then if we estimate the following regression model, we’d expect β = 0.

The problem here is that our substantive research hypothesis is also the one that we want to try to find evidence against. We could just proceed like usual and just say that if we fail to reject the null this is evidence in support of our theory, but the problem with doing this is that failure to reject the null is not the same thing as finding support for the null hypothesis.

There are a few ideas in the literature for how we should approach this instead. Many of these approaches are Bayesian, but most of my research relies on Frequentist statistics, so these approaches were a no-go for me. However, there is one really simple approach that is consistent with the Frequentist paradigm: equivalence testing . The idea is simple. Propose some absolute effect size that is of minimal interest and then test whether an observed effect is different from it. This minimum effect is called the “smallest effect size of interest” (SESOI). I read about the approach in an article by Harms and Lakens (2018) in the Journal of Clinical and Translational Research .

Say, for example, that we deemed a t-value of +/-1.96 (the usual threshold for rejecting the null hypothesis) as extreme enough to constitute good evidence of a non-zero effect. We could make the appropriate adjustments to our t-distribution to identify a new range of t-values that would allow us to reject the hypothesis that an effect is non-zero. This is illustrated in the below figure. We can now see a range of t-values in the middle where we’d have t-values such that we could reject the non-zero hypothesis at the p < 0.05 level. This distribution looks like it’s been inverted relative to the usual null distribution. The reason is that with this approach what we’re doing is conducting a pair of alternative one-tailed tests. We’re testing both the hypothesis that β / se(β) - 1.96 > 0 and β / se(β) + 1.96 < 0. In the Harms and Lakens paper cited above, they call this approach two one-sided tests or TOST (I’m guessing this is pronounced “toast”).

Something to pay attention to with this approach is that the observed t-statistic needs to be very small in absolute magnitude for us to reject the hypothesis of a non-zero effect. This means that the bar for testing a null research hypothesis is actually quite high. This is demonstrated using the following simulation in R. Using the {seerrr} package, I had R generate 1,000 random draws (each of size 200) for a pair of variables x and y where the former is a binary “treatment” and the latter is a random normal “outcome.” By design, there is no true causal relationship between these variables. Once I simulated the data, I then generated a set of estimates of the effect of x on y for each simulated dataset and collected the results in an object called sim_ests . I then visualized two metrics that that I calculated with the simulated results: (1) the rejection rate for the null hypothesis test and (2) the rejection rate for the two one-sided equivalence tests. As you can see, if we were to try to test a research null hypothesis the usual way, we’d expect to be able to fail to reject the null about 95% of the time. Conversely, if we were to use the two one-sided equivalence tests, we’d expect to reject the non-zero alternative hypothesis only about 25% of the time. I tested out a few additional simulations to see if a larger sample size would lead to improvements in power (not shown), but no dice.

The two one-sided tests approach strikes me as a nice method when dealing with a null research hypothesis. It’s actually pretty easy to implement, too. The one downside is that this test is under-powered. If the null is true, it will only reject the alternative 25% of the time (though you could select a different non-zero alternative which would possibly give you more power). However, this isn’t all bad. The flip side of the coin is that this is a really conservative test, so if you can reject the alternative that puts you on solid rhetorical footing to show the data really do seem consistent with the null.