Weekend batch
Avijeet is a Senior Research Analyst at Simplilearn. Passionate about Data Analytics, Machine Learning, and Deep Learning, Avijeet is also interested in politics, cricket, and football.
Free eBook: Top Programming Languages For A Data Scientist
Normality Test in Minitab: Minitab with Statistics
Machine Learning Career Guide: A Playbook to Becoming a Machine Learning Engineer
Formal hypothesis testing is perhaps the most prominent and widely-employed form of statistical analysis. It is sometimes seen as the most rigorous and definitive part of a statistical analysis, but it is also the source of many statistical controversies. The currently-prevalent approach to hypothesis testing dates to developments that took place between 1925 and 1940, especially the work of Ronald Fisher , Jerzy Neyman , and Egon Pearson .
In recent years, many prominent statisticians have argued that less emphasis should be placed on the formal hypothesis testing approaches developed in the early twentieth century, with a correspondingly greater emphasis on other forms of uncertainty analysis. Our goal here is to give an overview of some of the well-established and widely-used approaches for hypothesis testing. We will also provide some perspectives on how these tools can be effectively used, and discuss their limitations. We will also discuss some new approaches to hypothesis testing that may eventually come to be as prominent as these classical approaches.
A falsifiable hypothesis is a statement, or hypothesis, that can be contradicted with evidence. In empirical (data-driven) research, this evidence will always be obtained through the data. In statistical hypothesis testing, the hypothesis that we formally test is called the null hypothesis . The alternative hypothesis is a second hypothesis that is our proposed explanation for what happens if the null hypothesis is wrong.
The key element of a statistical hypothesis test is the test statistic , which (like any statistic) is a function of the data. A test statistic takes our entire dataset, and reduces it to one number. This one number ideally should contain all the information in the data that is relevant for assessing the two hypotheses of interest, and exclude any aspects of the data that are irrelevant for assessing the two hypotheses. The test statistic measures evidence against the null hypothesis. Most test statistics are constructed so that a value of zero represents the lowest possible level of evidence against the null hypothesis. Test statistic values that deviate from zero represent greater levels of evidence against the null hypothesis. The larger the magnitude of the test statistic, the stronger the evidence against the null hypothesis.
A major theme of statistical research is to devise effective ways to construct test statistics. Many useful ways to do this have been devised, and there is no single approach that is always the best. In this introductory course, we will focus on tests that starting with an estimate of a quantity that is relevant for assessing the hypotheses, then proceed by standardizing this estimate by dividing it by its standard error. This approach is sometimes referred to as “Wald testing”, after Abraham Wald .
As a basic example, let’s consider risk perception related to COVID-19. As you will see below, hypothesis testing can appear at first to be a fairly elaborate exercise. Using this example, we describe each aspect of this exercise in detail below.
The data shown below are simulated but are designed to reflect actual surveys conducted in the United States in March of 2020. Partipants were asked whether they perceive that they have a substantial risk of dying if they are infected with the novel coronavirus. The number of people stating each response, stratified on age, are shown below (only two age groups are shown):
High risk | Not high risk | |
---|---|---|
Age < 30 | 25 | 202 |
Age 60-69 | 30 | 124 |
Each subject’s response is binary – they either perceive themselves to be high risk, or not to be at high risk. When working with this type of data, we are usually interested in the proportion of people who provide each response within each stratum (age group). These are conditional proportions, conditioning on the age group. The numerical values of the conditional proportions are given below:
High risk | Not high risk | |
---|---|---|
Age < 30 | 0.110 | 0.890 |
Age 60-69 | 0.195 | 0.805 |
There are four conditional proportions in the table above – the proportion of younger people who perceive themselves to be at higher risk, 0.110=25/(25+202); the proportion of younger people who do not perceive themselves to be at high risk, 0.890=202/(25+202); the proportion of older people who perceive themselves to be at high risk 0.195=30/(30+124); and the proportion of older people who do not perceive themselves to be at high risk, 0.805=124/(30+124).
The trend in the data is that younger people perceive themselves to be at lower risk of dying than older people, by a difference of 0.195-0.110=0.085 (in terms of proportions). But is this trend only present in this sample, or is it generalizable to a broader population (say the entire US population)? That is the goal of conducting a statistical hypothesis test in this setting.
Corresponding to our data above is the unobserved population structure, which we can denote as follows
High risk | Not high risk | |
---|---|---|
Age < 30 | \(p\) | \(1-p\) |
Age 60-69 | \(q\) | \(1-q\) |
The symbols \(p\) and \(q\) in the table above are population parameters . These are quantitites that we do not know, and wish to assess using the data. In this case, our null hypothesis can be expressed as the statement \(p = q\) . We can estimate \(p\) using the sample proportion \(\hat{p} = 0.110\) , and similarly estimate \(q\) using \(\hat{q} = 0.195\) . However these estimates do not immediately provide us with a way of expressing the evidence relating to the hypothesis that \(p=q\) . This is provided by the test statistic.
As noted above, a test statistic is a reduction of the data to one number that captures all of the relevant information for assessing the hypotheses. A natural first choice for a test statistic here would be the difference in sample proportions between the two age groups, which is 0.195 - 0.110 = 0.085. There is a difference of 0.085 between the perceived risks of death in the younger and older age groups.
The difference in rates (0.085) does not on its own make a good test statistic, although it is a good start toward obtaining one. The reason for this is that the evidence underlying this difference in rates depends also on the absolute rates (0.110 and 0.195), and on the sample sizes (227 and 154). If we only know that the difference in rates is 0.085, this is not sufficient to evaluate the hypothesis in a statistical manner. A given difference in rates is much stronger evidence if it is obtained from a larger sample. If we have a difference of 0.085 with a very large sample, say one million people, then we should be almost certain that the true rates differ (i.e. the data are highly incompatiable with the hypothesis that \(p=q\) ). If we have the same difference in rates of 0.085, but with a small sample, say 50 people per age group, then there would be almost no evidence for a true difference in the rates (i.e. the data are compatiable with the hypothesis \(p=q\) ).
To address this issue, we need to consider the uncertainty in the estimated rate difference, which is 0.085. Recall that the estimated rate difference is obtained from the sample and therefore is almost certain to deviate somewhat from the true rate difference in the population (which is unknown). Recall from our study of standard errors that the standard error for an estimated proportion is \(\sqrt{p(1-p)/n}\) , where \(p\) is the outcome probability (here the outcome is that a person perceives a high risk of dying), and \(n\) is the sample size.
In the present analysis, we are comparing two proportions, so we have two standard errors. The estimated standard error for the younger people is \(\sqrt{0.11\cdot 0.89/227} \approx 0.021\) . The estimated standard error for the older people is \(\sqrt{0.195\cdot 0.805/154} \approx 0.032\) . Note that both standard errors are estimated, rather than exact, because we are plugging in estimates of the rates (0.11 and 0.195). Also note that the standard error for the rate among older people is greater than that for younger people. This is because the sample size for older people is smaller, and also because the estimated rate for older people is closer to 1/2.
In our previous discussion of standard errors, we saw how standard errors for independent quantities \(A\) and \(B\) can be used to obtain the standard error for the difference \(A-B\) . Applying that result here, we see that the standard error for the estimated difference in rates 0.195-0.11=0.085 is \(\sqrt{0.021^2 + 0.032^2} \approx 0.038\) .
The final step in constructing our test statistic is to construct a Z-score from the estimated difference in rates. As with all Z-scores, we proceed by taking the estimated difference in rates, and then divide it by its standard error. Thus, we get a test statistic value of \(0.085 / 0.038 \approx 2.24\) .
A test statistic value of 2.24 is not very close to zero, so there is some evidence against the null hypothesis. But the strength of this evidence remains unclear. Thus, we must consider how to calibrate this evidence in a way that makes it more interpretable.
By the central limit theorem (CLT), a Z-score approximately follows a normal distribution. When the null hypothesis holds, the Z-score approximately follows the standard normal distribution (recall that a standard normal distribution is a normal distribution with expected value equal to 0 and variance equal to 1). If the null hypothesis does not hold, then the test statistic continues to approximately follow a normal distribution, but it is not the standard normal distribution.
A test statistic of zero represents the least possible evidence against the null hypothesis. Here, we will obtain a test statistic of zero when the two proportions being compared are identical, i.e. exactly the same proportions of younger and older people perceive a substantial risk of dying from a disease. Even if the test statistic is exactly zero, this does not guarantee that the null hypothesis is true. However it is the least amount of evidence that the data can present against the null hypothesis.
In a hypothesis testing setting using normally-distrbuted Z-scores, as is the case here (due to the CLT), the standard normal distribution is the reference distribution for our test statistic. If the Z-score falls in the center of the reference distribution, there is no evidence against the null hypothesis. If the Z-score falls into either tail of the reference distribution, then there is evidence against the null distribution, and the further into the tails of the reference distribution the Z-score falls, the greater the evidence.
The most conventional way to quantify the evidence in our test statistic is through a probability called the p-value . The p-value has a somewhat complex definition that many people find difficult to grasp. It is the probability of observing as much or more evidence against the null hypothesis as we actually observe, calculated when the null hypothesis is assumed to be true. We will discuss some ways to think about this more intuitively below.
For our purposes, “evidence against the null hypothesis” is reflected in how far into the tails of the reference distribution the Z-score (test statistic) falls. We observed a test statistic of 2.24 in our COVID risk perception analysis. Recall that due to the “empirical rule”, 95% of the time, a draw from a standard normal distribution falls between -2 and 2. Thus, the p-value must be less than 0.05, since 2.24 falls outside this interval. The p-value can be calculated using a computer, in this case it happens to be approximately 0.025.
As stated above, the p-value tells us how likely it would be for us to obtain as much evidence against the the null hypothesis as we observed in our actual data analysis, if we were certain that the null hypothesis were true. When the null hypothesis holds, any evidence against the null hypothesis is spurious. Thus, we will want to see stronger evidence against the null from our actual analysis than we would see if we know that the null hypothesis were true. A smaller p-value therefore reflects more evidence against the null hypothesis than a larger p-value.
By convention, p-values of 0.05 or smaller are considered to represent sufficiently strong evidence against the null hypothesis to make a finding “statistically significant”. This threshold of 0.05 was chosen arbitrarily 100 years ago, and there is no objective reason for it. In recent years, people have argued that either a lesser or a greater p-value threshold should be used. But largely due to convention, the practice of deeming p-values smaller than 0.05 to be statistically significant continues.
Here is a restatement of the above discussion, using slightly different language. In our analysis of COVID risk perceptions, we found a difference in proportions of 0.085 between younger and older subjects, with younger people perceiving a lower risk of dying. This is a difference based on the sample of data that we observed, but what we really want to know is whether there is a difference in COVID risk perception in the population (say, all US adults).
Suppose that in fact there is no difference in risk perception between younger and older people. For instance, suppose that in the population, 15% of people believe that they have a substantial risk of dying should they become infected with the novel coronavirus, regardless of their age. Even though the rates are equal in this imaginary population (both being 15%), the rates in our sample would typically not be equal. Around 3% of the time (0.024=2.4% to be exact), if the rates are actually equal in the population, we would see a test statistic that is 2.4 or larger. Since 3% represents a fairly rare event, we can conclude that our observed data are not compatible with the null hypothesis. We can also say that there is statistically significant evidence against the null hypothesis, and that we have “rejected” the null hypothesis at the 3% level.
In this data analysis, as in any data analysis, we cannot confirm definitively that the alternative hypothesis is true. But based on our data and the analysis performed above, we can claim that there is substantial evidence against the null hypothesis, using standard criteria for what is considered to be “substantial evidence”.
A very common setting where hypothesis testing is used arises when we wish to compare the means of a quantitative measurement obtained for two populations. Imagine, for example, that we have two ways of manufacturing a battery, and we wish to assess which approach yields batteries that are longer-lasting in actual use. To do this, suppose we obtain data that tells us the number of charge cycles that were completed in 200 batteries of type A, and in 300 batteries of type B. For the test developed below to be meaningful, the data must be independent and identically distributed samples.
The raw data for this study consists of 500 numbers, but it turns out that the most relevant information from the data is contained in the sample means and sample standard deviations computed within each battery type. Note that this is a huge reduction in complexity, since we started with 500 measurements and are able to summarize this down to just four numbers.
Suppose the summary statistics are as follows, where \(\bar{x}\) , \(\hat{\sigma}_x\) , and \(n\) denote the sample mean, sample standard deviation, and sample size, respectively.
Type | \(\bar{x}\) | \(\hat{\sigma}_x\) | \(n\) |
---|---|---|---|
420 | 70 | 200 | |
403 | 90 | 300 |
The simplest measure comparing the two manufacturing approaches is the difference 420 - 403 = 17. That is, batteries of type A tend to have 17 more charge cycles compared to batteries of type B. This difference is present in our sample, but is it also true that the entire population of type A batteries has more charge cycles than the entire population of type B batteries? That is the goal of conducting a hypothesis test.
The next step in the present analysis is to divide the mean difference, which is 17, by its standard error. As we have seen, the standard error of the mean, or SEM, is \(\sigma/n\) , where \(\sigma\) is the standard deviation and \(n\) is the sample size. Since \(\sigma\) is almost never known, we plug in its estimate \(\hat{\sigma}\) . For the type A batteries, the estimated SEM is thus \(70/\sqrt{200} \approx 4.95\) , and for the type B batteries the estimated SEM is \(90/\sqrt{300} \approx 5.2\) .
Since we are comparing two estimated means that are obtained from independent samples, we can pool the standard deviations to obtain an overall standard deviation of \(\sqrt{4.95^2 + 5.2^2} \approx 7.18\) . We can now obtain our test statistic \(17/7.18 \approx 2.37\) .
The test statistic can be calibrated against a standard normal reference distribution. The probability of observing a standard normal value that is greater in magnitude than 2.37 is 0.018 (this can be obtained from a computer). This is the p-value, and since it is smaller than the conventional threshold of 0.05, we can claim that there is a statistically significant difference between the average number of charge cycles for the two types of batteries, with the A batteries having more charge cycles on average.
The analysis illustrated here is called a two independent samples Z-test , or just a two sample Z-test . It may be the most commonly employed of all statistical tests. It is also common to see the very similar two sample t-test , which is different only in that it uses the Student t distribution rather than the normal (Gaussian) distribution to calculate the p-values. In fact, there are quite a few minor variations on this testing framework, including “one sided” and “two sided” tests, and tests based on different ways of pooling the variance. Due to the CLT, if the sample size is modestly large (which is the case here), the results of all of these tests will be almost identical. For simplicity, we only cover the Z-test in this course.
The tests for comparing proportions and means presented above are quite similar in many ways. To provide one more example of a hypothesis test that is somewhat different, we consider a test for a correlation coefficient.
Recall that the sample correlation coefficient \(\hat{r}\) is used to assess the relationship, or association, between two quantities X and Y that are measured on the same units. For example, we may ask whether two biomarkers, serum creatinine and D-dimer, are correlated with each other. These biomarkers are both commonly used in medical settings and are obtained using blood tests. D-dimer is used to assess whether a person has blood clots, and serum creatinine is used to measure kidney performance.
Suppose we are interested in whether there is a correlation in the population between D-dimer and serum creatinine. The population correlation coefficient between these two quantitites can be denoted \(r\) . Our null hypothesis is \(r=0\) . Suppose that we observe a sample correlation coefficient of \(\hat{r}=0.15\) , using an independent and identically distributed sample of pairs \((x, y)\) , where \(x\) is a D-dimer measurement and \(y\) is a serum creatinine measurement. Are these data consistent with the null hypothesis?
As above, we proceed by constructing a test statistic by taking the estimated statistic and dividing it by its standard error. The approximate standard error for \(\hat{r}\) is \(1/\sqrt{n}\) , where \(n\) is the sample size. The test statistic is therefore \(\sqrt{n}\cdot \hat{r} \approx 1.48\) .
We now calibrate this test statistic by comparing it to a standard normal reference distribution. Recall from the empirical rule that 5% of the time, a standard normal value falls outside the interval (-2, 2). Therefore, if the test statistic is smaller than 2 in magnitude, as is the case here, its p-value is greater than 0.05. Thus, in this case we know that the p-value will exceed 0.05 without calculating it, and therefore there is no basis for claiming that D-dimer and serum creatinine levels are correlated in this population.
A p-value is the most common way of calibrating evidence. Smaller p-values indicate stronger evidence against a null hypothesis. By convention, if the p-value is smaller than some threshold, usually 0.05, we reject the null hypothesis and declare a finding to be “statistically significant”. How can we understand more deeply what this means? One major concern should be obtaining a small p-value when the null hypothesis is true. If the null hypothesis is true, then it is incorrect to reject it. If we reject the null hypothesis, we are making a false claim. This can never be prevented with complete certainty, but we would like to have a very clear understanding of how likely it is to reject the null hypothesis when the null hypothesis is in fact true.
P-values have a special property that when the null distribution is true, the probability of observing a p-value smaller than 0.05 is 0.05 (5%). In fact, the probability of observing a p-value smaller than \(t\) is equal to \(t\) , for any threshold \(t\) . For example, the probability of observing a p-value smaller than 0.1, when the null hypothesis is true, is 10%.
This fact gives a more concrete understanding of how strong the evidence is for a particular p-value. If we always reject the null hypothesis when the p-value is 0.1 or smaller, then over the long run we will reject the null hypothesis 10% of the time when the null hypothesis is true. If we always reject the null hypothesis when the p-value is 0.05 or smaller, then over the long run we will reject the null hypothesis 5% of the time when the null hypothesis is true.
The approach to hypothesis testing discussed above largely follows the framework developed by RA Fisher around 1925. Note that although we mentioned the alternative hypothesis above, we never actually used it. A more elaborate approach to hypothesis testing was developed somewhat later by Egon Pearson and Jerzy Neyman. The “Neyman-Pearson” approach to hypothesis testing is even more formal than Fisher’s approach, and is most suited to highly planned research efforts in which the study is carefully designed, then executed. While ideally all research projects should be carried out this way, in reality we often conduct research using data that are already available, rather than using data that are specifically collected to address the research question.
Neyman-Pearson hypothesis testing involves specifying an alternative hypothesis that we anticipate encountering. Usually this alternative hypothesis represents a realistic guess about what we might find once the data are collected. In each of the three examples above, imagine that the data are not yet collected, and we are asked to specify an alternative hypothesis. We may arrive at the following:
In comparing risk perceptions for COVID, we may anticipate that older people will perceive a 30% risk of dying, and younger people will anticipate a 5% risk of dying.
In comparing the number of charge cycles for two types of batteries, we may anticipate that batter type A will have on average 500 charge cycles, and battery type B will have on average 400 charge cycles.
In assessing the correlation between D-dimer and serum creatinine levels, we may anticipate a correlation of 0.3.
Note that none of the numbers stated here are data-driven – they are specified before any data are collected, so they do not match the results from the data, which were collected only later. These alternative hypotheses are all essentially speculations, based perhaps on related data or theoretical considerations.
There are several benefits of specifying an explicit alternative hypothesis, as done here, even though it is not strictly necessary and can be avoided entirely by adopting Fisher’s approach to hypothesis testing. One benefit of specifying an alternative hypothesis is that we can use it to assess the power of our planned study, which can in turn inform the design of the study, in particular the sample size. The power is the probability of rejecting the null hypothesis when the alternative hypothesis is true. That is, it is the probability of discovering something real. The power should be contrasted with the level of a hypothesis test, which is the probability of rejecting the null hypothesis when the null hypothesis is true. That is, the level is the probability of “discovering” something that is not real.
To calculate the power, recall that for many of the test statistics that we are considering here, the test statistic has the form \(\hat{\theta}/{\rm SE}(\hat{\theta})\) , where \(\hat{\theta}\) is an estimate. For example, \(\hat{\theta}\) ) may be the correlation coefficient between D-dimer and serum creatinine levels. As stated above, the power is the probability of rejecting the null hypothesis when the alternative hypothesis is true. Suppose we decide to reject the null hypothesis when the test statistic is greater than 2, which is approximately equivalent to rejecting the null hypothesis when the p-value is less than 0.05. The following calculation tells us how to obtain the power in this setting:
Under the alternative hypothesis, \(\sqrt{n}(\hat{r} - r)\) approximately follows a standard normal distribution. Therefore, if \(r\) and \(n\) are given, we can easily use the computer to obtain the probability of observing a value greater than \(2 - \sqrt{n}r\) . This gives us the power of the test. For example, if we anticipate \(r=0.3\) and plan to collect data for \(n=100\) observations, the power is 0.84. This is generally considered to be good power – if the true value of \(r\) is in fact 0.3, we would reject the null hypothesis 84% of the time.
A study usually has poor power because it has too small of a sample size. Poorly powered studies can be very misleading, but since large sample sizes are expensive to collect, a lot of research is conducted using sample sizes that yield moderate or even low power. If a study has low power, it is unlikely to reject the null hypothesis even when the alternative hypothesis is true, but it remains possible to reject the null hypothesis when the null hypothesis is true (usually this probability is 5%). Therefore the most likely outcome of a poorly powered study may be an incorrectly rejected null hypothesis.
Statistics By Jim
Making statistics intuitive
By Jim Frost 10 Comments
A test statistic assesses how consistent your sample data are with the null hypothesis in a hypothesis test. Test statistic calculations take your sample data and boil them down to a single number that quantifies how much your sample diverges from the null hypothesis. As a test statistic value becomes more extreme, it indicates larger differences between your sample data and the null hypothesis.
When your test statistic indicates a sufficiently large incompatibility with the null hypothesis, you can reject the null and state that your results are statistically significant—your data support the notion that the sample effect exists in the population . To use a test statistic to evaluate statistical significance, you either compare it to a critical value or use it to calculate the p-value .
Statisticians named the hypothesis tests after the test statistics because they’re the quantity that the tests actually evaluate. For example, t-tests assess t-values, F-tests evaluate F-values, and chi-square tests use, you guessed it, chi-square values.
In this post, learn about test statistics, how to calculate them, interpret them, and evaluate statistical significance using the critical value and p-value methods.
Each test statistic has its own formula. I present several common test statistics examples below. To see worked examples for each one, click the links to my more detailed articles.
T-value for 1-sample t-test | Take the sample mean, subtract the hypothesized mean, and divide by the . | |
T-value for 2-sample t-test | Take one sample mean, subtract the other, and divide by the pooled standard deviation. | |
F-value for F-tests and ANOVA | Calculate the ratio of two . | |
Chi-squared value (χ ) for a Chi-squared test | Sum the squared differences between observed and expected values divided by the expected values. |
In the formulas above, it’s helpful to understand the null condition and the test statistic value that occurs when your sample data match that condition exactly. Also, it’s worthwhile knowing what causes the test statistics to move further away from the null value, potentially becoming significant. Test statistics are statistically significant when they exceed a critical value.
All these test statistics are ratios, which helps you understand their null values.
When a t-value equals 0, it indicates that your sample data match the null hypothesis exactly.
For a 1-sample t-test, when the sample mean equals the hypothesized mean, the numerator is zero, which causes the entire t-value ratio to equal zero. As the sample mean moves away from the hypothesized mean in either the positive or negative direction, the test statistic moves away from zero in the same direction.
A similar case exists for 2-sample t-tests. When the two sample means are equal, the numerator is zero, and the entire test statistic ratio is zero. As the two sample means become increasingly different, the absolute value of the numerator increases, and the t-value becomes more positive or negative.
Related post : How T-tests Work
When an F-value equals 1, it indicates that the two variances in the numerator and denominator are equal, matching the null hypothesis.
As the numerator and denominator become less and less similar, the F-value moves away from one in either direction.
Related post : The F-test in ANOVA
When a chi-squared value equals 0, it indicates that the observed values always match the expected values. This condition causes the numerator to equal zero, making the chi-squared value equal zero.
As the observed values progressively fail to match the expected values, the numerator increases, causing the test statistic to rise from zero.
Related post : How a Chi-Squared Test Works
You’ll never see a test statistic that equals the null value precisely in practice. However, trivial differences been sample values and the null value are not uncommon.
Test statistics are unitless. This fact can make them difficult to interpret on their own. You know they evaluate how well your data agree with the null hypothesis. If your test statistic is extreme enough, your data are so incompatible with the null hypothesis that you can reject it and conclude that your results are statistically significant. But how does that translate to specific values of your test statistic? Where do you draw the line?
For instance, t-values of zero match the null value. But how far from zero should your t-value be to be statistically significant? Is 1 enough? 2? 3? If your t-value is 2, what does it mean anyway? In this case, we know that the sample mean doesn’t equal the null value, but how exceptional is it? To complicate matters, the dividing line changes depending on your sample size and other study design issues.
Similar types of questions apply to the other test statistics too.
To interpret individual values of a test statistic, we need to place them in a larger context. Towards this end, let me introduce you to sampling distributions for test statistics!
Performing a hypothesis test on a sample produces a single test statistic. Now, imagine you carry out the following process:
This process produces the distribution of test statistic values that occurs when the effect does not exist in the population (i.e., the null hypothesis is true). Statisticians refer to this type of distribution as a sampling distribution, a kind of probability distribution.
Why would we need this type of distribution?
It provides the larger context required for interpreting a test statistic. More specifically, it allows us to compare our study’s single test statistic to values likely to occur when the null is true. We can quantify our sample statistic’s rareness while assuming the effect does not exist in the population. Now that’s helpful!
Fortunately, we don’t need to collect many random samples to create this distribution! Statisticians have developed formulas allowing us to estimate sampling distributions for test statistics using the sample data.
To evaluate your data’s compatibility with the null hypothesis, place your study’s test statistic in the distribution.
Related post : Understanding Probability Distributions
Suppose our t-test produces a t-value of two. That’s our test statistic. Let’s see where it fits in.
The sampling distribution below shows a t-distribution with 20 degrees of freedom, equating to a 1-sample t-test with a sample size of 21. The distribution centers on zero because it assumes the null hypothesis is correct. When the null is true, your analysis is most likely to obtain a t-value near zero and less likely to produce t-values further from zero in either direction.
The sampling distribution indicates that our test statistic is somewhat rare when we assume the null hypothesis is correct. However, the chances of observing t-values from -2 to +2 are not totally inconceivable. We need a way to quantify the likelihood.
From this point, we need to use the sampling distributions’ ability to calculate probabilities for test statistics.
Related post : Sampling Distributions Explained
The significance level uses critical values to define how far the test statistic must be from the null value to reject the null hypothesis. When the test statistic exceeds a critical value, the results are statistically significant.
The percentage of the area beneath the sampling distribution curve that is shaded represents the probability that the test statistic will fall in those regions when the null is true. Consequently, to depict a significance level of 0.05, I’ll shade 5% of the sampling distribution furthest away from the null value.
The two shaded areas are equidistant from the null value in the center. Each region has a likelihood of 0.025, which sums to our significance level of 0.05. These shaded areas are the critical regions for a two-tailed hypothesis test. Let’s return to our example t-value of 2.
Related post : What are Critical Values?
In this example, the critical values are -2.086 and +2.086. Our test statistic of 2 is not statistically significant because it does not exceed the critical value.
Other hypothesis tests have their own test statistics and sampling distributions, but their processes for critical values are generally similar.
Learn how to find critical values for test statistics using tables:
Related post : Understanding Significance Levels
P-values are the probability of observing an effect at least as extreme as your sample’s effect if you assume no effect exists in the population.
Test statistics represent effect sizes in hypothesis tests because they denote the difference between your sample effect and no effect —the null hypothesis. Consequently, you use the test statistic to calculate the p-value for your hypothesis test.
The above p-value definition is a bit tortuous. Fortunately, it’s much easier to understand how test statistics and p-values work together using a sampling distribution graph.
Let’s use our hypothetical test statistic t-value of 2 for this example. However, because I’m displaying the results of a two-tailed test, I need to use t-values of +2 and -2 to cover both tails.
Related post : One-tailed vs. Two-Tailed Hypothesis Tests
The graph below displays the probability of t-values less than -2 and greater than +2 using the area under the curve. This graph is specific to our t-test design (1-sample t-test with N = 21).
The sampling distribution indicates that each of the two shaded regions has a probability of 0.02963—for a total of 0.05926. That’s the p-value! The graph shows that the test statistic falls within these areas almost 6% of the time when the null hypothesis is true in the population.
While this likelihood seems small, it’s not low enough to justify rejecting the null under the standard significance level of 0.05. P-value results are always consistent with the critical value method. Learn more about using test statistics to find p values .
While test statistics are a crucial part of hypothesis testing, you’ll probably let your statistical software calculate the p-value for the test. However, understanding test statistics will boost your comprehension of what a hypothesis test actually assesses.
Related post : Interpreting P-values
July 5, 2024 at 8:21 am
“As the observed values progressively fail to match the observed values, the numerator increases, causing the test statistic to rise from zero”.
Sir, this sentence is written in the Chi-squared Test heading. There the observed value is written twice. I think the second one to be replaced with ‘expected values’.
July 5, 2024 at 4:10 pm
Thanks so much, Dr. Raj. You’re correct about the typo and I’ve made the correction.
May 9, 2024 at 1:40 am
Thank you very much (great page on one and two-tailed tests)!
May 6, 2024 at 12:17 pm
I would like to ask a question. If only positive numbers are the possible values in a sample (e.g. absolute values without 0), is it meaningful to test if the sample is significantly different from zero (using for example a one sample t-test or a Wilcoxon signed-rank test) or can I assume that if given a large enough sample, the result will by definition be significant (even if a small or very variable sample results in a non-significant hypothesis test).
Thank you very much,
May 6, 2024 at 4:35 pm
If you’re talking about the raw values you’re assessing using a one-sample t-test, it doesn’t make sense to compare them to zero given your description of the data. You know that the mean can’t possibly equal zero. The mean must be some positive value. Yes, in this scenario, if you have a large enough sample size, you should get statistically significant results. So, that t-test isn’t tell you anything that you don’t already know!
However, you should be aware of several things. The 1-sample test can compare your sample mean to values other than zero. Typically, you’ll need to specify the value of the null hypothesis for your software. This value is the comparison value. The test determines whether your sample data provide enough evidence to conclude that the population mean does not equal the null hypothesis value you specify. You’ll need to specify the value because there is no obvious default value to use. Every 1-sample t-test has its subject-area context with a value that makes sense for its null hypothesis value and it is frequently not zero.
I suspect that you’re getting tripped up with the fact that t-tests use a t-value of zero for its null hypothesis value. That doesn’t mean your 1-sample t-test is comparing your sample mean to zero. The test converts your data to a single t-value and compares the t-value to zero. But your actual null hypothesis value can be something else. It’s just converting your sample to a standardized value to use for testing. So, while the t-test compares your sample’s t-value to zero, you can actually compare your sample mean to any value you specify. You need to use a value that makes sense for your subject area.
I hope that makes sense!
May 8, 2024 at 8:37 am
Thank you very much Jim, this helps a lot! Actually, the value I would like to compare my sample to is zero, but I just couldn’t find the right way to test it apparently (it’s about EEG data). The original data was a sample of numbers between -1 and +1, with the question if they are significantly different from zero in either direction (in which case a one sample t-test makes sense I guess, since the sample mean can in fact be zero). However, since a sample mean of 0 can also occur if half of the sample differs in the negative, and the other half in the positive direction, I also wanted to test if there is a divergence from 0 in ‘absolute’ terms – that’s how the absolute valued numbers came about (I know that absolute values can also be zero, but in this specific case, they were all positive numbers) And a special thanks for the last paragraph – I will definitely keep in mind, it is a potential point of confusion.
May 8, 2024 at 8:33 pm
You can use a 1-sample t test for both cases but you’ll need to set them up slightly different. To detect a positive or negative difference from zero, use a 2-tailed test. For the case with absolute values, use a one-tailed test with a critical region in the positive end. To learn more, read about One- and Two-Tailed Tests Explained . Use zero for the comparison value in both cases.
February 12, 2024 at 1:00 am
Very helpful and well articulated! Thanks Jim 🙂
September 18, 2023 at 10:01 am
Thank you for brief explanation.
July 25, 2022 at 8:32 am
the content was helpful to me. thank you
Run a free plagiarism check in 10 minutes, generate accurate citations for free.
Published on July 17, 2020 by Rebecca Bevans . Revised on June 22, 2023.
The test statistic is a number calculated from a statistical test of a hypothesis. It shows how closely your observed data match the distribution expected under the null hypothesis of that statistical test.
The test statistic is used to calculate the p value of your results, helping to decide whether to reject your null hypothesis.
What exactly is a test statistic, types of test statistics, interpreting test statistics, reporting test statistics, other interesting articles, frequently asked questions about test statistics.
A test statistic describes how closely the distribution of your data matches the distribution predicted under the null hypothesis of the statistical test you are using.
The distribution of data is how often each observation occurs, and can be described by its central tendency and variation around that central tendency. Different statistical tests predict different types of distributions, so it’s important to choose the right statistical test for your hypothesis.
The test statistic summarizes your observed data into a single number using the central tendency, variation, sample size, and number of predictor variables in your statistical model.
Generally, the test statistic is calculated as the pattern in your data (i.e., the correlation between variables or difference between groups) divided by the variance in the data (i.e., the standard deviation ).
Discover proofreading & editing
Below is a summary of the most common test statistics, their hypotheses, and the types of statistical tests that use them.
Different statistical tests will have slightly different ways of calculating these test statistics, but the underlying hypotheses and interpretations of the test statistic stay the same.
Test statistic | Null and alternative hypotheses | Statistical tests that use it |
---|---|---|
value | The means of two groups are equal The means of two groups are not equal | test |
value | The means of two groups are equal The means of two groups are not equal | test |
value | The variation among two or more groups is greater than or equal to the variation between the groups The variation among two or more groups is smaller than the variation between the groups | |
-value | Two samples are independent Two samples are not independent (i.e., they are correlated) | correlation tests |
In practice, you will almost always calculate your test statistic using a statistical program (R, SPSS, Excel, etc.), which will also calculate the p value of the test statistic. However, formulas to calculate these statistics by hand can be found online.
The t value of the regression test is 2.36 – this is your test statistic.
For any combination of sample sizes and number of predictor variables, a statistical test will produce a predicted distribution for the test statistic. This shows the most likely range of values that will occur if your data follows the null hypothesis of the statistical test.
The more extreme your test statistic – the further to the edge of the range of predicted test values it is – the less likely it is that your data could have been generated under the null hypothesis of that statistical test.
The agreement between your calculated test statistic and the predicted values is described by the p value . The smaller the p value, the less likely your test statistic is to have occurred under the null hypothesis of the statistical test.
Because the test statistic is generated from your observed data, this ultimately means that the smaller the p value, the less likely it is that your data could have occurred if the null hypothesis was true.
Test statistics can be reported in the results section of your research paper along with the sample size, p value of the test, and any characteristics of your data that will help to put these results into context.
Whether or not you need to report the test statistic depends on the type of test you are reporting.
Which statistics to report | |
---|---|
Correlation and regression tests | or regression coefficient for each predictor variable value for each predictor |
Tests of difference between groups | value for the test statistic |
By surveying a random subset of 100 trees over 25 years we found a statistically significant ( p < 0.01) positive correlation between temperature and flowering dates ( R 2 = 0.36, SD = 0.057).
In our comparison of mouse diet A and mouse diet B, we found that the lifespan on diet A ( M = 2.1 years; SD = 0.12) was significantly shorter than the lifespan on diet B ( M = 2.6 years; SD = 0.1), with an average difference of 6 months ( t (80) = -12.75; p < 0.01).
If you want to know more about statistics , methodology , or research bias , make sure to check out some of our other articles with explanations and examples.
Methodology
Research bias
A test statistic is a number calculated by a statistical test . It describes how far your observed data is from the null hypothesis of no relationship between variables or no difference among sample groups.
The test statistic tells you how different two or more groups are from the overall population mean , or how different a linear slope is from the slope predicted by a null hypothesis . Different test statistics are used in different statistical tests.
The formula for the test statistic depends on the statistical test being used.
Generally, the test statistic is calculated as the pattern in your data (i.e. the correlation between variables or difference between groups) divided by the variance in the data (i.e. the standard deviation ).
The test statistic you use will be determined by the statistical test.
You can choose the right statistical test by looking at what type of data you have collected and what type of relationship you want to test.
The test statistic will change based on the number of observations in your data, how variable your observations are, and how strong the underlying patterns in the data are.
For example, if one data set has higher variability while another has lower variability, the first data set will produce a test statistic closer to the null hypothesis , even if the true correlation between two variables is the same in either data set.
Statistical significance is a term used by researchers to state that it is unlikely their observations could have occurred under the null hypothesis of a statistical test . Significance is usually denoted by a p -value , or probability value.
Statistical significance is arbitrary – it depends on the threshold, or alpha value, chosen by the researcher. The most common threshold is p < 0.05, which means that the data is likely to occur less than 5% of the time under the null hypothesis .
When the p -value falls below the chosen alpha value, then we say the result of the test is statistically significant.
If you want to cite this source, you can copy and paste the citation or click the “Cite this Scribbr article” button to automatically add the citation to our free Citation Generator.
Bevans, R. (2023, June 22). Test statistics | Definition, Interpretation, and Examples. Scribbr. Retrieved August 12, 2024, from https://www.scribbr.com/statistics/test-statistic/
Other students also liked, understanding p values | definition and examples, choosing the right statistical test | types & examples, what is effect size and why does it matter (examples), what is your plagiarism score.
Statistical hypothesis testing (also 'confirmatory data analysis') is used in inferential statistics to either confirm or falsify a hypothesis based on empirical observations .
An example: It is assumed, that people in the US, over time, are getting older (on average). In this case, the hypothesis to be confirmed is: 'the average age of people in the US is rising'. This is called the alternative hypothesis , whereas the current opinion 'the average age of people in the US stays the same' is called the null hypothesis . The goal of a statistical test would be to either verify of falsify the alternative hypothesis.
In hypothesis testing, we differentiate between parametric and non-parametric tests. In parametric tests we compare location and dispersion parameters of two samples and check for compliance. Examples for parametric tests are the t-test , f-test and the χ2-test. In nonparametric tests on the other hand, no assumptions about probability distributions of the population which is being assessed are being made. Examples are the Kolmogorov-Smirnov test, the chi-square test and the Shapiro-Wilk test.
Performing hypothesis tests: In order to perform statistical hypothesis testing, we first have to collect the according empirical data (for example: age reached of 100 people, born in 1900 and 1920 respectively). Depending on the hypothesis made and the resulting test procedure, a mathematically defined test statistic (f-statistic, t-statistic, …) is deducted from the observed data. Based on this value, we can determine whether the null hypothesis can be rejected or not – accounting for a specified rate of reliability (1- error probability). The null hypothesis should only be rejected based on a very low probability of error (p≤5%). Since errors when verifying or falsifying hypotheses cannot be generally excluded, errors of the first kind (=a true null hypothesis is incorrectly rejected, also: type I error) and errors of the second kind (= a true alternative hypothesis is incorrectly rejected, also: type II errors) are usually explicitly specified.
Please note that the definitions in our statistics encyclopedia are simplified explanations of terms. Our goal is to make the definitions accessible for a broad audience; thus it is possible that some definitions do not adhere entirely to scientific standards.
IMAGES
COMMENTS
Step 5: Present your findings. The results of hypothesis testing will be presented in the results and discussion sections of your research paper, dissertation or thesis.. In the results section you should give a brief summary of the data and a summary of the results of your statistical test (for example, the estimated difference between group means and associated p-value).
Formulate the Hypotheses: Write your research hypotheses as a null hypothesis (H 0) and an alternative hypothesis (H A).; Data Collection: Gather data specifically aimed at testing the hypothesis.; Conduct A Test: Use a suitable statistical test to analyze your data.; Make a Decision: Based on the statistical test results, decide whether to reject the null hypothesis or fail to reject it.
In hypothesis testing, the goal is to see if there is sufficient statistical evidence to reject a presumed null hypothesis in favor of a conjectured alternative hypothesis.The null hypothesis is usually denoted \(H_0\) while the alternative hypothesis is usually denoted \(H_1\). An hypothesis test is a statistical decision; the conclusion will either be to reject the null hypothesis in favor ...
A statistical hypothesis test is a method of statistical inference used to decide whether the data sufficiently support a particular hypothesis. A statistical hypothesis test typically involves a calculation of a test statistic. Then a decision is made, either by comparing the test statistic to a critical value or equivalently by evaluating a p ...
Hypothesis testing is a tool for making statistical inferences about the population data. It is an analysis tool that tests assumptions and determines how likely something is within a given standard of accuracy. Hypothesis testing provides a way to verify whether the results of an experiment are valid. A null hypothesis and an alternative ...
A statistical hypothesis is an assumption about a population parameter.. For example, we may assume that the mean height of a male in the U.S. is 70 inches. The assumption about the height is the statistical hypothesis and the true mean height of a male in the U.S. is the population parameter.. A hypothesis test is a formal statistical test we use to reject or fail to reject a statistical ...
HYPOTHESIS TESTING. A clinical trial begins with an assumption or belief, and then proceeds to either prove or disprove this assumption. In statistical terms, this belief or assumption is known as a hypothesis. Counterintuitively, what the researcher believes in (or is trying to prove) is called the "alternate" hypothesis, and the opposite ...
Hypothesis testing is a form of inferential statistics that allows us to draw conclusions about an entire population based on a representative sample. You gain tremendous benefits by working with a sample. In most cases, it is simply impossible to observe the entire population to understand its properties.
Hypothesis testing is an act in statistics whereby an analyst tests an assumption regarding a population parameter. The methodology employed by the analyst depends on the nature of the data used ...
A hypothesis test is a statistical inference method used to test the significance of a proposed (hypothesized) relation between population statistics (parameters) and their corresponding sample estimators. In other words, hypothesis tests are used to determine if there is enough evidence in a sample to prove a hypothesis true for the entire population. The test considers two hypotheses: the ...
Components of a Formal Hypothesis Test. The null hypothesis is a statement about the value of a population parameter, such as the population mean (µ) or the population proportion (p).It contains the condition of equality and is denoted as H 0 (H-naught).. H 0: µ = 157 or H0 : p = 0.37. The alternative hypothesis is the claim to be tested, the opposite of the null hypothesis.
Significance tests give us a formal process for using sample data to evaluate the likelihood of some claim about a population value. Learn how to conduct significance tests and calculate p-values to see how likely a sample result is to occur by random chance. You'll also see how we use p-values to make conclusions about hypotheses.
Definition: statistical procedure. Hypothesis testing is a statistical procedure in which a choice is made between a null hypothesis and an alternative hypothesis based on information in a sample. The end result of a hypotheses testing procedure is a choice of one of the following two possible conclusions:
Statistics - Hypothesis Testing, Sampling, Analysis: Hypothesis testing is a form of statistical inference that uses data from a sample to draw conclusions about a population parameter or a population probability distribution. First, a tentative assumption is made about the parameter or distribution. This assumption is called the null hypothesis and is denoted by H0.
Hypothesis testing is a statistical method used to determine if there is enough evidence in a sample data to draw conclusions about a population. It involves formulating two competing hypotheses, the null hypothesis (H0) and the alternative hypothesis (Ha), and then collecting data to assess the evidence. ...
A statistical hypothesis test may return a value called p or the p-value. This is a quantity that we can use to interpret or quantify the result of the test and either reject or fail to reject the null hypothesis. This is done by comparing the p-value to a threshold value chosen beforehand called the significance level.
What is Hypothesis Testing? Hypothesis testing in statistics uses sample data to infer the properties of a whole population.These tests determine whether a random sample provides sufficient evidence to conclude an effect or relationship exists in the population. Researchers use them to help separate genuine population-level effects from false effects that random chance can create in samples.
Statistical tests are used in hypothesis testing. They can be used to: ... Test statistics | Definition, Interpretation, and Examples The test statistic is a number, calculated from a statistical test, used to find if your data could have occurred under the null hypothesis. 255.
Hypothesis tests # Formal hypothesis testing is perhaps the most prominent and widely-employed form of statistical analysis. It is sometimes seen as the most rigorous and definitive part of a statistical analysis, but it is also the source of many statistical controversies. The currently-prevalent approach to hypothesis testing dates to developments that took place between 1925 and 1940 ...
Test statistics represent effect sizes in hypothesis tests because they denote the difference between your sample effect and no effect —the null hypothesis. Consequently, you use the test statistic to calculate the p-value for your hypothesis test. The above p-value definition is a bit tortuous.
The test statistic is a number calculated from a statistical test of a hypothesis. It shows how closely your observed data match the distribution expected under the null hypothesis of that statistical test. The test statistic is used to calculate the p value of your results, helping to decide whether to reject your null hypothesis.
Hypothesis Testing Steps. There are 5 main hypothesis testing steps, which will be outlined in this section. The steps are: Determine the null hypothesis: In this step, the statistician should ...
Performing hypothesis tests: In order to perform statistical hypothesis testing, we first have to collect the according empirical data (for example: age reached of 100 people, born in 1900 and ...