A parametric test is a statistical method that makes specific assumptions about the data it analyzes, most importantly that the data follows a bell-shaped (normal) distribution. Tests like the t-test, ANOVA, and Pearson correlation all fall into this category. They work by using the mathematical properties of that distribution, specifically means and standard deviations, to determine whether differences or relationships in your data are real or just due to chance.
If you’re taking a statistics course or trying to analyze data for a project, understanding parametric tests matters because choosing the wrong type of test can lead to misleading results. Here’s what you need to know about how they work, when to use them, and what to do when your data doesn’t cooperate.
How Parametric Tests Work
Every parametric test starts with the same basic logic: it assumes that the population your data came from has a known shape, specifically a normal distribution. That’s the familiar bell curve where most values cluster around the average and fewer values appear at the extremes. The test then uses the mean and standard deviation of your sample to calculate how likely your results would be if there were no real effect or difference.
This reliance on distribution shape is what puts the “parametric” in parametric test. The word refers to parameters, the numerical characteristics (like the mean and standard deviation) that define a distribution’s shape. Because these tests assume a specific shape, they can extract more information from your data than methods that don’t make that assumption. The tradeoff is that if your data badly violates the assumption, the results may not be trustworthy.
The Three Core Assumptions
Before running a parametric test, your data needs to meet three conditions. Not perfectly, but approximately.
- Normal distribution. The data in each group should roughly follow a bell curve. Small departures from normality are usually fine, especially with larger samples, but severely skewed data or data with extreme outliers can distort your results.
- Equal variance. When comparing groups, the spread of data in each group should be similar. If one group’s values are tightly clustered and another’s are all over the place, the test may give unreliable conclusions.
- Independence. Each data point should be unrelated to the others. Measuring the same person twice, for example, creates paired data that requires a specific type of test designed for that situation. Ignoring this assumption is often the most damaging mistake.
Your data also needs to be measured on a continuous scale, meaning values like weight, temperature, blood pressure, or test scores. Parametric tests do not apply to strictly categorical data like yes/no responses or color preferences.
Common Parametric Tests and When to Use Them
T-Tests
The t-test compares means between two groups. It comes in three versions depending on your situation. A one-sample t-test checks whether your sample’s average differs from a known or hypothetical value, like testing whether the average height in your class differs from the national average. An independent samples t-test (sometimes called an unpaired t-test) compares the averages of two separate groups, such as a treatment group versus a control group. A paired samples t-test compares two measurements from the same subjects, like before-and-after scores in a study where the same people are measured at two time points.
ANOVA
ANOVA extends the logic of the t-test to three or more groups. If you wanted to compare average test scores across four different teaching methods, ANOVA is the appropriate choice. It tells you whether at least one group’s mean differs significantly from the others, though it won’t tell you which specific groups differ without follow-up testing. A one-way ANOVA uses one grouping variable (like teaching method), while a two-way ANOVA lets you examine two grouping variables at once (like teaching method and class size).
Pearson Correlation
Pearson correlation measures the strength and direction of a linear relationship between two continuous variables. It produces a value between -1 and +1. A value near +1 means both variables increase together, a value near -1 means one increases as the other decreases, and a value near 0 means no linear relationship exists.
Linear Regression
Linear regression goes a step beyond correlation by modeling the relationship between variables as a line, allowing you to predict one variable based on another. Like the other parametric methods, it assumes the data’s residuals (the gaps between predicted and actual values) are normally distributed.
Why Parametric Tests Are More Powerful
When your data genuinely meets the assumptions, parametric tests have higher statistical power than their non-parametric alternatives. Statistical power is the ability to detect a real effect when one exists. In practical terms, this means parametric tests are more likely to find a statistically significant result if a true difference or relationship is present in your data. Non-parametric tests, because they don’t use all the information embedded in the distribution’s shape, are less sensitive and more likely to miss real effects.
This power advantage is the reason parametric tests remain the standard tools in most fields. You wouldn’t want to use a less sensitive test if your data supports a more sensitive one.
How to Check Whether Your Data Qualifies
Checking normality is the most common hurdle. You can approach it visually, statistically, or both. Visual methods include histograms (does the shape look roughly bell-curved?), boxplots (are the data symmetric without extreme outliers?), and Q-Q plots, which plot your data against what a perfectly normal distribution would look like. If the points fall close to a straight diagonal line, your data is approximately normal.
For a more formal check, statistical tests can compare your data’s distribution to a theoretical normal curve. The Shapiro-Wilk test is the most widely recommended option and is built into most statistical software. If the test returns a non-significant result (typically p > 0.05), your data is consistent with normality. A significant result suggests the data departs from a normal distribution. Experts recommend using both visual and statistical approaches together rather than relying on either one alone.
For equal variance, Levene’s test is the standard check. It tests whether the variability in each group is roughly equivalent. Most software runs this automatically when you perform a t-test or ANOVA.
The Central Limit Theorem: A Safety Net
Here’s where things get more forgiving. The central limit theorem states that as your sample size grows, the distribution of sample means approaches a normal distribution regardless of the shape of the underlying data. The commonly cited threshold is a sample size of 30 or more. At that point, the sampling distribution is considered close enough to normal that parametric tests produce reliable results even if the raw data itself isn’t perfectly bell-shaped.
This means that with large enough samples, the normality assumption becomes much less of a concern. It doesn’t disappear entirely, as extreme skewness or heavy outliers can still cause problems, but moderate departures from normality are effectively neutralized. For small samples, however, the assumption matters much more, and you should check it carefully.
What to Use When Assumptions Are Violated
When your data doesn’t meet parametric assumptions and your sample is too small for the central limit theorem to help, non-parametric tests are the alternative. These tests make no assumptions about the distribution’s shape. They typically work by ranking the data from smallest to largest and analyzing the ranks rather than the raw values.
Each major parametric test has a non-parametric counterpart:
- Independent samples t-test → Mann-Whitney U test
- Paired samples t-test → Wilcoxon signed-rank test
- One-way ANOVA → Kruskal-Wallis test
- Pearson correlation → Spearman rank correlation
Non-parametric tests are valid for both normally and non-normally distributed data, which raises an obvious question: why not just use them all the time? The answer comes back to power. When the parametric assumptions hold, non-parametric tests are less likely to detect real effects. You’re giving up sensitivity for flexibility. If your data supports a parametric approach, that’s the stronger choice.
Choosing the Right Test
The decision tree is simpler than it looks. Start with your research question: are you comparing group averages, or measuring a relationship between variables? Then count your groups. Two groups point toward a t-test, three or more toward ANOVA, and relationships between continuous variables toward correlation or regression. Next, check your assumptions. If your data is continuous, approximately normal (or your sample exceeds 30), and your groups have similar spread, go parametric. If not, switch to the non-parametric equivalent.
One common mistake is treating ordinal data (like Likert scale ratings from 1 to 5) as continuous and running parametric tests on it. Whether this is acceptable is debated, but the safest approach with clearly ordinal data and small samples is to use non-parametric methods. With larger samples and many scale points, parametric tests on ordinal data often perform well in practice.