Must all raw variables be normal?

No; requirements depend on the chosen model and test.

Are parametric tests always more powerful?

No universal ordering applies; performance depends on the true data-generating process and alternatives.

Can I use a test without checking its assumptions?

Software permits it, but the result may lack its advertised interpretation.

AI Glossary / Statistics and probability / Intermediate

Parametric test

Definition

A parametric test is a statistical test whose reference distribution or inference is derived from a model with specified distributional or parameter assumptions. Common examples include t-tests, classical analysis-of-variance tests, and tests within binomial, Poisson, or regression models. The assumptions differ by procedure; parametric does not mean that every dataset must be normally distributed.

What it means

The test connects a claim about a population or model parameter with the variability expected in repeated samples under that claim. It uses the chosen model to determine how unusual the observed test statistic would be.
A parametric model can be useful when it captures the important features of the data and design. Its apparent precision can be misleading when key assumptions are wrong. The relevant question is not whether the label sounds rigorous, but whether the procedure answers the intended question under defensible conditions.

How it works

Specify the outcome, population, parameter, null hypothesis, and alternative. Then identify a sampling model that reflects the observation process. Continuous measurements, binary outcomes, counts, and repeated observations require different considerations.
Choose a test statistic and its null reference distribution. For example, a normal mean with known variance permits a z statistic; estimating the variance from an independent normal sample leads to a t statistic with the appropriate degrees of freedom.
Check the assumptions that matter for that particular calculation. These may concern independent observations, distributional form, equal variances, the conditional mean relationship, or the covariance structure. Regression normality assumptions, when required for exact inference, concern errors conditional on predictors rather than every raw variable being normal.
Some procedures are reasonably robust to certain deviations in suitable samples, but robustness is not unlimited. Severe outliers, heavy tails, small groups, dependence, or an incorrect model can change error rates and interpretation.
Use diagnostics and study design knowledge together. A normality test alone cannot certify the entire analysis, and choosing a method only after several preliminary significance tests can complicate inferential properties.
Report the effect estimate, uncertainty interval, p-value if relevant, and practical meaning. Consider sensitivity analyses or alternative models where assumptions are questionable. A nonparametric or resampling procedure still needs its own assumptions and may target a different quantity.
Distinguish exact finite-sample inference from an approximation. Large-sample methods may relax some distributional requirements, but large datasets do not automatically repair biased sampling or unmodeled clustering.

Three practical examples

Product evaluation: analysts use a suitable mean-comparison model to assess a prespecified difference in task scores.
AI reliability: a team tests a success-probability benchmark under a justified independent binomial model.
Operational research: a count model assesses event rates while checking whether the observed variation and exposure structure fit its assumptions.

Think of it this way

Think of using a calibrated measuring template. The template translates an observed discrepancy into evidence only when its shape matches the problem. A neatly calculated answer from the wrong template can be less useful than a less precise answer from a model that fits the situation.

Worked example

Consider a fictional process with independent normally distributed measurements and a genuinely known population standard deviation of 6 units. A sample of 36 observations has mean 52. Test H0: population mean equals 50 against a two-sided alternative.
The standard error is 6/sqrt(36) = 1. The z statistic is (52 - 50)/1 = 2. Under H0 and the stated assumptions, the two-sided p-value is 2[1 - Phi(2)], approximately 0.04550.
A corresponding 95% normal confidence interval is 52 plus or minus 1.959964 times 1, approximately 50.0400 to 53.9600. It narrowly excludes 50, consistent with the two-sided 5% test result.
This is deliberately a known-variance example. If 6 were instead the sample standard deviation, using the same normal reference would ignore variance-estimation uncertainty; a one-sample t procedure with 35 degrees of freedom would be the usual exact approach under independent normal sampling.
The observed mean difference is 2 units. Whether that difference matters operationally depends on requirements and costs, not on crossing the 0.05 threshold. The p-value is not the probability that the null hypothesis is true.

Illustrative case study

Illustrative fictional case: Fir Operations compares service-time measurements across several teams. An initial spreadsheet applies a standard test to every ticket as though tickets were independent.
Reviewers find repeated tickets from the same customers and substantial team differences. They specify a model and uncertainty calculation that reflect the grouping, then compare conclusions with an appropriate sensitivity analysis. This fictional case shows that dependence can be more consequential than whether a histogram looks bell-shaped.

Common misconceptions and mistakes

Do not equate parametric with normal in every application. Do not assume that nonparametric means assumption-free. Avoid treating a large sample as a cure for all design problems.
Do not replace a paired analysis with an independent-samples test when observations are linked. Do not claim that statistical significance establishes a large benefit, or that nonrejection proves equality. A method's name does not substitute for a clear target parameter and sampling justification.

Business use

Parametric tests help businesses assess explicit claims efficiently when the model and design are appropriate. Reports should make the assumptions understandable: what is independent, which parameter is tested, and why the reference calculation applies. Pair statistical evidence with effect size, uncertainty, diagnostic findings, and the consequences of the proposed decision.

Check your understanding

Why is a z reference used in the example?

Answer: The model explicitly assumes known population variance and independent normal measurements.

What changes if variance is estimated?

Answer: The appropriate reference must account for that additional uncertainty.

Does p approximately 0.0455 establish business value?

Answer: No.

Related terms

Learn next

Next study the assumptions behind a specific test rather than relying on broad labels. Then compare robust inference, alternative outcome models, and design-based procedures for the same business question.

Sources

Frank Harrell, Glossary of Statistical Terms, parametric test: https://hbiostat.org/glossary/ — primary terminology source. NIST/SEMATECH, Confidence Limits for the Mean: https://www.itl.nist.gov/div898/handbook/eda/section3/eda352.htm — one-sample t procedure and interval interpretation. NIST/SEMATECH, Quantitative Techniques: https://www.itl.nist.gov/div898/handbook/eda/section3/eda35.htm — testing and practical significance. The known-variance illustration and fictional case are original.

Previous: Parametric model Next: Paranoid algorithm