In statistics, assumption testing means checking that your data meets a test's requirements, things like normality, equal variances, and independence, before you trust the result. If you searched "assumption testing" hoping for product-discovery advice, this isn't that. This is the statistics version: the checks that decide whether your t-test, ANOVA, or regression is telling you the truth or quietly lying to you.

Most people run one quick test, see a p-value, and move on. That's where a lot of analyses go wrong. Below, you'll find the core assumptions, the exact test or plot that checks each one, how to read the result, a table of which assumptions each common test needs, and what to actually do when one fails.

What is assumption testing in statistics?

An assumption is a condition your data has to meet for a statistical test to work the way it's supposed to. A hypothesis is the claim you're actually testing (for example, "this drug lowers blood pressure"). Assumption testing checks the conditions. Hypothesis testing checks the claim. You do the first so you can believe the second.

Most of the common tests, like t-tests, ANOVA, and linear regression, are parametric, which means they assume your data follows a particular shape, usually something close to a normal distribution. Those assumptions are the price of admission. When they hold, parametric tests are powerful and efficient. When they're badly violated, the test can hand you a confident answer that's simply wrong.

A few assumptions show up again and again across tests:

Normality: the data (or the model's residuals) is roughly normally distributed.
Homogeneity of variance: the groups you're comparing have similar spread.
Independence: each observation is its own data point, not tangled up with the others.
Linearity: the relationship you're modeling is actually a straight line.
No multicollinearity: your predictors aren't near-duplicates of each other.

Nonparametric tests (Mann-Whitney, Kruskal-Wallis, and friends) make far fewer of these assumptions, which is one of your escape hatches when things go sideways. More on that later.

When should you check your assumptions?

Check them after you fit or choose your model, and before you report anything. For something like regression or ANOVA, you need the model first, because several assumptions are about the residuals (the leftover errors), and you can't look at residuals until the model exists.

Your assumptions will never be perfectly met, because real data is messy. So the question worth asking is whether any violation is big enough to change your conclusion, rather than whether the conditions are spotless (they won't be).

Mild violations are often forgivable. Thanks to the central limit theorem, a reasonably large sample pulls the sampling distribution of the mean toward normal even when the raw data isn't, so a bit of non-normality rarely hurts a t-test or ANOVA. Dependence between observations is the dangerous kind. It can wreck an analysis even when it's hard to spot, which is why statisticians tend to treat independence as the assumption you most need to get right by design.

The core statistical assumptions (and how to test each)

Each assumption below follows the same quick pattern: what it is, how to test it, and how to read the result. Pick the ones that apply to your test (the table in the next section maps that out).

Normality

What it is: the values, or more often the model's residuals, follow an approximately normal distribution.

How to test it: the Shapiro-Wilk test is the usual go-to, with the Kolmogorov-Smirnov test as an alternative. The visual check is a Q-Q plot, where normal data hugs the diagonal line. Skewness and kurtosis numbers add detail about the shape.

How to read it: a Shapiro-Wilk p-value below 0.05 flags a departure from normal. On a Q-Q plot, points drifting off the line at the ends mean heavy tails; a consistent curve means skew. One caveat worth remembering: Shapiro-Wilk gets touchy with large samples (it flags trivial wobbles) and sleepy with small ones. For the deeper version of this in a regression context, see our guide on normality in multiple regression.

Homogeneity of variance

What it is: the groups you're comparing (or the residuals across the range of a regression) have roughly equal spread. The fancy word for the regression version is homoscedasticity.

How to test it: Levene's test is the workhorse, and it holds up well even when the data isn't perfectly normal. Bartlett's test is more powerful but assumes normality, so it's pickier. For regression, a residuals-versus-fitted plot does the visual job: you want a shapeless cloud, not a fan or a funnel.

How to read it: a significant Levene's result (p below 0.05) says the variances differ enough to worry about. A fanning pattern in the residual plot is the visual tell. We go deeper in our piece on homoscedasticity in regression.

Independence and randomness

What it is: each observation stands on its own. One data point shouldn't predict the next.

How to test it: independence is mostly won or lost in how you collect data, so think about it before you ever open your dataset. After the fact, the Durbin-Watson statistic checks for autocorrelation in regression residuals, and a runs test checks whether a sequence looks random.

How to read it: Durbin-Watson runs from 0 to 4, and a value near 2 means no autocorrelation; drift toward 0 or 4 signals that consecutive residuals are related. Take this one seriously. Dependence is the violation most likely to quietly invalidate your results.

Linearity

What it is: for linear models, the relationship between your predictors and the outcome is actually a straight line. (Linearity is easy to skip and a common miss, so it’s worth a deliberate check.)

How to test it: plot residuals against fitted values, or against each predictor, and look for curvature. A plain scatterplot of outcome versus predictor works too. Partial residual plots isolate one predictor at a time.

How to read it: a clear curve or U-shape in the residuals means the straight-line model is missing something, and you may need a transformation or a curved term.

Multicollinearity

What it is: two or more predictors carry nearly the same information, which makes a regression's coefficients unstable and hard to interpret.

How to test it: the Variance Inflation Factor (VIF) is the standard measure, with condition indices as a backup.

How to read it: a common rule of thumb flags VIF above 10, though plenty of analysts start worrying around 5. High VIF doesn't bias your predictions, but it does make individual coefficients untrustworthy. Our multiple linear regression assumptions guide walks through fixes like dropping or combining predictors.

Which assumptions each statistical test needs

Not every test needs every assumption, and mixing them up is one of the most common mistakes in applied stats. Logistic regression, for instance, doesn't assume normal or equal-variance residuals the way linear regression does, and people trip over that constantly.

Here's the quick map:

🧪 Test	✅ Assumptions it needs	⚠️ Watch out for
Independent t-test	Independence, normality (within groups), equal variances	Use Welch's t-test if variances differ
One-way ANOVA	Independence, normally distributed residuals, homogeneity of variance	Independence matters most; normality least
Linear regression	Linearity, independent errors, equal variance, normal residuals, low multicollinearity	Check residual plots, not raw data
Logistic regression	Independence, linearity of the logit, low multicollinearity, enough events per predictor	No normality or equal-variance assumption

A quick sanity note: the famous People Also Ask question, "what are the assumptions for testing the difference between two means?", is really just asking about the t-test row above.

Check your assumptions, don't just test them

A single significance test is a blunt instrument for this job. Lean on it alone and you get false confidence in both directions.

A 2023 paper in Behavior Research Methods made this vivid: a Shapiro-Wilk test can label noisy, skewed, and even bimodal distributions all as "normal," because a p-value above 0.05 only means you failed to detect a problem, not that there isn't one. The author's point is simple and a little uncomfortable. You should check your assumptions (look at them), not just test them (run one function and move on).

💡 Note: A test tells you whether an assumption might be off. A plot tells you how it's off and how much. You want both.

Treat a violation as a spectrum, not a pass/fail stamp. A tiny deviation at a sample size of 5,000 might be statistically "significant" and practically meaningless. A big skew at a sample size of 30 might slip past the test and still mess up your result. Report what you checked, how you checked it, and why you made the call you did. Future-you (and your reviewers) will thank you.

What to do when an assumption is violated

A failed assumption just hands you a fork in the road. You've got four well-worn options, roughly in order of how much they change your approach.

Transform the data. A log, square-root, or Box-Cox transformation can pull a skewed variable toward normal or steady out uneven variance. Cheap to try, often enough on its own.
Use a robust or Welch version. Welch's t-test and Welch's ANOVA don't assume equal variances, so they're a clean swap when homogeneity fails. Heteroscedasticity-consistent standard errors do a similar job in regression.
Switch to a nonparametric test. When normality is genuinely broken, Mann-Whitney U replaces the t-test, Kruskal-Wallis replaces one-way ANOVA, and Spearman's correlation replaces Pearson's. They ask less of your data's shape.
Bootstrap. Resample your data thousands of times to build confidence intervals and tests empirically, without leaning on a distributional assumption at all. Handy when nothing off the shelf quite fits.

One caution: don't over-engineer this. Running a pre-test, then picking your main test based on the result, on the same data, can inflate your error rate. Decide your approach with as much care up front as you can, and use the checks to catch genuine problems rather than to audition tests.

How to run assumption checks (R, SPSS, or Julius)

You don't need a stats PhD to run these, but you do need to know which button to press.

In R, most checks are one line each: shapiro.test() for normality, car::leveneTest() for equal variances, car::vif() for multicollinearity, and plot(model) to get the four standard regression diagnostic plots, including the residuals-versus-fitted and Q-Q plots.

In SPSS, Analyze → Descriptive Statistics → Explore gives you Q-Q plots and the Shapiro-Wilk test, Levene's test rides along in the t-test and ANOVA dialogs, and collinearity diagnostics live in the Regression options. It's all there, just buried in menus.

Run every assumption check from one question

Writing the code or hunting through menus is exactly where non-statisticians get stuck. We built Julius so you don't have to. It's a data analysis tool that lets you ask questions in plain language and get real statistical output back, plots included.

Here's how Julius helps:

Bring any dataset, or no dataset: upload a CSV or spreadsheet, connect a source like PostgreSQL or Snowflake, or use data search to pull public data when you don't have a file yet.
Ask it to check your assumptions: tell Julius the test you're planning, and it runs the normality, variance, and multicollinearity checks for that test.
Get the plots and the readout: Julius returns the Q-Q plots, Levene's and Shapiro-Wilk results, and VIF values, with a plain-English explanation of what each one means.
Make it repeatable: save your assumption-check workflow and re-run it on next week’s data without starting over.

Ready to check your assumptions without writing a line of code? Try Julius for free today.

Frequently asked questions

What is the difference between an assumption and a hypothesis?

An assumption is a condition your data must meet for a test to be valid, like normality or independence. A hypothesis is the actual claim you're testing, like "these two groups have different means." You check assumptions so you can trust the result of your hypothesis test.

When should you check model assumptions?

Check your assumptions after fitting the model and before reporting results, since assumptions like normality and equal variance are usually assessed on the model's residuals. Focus on violations serious enough to change your conclusion, not on achieving a perfect fit that real data never delivers.

What are the assumptions for testing the difference between two means?

Testing the difference between two means with an independent t-test requires three things: independent observations, roughly normal data within each group, and similar variances between groups. If the variances differ, Welch's t-test drops the equal-variance requirement.

What happens if a statistical assumption is violated?

A violated assumption can bias your results or inflate your error rate, but you have options: transform the data, switch to a robust or Welch version of the test, use a nonparametric alternative, or bootstrap. The right fix depends on which assumption failed and how badly.

Do all statistical tests require normality?

No, not all statistical tests require normality. Nonparametric tests like Mann-Whitney U and Kruskal-Wallis make no normality assumption, and with larger samples the central limit theorem makes many parametric tests fairly tolerant of non-normal data anyway.