When Parametric Tests Fail: Violations of Assumptions
You'll learn: How to recognize when a t-test is no longer appropriate due to broken assumptions.
The Breaking Point
You've learned that t-tests rely on certain assumptions: normality of data (or sampling distribution), equal variances, and independence. But what happens when these assumptions don't hold? Using a t-test anyway is like following a recipe for bread when you're missing flour—you'll get something, but not what you intended.
Spotting the Red Flags
Non-normality with small samples: If your sample size is small (typically n < 30) and your data show strong skewness, heavy outliers, or non-normal distributions (confirmed through visual checks or tests like Shapiro-Wilk), the t-test's results become unreliable. The Central Limit Theorem hasn't kicked in yet to rescue you.
Extreme outliers: Even with larger samples, severe outliers can distort the mean and inflate the standard error, making your t-statistic misleading. Think of outliers as a single person shouting over a conversation—they dominate the signal.
Unequal variances: When comparing two groups with vastly different variances and unequal sample sizes, even Welch's t-test may struggle. The assumption violation undermines the test's ability to accurately estimate the sampling distribution.
Non-independence: If observations are related in ways you haven't accounted for (clustered data, time series correlations), the independence assumption fails completely, and your p-values become meaningless.
What to Do Instead
When these violations occur, switch to non-parametric alternatives (like the Wilcoxon tests you learned earlier), use bootstrap methods, transform your data, or apply robust statistical techniques designed to handle assumption violations.
Key Takeaway: Parametric t-tests fail when normality assumptions break down in small samples, when severe outliers distort results, or when variances are grossly unequal—recognizing these violations prevents drawing false conclusions from invalid tests.