Non-Parametric Tests: When Standard Statistical Tests Fail
After Hypothesis Testing Explained: H0, H1, p-value and Significance, the next interview question is whether the chosen test is valid for the data in front of you. Indian startup data - Swiggy delivery times, Zomato ratings, UPI transaction values - is almost NEVER normally distributed. This matters because blindly applying t-tests to such data produces misleading p-values, and interviewers expect you to check the distribution before choosing your test.
- Indian startup data - Swiggy delivery times, Zomato ratings, UPI transaction values - is almost NEVER normally distributed.
- Distribution plots show heavy right skew, zero-inflation, and multi-modality.
- Blindly applying t-tests to such data produces misleading p-values.
- Always check with a histogram or Shapiro-Wilk test before choosing your test.
- Use non-parametric tests for skewed data, small sample, ordinal data, non-normal distributions, non-normal differences, non-linear monotonic relationship, or small expected frequencies.
- Interview Favourite: Non-parametric tests - if interviewer gives you non-normal data and you apply t-test, you'll be penalised.
Why Standard Statistical Tests Fail on Startup Data
Reality Check: Indian startup data - Swiggy delivery times, Zomato ratings, UPI transaction values - is almost NEVER normally distributed. Distribution plots show heavy right skew, zero-inflation, and multi-modality.
Blindly applying t-tests to such data produces misleading p-values. Always check with a histogram or Shapiro-Wilk test before choosing your test.
Always check with a histogram or Shapiro-Wilk test before choosing your test.
How to Apply the Safeguard
The logic is simple: first identify the standard parametric test you would have used, then check whether the data violates the normality requirement or has ordinal, small-sample, skewed, or frequency issues. If it does, choose the matching non-parametric equivalent.
- For a one-sample t-test: use Wilcoxon Signed-Rank (1-sample) when the data is skewed, the sample is small, or the data is ordinal. Example: Is Swiggy median delivery time = 30 min?
- For a two-sample t-test: use Mann-Whitney U Test for non-normal distributions or ordinal data. Example: Comparing delivery times: Bengaluru vs Chennai.
- For a paired t-test: use Wilcoxon Signed-Rank (paired) for before/after on same users with non-normal differences. Example: User spend before/after app redesign.
- For one-way ANOVA: use Kruskal-Wallis H Test for 3+ groups and non-normal data. Example: Zomato ratings across North/South/West Indian cuisines.
- For Pearson Correlation: use Spearman Rank Correlation for a non-linear monotonic relationship or ordinal data. Example: Relationship between order value and customer rating.
- For two-proportion z-test: use Chi-square test / Fisher's Exact for small expected frequencies. Example: Conversion rate by device type (mobile/desktop).
Ordinal Data and Non-Normal Data in Practice
Ordinal data means categories with meaningful order. Customer satisfaction: 1-5 stars is an example, and Spearman rank correlation is a suitable analysis for ordinal data.
Skewed data needs extra care. Skewness > 0 means right-skewed, with Mean > Median > Mode and a long right tail. The key interview move is to avoid treating this like clean normal data unless the distribution check supports it.
Structuring a Non Interview Answer
"You are given non-normal Swiggy delivery time data and need to compare Bengaluru vs Chennai. Which test would you use, and what would you check first?"
Interview Favourite: Non-parametric tests - if interviewer gives you non-normal data and you apply t-test, you'll be penalised. Always check distribution first.
The most frequent error is blindly applying t-tests to skewed, small-sample, ordinal, or non-normal data. It costs points because such data produces misleading p-values, and the correct safeguard is to check with a histogram or Shapiro-Wilk test before choosing your test.
Conclusion
Non-parametric tests are the analyst's safeguard when real Indian startup data violates normality assumptions. The final takeaway is simple: check the distribution first, then choose the non-parametric equivalent that matches the business question.