Course contentsShow
← Data Science
Lesson 391 of 2,14510. Hypothesis Tests for MeansFree lesson

The Sign Test for Medians

Test whether a population median differs from a hypothesized value using only the signs of differences.

The Sign Test for Medians

You'll learn how to test whether a population median differs from a hypothesized value using only the direction (sign) of differences—no distributional assumptions required.


What is the Sign Test?

When your data violates the normality assumption for t-tests, or when you're working with ordinal data, the sign test offers a simple, robust alternative. Instead of analyzing the magnitude of differences, you only look at whether each observation is above or below the hypothesized median.

Think of it like a coin flip: if the true median equals your hypothesized value, roughly half your observations should fall above it and half below—just like getting heads or tails with a fair coin.

How It Works

  1. State your hypotheses: For example, H₀: median = 50 vs H₁: median ≠ 50
  2. Calculate differences: Subtract the hypothesized median from each observation
  3. Count the signs: Ignore zeros; count how many differences are positive (+) and how many are negative (−)
  4. Use binomial logic: Under H₀, positive and negative signs are equally likely (p = 0.5)
  5. Calculate a p-value: If you observe an extreme imbalance (e.g., 18 positives out of 20), that's evidence against H₀

When to Use It

  • Your data is severely skewed or has outliers
  • Sample size is small and normality can't be verified
  • You have ordinal data where "greater than" makes sense but exact distances don't
  • You want a quick, assumption-light test

The sign test trades statistical power for robustness—it requires larger samples to detect effects compared to the t-test, but it works when nothing else does.

Key Takeaway: The sign test evaluates whether a median differs from a target by counting positive vs. negative deviations, treating the data like a binomial experiment—perfect when parametric assumptions fail.