Statistical Testing and Hypothesis Testing

Statistical testing (or hypothesis testing) is a structured method used to make decisions or draw conclusions about a population based on data from a sample.

It helps answer questions like:

“Is there really a difference between two groups, or could the difference be due to random chance?”

In simple terms:

“We test whether the data provide enough evidence to support a claim.”

Key Idea

Statistical testing uses sample data to test a hypothesis about a population parameter (like the mean, proportion, or variance).
It’s a way to quantify uncertainty and make objective decisions rather than relying on guesswork.

Steps in Hypothesis Testing

1. State the hypotheses - Formulate the null (H₀) and alternative (H₁) hypotheses.
2. Choose significance level (α) - Decide how much risk of error you’ll accept (commonly 0.05).
3. Collect and summarise data - Compute sample statistics (mean, variance, etc.).
4. Calculate test statistic - Use an appropriate formula (z, t, χ², F, etc.) to measure difference.
5. Compute p-value - Probability of observing the result if H₀ is true.
6. Make a decision - If p-value < α → reject H₀; otherwise, fail to reject H₀.

1. Null and Alternative Hypotheses

We always start by assuming H₀ is true, then use data to decide whether there’s enough evidence to reject it.

2. Significance Level (α)

The significance level (α) represents the probability of rejecting H₀ when it’s actually true (Type I error).

Common choices:

• α = 0.05 (5%) → 95% confidence
• α = 0.01 (1%) → 99% confidence

Smaller α → stricter test (less chance of false positives).

3. Test Statistic

A test statistic measures how far your sample result is from what H₀ predicts — in standardised units (Z, T, F, or χ²).

4. P-value and Decision Rule

• p-value = Probability of observing a test statistic as extreme as (or more extreme than) the one from your data, if H₀ is true.
• Decision:
• • If p-value ≤ α → Reject H₀ (evidence supports H₁)
  • If p-value > α → Fail to reject H₀ (no strong evidence)

Smaller p-value → stronger evidence against H₀.

5. Types of Errors

In testing, we balance these errors by adjusting α, sample size, and test power.

6. Confidence Level & Power

• Confidence Level - 1 − α (probability of not making Type I error)
• Power of a Test - 1 − β (probability of correctly rejecting a false H₀)

A powerful test (power close to 1) is more likely to detect real effects.

Example (Two-tailed t-test)

Suppose you’re testing whether the average height of a sample differs from 170 cm.

1. H₀: μ = 170 , H₁: μ ≠ 170
2. α = 0.05
3. Compute t-statistic
4. Compare p-value to 0.05
5. If p < 0.05 → reject H₀ (significant difference)

The result tells you whether your sample mean differs significantly from 170 cm.

In Machine Learning

• Model validation - Testing if two models perform significantly differently
• Feature selection - Checking if a feature significantly impacts target
• A/B testing - Comparing conversion rates, engagement, etc.
• Error analysis - Checking if residuals are normally distributed
• Experiment design - Quantifying uncertainty and statistical significance