Independence and Why It Matters for Every Test You Run

Two events are independent if knowing that one occurred tells you nothing about whether the other will occur. This sounds like a narrow technical definition, but it is the load-bearing assumption of almost every classical statistical procedure. When it holds, standard formulas work. When it fails, the same formulas can give wildly wrong answers without any warning.

The Formal Definition

Events A and B are independent if and only if:

P(A and B) = P(A) x P(B)

Equivalently, P(A | B) = P(A): knowing B happened does not change the probability of A. Both conditions say the same thing in different notation.

A coin flip and a die roll are independent. The result of the coin tells you nothing about the die. P(heads and six) = 0.5 x 1/6 = 1/12, which you can verify directly.

A Concrete Case of Dependence

Draw a card from a standard 52-card deck without replacement. The probability the first card is an ace is 4/52. Given that the first card was an ace, the probability the second card is also an ace is 3/51, because one ace is gone and the deck is smaller. These events are dependent: knowing the first outcome changes the probability of the second.

If you incorrectly assumed independence and used 4/52 for both draws, you would calculate P(two aces) = (4/52) x (4/52) = 0.00592. The correct answer is (4/52) x (3/51) = 0.00452. The error is not enormous here, but in other contexts the gap is severe.

Why Independence Matters in Sampling

Standard errors, confidence intervals, and hypothesis tests all assume your observations are independent. The formula for the standard error of a mean is sigma / sqrt(n), where n is the sample size. This formula is derived under the assumption of independence.

Suppose you survey 200 students to estimate average study hours, but all 200 come from the same dormitory. Students in the same dorm share schedules, social norms, and culture. Their answers are correlated with each other. When you plug n = 200 into the standard error formula, you get a number that is too small: you are acting as if you have 200 independent data points when you effectively have fewer. Your confidence interval will be too narrow and you will be overconfident in your estimate.

The actual effective sample size in a clustered sample like this depends on how correlated observations within clusters are. This is measured by the intraclass correlation coefficient (ICC). High ICC means observations within a cluster are very similar, and your effective sample size is much smaller than your raw count.

Time Series and Autocorrelation

In time series data, observations are almost never independent. Today's stock price is correlated with yesterday's. This week's temperature is correlated with last week's. This correlation between a variable and its own past values is called autocorrelation, and it violates the independence assumption of standard regression.

Running OLS regression on autocorrelated data without correcting for it produces standard errors that are too small, t-statistics that are too large, and p-values that are too small. You will reject null hypotheses more often than you should. Correcting for autocorrelation requires either transforming the data or using regression methods that explicitly model the dependence structure.