P-Value¶
Formula¶
\[
p = P(T \ge t_{\text{obs}}\mid H_0)
\]
Parameters¶
- \(T\): test statistic under the null distribution
- \(t_{\text{obs}}\): observed statistic value
- \(H_0\): null hypothesis
What it means¶
The p-value is the probability, assuming the null hypothesis is true, of observing a test statistic at least as extreme as the observed one.
What it's used for¶
- Hypothesis testing and significance reporting.
- Screening evidence against a null model.
Key properties¶
- Smaller p-values indicate stronger incompatibility with \(H_0\).
- Depends on test definition and tail convention.
Common gotchas¶
- Not the probability that \(H_0\) is true.
- Not a measure of effect size or practical importance.
Example¶
A p-value of 0.01 means results this extreme would occur about 1% of the time under \(H_0\).
How to Compute (Pseudocode)¶
Input: observed test statistic t_obs, null distribution (exact/approximate/permutation)
Output: p-value
compute tail probability under H0 at least as extreme as t_obs
(tail definition depends on one-sided vs two-sided test)
return that probability
Complexity¶
- Time: Depends on how the null distribution is obtained (closed-form CDF lookup, numerical integration, or resampling/permutation)
- Space: Depends on whether null samples/statistics are stored or streamed
- Assumptions: Test statistic and tail convention are fixed before interpretation; null-distribution estimation method determines practical cost