B Review on hypothesis testing
The process of hypothesis testing has an interesting analogy with a trial that helps on understanding the elements present in a formal hypothesis test in an intuitive way.
Hypothesis testing | Trial |
---|---|
Null hypothesis \(H_0\) | Accused of comitting a crime. It has the “presumption of innocence”, which means that it is not guilty until there is enough evidence to supporting its guilt |
Sample \(X_1,\ldots,X_n\) | Collection of small evidences supporting innocence and guilt. These evidences contain a certain degree of uncontrollable randomness because of how they were collected and the context regarding the case |
Statistic \(T_n\) | Summary of the evicences presented by the prosecutor and defense lawyer |
Distribution of \(T_n\) under \(H_0\) | The judge conducting the trial. Evaluates the evidence presented by both sides and presents a verdict for \(H_0\) |
Significance level \(\alpha\) | \(1-\alpha\) is the strength of evidences required by the judge for condemning \(H_0\). The judge allows evidences that on average condemn \(100\alpha\%\) of the innocents, due to the randomness inherent to the evidence collection process. \(\alpha=0.05\) is considered a reasonable level |
\(p\)-value | Decision of the judge that measures the degree of compatibility, in a scale \(0\)–\(1\), of the presumption of innocence with the summary of the evidences presented. If \(p\)-value\(<\alpha\), \(H_0\) is declared guilty. Otherwise, is declared not guilty |
\(H_0\) is rejected | \(H_0\) is declared guilty: there are strong evidences supporting its guilt |
\(H_0\) is not rejected | \(H_0\) is declared not guilty: either is innocent or there are no enough evidences supporting its guilt |
More formally, the \(p\)-value of an hypothesis test about \(H0\) is defined as:
The \(p\)-value is the probability of obtaining a statistic more unfavourable to \(H_0\) than the observed, assuming that \(H_0\) is true.
Therefore, if the \(p\)-value is small (smaller than the chosen level \(\alpha\)), it is unlikely that the evidence against \(H_0\) is due to randomness. As a consequence, \(H_0\) is rejected. If the \(p\)-value is large (larger than \(\alpha\)), then it is more possible that the evidences against \(H_0\) are merely due to the randomness of the data. In this case, we do not reject \(H_0\).
If \(H_0\) holds, then the \(p\)-value (which is a random variable) is distributed uniformly in \((0,1)\). If \(H_0\) does not hold, then the distribution of the \(p\)-value is not uniform but concentrated at \(0\) (where the rejections of \(H_0\) take place).