Imagine that I am studying racial discrimination. In particular, whether a white candidate (for a job) is chosen more often than an equally qualified minority candidate.

I ask people to choose between a white candidate and an equally qualified minority candidate. Suppose that white candidates are chosen 54% of the time, and minority candidate is chosen 46% of the time. Should the hypothesis test for whether white people are chosen as often as black people be

$H_0: \pi = 0.50$ or $H_0: \pi = 0.46$?


I refer to What follows if we fail to reject the null hypothesis? to explain that, in hypothesis testing, the objective is to show that your data 'rejects' $H_0$ and 'supports $H_1$.

You want to 'show' that '... whether a white candidate (for a job) is chosen more often than an equally qualified minority candidate'.

So if $\pi$ is the fraction of white candidates that are chosen then $H_1: \pi > 0.5$ versus $H_0: \pi = 0.5$.

This is done before you have seen the data. This was also argued by @Greg Snow

If you draw a random sample of size $n$ ($n$ large enough), then for each sample $s$ you observe a fraction $p_s$ in that sample. Obviously, in another sample you will observe another $p_s$ so this 'sample fraction' changes from sample to sample and is therefore a random variable. If $n$ is large enough, then this random variable 'sample fraction' will be normally distributed with mean the population fraction $\pi$ and standard deviation $\sqrt{\pi(1-\pi)/n}$.

So if you choose a significance level (e.g.) $\alpha=0.05$ then you find that $P(p_s \ge \pi + 1.645 \sqrt{\pi(1-\pi)/n})=0.05$. So the rejection region with $\pi=0.5$ (if $H_0$ is true) and e.g. $n=100$ would be $p_s \ge 0.5+1.645\sqrt{0.25/100}$ or $p_s \ge 0.58$.

All this reasoning is done without looking at the data, you only have to fix $\alpha$ and the sample size $n$.

Only in the final step you look at your specific sample and in your your sample the fraction of white that are chosen is 0.54. Si the value of $p_s$ for your sample ($\bar{p}_s=0.54$) is not in the critical region ($p_s \ge 0.58$).

Note that this decision depends on the sample size and on the $\alpha$.

So it is very important to distinguish between the population fraction $\pi$ which is the one you use in your hypothesis and the outcome of your specific sample fraction $\bar{p}_s$. It is also important to distinguish the fraction of your sample $\bar{p}_s$ (which is only one number, namely the fraction in the sample that you have) from the outcome of any possible sample of the same size , this latter is a random variable $p_s$ because its value depends on the sample, so $p_s$ is not one value but a random variable (thus a distribution).


In practice the null hypothesis should be chosen before looking at the data (often before collecting any data). So if the 46% is representative of the data that you are using for the test, then it is not appropriate for the null hypothesis. On the other hand, if 46% is a historical value and you want to see if new data indicates that a change has occurred vs. the status quo being maintained, then the null comparing to 0.46 is appropriate.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.