Statistical hypothesis testing

From Wikipedia, the free encyclopedia

One may be faced with the problem of making a definite decision with respect to an uncertain hypothesis which is known only through its observable consequences. A statistical hypothesis test, or more briefly, hypothesis test, is an algorithm to state the alternative (for or against the hypothesis) which minimizes certain risks.

This article describes the commonly used frequentist treatment of hypothesis testing. From the Bayesian point of view, it is appropriate to treat hypothesis testing as a special case of normative decision theory (specifically a model selection problem) and it is possible to accumulate evidence in favor of (or against) a hypothesis using concepts such as likelihood ratios known as Bayes factors.

There are several preparations we make before we observe the data.

The hypothesis must be stated in mathematical/statistical terms that make it possible to calculate the probability of possible samples assuming the hypothesis is correct. For example: The mean response to treatment being tested is equal to the mean response to the placebo in the control group. Both responses have the normal distribution with this unknown mean and the same known standard deviation ... (value).
A test statistic must be chosen that will summarize the information in the sample that is relevant to the hypothesis. Such a statistic is known as a sufficient statistic. In the example given above, it might be the numerical difference between the two sample means, m₁ − m₂.
The distribution of the test statistic is used to calculate the probability sets of possible values (usually an interval or union of intervals). In this example, the difference between sample means would have a normal distribution with a standard deviation equal to the common standard deviation times the factor $\sqrt{\frac{1}{n_1} + \frac{1}{n_2}}$ where n₁ and n₂ are the sample sizes.
Among all the sets of possible values, we must choose one that we think represents the most extreme evidence against the hypothesis. That is called the critical region of the test statistic. The probability of the test statistic falling in the critical region when the hypothesis is correct is called the alpha value (or size) of the test.
The probability that a sample falls in the critical region when the parameter is $θ$ , where $θ$ is for the alternative hypothesis, is called the power of the test at $θ$ . The power function of a critical region is the function that maps $θ$ to the power of $θ$ .

After the data is available, the test statistic is calculated and we determine whether it is inside the critical region.

If the test statistic is inside the critical region, then our conclusion is one of the following:

The hypothesis is incorrect, therefore reject the null hypothesis. (Therefore the critical region is sometimes called the rejection region, while its complement is the acceptance region.)
An event of probability less than or equal to alpha has occurred.

The researcher has to choose between these logical alternatives. In the example we would say: the observed response to treatment is statistically significant.

If the test statistic is outside the critical region, the only conclusion is that

There is not enough evidence to reject the hypothesis.

This is not the same as evidence in favor of the hypothesis. That we cannot obtain using these arguments, since lack of evidence against a hypothesis is not evidence for it. On this basis, statistical research progresses by eliminating error, not by finding the truth.

1 Common test statistics
2 Criticism
3 See also
4 External links

[edit] Common test statistics

Name	Formula	Assumptions
One-sample z-test	$z=\frac{\overline{x} - \mu_0}{\frac{\sigma}{\sqrt{n}}}$	(Normal distribution or n ≥ 30) and σ known. (z is the distance from the mean in standard deviations. It is possible to calculate a minimum proportion of a population that falls within n standard deviations(see Chebyshev's inequality).
Two-sample z-test	$z=\frac{(\overline{x}_1 - \overline{x}_2) - (\mu_1 - \mu_2)}{\sqrt{\frac{\sigma_1^2}{n_1} + \frac{\sigma_2^2}{n_2}}}$	Normal distribution and independent observations and (σ₁ AND σ₂ known)
One-sample t-test	$t=\frac{\overline{x} - \mu_0}{\frac{s}{\sqrt{n}}},$ $d f = n - 1$	(Normal population or n > 30) and σ unknown
Two-sample pooled t-test	$t=\frac{(\overline{x}_1 - \overline{x}_2) - (\mu_1 - \mu_2)}{s_p\sqrt{\frac{1}{n_1} + \frac{1}{n_2}}},$ $s_p^2=\frac{(n_1 - 1)s_1^2 + (n_2 - 1)s_2^2}{n_1 + n_2 - 2},$ $d f = n 1 + n 2 - 2$	(Normal populations or n₁ + n₂ > 40) and independent observations and σ₁ = σ₂ and (σ₁ and σ₂ unknown)
Two-sample unpooled t-test	$t=\frac{(\overline{x}_1 - \overline{x}_2) - (\mu_1 - \mu_2)}{\sqrt{\frac{s_1^2}{n_1} + \frac{s_2^2}{n_2}}},$ $df=\frac{(n_1 - 1)(n_2 - 1)}{(n_2 - 1)c^2 + (n_1 - 1)(1 - c^2)},$ $c=\frac{\frac{s_1^2}{n_1}}{\frac{s_1^2}{n_1} + \frac{s_2^2}{n_2}}$ or $d f = min{n 1, n 2}$	(Normal populations or n₁ + n₂ > 40) and independent observations and σ₁ ≠ σ₂ and (σ₁ and σ₂ unknown)
Paired t-test	$t=\frac{\overline{d} - d_0}{s_d},$ $d f = n - 1$	(Normal population of differences or n > 30) and σ unknown
One-proportion z-test	$z=\frac{\hat{p} - p}{\sqrt{\frac{p(1-p)}{n}}}$	np > 10 and n(1 − p) > 10
Two-proportion z-test, equal variances	$z=\frac{(\hat{p}_1 - \hat{p}_2) - ({p}_1 - {p}_2)}{\sqrt{\hat{p}(1 - \hat{p})(\frac{1}{n_1} + \frac{1}{n_2})}}$ $\hat{p}=\frac{x_1 + x_2}{n_1 + n_2}$	n₁p₁ > 5 AND n₁(1 − p₁) > 5 and n₂p₂ > 5 and n₂(1 − p₂) > 5 and independent observations
Two-proportion z-test, unequal variances	$z=\frac{(\hat{p}_1 - \hat{p}_2) - (p_1 - p_2)}{\sqrt{\frac{\hat{p}_1(1 - \hat{p}_1)}{n_1} + \frac{\hat{p}_2(1 - \hat{p}_2)}{n_2}}}$	n₁p₁ > 5 and n₁(1 − p₁) > 5 and n₂p₂ > 5 and n₂(1 − p₂) > 5 and independent observations

The statistics for some other tests have their own page on Wikipedia, including the Wald test and the likelihood ratio test.

[edit] Criticism

Some statisticians have commented that pure "significance testing" has what is actually a rather strange goal of detecting the existence of a "real" difference between two populations. In practice a difference can almost always be found given a large enough sample, what is typically the more relevant goal of science is a determination of causal effect size. The amount and nature of the difference, in other words, is what should be studied. Many researchers also feel that hypothesis testing is something of a misnomer. In practice a single statistical test in a single study rarely "proves" anything.

"A little thought reveals a fact widely understood among statisticians: The null hypothesis, taken literally (and that's the only way you can take it in formal hypothesis testing), is always false in the real world.... If it is false, even to a tiny degree, it must be the case that a large enough sample will produce a significant result and lead to its rejection. So if the null hypothesis is always false, what's the big deal about rejecting it?" (Cohen 1990)

(The above criticism only applies to point hypothesis tests. If one were testing, for example, whether a parameter is greater than zero, it would not apply.)

"... surely, God loves the .06 nearly as much as the .05." (Rosnell and Rosenthal 1989)

"How has the virtually barren technique of hypothesis testing come to assume such importance in the process by which we arrive at our conclusions from our data?" (Loftus 1991)

"Despite the stranglehold that hypothesis testing has on experimental psychology, I find it difficult to imagine a less insightful means of transiting from data to conclusions." (Loftus 1991)

Even when you reject null hypothesis, effect sizes should be taken into consideration. If the effect is statistically significant but the effect size is very small, then it is a stretch to consider the effect theoretically important.