Critical value vs test statistic

By Jude Wallis · Published

A test statistic is computed from your sample and measures how far the result sits from what the null hypothesis claims. A critical value is a cutoff fixed by the significance level and the reference curve before any data arrive. Reject H0H_0 when the statistic lands at or beyond the cutoff.

AP Statistics: Unit 3 (topics 3.6 p-Values, 3.7 Carrying Out a Test for a Population Proportion, 4.5 Carrying Out a Test for a Population Mean or Population Mean Difference). Test statistics are computed in the carrying-out topics of the Fall 2026 AP Statistics course, 3.7 for a proportion and 4.5 for a mean. The topic list has a topic titled p-Values (3.6) and none titled critical value or rejection region. Critical values still appear in every confidence interval as z or t.

Critical value vs test statistic: the short answer

The test statistic comes from the data. The critical value comes from a table. A significance test compares the two.

For a zz or tt test, the test statistic counts how many standard errors the sample result sits from the null value. A new sample gives a new test statistic. The critical value is the boundary of the rejection region: the point on the reference curve that cuts off area α\alpha (alpha, the significance level) in the tail or tails named by the alternative hypothesis. It is settled by α\alpha, the direction of HaH_a, and the curve, so it is known before the sample is drawn and stays put when the data change.

The decision rule joins them: reject H0H_0 when the test statistic is at or beyond the critical value, and fail to reject when it falls short.

What the test statistic is

A test statistic is one number computed from the sample whose behavior is known when H0H_0 is true. For proportions and means it has a single form:

test statistic=statistic−null valuestandard error\text{test statistic} = \frac{\text{statistic} - \text{null value}}{\text{standard error}}

In a one-proportion test that is z=p^−p0p0(1−p0)/nz = \dfrac{\hat{p} - p_0}{\sqrt{p_0(1-p_0)/n}}, where p^\hat{p} ("p-hat") is the sample proportion and p0p_0 is the value the null claims. In a one-sample test for a mean it is t=xˉ−μ0s/nt = \dfrac{\bar{x} - \mu_0}{s/\sqrt{n}}, where xˉ\bar{x} ("x-bar") is the sample mean and ss is the sample standard deviation.

Every input except the null value comes from the sample. That is why the test statistic carries the evidence: a value near 0 says the sample looks like what H0H_0 predicts, and a value far from 0 says it does not. The full recipe is in what is a standardized test statistic.

What the critical value is

A critical value is a point on the reference distribution chosen so that a stated area lies beyond it. In a test, that area is α\alpha. Three choices fix it, and none of them involves the sample result:

  1. The significance level α\alpha.
  2. Whether HaH_a is one-sided or two-sided.
  3. The reference curve, including its degrees of freedom when it has them.

On the standard normal curve at α=0.05\alpha = 0.05, a right-tailed test has critical value 1.645 and a two-sided test has critical values ±1.960\pm 1.960. On a tt curve with 24 degrees of freedom the two-sided cutoff rises to 2.064. A chi-square test uses the upper tail only, so with 2 degrees of freedom at α=0.05\alpha = 0.05 the single cutoff is 5.991. These come from the z-table, the t-table, and the chi-square table.

Critical values have a second job in confidence intervals, where z∗z^* or t∗t^* multiplies the standard error to give the margin of error. That use has no test statistic at all.

The differences side by side

FeatureTest statisticCritical value
Comes fromThe sample data and the null valueA table or technology
Depends on the sample resultYesNo
Depends on α\alphaNoYes
Known before data are collectedNoYes
What it measuresDistance of the sample from the null claimWhere the rejection region begins
Changes with a new sampleYesOnly if the degrees of freedom change
Used in a confidence intervalNoYes, as z∗z^* or t∗t^*

One caution on the row about a new sample. For a tt test the degrees of freedom depend on the sample size, so a sample of a different size has a different critical value. The data values themselves never move it.

When to use which

You need a test statistic in every significance test. It is the step where the data enter.

You need a critical value in two situations. The first is a confidence interval, where it sets the margin of error. The second is a test decided by the rejection-region method: find the cutoff for your α\alpha, then check which side of it the test statistic falls on.

The same decision can be reached by converting the test statistic into a p-value and comparing that with α\alpha. The two methods always agree, because a test statistic is at or beyond the critical value exactly when its p-value is at or below α\alpha. The Fall 2026 AP Statistics topic list has a topic titled p-Values (3.6) and none titled critical value or rejection region. The p-value also reports more: it says how far past the boundary the statistic landed, and the critical value only marks the boundary. The comparison of those two thresholds is in p-value vs alpha.

The classic mix-up and how to avoid it

The common error is carrying one critical value into every problem, usually 1.96. That number is right for a two-sided test at α=0.05\alpha = 0.05 on the standard normal curve and wrong almost everywhere else. A test statistic of 2.00 rejects against 1.960 on the normal curve. The same 2.00 from a mean with 24 degrees of freedom fails to reject, because the two-sided cutoff there is 2.064.

A second error is comparing unlike things: setting a test statistic against α\alpha, or a p-value against a critical value. A test statistic and a critical value live on the same axis, the horizontal axis of the reference curve. A p-value and α\alpha are both areas under that curve. Compare a point with a point or an area with an area.

Two checks prevent both. Before looking up a cutoff, name the curve, its degrees of freedom, and the tail or tails. Then ask where each number came from: if it used the sample, it is the test statistic, and if it used only α\alpha and the curve, it is the critical value.

A proportion test decided with a critical value

A school claims that half of its students favor a later start time. In a random sample of 400 students, 220 favor it. Test H0:p=0.50H_0: p = 0.50 against Ha:p>0.50H_a: p > 0.50 at α=0.05\alpha = 0.05 using a critical value, and identify which number came from the data.

  1. Check the large counts condition with the null value: np0=400(0.50)=200np_0 = 400(0.50) = 200 and n(1−p0)=200n(1-p_0) = 200, both at least 10. The sample is random, and 400 students is assumed to be under 10% of the school's population.

  2. Find the critical value first, since it needs no data. The test is right-tailed at α=0.05\alpha = 0.05 on the standard normal curve, so the critical value is 1.645. Reject H0H_0 if z≥1.645z \ge 1.645.

  3. Compute the sample proportion: p^=220400=0.55\hat{p} = \frac{220}{400} = 0.55.

  4. Compute the standard error under H0H_0: 0.50×0.50400=0.000625=0.025\sqrt{\frac{0.50 \times 0.50}{400}} = \sqrt{0.000625} = 0.025.

  5. Compute the test statistic: z=0.55−0.500.025=2.00z = \frac{0.55 - 0.50}{0.025} = 2.00.

  6. Compare: 2.00≥1.6452.00 \ge 1.645, so the test statistic is in the rejection region. Reject H0H_0.

  7. Confirm with the p-value: the area to the right of z=2.00z = 2.00 is 0.0228, which is below 0.05. The two methods agree.

The critical value is 1.645 and was fixed by α=0.05\alpha = 0.05, the right-tailed alternative, and the normal curve. The test statistic is z=2.00z = 2.00 and came from the sample. Since 2.00≥1.6452.00 \ge 1.645, reject H0H_0: there is convincing evidence that more than half of the students favor a later start time.

The same test statistic against two different cutoffs

A random sample of 25 exam scores has xˉ=83\bar{x} = 83 and s=7.5s = 7.5, and the scores show no strong skew or outliers. Test H0:μ=80H_0: \mu = 80 against Ha:μ≠80H_a: \mu \neq 80 at α=0.05\alpha = 0.05. Then show what goes wrong if the normal critical value 1.960 is used.

  1. Compute the standard error: sn=7.525=7.55=1.5\frac{s}{\sqrt{n}} = \frac{7.5}{\sqrt{25}} = \frac{7.5}{5} = 1.5.

  2. Compute the test statistic: t=83−801.5=2.00t = \frac{83 - 80}{1.5} = 2.00.

  3. Find the correct critical value. The reference curve is tt with n−1=24n - 1 = 24 degrees of freedom, and the test is two-sided at α=0.05\alpha = 0.05, so the cutoffs are ±2.064\pm 2.064.

  4. Compare: 2.00<2.0642.00 < 2.064, so the test statistic is short of the rejection region. Fail to reject H0H_0.

  5. Confirm with the p-value: on 24 degrees of freedom the two-sided p-value for t=2.00t = 2.00 is 0.0569, which is above 0.05.

  6. Now use the wrong cutoff. Against 1.960, the statistic 2.00 appears to reject, and the matching normal p-value would be 0.0455. The arithmetic for the test statistic was identical; only the cutoff changed the verdict.

t=2.00t = 2.00 with 24 degrees of freedom. The correct critical value is 2.064, so the test fails to reject H0H_0 (p-value 0.0569): the sample does not give convincing evidence that the mean differs from 80. Using 1.960 would have rejected in error. The test statistic came from the data, and the critical value had to come from the right curve.

Frequently asked questions

Is the critical value the same as the test statistic?

No. The test statistic is calculated from the sample and changes with every new sample. The critical value is a cutoff determined by the significance level, the direction of the alternative, and the reference curve. A test compares the first against the second.

What happens if the test statistic equals the critical value?

The rejection region includes its boundary, so you reject H0H_0. At that point the p-value equals α\alpha exactly, and the p-value rule, reject when the p-value is at or below α\alpha, gives the same decision.

Does a bigger sample change the critical value?

For a zz test, no: 1.645 and 1.960 hold at every sample size. For a tt test the degrees of freedom rise with nn, so the critical value falls slowly toward the normal value. A bigger sample mainly changes the test statistic, because it shrinks the standard error.