Further hypothesis tests and errors
In a minute
HL adds tests for a mean (z or t), a proportion (binomial), a Poisson rate and (). You also find and the probabilities of and .
You'll be able to
- Find and critical regions for a test on a normal mean, and decide a test with them.
- Test a mean with a z-test when is known and a when it is not, including .
- Test a proportion with the and a mean rate with the , one-tailed, and choose the critical region.
- Use your GDC to test : and interpret the result.
- Describe Type I and Type II errors in context and calculate their probabilities.
Before you start3 quick questions on what this builds on
Revise first: 4.1 Sampling, data quality and outliers, 4.4 Correlation and regression (y on x), 4.7 Discrete random variables and expected value, 4.8 Binomial distribution, 4.9 Normal distribution, 4.11 Chi-squared tests and the t-test, 4.14 Linear combinations and unbiased estimates, 4.15 Sample mean and the central limit theorem, 4.16 Confidence intervals for the mean and 4.17 Poisson distribution
Critical values and critical regions
In 4.11 you decided with a p-value. Another route: fix in advance which results reject , then see where your result lands.
To test the mean of a with known , use as the test statistic. If : is true, the is:
The critical region is the tail points to, with area equal to the ; a puts half in each tail. Find the boundary with the .
Standardising, , gives critical values of instead, such as for a lower-tail test. The decision is the same.
Try it first, then reveal one step at a time.
- Under , , a standard deviation of g.
- is "less than": inverse normal with area to the left, , , gives .
- Critical region: g. A sample mean in it is evidence that the bags are underfilled on average.
Try it first, then reveal one step at a time.
- : ; : , two-tailed. Under , , since .
- Put in each tail. Inverse normal with area to the left gives ; area gives .
- Critical region: or . A sample mean of mm would not be evidence of a change.
Testing a population mean: z or t
is known only when the question gives the population value. A standard deviation from the sample is not . The t-test uses with .
Try it first, then reveal one step at a time.
- : ; : , where is the population mean mass in grams.
- is known, so use a z-test: data list, , , . The GDC gives , , p-value .
- , so reject : there is evidence at the level that the mean mass is below g. At it would not be enough.
Try it first, then reveal one step at a time.
- : ; : , where is the population mean waiting time in minutes.
- No population is given, so use a one-sample t-test with and .
- The GDC gives , , and p-value .
- , so reject : there is evidence at the level that the mean wait is longer than minutes.
You won't be asked for a critical region in a t-test: decide with the p-value.
Matched pairs
Sometimes each subject is measured twice, such as before and after a course. The lists are linked, so the two-sample t-test from 4.11 does not apply.
Try it first, then reveal one step at a time.
- Matched pairs: let , so a faster time after gives . : ; : , where is the population mean difference.
- Differences: . One-sample t-test on them with : , , p-value .
- : reject . There is evidence at the level that the plan reduces the mean time. With no comparison group, other causes (such as cooler weather) can't be ruled out.
Try it first, then reveal one step at a time.
- (a) Each student sits both tests, so the scores pair up. Each student's difference removes the big variation in spelling ability.
- (b) Let . "Changes" has no direction, so uses .
- (a) Both scores come from the same student, so the data are matched pairs: analyse the differences as one sample. (b) : ; : , where is the population mean difference, silent minus music.
Tests with the binomial and Poisson distributions
When the test counts things, the count is a . Test a proportion with the , and a mean rate with the .
Try it first, then reveal one step at a time.
- Let be the proportion of all the company's seeds that germinate. : ; : .
- Under , . p-value .
- : reject . There is evidence at the level that fewer than of the seeds germinate.
Try it first, then reveal one step at a time.
- : ; : , where is the proportion of all booked passengers who don't turn up. Under , . A rise means large , so find the smallest with .
- Use : , too big; .
- Critical region: . If or more of the don't turn up, the airline rejects .
Try it first, then reveal one step at a time.
- (a) Over weeks, : ; : , where is the mean number of accidents in weeks. So .
- A fall means small : find the largest with . and , so the critical region is .
- (b) is not in the critical region: do not reject . There is insufficient evidence at the level that the accident rate has fallen.
Testing for correlation: is ρ = 0?
In 4.4 you found for a sample. A sample can show some even when the has none. This test asks whether is far enough from to count as evidence.
: , and is , or , following the claim. The test assumes the come from a bivariate normal distribution. Your GDC gives and the p-value.
Try it first, then reveal one step at a time.
- : ; : , where is the population correlation between spend and sales.
- Linear regression t-test with : the GDC gives and p-value .
- : reject . There is evidence at the level of positive correlation between spend and sales. It doesn't show that advertising causes the sales.
Try it first, then reveal one step at a time.
- : ; : . "Correlated" has no direction, so choose on the GDC.
- The GDC gives and p-value .
- : do not reject . There is insufficient evidence at the level of correlation between distance and price, though at there would be.
Type I and Type II errors
A test decides from a sample, so it can go wrong in two ways.
So for a Type II probability: find the critical region with the distribution, switch to the true parameter value, then find the probability of landing outside the region.
Try it first, then reveal one step at a time.
- (a) Under , , since . Inverse normal, area : , so the critical region is hours.
- (b) is continuous, so exactly.
- (c) Now . Type II means is outside the critical region: . Use the stored , not .
- Interpret: if the bulbs really average hours, this test misses it about time in .
Try it first, then reveal one step at a time.
- (a) Type I: when . , below the level as the discrete rule requires.
- (b) Type II: when the true is . With , .
- Interpret: even though the no-show rate has more than doubled, the test misses it about of the time.
Common mistakes
- Writing hypotheses about the sample, such as : . Use population parameters: , , or .
- Stating "reject " with no reason. Compare first (p-value with level, or result with critical region), then decide.
- Picking a discrete critical region whose probability is closest to but above it. It must stay below the significance level.
- Rounding the critical value before a Type II calculation. Store the full GDC value and round only the final answer.
Exam tips
- **State** hypotheses with the parameter (, , or ) and say in words what it is, such as "the population mean waiting time".
- **Find** a discrete critical region: show the tail probabilities either side of the boundary, such as and .
- A conclusion needs its reason: the comparison, the decision and the meaning in context, such as ", so reject : evidence that fewer than germinate". Never write "proves".
- Write down the test and what you enter in the GDC: the distribution, its parameters and the direction of .
Key takeaways
Practise this
35 past-paper questions on Further hypothesis testing in the AI HL bank, each with a mark scheme and a worked explanation.
Also: Type I and Type II errors (11)