FRQ 3 is the only free-response question built entirely around one complete inference procedure. You get a scenario with data. Then you either run a significance test or build a confidence interval, and you use the result to answer a question in context. The question is worth 10 points. Section II gives you 90 minutes for four 10-point questions, and the whole section counts for 50% of your exam score. That makes this one question about 12.5% of your total, so budget roughly 22 minutes for it.
The question is split into lettered parts. Readers score each part separately against a scoring guide, so a slip in part (c) does not erase the credit you earned in parts (a) and (b). A later part can usually still earn full credit if it uses your earlier wrong number correctly, so never leave a part blank because an earlier answer felt shaky.
Most of the points come from four moves. You name the parameter and the procedure, check the conditions, do the calculation, and state a conclusion in context. The last part usually adds a follow-up, such as naming the error you could have made or judging practical significance, and that earns points too. Readers reward clear statistical language more than arithmetic. A correct z-statistic with no context earns less than students expect. A conclusion that ties the p-value to α and to the scenario earns points reliably. The data will come from the proportions unit (20% of the course) or the means unit (15%), so know the test and the interval versions for both.
FRQ 3 can also be a chi-square test for homogeneity or independence, which sits in the proportions unit. Then there is no parameter symbol to define. You state the hypotheses in words, such as "there is no association between grade level and preferred lunch option." The condition check is that every expected count is at least 5. You report the χ² statistic with its degrees of freedom and a p-value, and there is no interval to build, because chi-square has no interval version.
The rubric, in plain terms
10 points, earned line by line. Readers score what is on the page, nothing else.
Parameter and procedure2 pts
One point for defining the parameter in words that name the population or treatment and the variable. Examples are "the true proportion of borrowers who would return books on time with text reminders" or "the true mean difference in time, old shoe minus new shoe." For a test, the hypotheses must use that parameter, not sample statistics. A second point for naming the correct procedure, such as a two-sample z-test for a difference in proportions or a paired t-interval for a mean difference. For a chi-square test there is no parameter, so the first point goes to hypotheses stated in words about association or about distributions, and the second to naming a chi-square test for homogeneity or for independence.
Conditions2 pts
One point for a randomness check that matches the design. Cite random sampling for a survey and random assignment for an experiment. Saying only "random: yes" with no reason does not earn it. When you sample without replacement, also state that n is less than 10% of the population, so the observations are close to independent. A randomized experiment does not need this 10% check. A second point for the correct shape check. For proportions, show that the success and failure counts are all at least 10. For means, either cite n of at least 30 or, for a small sample, describe a graph of the data showing no strong skew and no outliers. For a chi-square test, show that every expected count is at least 5.
Calculation2 pts
For a test, one point for a correct test statistic with the formula or values shown, and one point for the correct p-value with the right tail or tails. For an interval, one point for the correct standard error and critical value (t* with the right degrees of freedom when using t), and one point for the correct interval endpoints. For a chi-square test, the points go to the χ² statistic with the correct degrees of freedom and to the p-value. Answers consistent with a wrong but reasonable earlier step can still earn credit.
Conclusion or interpretation in context2 pts
For a test, one point for comparing the p-value to α and making the right decision about H0. A second point for stating the conclusion as evidence (or lack of evidence) for Ha, in the words of the scenario. This works the same way for a chi-square test. For an interval, one point for an interpretation that says the interval captures the parameter, with the confidence level and context. A second point for using the interval to answer the question asked, for example by noting whether 0 is a plausible value. False statements, such as calling the p-value the chance that H0 is true, cost the point.
Follow-up reasoning2 pts
Each follow-up the question asks for is worth one point. These include naming the error you could have made and describing it in context, judging a stated claim against the interval, or explaining statistical versus practical significance. The answer must use the scenario's numbers or context, not a textbook definition alone.
Practice prompts
Prompt 1
Text reminders versus email reminders at a library
The Eastbrook Public Library wants more borrowers to return books on time. Staff chose 400 borrowers who had returned at least one book late in the past year. They randomly assigned 200 of them to get a text-message reminder two days before each due date. The other 200 got the library's usual email reminder. During the next month, 142 of the 200 borrowers in the text group returned their books on time, compared with 121 of the 200 in the email group.
(a)Define the parameters of interest and state the hypotheses for a test of whether text reminders lead to a higher on-time return rate than email reminders. Name the inference procedure you would use. [2 pts]
(b)Check the conditions for the procedure you named in part (a). [2 pts]
(c)Calculate the test statistic and the p-value. [2 pts]
(d)Using a significance level of α = 0.05, state your conclusion in context. [2 pts]
(e)Based on your conclusion in part (d), which type of error could the library have made: Type I or Type II? Describe that error in context, and give one practical consequence of it for the library. [2 pts]
Prompt 2
Does a new racing shoe make runners faster?
A cross-country coach wants to know whether a new racing shoe makes her runners faster. Twelve runners from her team volunteered. Each runner ran the same 3-kilometer course twice, one week apart, once in their old shoes and once in the new shoes. For each runner, a coin flip decided which pair they wore first. The list below gives each runner's difference in seconds, old-shoe time minus new-shoe time. A positive difference means the runner was faster in the new shoes.
Mean difference = 4.0 seconds. Standard deviation of the differences = 5.51 seconds. The runners' typical 3-kilometer time is about 690 seconds.
(a)Identify an appropriate inference procedure for estimating the effect of the new shoes, and define the parameter in context. [2 pts]
(b)Check the conditions for the procedure you identified in part (a). [2 pts]
(c)Construct a 95% confidence interval for the parameter you defined in part (a). [2 pts]
(d)Interpret your interval in context. Then use it to decide whether the data give convincing evidence that the new shoes reduce 3-kilometer times on average. [2 pts]
(e)(i) The shoe company advertises that its shoe cuts average 3-kilometer times by more than 10 seconds. Does your interval support this claim? Explain. (ii) The coach says, "The result is statistically significant, so the shoes are worth buying for the whole team." Using the data, give one reason why statistical significance alone does not settle this. [2 pts]
Scored examples
Two responses per prompt — one near the top of the rubric, one in the middle — with every point accounted for. The mid response earns its score through the mistakes students actually make.
Text reminders versus email reminders at a library · high-scoring response10/10
(a) Let p_T = the true proportion of borrowers like these who would return books on time if they all got text reminders. Let p_E = the true proportion who would return books on time if they all got email reminders.
H0: p_T − p_E = 0
Ha: p_T − p_E > 0
Procedure: two-sample z-test for a difference in proportions.
(b) Random: the 400 borrowers were randomly assigned to the text or email group. No 10% check is needed because this is an experiment, not a sample from a population.
Large counts: the pooled proportion is 263/400 = 0.6575. Each group's expected counts are 200(0.6575) = 131.5 on time and 200(0.3425) = 68.5 late. All are at least 10, so the sampling distribution of p̂_T − p̂_E is approximately Normal.
(d) Because the p-value of 0.0135 is less than α = 0.05, we reject H0. There is convincing evidence that text reminders lead to a higher on-time return rate than email reminders for borrowers like these, who had returned a book late in the past year.
(e) We rejected H0, so the only error we could have made is a Type I error. That would mean concluding that texts raise the on-time rate when really they work no better than email. The library might then pay for a texting service that gets no extra books back on time.
How it scored
2/2a — Both parameters say who they describe and what is measured, and the hypotheses use those parameters with the correct one-sided direction. The procedure is named exactly.
2/2b — The random check cites random assignment, which matches the design. The large-counts check uses the pooled proportion and shows all four expected counts, which is the strongest version of this check for a test.
2/2c — The pooled standard error, z = 2.21 and the one-sided p-value of 0.0135 are all correct, and the work is shown.
2/2d — The p-value is compared to α, the decision is correct, and the conclusion is worded as evidence for Ha in context. Limiting it to borrowers like these is a careful touch, because the 400 borrowers were chosen, not randomly sampled.
2/2e — The response links the error type to the decision: rejecting H0 makes only a Type I error possible. The description and the consequence are both specific to the library.
Text reminders versus email reminders at a library · mid-scoring response5/10
(a) p = proportion of people who return their books on time
H0: p1 = p2
Ha: p1 > p2
2-prop z test
(b) Random: the borrowers were randomly assigned to groups.
10%: 200 is less than 10% of all library users.
Normal: n = 200 ≥ 30, so it is Normal.
(c) p̂1 = .71, p̂2 = .605, p̂c = .6575
z = 2.21, p = 0.0135
(d) The p-value means there is only a 1.35% chance that the null hypothesis is true. Since .0135 < .05 we reject H0. Text reminders work better.
(e) Type II error, because any test can be wrong. A Type II error would be the library thinking texts work better when they really don't, so they would waste money on texts.
How it scored
1/2a — The procedure point is earned, and the hypotheses have the right direction. The parameter point is lost because p1 and p2 are never tied to a group. One "p" for "people" does not say which proportion belongs to texts and which to email, and it does not say it is the true value.
1/2b — Random assignment is cited, so the random point is earned. The 10% check is not needed for an experiment, but it costs nothing here. The Normal point is lost: n ≥ 30 is a rule for means. For proportions you must show the success and failure counts are all at least 10.
2/2c — The statistic and the p-value are both correct, and the sample proportions shown are enough to follow the work.
1/2d — Comparing 0.0135 to 0.05 and rejecting H0 earns one point. The second point is lost for two reasons. The first sentence is a false statement, because a p-value is computed by assuming H0 is true, so it cannot be the chance that H0 is true. And "text reminders work better" states a certainty instead of convincing evidence.
0/2e — After rejecting H0, a Type II error is impossible, so naming it loses the first point. The description the student gives (thinking texts work when they don't) is actually a Type I error, so the error and its description contradict each other. The second point is lost too.
Does a new racing shoe make runners faster? · high-scoring response9/10
(a) Paired t-interval for a mean difference. μ_d = the true mean difference in 3 km time (old shoe minus new shoe), in seconds, for runners like these.
(b) Random: the 12 runners are a random sample from the team.
Normal: n = 12 is less than 30, so I checked the data. A dotplot of the differences runs from −5 to 12. It is roughly symmetric with no outliers and no strong skew, so a t procedure is reasonable.
(d) We are 95% confident that the interval from 0.50 to 7.50 seconds captures the true mean difference in 3 km time (old minus new) for runners like these. Every value in the interval is positive, so 0 is not a plausible value. There is convincing evidence that the new shoes reduce times on average.
(e)(i) No. Every plausible value in the interval is at most 7.5 seconds, which is below 10. The data do not support a mean savings of more than 10 seconds.
(ii) The true savings could be as small as 0.5 seconds on a 690-second race, which is under a 0.1% improvement. A gain that small might not change any race result, so it may not be worth the cost for the whole team.
How it scored
2/2a — The procedure is exactly right for paired data. The parameter is a mean difference, with the direction of subtraction and the units stated.
1/2b — The Normal check is excellent: it notes n < 30 and then describes the actual distribution of the differences. The random point is lost because the runners volunteered and were never randomly sampled. The randomness in this study is the coin flip that set the shoe order, and that is what the response needed to cite.
2/2c — The degrees of freedom, t*, standard error and endpoints are all correct.
2/2d — The interpretation names the confidence level, the parameter and the context. Using the all-positive interval to rule out 0 answers the evidence question correctly.
2/2e — Part (i) uses the whole interval, not just one value, to judge the claim. Part (ii) puts the low end of the interval next to the 690-second race time, which is exactly the practical-significance argument the question wants.
Does a new racing shoe make runners faster? · mid-scoring response3/10
(a) Paired t-interval. μ = the mean 3 km time of the runners.
(b) Random: a coin flip decided which shoe was first.
Normal: the sample is big enough for the Central Limit Theorem, so the sampling distribution is Normal.
(d) We are 95% confident that 95% of the runners ran between 0.88 and 7.12 seconds faster in the new shoes. 0 is not in the interval, so the new shoes reduce times.
(e)(i) No, because 10 is not in the interval, so the company's claim is not plausible.
(ii) It is statistically significant, which means the effect is big and real, so the coach is right.
How it scored
1/2a — The procedure point is earned. The parameter point is lost because "the mean 3 km time" describes a single time, not the mean of the paired differences that the interval actually estimates.
1/2b — Citing the coin flip earns the random point, because that randomization is the design's source of randomness. The Normal point is lost: 12 differences are not enough for the Central Limit Theorem. The response needed to look at the 12 differences and report no strong skew or outliers.
0/2c — Both points are lost. The first point needs the standard error and the critical value together. The standard error of 1.591 is right, but z* = 1.96 is the wrong critical value. With an unknown population standard deviation and 11 degrees of freedom, the critical value is t* = 2.201. The second point is lost because the endpoints (0.88, 7.12) are wrong, so the interval is too narrow. This is a mistake inside the calculation itself, not an earlier part's error carried forward, so no consistency credit applies.
0/2d — The interpretation point is lost, because the interval estimates the mean difference, not the range that 95% of individual runners fall in. The evidence point is lost too. Checking whether 0 is in the interval is the right move, but "so the new shoes reduce times" states a certainty instead of convincing evidence. The library response lost its part (d) point for the same reason.
1/2e — Part (i) earns its point, though saying every value in the interval is below 10 would be stronger than "10 is not in it." Part (ii) earns nothing. Statistical significance says the effect is probably not zero, not that it is big. The interval allows a gain as small as about 1 second on a 690-second race.
Where points die
Writing hypotheses with sample statistics, or using a symbol like p that is never tied to a group. Hypotheses are always about parameters, the true values you cannot see. Write "p_T = the true proportion of borrowers who would return on time with text reminders" before you write H0. With two groups, define both symbols, and with paired data, define μ_d with the direction of subtraction.
Checking Normality with "n ≥ 30" for every problem, including proportions and small samples. Match the check to the procedure. For proportions, show that the success and failure counts are all at least 10. For means with n under 30, look at the data themselves. A dotplot or boxplot should show no strong skew and no outliers. A Normal probability plot should look roughly straight, because skew makes it curve and outliers peel away at the ends.
Saying the p-value is the probability that H0 is true, or that α is the chance your conclusion is wrong. The p-value is the probability of getting a result at least as extreme as yours if H0 were true. α is the Type I error risk you choose in advance. Setting α = 0.05 means that when H0 is true, you will wrongly reject it 5% of the time.
Naming the wrong error type, or describing a Type I error and calling it Type II. Look at your decision first. If you rejected H0, only a Type I error is possible: you found an effect that is not really there, like concluding a fertilizer raises yield when it does nothing. If you failed to reject H0, only a Type II error is possible: a real effect went undetected.
Treating "statistically significant" as "large" or "important". Significance only means the result would be unlikely if there were no effect. A huge sample can make a tiny effect significant, like a 0.5-point gain on a 100-point test. To judge practical importance, compare the size of the effect, or the low end of the interval, to the scale of the context.
Interpreting a confidence interval as describing individuals, or saying there is a 95% chance the parameter is in this interval. Say: "We are 95% confident that the interval from __ to __ captures the true [parameter in context]." The interval estimates a mean or a proportion, not where 95% of individual values fall. The 95% describes how often the method works across repeated samples.
Score your own response
0 / 10
Be the harshest reader in the room: award a point only when your response does what the line says, on the page — not what you meant to write.