Skip to content

AP Statistics · technique

How to answer the free-response section

On this page · 5 sections

The free-response section of the AP Statistics exam gives you four questions in 90 minutes, and it counts for 50% of your score. Each question is worth 10 points. Starting with the May 2027 exam, the section has no Investigative Task, and all four questions are worth the same 10 points. That works out to about 22 minutes per question if you split the time evenly.

Each question has a fixed job, named by the skill categories it targets. FRQ 1 is Multi-Focus on Practices 1 and 2, which means choosing a method and analyzing data. Expect to describe or compare distributions, read a graph or table, or explain how a sample or experiment should be set up. FRQ 2 is Multi-Focus on Practices 3 and 4, which pairs probability and simulation with an argument about a claim. FRQ 3 is the Inference question. You carry out a significance test or build a confidence interval for proportions or means. It can also be a chi-square test for goodness of fit, homogeneity or independence, because chi-square sits in the inference for proportions unit. FRQ 4 is Multi-Focus on Practices 2, 3 and 4, so it mixes data analysis, probability and argument in a single scenario.

Readers score each question part by part against a point rubric. A correct number with no explanation rarely earns full credit, because the rubric rewards communication as much as arithmetic. You earn points by naming the method, showing the setup, checking conditions with real numbers from the problem, and stating a conclusion in the context of the problem. Parts are scored separately, so a mistake in part (a) does not wipe out part (c). If you use a wrong earlier answer correctly in a later part, you can usually still earn the later point. You may use a graphing calculator, and the exam gives you a formula sheet and tables.

The steps

  1. 1

    1. Read all four questions before you write

    Spend the first two or three minutes skimming every question and every part. Note which unit each one comes from and which one you can do fastest. Start with your strongest question, because a confident first answer saves time for the harder ones. Watch the clock at roughly 22, 45 and 67 minutes, which are the ends of each quarter of the section. That pacing is our suggestion, not a College Board rule. If a part has you stuck for more than two minutes, write what you can and move on. A blank part scores zero, while a correct setup can still earn a point.

  2. 2

    2. Find the verb and the context in every part

    The verb tells you what a full answer looks like. Identify wants a short, specific answer. Describe wants the features of something, such as the shape, center and variability of a distribution. Compare wants explicit comparison words like greater than or less variable than, not two separate descriptions sitting side by side. Interpret wants you to say what a number means in the setting of the problem. Justify wants a claim tied to evidence, usually a value you calculated. Before you write, underline the people, units and variable in the stem. Every answer should use them, because a conclusion about "the mean" earns less than one about "the mean commute time of district employees, in minutes."

  3. 3

    3. Describe and compare distributions with numbers

    On FRQ 1 and often FRQ 4, you will describe or compare distributions from a dotplot, histogram, boxplot or table. Cover four things: shape, center, variability, and any unusual features such as outliers or gaps. Back each one with a value, such as "the median for Plant A is about 41 grams, which is higher than the median of 35 grams for Plant B." If the data are skewed or have outliers, use the median and IQR, because they are resistant to extreme values. The mean and standard deviation get pulled toward the tail. When you call a point an outlier, show the 1.5 × IQR fence that it falls beyond, so the reader sees a test and not a guess.

  4. 4

    4. Write regression answers in context

    When a question gives you a least-squares line, such as predicted price = 4.20 + 0.85(weight), each interpretation has a set form. For the slope, say the predicted y changes by that amount for each one-unit increase in x, and name the units and variables: "for each additional gram of weight, the predicted price increases by $0.85." For the y-intercept, give the predicted y when x is 0, and say whether x = 0 makes sense in the setting. A residual is actual y minus predicted y, so a negative residual means the line overestimated that point. For r², say it is the percent of the variation in y explained by the linear model with x, such as "72% of the variation in price is explained by the linear model using weight." To judge whether a line fits, look at the residual plot. A random scatter with no curve supports a linear model, but a clear curved pattern means a line is not a good fit.

  5. 5

    5. Run the inference procedure in four labeled parts

    FRQ 3, and some parts of other questions, ask for a full test or interval. Write it in four labeled chunks so the reader can find each point. First, state the hypotheses in symbols and define the parameter in words, such as p₁ − p₂, where each p is the true proportion of a named population. Second, name the procedure, such as a two-sample z-test for a difference in proportions, and check its conditions with the actual numbers. Third, do the mechanics: the test statistic, its formula with values plugged in, and the p-value or interval. Fourth, conclude by comparing the p-value to α, then state the result about the parameter in context. Use the language of evidence, not proof. For a confidence interval, the fourth part is an interpretation in context instead, such as "We are 95% confident that the interval from 0.04 to 0.19 captures the true difference in the proportions." If the question asks about a claimed value, say whether the interval contains it. A value outside the interval is not plausible, while a value inside it is still plausible.

  6. 6

    6. Check conditions by showing the numbers

    Writing "conditions are met" earns nothing. For each condition, name it and show why it holds in this problem. Randomness: say where the random sample or random assignment came from. Large counts for proportions: show that the successes and failures are each at least 10, with the actual values. For a confidence interval use the observed counts n·p̂ and n(1−p̂). For a test use the counts expected under H₀ (n·p₀, or the pooled proportion for two samples). Large counts for chi-square: find the expected count in every cell and show that all of them are at least 5, by writing down the smallest one. Normality for means: say the sample size is at least 30, or that a graph of the sample shows no strong skew or outliers. The 10% condition applies when you sample without replacement: show that n is less than 10% of the population. If a condition fails, say so and explain what that means for the result instead of pretending it passed.

  7. 7

    7. Show probability work in words, not calculator syntax

    A line like normalcdf(47, 1E99, 42, 3.5) does not earn full credit, because it doesn't tell the reader what the numbers mean. Name the distribution and its parameters, such as "Normal with mean 42 and standard deviation 3.5" or "binomial with n = 20 and p = 0.08." Write the event with the correct boundary, such as P(X ≥ 3). For a binomial count, at least 3 means 1 − P(X ≤ 2), and an off-by-one error here is one of the most common ways to lose a point. Then give the answer. For a simulation, say which digits stand for which outcome, how many digits make one trial, and what you record at the end of each trial.

  8. 8

    8. Finish with one answer per part, then reread

    Give one answer to each part. If you write two methods or two conclusions that disagree, readers generally score the weaker one, so cross out anything you do not want graded. Use leftover time to check three things. Does every conclusion use the context? Does every test conclusion link the p-value to α? Does any causal claim come from a randomized experiment and not an observational study? Fixing one missing "in context" phrase can be worth a full point.

Worked through

Here is an original practice question in the style of FRQ 3, the Inference question. The 10-point breakdown is our own illustration of how points are typically spread, not an official rubric.

The question. A clinic recruited 200 volunteers who take daily blood-pressure medication. Researchers randomly assigned 100 of them to receive a daily text reminder and the other 100 to receive no reminder. After 30 days, 74 of the reminder group and 61 of the no-reminder group had taken every dose. (a) Do the data give convincing evidence, at the α = 0.05 level, that text reminders increase the proportion of patients like these who take every dose? (b) Describe a Type I error in context, and give one consequence of making it.

Model response, part (a).

Hypotheses. Let p_R be the true proportion of patients like these who would take every dose with reminders, and p_N the true proportion without reminders. H₀: p_R − p_N = 0. Hₐ: p_R − p_N > 0.

Procedure and conditions. I will use a two-sample z-test for a difference in proportions. Random: the patients were randomly assigned to the two groups. Large counts: the pooled proportion is (74 + 61)/200 = 0.675. The expected counts are 100(0.675) = 67.5 successes and 100(0.325) = 32.5 failures in each group, and all are at least 10. The 10% condition does not apply, because this is an experiment with random assignment, not a sample drawn from a population.

Mechanics. The sample proportions are 0.74 and 0.61, so the difference is 0.13. The standard error is √[0.675 × 0.325 × (1/100 + 1/100)] ≈ 0.0662. The test statistic is z = (0.13 − 0)/0.0662 ≈ 1.96. The p-value is P(Z ≥ 1.96) ≈ 0.025.

Conclusion. Because the p-value of 0.025 is less than α = 0.05, I reject H₀. There is convincing evidence that text reminders increase the true proportion of patients like these who take every dose. Because the treatments were randomly assigned, the reminders can be said to cause the increase. The volunteers were not a random sample, so the result applies to patients like these volunteers, not to every patient.

Model response, part (b). A Type I error would mean concluding that reminders raise the proportion who take every dose when in reality they do not. One consequence is that the clinic could pay for a reminder system that does not help patients take their medication.

Where the points come from. Hypotheses earn 2 points: one for correct symbols with the right direction, and one for defining the parameters in words as true proportions of a named group. The procedure and conditions earn 2 points: one for naming the correct test, and one for checking random assignment and large counts with real numbers. The mechanics earn 2 points: one for a correct test statistic with work shown, and one for the p-value. The conclusion earns 2 points: one for linking the p-value to α, and one for stating the result about the true proportions in context. Part (b) earns 2 points: one for the error in context and one for a consequence that follows from it.

Where a mid-range answer loses points. Here is how a typical 3-point response adds up, line by line.

Hypotheses, 0 of 2. It writes 0.74 > 0.61, using sample values instead of parameter symbols, and never defines a parameter in words, so it loses both points.

Procedure and conditions, 1 of 2. It names the two-sample z-test for a difference in proportions, which keeps one point. It then writes "conditions met" without showing random assignment or any counts, which loses the other.

Mechanics, 2 of 2. It gets z ≈ 1.96 with the formula shown and a p-value of about 0.025, so it keeps both points.

Conclusion, 0 of 2. It writes "we accept that reminders work." That never mentions α, and "accept" says nothing about the true proportions, so it loses both points.

Part (b), 0 of 2. It says a Type I error is rejecting H₀ when H₀ is true, but never says what that means for these patients and gives no consequence, so it loses both points.

That is 0 + 1 + 2 + 0 + 0 = 3 points out of 10. Every one of those losses is a writing problem, not a math problem.

Common mistakes

  • Writing hypotheses with sample statistics such as x̄ or p̂, or with no parameter defined in words. Hypotheses are always about the true population value, such as the true mean or true proportion of a named group.

  • Saying "conditions are met" without showing them. Each condition needs evidence from the problem: where the randomness came from, the actual expected counts for large counts, or the sample size or a graph for Normality.

  • Concluding "accept H₀" or "this proves the claim." A large p-value means you fail to reject H₀ because the evidence is not convincing. A small p-value gives convincing evidence for Hₐ, but it still does not prove it.

  • Interpreting a p-value as the probability that H₀ is true, or a 95% confidence level as meaning 95% of the data fall in the interval. The p-value assumes H₀ is true, and the confidence level describes how often the method captures the true parameter.

  • Describing two distributions separately instead of comparing them, or using the mean and standard deviation for skewed data with outliers. Use comparison words with numbers, and switch to the median and IQR when the data are skewed.

  • Claiming cause and effect from an observational study, or generalizing an experiment on volunteers to the whole population. Random assignment supports causation, while random sampling supports generalizing to a population.

Checklist

  • I can name the job of each of the four free-response questions and budget about 22 minutes for each.
  • I can write hypotheses in symbols and define the parameter in words, using the true population value and not a sample statistic.
  • I can name the correct inference procedure for one or two proportions, one or two means, paired data, or a chi-square test.
  • I can check every condition with actual numbers from the problem: at least 10 successes and 10 failures for z procedures on proportions, and all expected counts at least 5 for a chi-square test.
  • I can write a test conclusion that links the p-value to α and states the result about the parameter in context.
  • I can interpret a confidence interval and a confidence level without saying that 95% of the data fall inside.
  • I can interpret the slope, y-intercept, a residual and r² of a least-squares line in context, and use a residual plot to judge whether a line fits.
  • I can show a Normal or binomial probability with the distribution, its parameters and the event written out, not just calculator syntax.
  • I can compare two distributions on shape, center, variability and outliers using comparison words and values.

More AP Statistics technique