Course guides

Which Statistical Test to Use: A Chart, a Flowchart and 10 Worked Examples

Every inference question on a stats final starts with one choice: z or t, one sample or two, paired or independent, interval or test. Four questions settle it. Here are the chart, the flowchart, the conditions and 10 prompts with answers.

By the Lykke team18 min read

Key takeaways

  • Classify the response first: a category or a yes/no points to proportions or chi-square, and a number with units points to means or regression.
  • Two lists of numbers from the same subjects, or from matched pairs, are paired: subtract within each pair and run a one-sample t on the differences.
  • Proportions use z. Means use t, because the population standard deviation is almost never known.
  • "Estimate" asks for a confidence interval and "convincing evidence" asks for a test. Check the random, 10% and shape conditions before you calculate.
  • AP Statistics dropped chi-square goodness of fit and slope inference for 2026–27. If your college syllabus lists them, study them anyway.
In this guide
  1. How to choose a statistical test in four questions
  2. When to use which statistical test: the chart
  3. Which statistical test to use: a flowchart
  4. Confidence interval or hypothesis test?
  5. Check the conditions before you calculate
  6. 10 practice prompts: name the procedure, then say why
  7. The traps that cost the most points
  8. How to practice choosing the right test before your final
  9. Beyond the intro final: ANOVA, nonparametric tests and AP's 2026–27 changes
  10. FAQ
  11. Sources

Every inference question on an intro statistics final starts with the same decision: which statistical test to use. Once you know the procedure, a calculator handles most of the arithmetic. Pick the wrong procedure, though, and the hypotheses, the conditions and the conclusion all go wrong with it.

Four questions make the choice for you. Is the response variable categorical or quantitative? How many samples or groups are there, or are two variables measured on each subject? If there are two sets of numbers, are they paired or independent? And does the question want an estimate (a confidence interval) or a verdict on a claim (a hypothesis test)? Answer those, check the conditions, and the procedure has a name.

Below are the decision chart, a text flowchart, the conditions for each procedure, 10 practice prompts with answers and the traps that cost points. The guide works for a standard intro course and for AP Statistics, which covers introductory college-level material, and it flags what the 2026–27 AP revision dropped. For studying the rest of the course, see how to pass statistics.

How to choose a statistical test in four questions

College Board's sample inference question for the revised AP Statistics exam shows where the points sit. It is scored out of 10. Three points go to identifying the variables, classifying them as categorical or quantitative and naming the procedure; two go to the conditions; one goes to the test statistic and p-value together (AP Statistics Course and Exam Description, sample Question 3). Your professor's rubric may weigh things differently, but every later step rests on the first three.

1. Is the response categorical or quantitative?

Ask what gets recorded for one subject. A category, such as a major, a brand or yes/no, makes the variable categorical, so you'll work with counts and proportions. A number with units, such as minutes, dollars or exam points, makes it quantitative, so you'll work with means, or with a regression line when two such variables are measured together. The free AP Statistics section on variables drills the classification, including numbers that are really labels, like jersey numbers.

Watch the word "percent." "The percent of students who passed" describes a categorical variable, because each student either passed or didn't. "The percent of each paycheck a student saves" is a number recorded for every student, which makes it quantitative: you'd average it.

2. How many samples, groups or variables?

  • One sample: estimate one value, or test it against a claimed value.
  • Two groups: two separate samples, or two treatment groups in a randomized experiment.
  • Three or more groups: chi-square for a categorical response; ANOVA for a quantitative one, if your course teaches it.
  • Two variables on each subject: two categorical variables go in a two-way table (chi-square), and two quantitative ones go on a scatterplot (regression).

3. Paired or independent?

Two sets of measurements are paired when each value has one specific partner in the other set: the same subject measured twice, or two subjects matched on purpose. The midterm and final scores of the same 30 students are paired by student; scores from 30 students in one section and 30 in another are independent. Paired data collapse to a single sample: subtract within each pair and analyze the differences (OpenStax, 10.4). NIST's statistics handbook treats before-and-after measurements on the same objects the same way (NIST/SEMATECH, 7.3.1.1).

4. Estimate or test?

"Estimate," "how large" or "what values are plausible" asks for a confidence interval. "Is there convincing evidence," "do the data suggest" or a stated significance level asks for a test. The interval-or-test section below covers the gray areas.

Then check the conditions for that procedure, and only then calculate.

When to use which statistical test: the chart

Write the procedure's full name: how many samples, z or t, interval or test, and the parameter. The names below follow College Board's course description. The goodness-of-fit, slope and ANOVA rows come from OpenStax's Introductory Statistics and Penn State's STAT 200 notes, because the AP course doesn't include them.

What you haveParameterProcedureStatistic
One sample, categorical response with two outcomespOne-sample z-interval or z-test for a population proportionz
Two independent samples or treatment groups, categorical response with two outcomesp₁ − p₂Two-sample z-interval or z-test for the difference between two population proportionsz (the test pools the two samples)
One sample, one categorical variable compared with claimed percentagesthe whole distributionChi-square goodness-of-fit testχ², df = categories − 1
One sample, two categorical variablesassociationChi-square test for independenceχ², df = (rows − 1)(columns − 1)
Two or more samples or treatment groups, one categorical variablethe distributionsChi-square test for homogeneityχ², df = (rows − 1)(columns − 1)
One sample, quantitative responseμOne-sample t-interval or t-test for a population meant, df = n − 1
Paired data, quantitative responseμ_dOne-sample t-interval or t-test for a population mean difference (paired t)t, df = number of pairs − 1
Two independent samples or treatment groups, quantitative responseμ₁ − μ₂Two-sample t-interval or t-test for the difference between two population meanst, df from your calculator
Two quantitative variables measured on each subjectβ₁ (the slope)t-interval or t-test for the slope of the regression linet, df = n − 2
Three or more groups, quantitative responseall means equalOne-way ANOVAF

Proportions use z. Means use t, because you estimate the spread with the sample standard deviation s: a z procedure for a mean needs the population standard deviation, which OpenStax notes is rarely known (9.3). In a two-sample t, keep each sample's own variance. OpenStax's instruction is not to pool them (10.1), and the standard error on the AP formula sheet is unpooled too. To see why t has fatter tails than z, the free t-distribution demo lets you shrink the sample size and compare the chance of a value beyond 2 under t with the normal curve's 2.28%.

Which statistical test to use: a flowchart

Work down from the top. Every branch ends at a row of the chart.

  1. Is the response categorical?
    • One sample, two outcomes → one-sample z for p.
    • Two independent groups, two outcomes → two-sample z for p₁ − p₂.
    • One variable with three or more categories, checked against claimed percentages → chi-square goodness of fit.
    • A two-way table from one sample → chi-square test for independence.
    • A two-way table from separate samples or treatment groups → chi-square test for homogeneity.
  2. Is the response quantitative?
    • One sample → one-sample t for μ.
    • Two linked measurements per subject or per matched pair → paired t on the differences.
    • Two independent groups → two-sample t for μ₁ − μ₂.
    • Three or more groups → one-way ANOVA, if your course covers it.
  3. Are both variables quantitative and measured on the same subjects? → t procedure for the regression slope.
  4. Interval or test? "Estimate" → interval. "Convincing evidence" or a given α → test. Chi-square tests and ANOVA come only as tests in an intro course.
  5. Conditions: random, 10% and the shape check for that procedure. If one fails, say so in your answer instead of skipping it.

Confidence interval or hypothesis test?

The wording of the question usually decides:

The question saysUse
"Estimate," "how much," "how large is the difference," "construct a 95% interval"Confidence interval
"Is there convincing evidence," "does this suggest," "at α = 0.05," "test the claim"Hypothesis test
"Does the interval support the claim that…?"An interval, read as a test

The two are linked. NIST's handbook states that for every hypothesis test there is an equivalent statement about whether the hypothesized value falls inside a confidence interval (NIST/SEMATECH, 7.1.5). So if a 95% interval for a difference in means leaves out 0, a two-sided test at α = 0.05 would reject "no difference." College Board's course description frames an interval the same way: a range of plausible values that can serve as evidence about a claim. The AP course's free section on justifying a claim with an interval practices exactly that move, and the free confidence interval simulator shows what a confidence level means: set a true proportion, draw sample after sample, and count how many intervals capture it.

Check the conditions before you calculate

In College Board's course description, every procedure gets three checks, and only the third one changes (Units 3 and 4). The table adds the slope test from Penn State's notes.

  1. Random: a random sample, or random assignment in an experiment. The two-sample procedures need two independent random samples or a randomized experiment.
  2. 10% condition: when you sample without replacement, each sample is at most 10% of its population. The course description notes that this check isn't needed when the data come from a randomized experiment.
  3. The shape check, which depends on the procedure:
ProcedureShape check
One-sample z-interval for pat least 10 successes and 10 failures in the sample
One-sample z-test for pnp₀ and n(1 − p₀) at least 10, using the claimed value p₀
Two-sample z-interval for p₁ − p₂at least 10 successes and 10 failures in each sample
Two-sample z-test for p₁ − p₂the same four counts, computed with the pooled proportion p̂c
Any chi-square testevery expected count at least 5 (expected counts, not observed ones)
t for a mean or a mean differencethe population is roughly normal, or n ≥ 30, or the sample shows no strong skew or outliers; for paired data, check the differences
Two-sample tboth samples 30 or more, or both populations roughly normal; otherwise neither sample shows strong skew or outliers
t for a slopea linear pattern, independent errors, roughly normal residuals and equal spread in the residual plot

Check the cutoffs your own course uses. OpenStax's textbook uses 5 where AP uses 10: np and nq above five for one proportion (9.3), and at least five successes and five failures in each sample for two (10.3). The slope conditions come from Penn State's STAT 200 notes (Lesson 12), since the revised AP course no longer covers slope inference.

10 practice prompts: name the procedure, then say why

These prompts were written for this guide, not taken from real studies. Cover each answer and commit to a procedure before you read it.

1. "A random sample of 400 of a university's 30,000 students finds that 148 work at least 10 hours a week at a paid job. Estimate the proportion of all its students who do." One-sample z-interval for a population proportion. Each student is a yes or a no, there is one sample, and "estimate" asks for an interval. Checks: 148 successes and 252 failures are both at least 10, and 400 is under 10% of 30,000.

2. "A dining hall says 60% of students eat breakfast there at least once a week. In a random sample of 150 students, 78 do. At α = 0.05, is there convincing evidence that the true proportion is lower?" One-sample z-test for a population proportion, with H₀: p = 0.60 and Hₐ: p < 0.60. One yes/no variable, one sample, a claim and a significance level. In a test, the counts check uses the claimed value: 150 × 0.60 = 90 and 150 × 0.40 = 60, both at least 10.

3. "Researchers randomly assign 200 seeds to a new coating and 200 to no coating. 162 coated and 143 uncoated seeds sprout. Is there convincing evidence that the coating changes the sprouting rate?" Two-sample z-test for the difference between two population proportions, two-sided. Sprouted or not is categorical, and there are two treatment groups. Checks: random assignment; with the pooled proportion 305/400 = 0.7625, each group's expected counts are 152.5 sprouts and 47.5 failures, all at least 10; no 10% check for an experiment.

4. "The one-way commute times of 25 randomly selected employees have a mean of 27.4 minutes and a standard deviation of 8.1 minutes. Estimate the mean commute for all employees." One-sample t-interval for a population mean, df = 24. Minutes are quantitative, there is one sample, and you have s, not σ, so it's t, not z. Because 25 is under 30, look at a dotplot or boxplot of the 25 times for strong skew or outliers before you trust the interval.

5. "Fifteen runners each run a timed mile in two shoe models, in random order. Do the mean times differ between the models?" Paired t-test: a one-sample t-test for the population mean difference μ_d on the 15 differences (model A minus model B), df = 14. Each runner contributes two linked times. A two-sample t would assume the two lists are independent, and they aren't: a fast runner is fast in both shoes. Check the 15 differences, not the raw times, for strong skew or outliers.

6. "Independent random samples of 45 first-year students who live on campus and 38 who commute report last night's hours of sleep. Is there convincing evidence of a difference in mean sleep?" Two-sample t-test for the difference between two population means. Hours are quantitative, the groups are different people, and nothing links a student in one group to a student in the other. Both samples have at least 30, so the shape check passes, and your calculator finds the degrees of freedom, somewhere between 37 and 81. The free Welch's t-test demo lets you move each group's mean, spread and size and watch the sampling distribution of the difference respond.

7. "A candy company says its bags are 25% red, 25% orange, 20% yellow, 15% green and 15% purple. Do the colors of 300 randomly selected candies contradict the claim?" Chi-square goodness-of-fit test, df = 5 − 1 = 4. One categorical variable with five categories, one sample, compared with claimed percentages. The expected counts are 75, 75, 60, 45 and 45, all at least 5 (OpenStax, 11.2). The free goodness-of-fit demo also uses candy colors: drag the observed counts and watch χ² and the p-value move.

8. "A random sample of 500 students records each student's class year (four years) and where they study most (library, home or café). Is study spot associated with class year?" Chi-square test for independence, df = (4 − 1)(3 − 1) = 6: one sample, two categorical variables. Now change the design: "Separate random samples of 150 students at each of three colleges are asked where they study most." That's a chi-square test for homogeneity, df = (3 − 1)(3 − 1) = 4: three samples, one variable. The tables look alike and the arithmetic is identical, so the way the data were collected decides the name. The AP course's free section on setting up chi-square tests includes a misconception check on exactly this.

9. "For 30 randomly selected used cars of one model, you know each car's age and price. Is there convincing evidence of a linear relationship between age and price?" t-test for the slope of the population regression line: H₀: β₁ = 0, t = b₁ / SE(b₁), df = 30 − 2 = 28 (Penn State STAT 200, 12.3.4 and 12.3.5). Both variables are quantitative and measured on the same cars. Check the scatterplot and the residual plot before you test. The free slope sampling demo draws samples of 15 points from a population line you set and shows how much the fitted slope varies from sample to sample.

10. "Exam scores are compared for students taught in three formats: in person, online and hybrid. Do the mean scores differ?" One-way ANOVA, if your course teaches it. The response is quantitative and there are three groups. Running a two-sample t test for each pair would raise the chance of a Type I error (Penn State STAT 200, Lesson 10). AP Statistics doesn't include ANOVA, so expect this one only in a college course.

The traps that cost the most points

  • Treating paired data as independent. Look for "before and after," "each subject," "the same," "twins," "couples" or "matched." If each number has a partner, analyze the differences.
  • Calling a measured percent a proportion. When every subject has its own percentage (body fat, share of income saved, battery charge left), it's a measurement and gets a t procedure.
  • Using z for a mean. Unless the problem hands you σ, use t at any sample size.
  • Checking the wrong counts. Chi-square conditions use expected counts, not observed ones; College Board's course description calls that out as the precise wording. A one-proportion test checks with p₀, an interval with p̂, and a two-proportion test with the pooled p̂c.
  • Naming the procedure loosely. In College Board's sample scoring, "two-sample z-test" on its own didn't earn the procedure point, and neither did "two-proportion z-interval" for a question that asked for a test. "Two-sample z-test for population proportions" did.
  • Writing "the variables are independent" after a large p-value. The course description counts that as accepting the null hypothesis. Write that the data don't give convincing evidence of an association. For more wording traps, see the p-value mistakes in the statistics course guide.

How to practice choosing the right test before your final

Choosing the test is a separate skill from carrying it out, and research on statistics word problems suggests how to practice it. In a set of experiments by Quilici and Mayer, students who studied example problems grouped by test (t test, correlation or chi-square) were more likely than students who saw no examples to sort new problems by their structure rather than their cover story. In the third experiment, students who studied and practiced four worked examples that stressed structure picked the right test for new problems more often than students whose examples stressed surface details, and the gain was strong for lower-ability students but not for higher-ability ones (Quilici & Mayer, 1996). So practice classifying problems by structure, across chapters, before you practice solving them.

A drill for the week before the final:

  1. Pull 20 inference problems from your homework, old exams and the textbook's review sets, and cover the chapter titles.
  2. For each one, write five labels and nothing else: variable type, number of samples or groups, paired or independent, interval or test, and the procedure's full name.
  3. Check them against the solutions. For every miss, write down the cue you skipped ("each runner ran in both shoes" → paired).
  4. Solve, in full and with conditions, only the problems you misclassified.
  5. Two days later, repeat with the problems in a new order. Turning the cues into flashcards works too, and active recall explains why answering from memory beats rereading your chart.

For full write-ups, the AP course's free inference free-response practice walks through a complete answer with a 10-point rubric. Its sections on setting up tests for one proportion, two proportions, a mean or a paired mean difference and two means each have exam-style free-response questions with model answers. Classes that use OpenStax's book can practice on Lykke's Introductory Statistics course page instead, free to open, whose sections match the book's chapters; start with hypothesis testing with two samples and the chi-square distribution. Signed out, a course page shows its first two sections; signing in on any plan opens the rest.

If statistics is one of several finals, split your study time across every exam before you start the drill.

Beyond the intro final: ANOVA, nonparametric tests and AP's 2026–27 changes

Your college course may go further than AP does. Here is what to recognize, even if you never calculate it.

One-way ANOVA

For a quantitative response across three or more groups, one-way ANOVA tests whether all the population means are equal; the alternative is that at least two of them differ (OpenStax, 13.1). It assumes normal populations with equal standard deviations, plus random, independent samples. Penn State's STAT 200 explains why you don't just compare every pair: k groups make k(k − 1)/2 pairs, so five groups would take 10 t tests and a higher chance of a Type I error. A significant F says some means differ, not which ones; Penn State's lesson follows it with Tukey's pairwise comparisons. Lykke's Introductory Statistics course page has a section on the F distribution and one-way ANOVA; it's reading only, with no quiz.

Nonparametric tests

When the response is ordinal (rankings, or ratings on an ordered scale from "strongly disagree" to "strongly agree") or a small sample is clearly far from normal, courses that teach rank-based methods swap in a counterpart. UCLA's Office of Advanced Research Computing pairs them like this, as general guidelines rather than hard rules (UCLA OARC):

DesignTest from the chartCounterpart
Two independent groups, measured responseTwo-sample tWilcoxon–Mann–Whitney
Two matched groups, measured responsePaired tWilcoxon signed-ranks
Two or more independent groups, measured responseOne-way ANOVAKruskal–Wallis
Two independent groups, categorical responseChi-squareFisher's exact test
Two matched groups, categorical responsenot in the chartMcNemar

The last row hides a trap even for an intro final. If the same people answer a yes/no question twice, the two proportions don't come from independent samples, so the two-sample z procedure doesn't apply.

What AP's 2026–27 revision dropped

College Board's revision of AP Statistics for 2026–27 removed the chi-square goodness-of-fit test and inference for regression slopes (the old Unit 9), and the first exam on the revised course is in May 2027 (AP Central). OpenStax's textbook still teaches goodness of fit, and Penn State's STAT 200 notes still teach the slope test, so if your syllabus lists them, they're fair game on your final.

Frequently asked questions

When do you use a t test instead of a z test?

Use z for proportions and t for means. A z procedure for a mean requires the population standard deviation σ, and in practice you almost never have it, so you estimate the spread with the sample standard deviation s and the statistic follows a t distribution (df = n − 1 for one sample). Proportion procedures use z, as long as the large-counts condition holds. If a problem really does give you σ, a z procedure for the mean is allowed; a standard deviation calculated from the data is s, not σ.

What is the difference between a paired t test and an independent t test?

A paired t test is for measurements that come in linked pairs, such as the same people before and after, or deliberately matched subjects. You subtract within each pair and run a one-sample t on the differences, with degrees of freedom equal to the number of pairs minus 1. An independent (two-sample) t test compares two separate groups with no individual links, and your calculator or software supplies its degrees of freedom. Ask whether each value in one list has one specific partner in the other.

When would you use a chi-square test?

When the data are counts in categories and the question is about a whole distribution rather than one proportion. Goodness of fit compares one categorical variable with claimed percentages, the test for independence asks whether two categorical variables recorded on one sample are associated, and the test for homogeneity asks whether several samples or treatment groups share one distribution. All three need every expected count to be at least 5, and none of them has a confidence interval version in an intro course.

What is the difference between the chi-square test of independence and homogeneity?

The design. Independence starts from one random sample, records two categorical variables on each subject and asks whether they're associated. Homogeneity starts from two or more separate samples or treatment groups, records one categorical variable and asks whether its distribution is the same in each. The expected counts, the statistic and the degrees of freedom, (rows − 1)(columns − 1), are computed the same way, so what changes is the wording of your hypotheses and your conclusion.

Which statistical test should I use to compare two groups?

It depends on the response. For a yes/no response, use a two-sample z procedure for the difference between two proportions. For a measured response, use a two-sample t procedure for the difference between two means, unless the groups are linked pair by pair, in which case run a paired t on the differences. If the response is ordinal, or the samples are small and clearly non-normal, courses that teach rank-based methods use the Wilcoxon–Mann–Whitney test instead.

What test do I use to compare three or more groups?

For a categorical response, a chi-square test for homogeneity. For a quantitative response, a one-way ANOVA, which tests whether all the group means are equal. Don't run a t test for every pair: Penn State's STAT 200 notes point out that five groups would take 10 tests and raise the chance of a Type I error. AP Statistics doesn't cover ANOVA, so this applies to college courses that teach it.

Is the chi-square goodness-of-fit test still on the AP Statistics exam?

No. College Board's revision of AP Statistics for the 2026–27 school year removed the chi-square goodness-of-fit test and inference for regression slopes, and the first exam on the revised course is in May 2027. The chi-square tests for independence and homogeneity remain. College courses set their own syllabus, and textbooks such as OpenStax's still include goodness of fit, so go by what your syllabus lists.

Sources

  1. AP Statistics Course and Exam Description, Effective Fall 2026 — College Board
  2. AP Statistics: Future Revisions — College Board, AP Central
  3. 9.3 Probability Distribution Needed for Hypothesis Testing — OpenStax, Introductory Statistics 2e
  4. 10.1 Two Population Means with Unknown Standard Deviations — OpenStax, Introductory Statistics 2e
  5. 10.3 Comparing Two Independent Population Proportions — OpenStax, Introductory Statistics 2e
  6. 10.4 Matched or Paired Samples — OpenStax, Introductory Statistics 2e
  7. 11.2 Goodness-of-Fit Test — OpenStax, Introductory Statistics 2e
  8. 13.1 One-Way ANOVA — OpenStax, Introductory Statistics 2e
  9. 7.1.5. What is the relationship between a test and a confidence interval? — NIST/SEMATECH e-Handbook of Statistical Methods
  10. 7.3.1.1. Analysis of paired observations — NIST/SEMATECH e-Handbook of Statistical Methods
  11. Lesson 10: One-Way ANOVA — Penn State Department of Statistics, STAT 200: Elementary Statistics
  12. Lesson 12: Correlation & Simple Linear Regression — Penn State Department of Statistics, STAT 200: Elementary Statistics
  13. What statistical analysis should I use? — UCLA Office of Advanced Research Computing (OARC) Statistical Methods and Data Analytics
  14. Role of examples in how students learn to categorize statistics word problems — Journal of Educational Psychology, Quilici & Mayer, 1996

This guide was researched from the sources above, drafted with AI assistance and checked against those sources before it was published. Dates, deadlines and offers change: check the official page before you act. Found something wrong or out of date? Email support@getlykke.com.

Keep reading

Lykke · the personal assistant for students

One Memory. Every AI.

The free Lykke Chrome extension syncs your Canvas courses into one memory, and ChatGPT, Claude and Muse can answer from your classes and deadlines. It only reads Canvas, it doesn't schedule anything on its own, and it never turns anything in for you.