Eduovisual
Biostatistics & Population Health
Paired vs unpaired t-test: selection and interpretation
— Same patients measured before vs after an intervention (BP before/after lisinopril, HbA1c pre/post lifestyle program)
— Matched-pair designs: twins, left vs right eye, case-control matched on age/sex
— Crossover trials where each subject receives both treatments
— Two distinct cohorts compared at one time point (drug vs placebo arms, smokers vs nonsmokers)
— Randomized parallel-group trials with no within-subject linkage
— Cross-sectional comparisons of independent populations
— Continuous outcome
— Approximate normality (or n large enough for CLT, typically >30)
— For unpaired: roughly equal variances (else Welch's t)
— For paired: normality applies to the differences, not raw values
Board pearl: If a stem mentions "each patient served as their own control," the answer is paired t-test — this phrase is the single most reliable trigger on Step 3 biostatistics items.
— "Twenty-four patients had their fasting glucose measured at baseline and 12 weeks after metformin initiation."
— "Each subject's symptom score was recorded before and after cognitive behavioral therapy."
— "Patients with bilateral cataracts received drug A in one eye and placebo in the contralateral eye."
— "Cases were matched 1:1 to controls on age, sex, and BMI."
— Crossover wording: "received treatment A for 4 weeks, washout, then treatment B."
— "Group 1 (n=50) received atorvastatin; group 2 (n=50, a separate cohort) received placebo."
— "Mean systolic BP was compared between men and women."
— "Patients were randomized to two parallel arms."
— No mention of repeat measurement or matching
— A randomized trial that also collects baseline and follow-up — the between-group comparison is unpaired (or ANCOVA), while a within-arm pre/post comparison is paired
— "Identical twins discordant for smoking" = paired (matched)
— Two unrelated groups of the same size are still unpaired — equal n ≠ paired
Key distinction: Pairing is defined by data structure (linkage between observations), not by sample size, randomization status, or whether the groups look demographically similar. Always ask: "Is there a natural one-to-one correspondence between a value in group 1 and a value in group 2?" If yes → paired; if no → unpaired. This single question resolves nearly every Step 3 t-test selection item you will encounter on the exam.
— Outcome variable type: Continuous (BP, weight, lab value, score) → t-test family is on the table. Categorical → chi-square or Fisher's. Time-to-event → log-rank.
— Number of groups: Two → t-test. Three or more → ANOVA (not t-test).
— Linkage: Repeated measurements on same subjects, or matched pairs → paired. Independent cohorts → unpaired.
— Distribution: Roughly symmetric/normal → parametric t-test. Skewed or ordinal → Wilcoxon signed-rank (paired) or Mann-Whitney U / Wilcoxon rank-sum (unpaired).
— n < 15 per group with non-normal data → strongly prefer nonparametric
— n > 30 per group → Central Limit Theorem rescues normality assumption for the sampling distribution of the mean, so t-test is generally acceptable
— If SDs in the two groups are similar → Student's t-test (pooled variance)
— If SDs differ markedly (rule of thumb: ratio > 2) → Welch's t-test (unequal variances)
— Modern practice often defaults to Welch's regardless
— Three arms analyzed by t-test → should be ANOVA with post-hoc correction
— Ordinal Likert outcome analyzed by t-test → consider nonparametric
— Repeated measures analyzed as independent samples → inflated p-value, wasted power
Board pearl: When a Step 3 stem gives you a small sample (n=10–15) with a skewed or ordinal outcome and pre/post data, the answer is Wilcoxon signed-rank, not paired t-test. Recognizing the nonparametric paired analog is a frequent next-step distractor.
— Step 1: For each subject, compute the difference: dᵢ = post − pre (or treatment − control eye)
— Step 2: Calculate mean difference d̄ and standard deviation of differences s_d
— Step 3: Test statistic: t = d̄ / (s_d / √n) where n = number of pairs
— Step 4: Degrees of freedom = n − 1 (one fewer than for unpaired with same total observations)
— Step 5: Compare t to critical value or compute p-value
— H₀: mean difference = 0 (no change from intervention)
— H₁: mean difference ≠ 0 (two-sided) or > 0 / < 0 (one-sided)
Step 3 management: When interpreting a single-arm pre/post study with a significant paired t-test, recommend a randomized controlled trial before changing practice — pairing controls within-subject variability but not external confounders.
— Step 1: Compute means x̄₁ and x̄₂, and SDs s₁ and s₂ for the two independent groups
— Step 2: Compute pooled SD (Student's): s_p = √[((n₁−1)s₁² + (n₂−1)s₂²) / (n₁+n₂−2)]
— Step 3: Test statistic: t = (x̄₁ − x̄₂) / [s_p × √(1/n₁ + 1/n₂)]
— Step 4: Degrees of freedom = n₁ + n₂ − 2 (Student's) or Welch-Satterthwaite approximation (Welch's)
— Step 5: Compare to critical value or compute p
— H₀: μ₁ = μ₂ (population means equal)
— H₁: μ₁ ≠ μ₂
— Independence of observations within and between groups
— Approximate normality of each group (or large n)
— Variance homogeneity (Student's only; check with Levene's test or rule-of-thumb SD ratio)
— Given means, SDs, and ns: estimate whether result is significant using "rule of 2 SE" — if difference > 2 × SE, p < 0.05
— Recognize that a p-value < 0.05 does not equal clinical significance — a 1 mmHg BP difference can be statistically significant in a huge trial but clinically trivial
— Distinguish statistical power issues: nonsignificant result in small n may reflect type II error, not equivalence
Key distinction: A paired design with the same total observations as an unpaired design generally has more power because within-subject correlation reduces error variance — this is why crossover trials need fewer patients than parallel-group trials for the same effect size.
1. Outcome type?
— Continuous → proceed to step 2
— Binary categorical → chi-square / Fisher's exact / z-test for proportions
— Time-to-event → Kaplan-Meier with log-rank
— Ordinal → Mann-Whitney / Wilcoxon
2. How many groups?
— Two → t-test family
— ≥3 → ANOVA (one-way, repeated-measures, or Kruskal-Wallis if nonparametric)
3. Are observations independent or linked?
— Independent cohorts → unpaired t-test (Student's or Welch's)
— Same subjects measured twice, or matched pairs → paired t-test
4. Are normality/variance assumptions met?
— Yes → parametric t-test
— No (small n + skew, or ordinal) → nonparametric equivalent:
▪ Paired → Wilcoxon signed-rank
▪ Unpaired → Wilcoxon rank-sum / Mann-Whitney U
5. Adjusting for covariates needed? → linear regression or ANCOVA instead
— Paired designs require fewer subjects to detect the same effect due to lower variance
— A general rule: n_paired ≈ n_unpaired × (1 − ρ), where ρ is the within-pair correlation
— If ρ is high (e.g., BP repeated in same person, ρ ≈ 0.7), paired design can need <40% as many subjects
— Using a t-test for more than two groups → inflated type I error; must use ANOVA with post-hoc Bonferroni or Tukey
— Failing to recognize matched case-control as paired data
Board pearl: The phrase "which test is most appropriate?" appears in nearly every Step 3 biostatistics item — answer by walking the algorithm: outcome type → number of groups → linkage → distribution. This 4-step heuristic resolves >90% of selection questions on the exam.
| • Scenario: A family physician enrolls 10 hypertensive patients in a 12-week lifestyle program. Systolic BP (mmHg) is measured at baseline and 12 weeks. | |||
| Patient | Baseline | 12 wk | Δ (post−pre) |
| 1 | 150 | 140 | −10 |
| 2 | 158 | 145 | −13 |
| 3 | 142 | 138 | −4 |
| 4 | 165 | 150 | −15 |
| 5 | 148 | 142 | −6 |
| 6 | 155 | 148 | −7 |
| 7 | 160 | 152 | −8 |
| 8 | 145 | 140 | −5 |
| 9 | 152 | 144 | −8 |
| 10 | 156 | 150 | −6 |
| • Calculations: | |||
| — Mean difference d̄ = −8.2 mmHg | |||
| — SD of differences s_d ≈ 3.5 mmHg | |||
| — SE = s_d / √n = 3.5 / √10 ≈ 1.11 | |||
| — t = −8.2 / 1.11 ≈ −7.4 | |||
| — df = n − 1 = 9; critical t at α=0.05 two-tailed ≈ 2.26 | |||
| — | t | = 7.4 >> 2.26 → reject H₀, p < 0.001 | |
| • 95% CI for mean change: −8.2 ± 2.26 × 1.11 = (−10.7, −5.7) mmHg — excludes zero, consistent with significant within-subject reduction. | |||
| • Clinical interpretation: SBP dropped on average ~8 mmHg over 12 weeks, statistically unlikely due to chance. However, without a control group, regression to the mean and the Hawthorne effect remain alternative explanations. A randomized trial is needed before claiming the program caused the reduction. | |||
| • What would the unpaired t-test (wrongly applied here) look like? Using group SDs (~6–8 mmHg) and ignoring within-subject correlation, the SE balloons, | t | drops, and p might exceed 0.05 — a false-negative purely due to wrong test choice. | |
| Step 3 management: When a stem provides pre/post data on the same patients, calculate the mean of differences first — this is the single most common computational shortcut tested. |
| • Scenario: A randomized trial compares LDL reduction (mg/dL) at 12 weeks between two independent groups of 50 patients each: drug A vs placebo. | ||
| — Drug A: mean LDL reduction = 42, SD = 18 | ||
| — Placebo: mean LDL reduction = 8, SD = 16 | ||
| • Calculations (Student's t, assuming equal variances): | ||
| — Difference in means: 42 − 8 = 34 mg/dL | ||
| — Pooled SD: s_p = √[((49)(18²) + (49)(16²)) / 98] = √[(15876 + 12544)/98] = √290 ≈ 17.0 | ||
| — SE of difference: s_p × √(1/50 + 1/50) = 17.0 × √0.04 = 17.0 × 0.2 = 3.4 | ||
| — t = 34 / 3.4 = 10.0 | ||
| — df = 98; critical t ≈ 1.98 for α = 0.05 two-tailed | ||
| — | t | = 10 >> 1.98 → reject H₀, p < 0.0001 |
| • 95% CI for difference: 34 ± 1.98 × 3.4 = (27.3, 40.7) mg/dL — robustly excludes zero. | ||
| • Welch's variant: Because SDs are similar (18 vs 16), Student's and Welch's yield nearly identical results. If SDs differed widely (e.g., 18 vs 40), Welch's would be preferred and df would be reduced. | ||
| • Interpretation: | ||
| — Drug A reduces LDL by ~34 mg/dL more than placebo, with high statistical confidence | ||
| — Clinical significance: A 34 mg/dL LDL drop translates to meaningful ASCVD risk reduction — both statistically and clinically robust | ||
| — Randomization supports a causal interpretation, unlike the unpaired pre/post design | ||
| • Why this is unpaired: No subject appears in both arms; there is no natural one-to-one linkage between an A-patient and a placebo-patient. Even if n is equal in both arms, the data are independent samples. | ||
| CCS pearl: When designing trial follow-up, an unpaired comparison at a single endpoint is straightforward but discards baseline information — using ANCOVA with baseline as covariate typically increases power further and is preferred for parallel-group RCTs with baseline measurements. |
— Inspect histogram / Q-Q plot for skew
— Run Shapiro-Wilk test of normality (though it's underpowered at small n)
— When in doubt, use the nonparametric analog
— Paired t-test → Wilcoxon signed-rank test: ranks the absolute values of within-subject differences, then tests if the sum of positive ranks differs from negative ranks
— Unpaired t-test → Wilcoxon rank-sum test (= Mann-Whitney U): pools all observations, ranks them, compares rank sums between groups
— Sign test: even more conservative paired alternative; uses only direction (+/−) of changes, ignoring magnitude
— Ordinal outcomes (Likert pain scores 0–10, NYHA class)
— Heavily skewed continuous data (e.g., CRP, ferritin, hospital length of stay)
— Outliers that disproportionately drive the mean
— Very small samples where normality cannot be assessed
Board pearl: A stem with "pain scale 0–10" or "ordinal symptom score" measured before and after treatment in 12 patients should trigger Wilcoxon signed-rank, NOT paired t-test — ordinal data violate the continuous-outcome assumption regardless of sample size.
— Common in chronic stable conditions (HTN, migraine prophylaxis, GERD)
— Analyze with paired t-test on the within-subject difference (Treatment A response − Treatment B response)
— Must check for carryover effects and period effects; if present, use only first-period data (effectively unpaired)
— Cases matched 1:1 (or 1:k) to controls on confounders like age, sex, smoking
— For continuous outcomes → paired t-test on within-pair differences
— For binary outcomes → McNemar's test (paired analog of chi-square)
— Failure to account for matching → loss of efficiency and biased SE estimates
— Monozygotic twin pairs discordant for an exposure (e.g., smoking) → paired analysis on the continuous outcome (lung function)
— Pairing controls for genetic and shared-environmental confounding
— Units of randomization are clinics, schools, or hospitals — not individuals
— Within-cluster correlation requires mixed-effects models or cluster-adjusted t-tests; naive t-tests underestimate variance and inflate type I error
— Cannot use paired t-test (only handles 2 measurements)
— Use repeated-measures ANOVA or linear mixed-effects models
Key distinction: Matched ≠ paired only when matching is frequency matching (group-level) rather than individual pair matching. Individual matching → paired analysis; frequency matching → adjust in regression but analyze as unpaired.
— Using unpaired t-test on paired data → inflated variance → reduced power → type II error → falsely concluding "no effect" when one exists
— Using paired t-test on independent data → artificially deflated variance → spurious significance → type I error
— Each t-test at α = 0.05 carries 5% type I risk; running 20 independent t-tests on the same dataset gives ~64% chance of at least one false positive (1 − 0.95²⁰)
— Step 3 favorite: secondary outcomes / subgroup analyses without correction
— Corrections:
▪ Bonferroni: α/k, very conservative
▪ Holm-Bonferroni: sequentially less conservative
▪ False discovery rate (Benjamini-Hochberg): preferred for many comparisons (genomics, biomarker panels)
Step 3 management: When a study reports a nonsignificant t-test with a small sample, the correct interpretation is "insufficient evidence to detect a difference," not "no difference exists" — this distinction is heavily tested on Step 3 evidence-based-medicine items.
— 3 or more groups are compared (e.g., low-dose vs mid-dose vs high-dose vs placebo)
— One-way ANOVA for independent groups; repeated-measures ANOVA for paired/longitudinal
— Significant ANOVA → post-hoc pairwise comparisons with correction
— Comparing means between groups while adjusting for a continuous covariate (e.g., baseline value, age, BMI)
— Most powerful approach for parallel-group RCTs with baseline measurements — preferred over both unpaired t-test on follow-up alone and paired t-test on change scores
— Multiple covariates need adjustment
— Predictor of interest is continuous (e.g., dose-response)
— t-test is mathematically a special case: an unpaired t-test = linear regression with one binary predictor
— Repeated measures across ≥3 time points
— Clustered data (patients within clinics)
— Missing data at random
— Complex trial designs (adaptive, Bayesian, factorial) require statistician involvement at protocol design, not after data collection
— Pre-specified analysis plans reduce p-hacking and selective reporting
— IRB review increasingly requires statistical justification of sample size
CCS pearl: If a study question stem describes a continuous outcome with baseline and follow-up in randomized arms, the most rigorous analysis is ANCOVA with baseline as covariate — not a paired t-test within each arm and not an unpaired t-test on change scores alone. This is a frequent Step 3 advanced-tier distractor.
— Z-test: Used when population SD is known (rarely in clinical research) or n is very large. Functionally similar to t-test as df → ∞.
— One-sample t-test: Compares a single group's mean to a hypothesized population value (e.g., is mean HbA1c in our clinic different from the national average of 7.5%?). Not a comparison between two groups.
— ANOVA (one-way): Extension to ≥3 independent groups; reduces to unpaired t-test when only 2 groups
— Repeated-measures ANOVA: Extension of paired t-test to ≥3 time points or conditions
— ANCOVA: t-test or ANOVA with covariate adjustment
— Linear regression: Generalizes all of the above; unpaired t-test ≡ regression of outcome on a single binary predictor
— Wilcoxon signed-rank (paired)
— Mann-Whitney U / Wilcoxon rank-sum (unpaired)
— Kruskal-Wallis (≥3 independent groups)
— Friedman test (≥3 paired conditions)
— Normal continuous, 2 independent → unpaired t-test
— Normal continuous, 2 paired → paired t-test
— Skewed continuous or ordinal, 2 independent → Mann-Whitney
— Skewed continuous or ordinal, 2 paired → Wilcoxon signed-rank
— Normal continuous, 3+ groups → ANOVA
— Need covariate adjustment → ANCOVA / regression
— Choosing one-sample t-test when there are clearly two groups
— Choosing ANOVA when there are only two groups (technically valid but t-test is preferred)
— Choosing paired t-test for cross-sectional matched data without true 1:1 linkage
Board pearl: A one-sample t-test stem typically reads "compared to the known national mean of X" — recognizing this single phrase distinguishes it from two-sample tests.
— Chi-square test of independence: Two categorical variables, expected cell counts ≥5
— Fisher's exact test: Same as chi-square but for small expected cell counts (<5) — common in rare disease 2×2 tables
— McNemar's test: Paired binary data (e.g., diagnostic test agreement, before/after categorical outcomes)
— z-test for two proportions: Comparing two independent percentages
— Kaplan-Meier + log-rank test: Compares survival curves between groups
— Cox proportional hazards regression: Hazard ratios with covariate adjustment
— Pearson correlation: Linear association between two continuous variables (parametric)
— Spearman rank correlation: Same but nonparametric / for ordinal data
— Sensitivity, specificity, PPV, NPV, likelihood ratios
— ROC curves with AUC comparisons
— "Percentage achieving HbA1c <7" — that's a proportion → chi-square or z-test, NOT t-test
— "Mortality at 30 days" — binary outcome → chi-square, log-rank, or logistic regression
— "Median hospital length of stay" — highly skewed → Mann-Whitney, not unpaired t-test
— Outcome reported as "%" → categorical
— Outcome reported as "median (IQR)" → likely skewed continuous → nonparametric
— Outcome reported as "mean (SD)" → continuous, t-test family appropriate
Key distinction: "Mean survival" in months → t-test family; "Median survival" with censoring → log-rank. The choice of summary statistic in the stem (mean vs median vs %) is itself a strong clue to the correct test category — a high-yield Step 3 shortcut.
— State the test used and why (paired vs unpaired, parametric vs nonparametric)
— Report mean ± SD (parametric) or median (IQR) (nonparametric) for descriptive stats
— Report the mean difference with 95% CI, not just the p-value
— Report exact p-values (e.g., p = 0.03) rather than "p < 0.05"
— Include effect size (Cohen's d) for interpretation across studies
— Acknowledge assumption checks (normality, variance homogeneity)
— Train yourself to ask: "Is this the right test for this design?"
— Check whether multiple testing has been adjusted for
— Look for pre-specified primary outcomes vs post-hoc exploratory tests
— Assess clinical significance, not just statistical significance — compare to minimal clinically important difference (MCID) for the outcome
— RCTs should report flow diagrams and intent-to-treat analyses
— Observational studies should describe matching strategies explicitly when paired analyses are used
— Share analysis code and de-identified data when possible
— Pre-register hypotheses on platforms like ClinicalTrials.gov
— Use Bonferroni or FDR corrections for secondary outcomes
— Translate "p = 0.001" into "very unlikely due to chance" — not "the treatment definitely works for everyone"
— Explain CI as a plausible range of effect, not a guarantee
— Distinguish average effect (mean difference) from individual variability (some patients won't benefit)
Step 3 management: When counseling a patient about a new therapy based on a trial, explain the absolute effect with CI ("on average BP drops 8 mmHg, between 5 and 11") rather than a p-value — this reflects shared decision-making and supports informed consent.
— Re-read landmark trials critically — identify the primary statistical test and check appropriateness
— Subscribe to JAMA's "Users' Guides to the Medical Literature" and NEJM's statistics primers
— Practice computing simple t-statistics by hand to build intuition (mean diff / SE)
— Use journal clubs to debate test selection in published studies
— Was the primary outcome pre-specified?
— Was the sample size justified with a power calculation?
— Were assumption checks performed (normality, variance)?
— Was the correct test applied for the design?
— Were multiple comparisons addressed?
— Was intent-to-treat analysis used?
— USPSTF and ACP guidelines increasingly cite Bayesian and network meta-analyses — broaden beyond frequentist t-tests
— Familiarize with non-inferiority and equivalence trial designs — these reframe how t-tests and CIs are interpreted
— Recognize adaptive trial designs that modify sample size mid-trial — special methodologies apply
— When reviewing a study, ask: "If the design were paired and I used unpaired, would my conclusion change?"
— Audit your own QI projects for appropriate test choice — pre/post quality improvement data are almost always paired by patient or by month
— Use online calculators (e.g., GraphPad, BMJ statistics) to simulate data and see how pairing affects p-values
— Practice with publicly available datasets (NHANES, MIMIC) to internalize the difference
Board pearl: QI projects that report "average wait time before vs after a workflow change" are paired by time period or unit, not by patient — analyze as paired t-test or interrupted time series, not unpaired.
— Choosing the wrong test can produce false-positive trial results, leading to adoption of ineffective or harmful therapies (e.g., historical examples of overstated benefit from underpowered pre/post studies without controls)
— Conversely, false-negative results can deprive patients of effective interventions
— Institutional review boards (IRBs) and FDA review require pre-specified statistical analysis plans to mitigate p-hacking
— Patients should understand whether they will receive both interventions (crossover, paired) or just one (parallel, unpaired)
— Carryover risks in crossover designs (residual drug effects, persistent physiologic changes) must be disclosed
— Equipoise must exist between arms in parallel-group trials
— Selective reporting of significant secondary outcomes from unplanned t-tests is a form of research misconduct
— CONSORT guidelines mandate reporting all pre-specified analyses, not just favorable ones
— Reanalysis with different tests post hoc to achieve significance ("p-hacking") undermines integrity
— When discharging a patient on a new medication justified by a single trial, the physician should appraise whether the trial's statistical analysis was appropriate — not just rely on the abstract conclusion
— Hand-offs should communicate uncertainty, not false precision
— Conflicts of interest and funding sources must be transparent — industry-sponsored trials with creative test selection have a documented bias toward positive results
— A statistically significant result in a homogeneous trial population may not generalize to diverse real-world patients — particularly relevant for elderly, pregnant, racial/ethnic minorities
Step 3 management: Before adopting a new therapy from a single trial, verify the trial's primary endpoint, statistical test choice, and pre-specification — uncritical adoption based on a misanalyzed study is both an ethical and patient safety failure, central to evidence-based practice.
— "Same patients before and after" → paired t-test
— "Two independent groups" → unpaired (independent-samples) t-test
— "Matched case-control, continuous outcome" → paired t-test
— "Crossover trial" → paired t-test
— "Twins discordant for exposure" → paired t-test
— "Ordinal pain scale before/after" → Wilcoxon signed-rank
— "Skewed lab value, two cohorts" → Mann-Whitney U / Wilcoxon rank-sum
— "Three dose groups, continuous outcome" → one-way ANOVA
— "Same patients, 4 time points" → repeated-measures ANOVA or mixed model
— "Percentage achieving target, two groups" → chi-square or z-test for proportions
— "Paired diagnostic test agreement (binary)" → McNemar's test
— "Survival between two groups" → log-rank test
— "Correlation between two continuous vars" → Pearson (or Spearman if skewed)
— Paired df = n − 1 (pairs)
— Unpaired df = n₁ + n₂ − 2 (Student's)
— Critical t at α=0.05, large df ≈ 1.96 (same as z)
— 95% CI: estimate ± ~2 × SE
— Effect size benchmarks (Cohen's d): 0.2 small, 0.5 medium, 0.8 large
— Power increases with: larger n, larger effect size, lower variance, higher α
— Paired designs typically yield 2–4× more power than unpaired for same n when correlation is high
— t-distribution has heavier tails than normal at low df, converges to normal as df → ∞
— CLT rescues t-test at n ≥ 30 even when raw data are non-normal
— Equal n in two groups ≠ paired design
— Significant p ≠ clinically important effect
— Nonsignificant p ≠ proven equivalence
Board pearl: The single phrase "each patient acted as their own control" appears in roughly half of all Step 3 paired t-test stems — instant recognition saves time.
Step 3 management: Before answering any biostatistics stem, identify outcome type, group number, and linkage in that order — this triage isolates the correct test family within seconds and is the most efficient exam-day heuristic for this category.
Choose a paired t-test when the same subjects (or matched pairs) generate two linked continuous, approximately normal measurements, and an unpaired (independent-samples) t-test when two separate groups are compared — getting this distinction right preserves statistical power, prevents false conclusions, and protects patients from adopting incorrectly analyzed evidence.
Board pearl: "Each patient served as their own control" is the highest-yield paired-design trigger phrase on Step 3 biostatistics items — recognizing it instantly resolves the most common selection question on the exam and is worth memorizing verbatim.