A pharmacist is asked whether a direct oral anticoagulant causes more major bleeding than warfarin in patients on hemodialysis. No randomized trials address this question. Which available study design provides the highest level of evidence?

Prospective cohort study comparing rates of bleeding

Case series of dialysis patients with bleeding events

Case-control study of patients with major bleeding

Cross-sectional survey of anticoagulant prescribing

A P&T committee must choose between two conflicting guidelines on the same anticoagulation question when it writes a hospital protocol. Which feature most strongly indicates that a guideline is trustworthy?

Systematic review, graded evidence, and managed conflicts

Consensus of senior experts from a single specialty society

Publication in the journal with the higher impact factor

The more recent publication date of the two guidelines

While preparing a drug monograph, a pharmacist suspects publication bias because only positive trials of a new drug appear in the literature. Which resource is most useful for finding completed trials of the drug whose results were never published in a journal?

Cochrane Database of Systematic Reviews

In a meta-analysis of 15 trials of a supplement for knee osteoarthritis pain, the funnel plot shows small trials clustered on the side favoring the supplement and almost no small trials showing no effect or harm; the large trials cluster near no effect. Which is the most likely explanation?

Publication bias favoring small positive trials

Regression to the mean in baseline pain scores

Confounding by indication in the included trials

Random-effects weighting of the largest trials

A noninferiority trial compares a new oral anticoagulant with enoxaparin/warfarin for VTE treatment; 14% of patients discontinued or crossed over to the other treatment. Which analysis approach is most appropriate for the noninferiority conclusion?

Both ITT and per-protocol, with noninferiority in both

Per-protocol alone, because it reflects treatment received

As-treated alone, grouping patients by the drug they took

ITT alone, because it is the conservative approach

A formulary committee must compare two CGRP monoclonal antibodies for migraine prevention. No head-to-head trials exist, but each drug has several placebo-controlled RCTs in similar populations. Which method provides the most valid comparison of the two drugs?

Network meta-analysis using indirect comparison

Comparing each drug’s effect size across separate trials

Cumulative meta-analysis of each drug’s trials by year

Pooling all placebo-controlled trials in one meta-analysis

Investigators are planning several drug comparisons. For which study question is a crossover design most appropriate?

Two antihypertensives compared by 24-hour BP in stable hypertension

Two antivirals compared by sustained virologic response in hepatitis C

Two antibiotics compared by clinical cure of community-acquired pneumonia

Two antiplatelet drugs compared by 1-year mortality after acute MI

A guideline panel using GRADE rates the evidence from several well-conducted observational studies (no RCTs available) for an intervention as low certainty. Which additional finding would justify upgrading the certainty of this evidence?

A relative risk of 0.2, consistent across studies

A pooled p-value for the main effect below 0.001

Funding of every study by nonprofit organizations

A pooled sample size of more than 50,000 patients

In a meta-analysis of 18 trials of an antihypertensive drug, the size of the blood pressure reduction varies widely. The authors hypothesize that trials using higher mean doses produced larger reductions. Which analysis best tests this hypothesis?

Meta-regression of effect size on mean dose

Cumulative meta-analysis ordered by publication year

Funnel plot with Egger’s regression test

Sensitivity analysis removing the largest trial

In a 2-year randomized trial, major cardiovascular events occurred in 12.0% of patients on placebo and 8.8% of patients on a new drug. A journal-club handout lists the number needed to treat as 31 and describes it as the number of patients treated for 2 years to prevent one event. What NNT should be reported?

32, since an NNT is rounded up to the next whole patient

31, because 31.25 rounds down to the nearest whole number

4, the reciprocal of the 27% relative risk reduction

3, the reciprocal of the 0.32 difference in event rates

A 52-week randomized trial compares time to first moderate or severe COPD exacerbation between two inhaler regimens. Patients who withdraw early are censored at their last visit. Which statistical method is most appropriate for the primary comparison?

Kaplan-Meier curves compared with the log-rank test

Chi-square test of the proportion with an exacerbation

Unpaired t-test of mean days to first exacerbation

Mann-Whitney U test of days to first exacerbation

A health system rolls out a new inpatient VTE prophylaxis guideline in March. Hospital-acquired VTE rates had already been falling slowly for 2 years. Which evaluation design best estimates the guideline’s effect while accounting for this pre-existing trend?

Interrupted time series of monthly VTE incidence

Count of voluntary VTE incident reports after rollout

Survey of prescriber satisfaction with the guideline

Comparison of the month before and month after

A meta-analysis of 9 RCTs of a drug for major depression reports a pooled standardized mean difference of −0.30 (95% CI −0.45 to −0.15) with I² = 85%; individual trial estimates range from −0.90 to +0.10. Which is the most appropriate next step in interpreting these results?

Explore sources of heterogeneity before using the estimate

Discard the meta-analysis because I² shows poor study quality

Re-run a fixed-effect model to obtain a narrower CI

Accept the pooled estimate because its CI excludes zero

A systematic review of menopausal hormone therapy and coronary heart disease finds that large prospective cohort studies show a 35–50% lower risk, whereas large RCTs show no benefit (HR 1.0; 95% CI 0.9–1.1). Which conclusion is most appropriate?

Rely on the RCTs; the cohort result reflects bias

Favor the cohorts because they reflect real-world use

Pool all the studies together to maximize precision

Average the two designs’ estimates with equal weights

A meta-analysis of 12 trials shows substantial heterogeneity (I² = 70%). The authors re-run the pooled analysis with a random-effects model instead of a fixed-effect model. Which change is expected?

A wider CI around the pooled estimate

A narrower CI around the pooled estimate

More weight given to the largest single trial

A drop in the I² statistic to near zero

A meta-analysis of 10 RCTs finds that a drug reduces postoperative nausea. Three of the trials had inadequate allocation concealment and a high risk of bias, and reviewers want to know whether the overall conclusion depends on those trials. Which analysis is most appropriate?

Sensitivity analysis without high-risk trials

Calculation of the I² statistic for heterogeneity

Subgroup analysis by type of surgical procedure

Egger’s regression test for funnel plot asymmetry

BCPS Biostatistics Practice Questions and Review

1Read any trial in this order

The question
Population, intervention, comparator and outcome. If you cannot state them in one sentence, the options will mislead you.
The design
Randomized or observational, parallel or not, superiority or noninferiority. The design limits every conclusion that follows.
The primary endpoint
One prespecified outcome. Secondary endpoints and subgroups generate hypotheses, and a subgroup effect in a trial that missed its primary endpoint is the classic over-reading.
The analysis population
Who was counted, and in which group. For noninferiority this choice matters in a specific way, covered on noninferiority and intention-to-treat.
The effect and its interval
Size first, then precision. A p value without an interval is half a result; see p values and confidence intervals.
The absolute numbers
Convert relative effects into absolute ones before you call a result meaningful, using the method on NNT and absolute risk reduction.

2Study designs and the measure each one yields

Randomized controlled trial: Investigators assign the exposure at random, which balances known and unknown factors between groups. Reports relative risk, hazard ratios or mean differences.
Prospective cohort study: Follows exposed and unexposed people forward from now and counts new outcomes in each group, so it yields incidence and relative risk.
Retrospective cohort study: Same logic as a cohort, but assembled from records that already exist. Faster, and limited by what was recorded.
Case-control study: Starts from people with and without the outcome and looks back for exposure. Incidence cannot be measured, so the result is an odds ratio.
Nested case-control study: A case-control study drawn from inside a defined cohort, which keeps cases and controls from the same source population.
Cross-sectional study: Measures exposure and outcome at a single point in time. Good for prevalence, weak for cause and effect.

3Choosing the test for independent groups

The usual first choice when the groups are separate people, by type of outcome data.
Outcome data	Groups compared	Usual test
Continuous, roughly normal	Two	Unpaired (Student’s) t-test
Continuous, roughly normal	Three or more	Analysis of variance (ANOVA)
Ordinal, or continuous but skewed	Two	Mann-Whitney U (Wilcoxon rank-sum)
Ordinal, or continuous but skewed	Three or more	Kruskal-Wallis
Nominal, adequate expected counts	Two or more	Chi-square
Nominal, small expected counts	Two	Fisher exact test

Paired measurements and time-to-event outcomes with censoring need different tools; the case round includes both, so work them there before you look them up.

4Sensitivity, specificity and the values that move

Sensitivity is the share of people with the disease who test positive; specificity is the share of people without it who test negative. Both describe the test itself and stay put when you carry the test from one setting to another.

Predictive values describe the test in a population. Positive and negative predictive value shift with prevalence: the same assay gives a lower positive predictive value in a low-prevalence clinic than on a ward where most patients have the disease. Mixing up the two pairs is the classic trap, so build the two-by-two table every time rather than reasoning from memory.

Do the setup on scratch paper: the arithmetic is quick, and placing each count in the right cell is the skill. The practice set filtered to biostatistics mixes more diagnostic-test and design items with clinical cases.

5Pooled evidence and where answers live

Meta-analysis questions come in a cluster of their own: whether studies are similar enough to pool, whether small studies distort the picture, and how much a pooled estimate should be trusted. Each analytic tool answers one narrow question. Learn which question that is before you learn how the tool works, and most of these questions become recognition tasks.

Drug information resources are sorted into tiers. Primary literature is the original studies. Secondary resources index or abstract it, as PubMed does. Tertiary resources (textbooks, compendia, review databases) digest it into recommendations. A question about missing or contradictory evidence usually turns on which tier can see the gap.

Antimicrobial trials are a frequent home for noninferiority designs, so these reading skills come back on the infectious diseases guide. Reporting standards for randomized trials are set out in the CONSORT statement, updated in 2025, which is a useful checklist when a stem describes a trial.

6Appraisal drills

These cover design, test choice, diagnostic accuracy and pooled evidence. Set up each two-by-two table or design label on paper first, then commit to an answer.

Study cases written by our editors, not BPS items. The real BCPS exam is computer-based multiple choice, answered in order with no going back.

Case 1 of 19 · 0 answered

How case rounds work

Pick an option to see the verdict and a note on every option. Keys while the round has focus: 1–5 or A–E answer, ← → move between cases. Nothing is saved or sent anywhere.

Case 01
A pharmacist is asked whether a direct oral anticoagulant causes more major bleeding than warfarin in patients on hemodialysis. No randomized trials address this question. Which available study design provides the highest level of evidence?
Pick an option: the verdict and a note on every option open here.
Best answer: A.
Without RCTs, a prospective cohort that follows exposed and unexposed patients forward ranks highest because it establishes temporality and directly measures incidence in each group. A case-control study is tempting, but it is more prone to recall and selection bias and yields only odds ratios, so it ranks below a cohort. Cross-sectional surveys cannot link exposure to outcomes over time, and case series lack a comparison group.
A
Correct: without trials, a prospective cohort ranks highest because it follows both groups forward and measures bleeding incidence directly.
B
A case series has no comparison group, so it cannot show whether DOACs bleed more than warfarin.
C
Case-control designs suit rare outcomes but are more prone to recall and selection bias and yield only odds ratios, ranking below a cohort.
D
A cross-sectional survey describes prescribing at one moment and cannot link the drug to later bleeding.
Case 02
A P&T committee must choose between two conflicting guidelines on the same anticoagulation question when it writes a hospital protocol. Which feature most strongly indicates that a guideline is trustworthy?
Pick an option: the verdict and a note on every option open here.
Best answer: D.
National Academy of Medicine (IOM) standards for trustworthy guidelines require a systematic evidence review, explicit grading of evidence and recommendation strength, a multidisciplinary panel, and conflict-of-interest management. Expert consensus from one specialty is tempting, but without a systematic review and diverse input it carries a high risk of bias. Recency and journal prestige say nothing about methodological rigor.
A
Expert consensus from one society, without a systematic review or diverse input, carries a high risk of bias.
B
Journal impact factor reflects citation patterns, not the rigor with which a guideline was made.
C
A newer date may mean more current data, but recency alone says nothing about methodological quality.
D
Correct: National Academy of Medicine standards define a trustworthy guideline by systematic review, graded evidence, and managed conflicts of interest.
Case 03
While preparing a drug monograph, a pharmacist suspects publication bias because only positive trials of a new drug appear in the literature. Which resource is most useful for finding completed trials of the drug whose results were never published in a journal?
Pick an option: the verdict and a note on every option open here.
Best answer: D.
ClinicalTrials.gov registers trials before they start and requires results posting for many applicable trials, so it reveals completed studies that never reached a journal. PubMed indexes only published literature, so it reproduces the bias. Cochrane reviews and UpToDate are secondary and tertiary syntheses that depend on what has been published and on their own search dates.
A
Cochrane reviews synthesize trials but rely largely on what has been published and on their own search dates.
B
UpToDate is a tertiary summary built from published literature, so it repeats the bias.
C
PubMed indexes journal articles, so trials that were never published simply are not there.
D
Correct: trials are registered before they start and many must post results, so ClinicalTrials.gov exposes completed studies that never reached a journal.
Case 04
In a meta-analysis of 15 trials of a supplement for knee osteoarthritis pain, the funnel plot shows small trials clustered on the side favoring the supplement and almost no small trials showing no effect or harm; the large trials cluster near no effect. Which is the most likely explanation?
Pick an option: the verdict and a note on every option open here.
Best answer: A.
Funnel-plot asymmetry, with small positive trials present and small null trials missing, suggests publication bias (a small-study effect), and the large trials near no effect suggest the true benefit is small. Confounding by indication is a problem of observational studies, not randomized trials. Model weighting does not create asymmetry in where the trials fall, and regression to the mean affects both arms of a randomized trial.
A
Correct: small positive trials present and small null trials missing is the classic funnel asymmetry of publication bias; the large trials hint the true effect is small.
B
Regression to the mean affects both randomized arms equally, so it cannot skew which trials appear in the plot.
C
Confounding by indication is a problem of observational studies, not of randomized trials.
D
The pooling model changes study weights, not where each trial sits on the funnel plot.
Case 05
A noninferiority trial compares a new oral anticoagulant with enoxaparin/warfarin for VTE treatment; 14% of patients discontinued or crossed over to the other treatment. Which analysis approach is most appropriate for the noninferiority conclusion?
Pick an option: the verdict and a note on every option open here.
Best answer: B.
In noninferiority trials, ITT is anticonservative because nonadherence and crossover make the groups look more alike and favor a finding of noninferiority, so the conclusion should hold in both the ITT and per-protocol analyses. ‘ITT is conservative’ is the classic trap; it is true only for superiority trials. Per-protocol or as-treated analyses alone lose the protection of randomization.
A
Per-protocol alone loses the protection of randomization.
B
Correct. In noninferiority trials ITT can make groups look alike, so the conclusion should hold in both ITT and per-protocol.
C
As-treated grouping breaks randomization and invites bias.
D
ITT is conservative for superiority, not for noninferiority, where crossover pushes results toward a false finding of no difference.
Case 06
In a 12-week randomized trial of a new analgesic versus placebo, 30% of patients in the drug group and 8% in the placebo group withdrew because of adverse effects. The safety analysis included only patients who completed the study and found similar adverse-effect rates in the two groups. Which bias most threatens the conclusion that the drug is as well tolerated as placebo?
Pick an option: the verdict and a note on every option open here.
Best answer: C.
Unequal, outcome-related withdrawal removed the patients who did not tolerate the drug, so analyzing completers only makes the drug look as safe as placebo; this is attrition bias. Selection bias arises from how patients are assigned or enrolled, which randomization protects against. Detection bias comes from unequal outcome assessment, and performance bias from unequal care or exposure apart from the intervention.
A
Randomization protects against selection bias in how patients are assigned; the problem here arose after enrollment.
B
Detection bias comes from assessing outcomes unequally between groups, not from losing patients.
C
Correct: withdrawals for adverse effects removed the patients who did not tolerate the drug, so a completers-only analysis hides the harm.
D
Performance bias reflects unequal care apart from the intervention, which this scenario does not describe.
Case 07
A formulary committee must compare two CGRP monoclonal antibodies for migraine prevention. No head-to-head trials exist, but each drug has several placebo-controlled RCTs in similar populations. Which method provides the most valid comparison of the two drugs?
Pick an option: the verdict and a note on every option open here.
Best answer: A.
With no head-to-head trial, a network (adjusted indirect) meta-analysis compares the two drugs through their common comparator, placebo, so randomization is kept within each trial. It is valid only when the trials are similar enough in patients, doses and outcomes (transitivity). Setting each drug’s effect size side by side across separate trials is a naive indirect comparison: it breaks randomization and mixes in every difference between the trials. Pooling all placebo trials gives a class effect, and a cumulative meta-analysis only tracks one drug’s evidence over time.
A
Correct. A network meta-analysis links both antibodies through placebo, their shared comparator, and keeps randomization within each trial, if the trials are similar enough.
B
Setting effect sizes side by side across separate trials is a naive indirect comparison. It breaks randomization, and trial differences confound it.
C
A cumulative meta-analysis shows how one drug’s pooled effect changed as trials were added over time. It does not compare the two drugs.
D
Pooling every placebo-controlled trial yields a class effect against placebo. It cannot say whether one antibody works better than the other.
Case 08
In a crossover study, 30 patients receive drug A for 6 weeks and, after a washout, drug B for 6 weeks; LDL-C is measured at the end of each period. The differences are normally distributed. Which test is most appropriate to compare the effect of the two drugs?
Pick an option: the verdict and a note on every option open here.
Best answer: D.
Each patient serves as their own control in a crossover design, so the paired measurements of continuous, normally distributed data are compared with a paired t-test. The unpaired t-test is the classic error because it treats the two measurements as independent groups and loses power. The Wilcoxon rank-sum (Mann-Whitney) test is the nonparametric test for independent samples, and chi-square is for nominal data.
A
The rank-sum test is the nonparametric test for independent groups, and these data are paired and normally distributed.
B
Chi-square compares categorical counts, not a continuous measure such as LDL-C.
C
The unpaired t-test treats the two measurements as independent groups, ignoring the crossover pairing and losing power.
D
Correct: each patient serves as their own control, so normally distributed paired differences call for a paired t-test.
Case 09
A rapid Clostridioides difficile test is evaluated in 200 patients with diarrhea; 50 have infection by the reference standard. The rapid test is positive in 45 of the 50 infected patients and in 15 of the 150 uninfected patients. What is the positive predictive value of the rapid test in this population?
Pick an option: the verdict and a note on every option open here.
Best answer: C.
PPV is true positives divided by all positive tests: 45 ÷ (45 + 15) = 75%. Sensitivity (45 ÷ 50) and specificity (135 ÷ 150) are both 90%, the tempting confusion. The NPV is 135 ÷ 140 = 96.4%, and 30% (60 ÷ 200) is simply the proportion of patients who tested positive.
A
96.4% is the negative predictive value: 135 true negatives out of 140 negative results.
B
30% is just the share of all 200 patients who tested positive (60 ÷ 200).
C
Correct: PPV = true positives ÷ all positive results = 45 ÷ (45 + 15) = 75%.
D
90% is both the sensitivity (45 ÷ 50) and the specificity (135 ÷ 150), not a predictive value.
Case 10
Investigators are planning several drug comparisons. For which study question is a crossover design most appropriate?
Pick an option: the verdict and a note on every option open here.
Best answer: A.
A crossover design needs a stable, chronic condition and a reversible, short-term outcome, so that each patient can receive both treatments in sequence with a washout; blood pressure in stable hypertension fits. Curable conditions such as pneumonia or hepatitis C change after the first period, so the second treatment period is not comparable. Mortality is irreversible and cannot be measured twice in the same patient.
A
Correct: stable chronic hypertension and a reversible short-term outcome let each patient take both drugs in turn with a washout.
B
Hepatitis C can be cured in the first period, so the second period would not start from the same disease state.
C
Pneumonia resolves after the first treatment, leaving nothing comparable to treat in the second period.
D
Death is irreversible and cannot be measured twice in the same patient, so mortality rules out a crossover.
Case 11
A guideline panel using GRADE rates the evidence from several well-conducted observational studies (no RCTs available) for an intervention as low certainty. Which additional finding would justify upgrading the certainty of this evidence?
Pick an option: the verdict and a note on every option open here.
Best answer: D.
Under GRADE, observational evidence starts as low certainty and can be upgraded for a large magnitude of effect (such as RR below 0.5, or 0.2 for very large), a dose-response gradient, or plausible confounding that would reduce the observed effect. A large sample and a small p-value improve precision but do not overcome the confounding risk that keeps observational evidence low. Funding source is assessed under risk of bias and is not an upgrading criterion.
A
A very small p-value reflects precision, not freedom from the confounding that keeps observational evidence at low certainty.
B
Funding source is judged within risk of bias and is not one of GRADE’s upgrading criteria.
C
A huge pooled sample narrows the interval but cannot remove the confounding inherent in observational designs.
D
Correct: a large, consistent effect such as RR 0.2 is a GRADE reason to upgrade observational evidence, since confounding is unlikely to explain it all.
Case 12
In a meta-analysis of 18 trials of an antihypertensive drug, the size of the blood pressure reduction varies widely. The authors hypothesize that trials using higher mean doses produced larger reductions. Which analysis best tests this hypothesis?
Pick an option: the verdict and a note on every option open here.
Best answer: B.
Meta-regression models the trial-level effect size against a study-level covariate such as mean dose, so it can show whether dose explains part of the heterogeneity. Egger’s test also uses regression, which makes it tempting, but it regresses effect on precision to detect small-study or publication bias. Removing one trial tests robustness, and cumulative meta-analysis shows how the evidence changed over time.
A
A cumulative meta-analysis shows how the pooled estimate changed over time, not whether dose explains the spread.
B
Correct: meta-regression models trial-level effect size against a study-level covariate such as mean dose.
C
Egger’s test also uses regression, but it regresses effect on precision to detect small-study or publication bias.
D
Dropping the largest trial tests robustness to one study and cannot test a dose relationship.
Case 13
In a 2-year randomized trial, major cardiovascular events occurred in 12.0% of patients on placebo and 8.8% of patients on a new drug. A journal-club handout lists the number needed to treat as 31 and describes it as the number of patients treated for 2 years to prevent one event. What NNT should be reported?
Pick an option: the verdict and a note on every option open here.
Best answer: C.
The absolute risk reduction is 12.0% − 8.8% = 3.2%, and the NNT is 1 ÷ 0.032 = 31.25. By convention, an NNT is rounded up to the next whole number so that the benefit is not overstated. The Cochrane Handbook (chapter 15) states this rule. So 32 patients need to be treated for 2 years to prevent one event. Using the relative risk reduction (3.2 ÷ 12.0 ≈ 27%) or a misplaced decimal gives NNTs of 4 or 3, which hugely exaggerate the benefit.
A
Rounding down to 31 overstates the benefit. By convention, an NNT is always rounded up.
B
An NNT is the reciprocal of the absolute risk reduction (3.2%), not of the relative risk reduction (about 27%).
C
Correct. ARR = 12.0% − 8.8% = 3.2%, and 1 ÷ 0.032 = 31.25, which is rounded up to 32 over 2 years.
D
3.2% is 0.032, not 0.32. Misplacing the decimal turns an NNT of 32 into 3.
Case 14
A 52-week randomized trial compares time to first moderate or severe COPD exacerbation between two inhaler regimens. Patients who withdraw early are censored at their last visit. Which statistical method is most appropriate for the primary comparison?
Pick an option: the verdict and a note on every option open here.
Best answer: B.
Time-to-event data with censoring are analyzed with Kaplan-Meier estimates compared by the log-rank test, or a Cox model to give a hazard ratio. A chi-square test of proportions ignores when events happened and how long censored patients were followed. t-tests and Mann-Whitney tests cannot handle censored observations, because patients without an event have no true time to exacerbation.
A
Comparing proportions ignores when exacerbations happened and how long censored patients were followed.
B
Correct: time-to-event data with censoring are summarized by Kaplan-Meier curves and compared with the log-rank test (or a Cox model for a hazard ratio).
C
A t-test needs a true event time for every patient, and censored patients never had one.
D
Mann-Whitney is nonparametric but still cannot handle censored observations.
Case 15
A health system rolls out a new inpatient VTE prophylaxis guideline in March. Hospital-acquired VTE rates had already been falling slowly for 2 years. Which evaluation design best estimates the guideline’s effect while accounting for this pre-existing trend?
Pick an option: the verdict and a note on every option open here.
Best answer: A.
An interrupted time series analyzes many data points before and after the intervention, so it can separate a change in level or slope caused by the guideline from the downward trend already underway. A simple one-month before-and-after comparison would credit the secular decline to the guideline and is vulnerable to random month-to-month variation. Voluntary reports undercount events, and satisfaction surveys do not measure patient safety.
A
Correct: many monthly points before and after rollout let an interrupted time series separate a change in level or slope from the decline already underway.
B
Voluntary incident reports undercount VTE and cannot show a trend-adjusted effect.
C
Prescriber satisfaction is an experience measure and says nothing about VTE rates.
D
A one-month before-and-after comparison credits the existing downward trend to the guideline and is swayed by month-to-month noise.
Case 16
A meta-analysis of 9 RCTs of a drug for major depression reports a pooled standardized mean difference of −0.30 (95% CI −0.45 to −0.15) with I² = 85%; individual trial estimates range from −0.90 to +0.10. Which is the most appropriate next step in interpreting these results?
Pick an option: the verdict and a note on every option open here.
Best answer: B.
An I² of 85% means most of the variation between trials reflects real differences rather than chance, so the pooled value may not apply to any one population; subgroup or meta-regression analyses (dose, severity, design) should explore why. A CI that excludes zero does not make an average of very different effects meaningful. A fixed-effect model ignores the heterogeneity and gives a falsely precise interval, and I² measures inconsistency, not study quality.
A
I² measures inconsistency between trials, not study quality, so it is no reason to discard the analysis.
B
Correct: an I² of 85% with estimates from −0.90 to +0.10 means the average may fit no real population, so subgroups or meta-regression should seek the cause.
C
A fixed-effect model ignores the heterogeneity and produces a falsely precise interval.
D
A CI excluding zero does not make an average of widely different effects meaningful.
Case 17
A systematic review of menopausal hormone therapy and coronary heart disease finds that large prospective cohort studies show a 35–50% lower risk, whereas large RCTs show no benefit (HR 1.0; 95% CI 0.9–1.1). Which conclusion is most appropriate?
Pick an option: the verdict and a note on every option open here.
Best answer: B.
Randomization balances unmeasured factors, so when well-conducted RCTs conflict with observational studies, the RCTs take precedence; women who chose hormone therapy were healthier (healthy-user bias), which explains the apparent cohort benefit. Pooling or averaging across designs carries the bias into the estimate. ‘Real-world’ generalizability does not make up for confounding in the observational data.
A
Real-world generalizability cannot make up for healthy-user confounding in the cohort data.
B
Correct: randomization balances unmeasured factors, and the healthier women who chose hormone therapy explain the apparent cohort benefit.
C
Pooling biased cohorts with the RCTs carries the bias into a falsely precise estimate.
D
An equal-weight average of the two designs simply splits the difference with a biased result.
Case 18
A meta-analysis of 12 trials shows substantial heterogeneity (I² = 70%). The authors re-run the pooled analysis with a random-effects model instead of a fixed-effect model. Which change is expected?
Pick an option: the verdict and a note on every option open here.
Best answer: A.
A random-effects model adds a between-study variance term, so with heterogeneity present the pooled CI widens, and study weights become more equal (relatively less weight to the largest trial). It does not remove heterogeneity; I² describes the data and does not depend on the model chosen. A narrower interval is what a fixed-effect model gives, because it ignores between-study variance.
A
Correct: the random-effects model adds between-study variance, so with heterogeneity present the pooled interval widens.
B
A narrower interval is what a fixed-effect model gives, because it ignores between-study variance.
C
Random-effects weights become more equal, so the largest trial gets relatively less weight, not more.
D
I² describes the data and does not depend on the pooling model chosen.
Case 19
A meta-analysis of 10 RCTs finds that a drug reduces postoperative nausea. Three of the trials had inadequate allocation concealment and a high risk of bias, and reviewers want to know whether the overall conclusion depends on those trials. Which analysis is most appropriate?
Pick an option: the verdict and a note on every option open here.
Best answer: B.
A sensitivity analysis repeats the pooling without the high-risk-of-bias trials; if the result holds, the conclusion is robust to their inclusion. Subgroup analysis explores effect modification by patient or procedure factors, not study quality. Egger’s test looks for small-study or publication bias, and I² measures inconsistency, so neither answers whether the biased trials drive the result.
A
I² quantifies inconsistency among trials but does not show whether the biased trials drive the result.
B
Correct: repeating the pooled analysis without the three high-risk trials shows whether the conclusion holds without them.
C
Subgroup analysis by surgery type explores effect modification by clinical factors, not study quality.
D
Egger’s test looks for small-study or publication bias rather than the influence of poorly concealed trials.

7References

BPS, BCPS Examination Specifications (effective September 2024). Domain 3B and its four lines. As of October 27, 2026.
CONSORT statement (2025 update). Reporting standard for randomized trials, with a noninferiority extension.

8Concepts in this domain

Next step

One more round before you close the tab

Short sessions add up. Pick up a case set on your phone or right here.

Get the app Practice free on this site

Pharmacotherapy Exam Prep 2026 is an app for iPhone and iPad only; there is no Android version. On any other phone or a computer, use the free practice set on this site in your browser.