Mastering Statistics and Critical Appraisal for MRCP Success

Every MRCP candidate eventually meets the question that begins: A new randomised trial reports a hazard ratio of 0.72 (95% CI 0.55–0.94)... For many, this is the moment a confident clinician freezes. Yet evidence-based medicine (EBM) questions are among the most predictable, formulaic marks in both Part 1 and Part 2 of the MRCP. As examiners, we watch candidates throw these marks away every single diet because they never systematically learned the core toolkit. This guide fixes that: study designs, statistical tests, diagnostic metrics, survival analysis, forest plots, screening bias — plus exam-style practice questions at the end.

Where EBM Appears in the MRCP

EBM is not a niche topic — it threads through both written papers:

• MRCP Part 1: basic and clinical science sections test study design, bias, statistical tests, error types, and screening principles.

• MRCP Part 2 (written): expect abstract-style stems, drug trial data, diagnostic test performance, forest plots, Kaplan–Meier curves, and hazard ratios.

Typically 3–8 questions per paper draw directly on EBM. For a candidate hovering near the pass mark, this topic alone can decide the diet.

1. Study Designs: The Hierarchy and the Discriminators

You must recognise a study type from a two-line vignette. Learn the classic trigger phrases:

Examiner tips:

• Case–control studies can only produce an odds ratio — if a stem gives incidence in exposed and unexposed groups, it is a cohort and you can calculate risk.

• Randomisation eliminates confounding by known and unknown variables; blinding reduces measurement (observer) bias.

• Intention-to-treat analysis (analyse patients in the group they were randomised to, regardless of adherence) preserves the benefits of randomisation and avoids attrition bias.

The Bias Checklist

• Selection bias – systematic difference in how subjects enter groups.

• Recall bias – classic in case–control studies; cases ruminate more about past exposures.

• Measurement/observer bias – fixed by blinding.

• Confounding – a third variable linked to both exposure and outcome; fixed by randomisation, stratification, or multivariable adjustment.

• Attrition bias – differential dropout; handled by intention-to-treat.

• Publication bias – negative trials go unpublished; detected with a funnel plot.

2. Choosing the Correct Statistical Test

MRCP loves matching a test to a scenario. First classify the data: continuous and normal (parametric) versus skewed/ordinal (non-parametric), or categorical.

Trigger words that demand a non-parametric test: median, skewed, ordinal scale, small sample. If the stem reports medians or a non-normal distribution, the parametric answer is the distractor.

3. Measures of Effect: RR, OR, ARR, RRR and NNT

Learn these formulas cold — they are pure free marks:

• Relative risk (RR) = risk in exposed ÷ risk in unexposed

• Odds ratio (OR) = (a×d) ÷ (b×c); approximates RR when the disease is rare

• Absolute risk reduction (ARR) = control event rate − treatment event rate

• Relative risk reduction (RRR) = ARR ÷ control event rate

• Number needed to treat (NNT) = 1 ÷ ARR (as a decimal)

Worked Example

A trial reports cardiovascular events in 20% of placebo patients and 14% of drug-treated patients.

• ARR = 20% − 14% = 6% (0.06)

• RRR = 6 ÷ 20 = 30%

• NNT = 1 ÷ 0.06 ≈ 17 patients to prevent one event

Examiner trap: RRR always looks more impressive than ARR. Drug reps quote RRR; sensible clinicians (and MRCP answers) demand the absolute figures.

4. Diagnostic Test Performance

Build the 2×2 table every time:

• Sensitivity = a ÷ (a + c) — of those with disease, how many test positive

• Specificity = d ÷ (b + d) — of those without disease, how many test negative

• PPV = a ÷ (a + b) — of positive tests, how many truly have disease

• NPV = d ÷ (c + d)

• LR+ = sensitivity ÷ (1 − specificity) — a high LR+ (e.g. 10) markedly raises disease probability

• LR− = (1 − sensitivity) ÷ specificity

• Post-test odds = pre-test odds × likelihood ratio

The two rules examiners test relentlessly:

1. Prevalence changes PPV and NPV — never sensitivity or specificity. In a low-prevalence population, even an excellent test has a poor PPV. Example: prevalence 2%, sensitivity 95%, specificity 90% → PPV = 0.019 ÷ (0.019 + 0.098) ≈ 16%. Most positives are false positives.

2. SnNOut / SpPIn: a highly Snsitive test, when Negative, rules Out; a highly Specific test, when Positive, rules In. A sensitive test (e.g. D-dimer) is chosen to exclude disease; a specific test confirms it.

An ROC curve plots sensitivity against 1−specificity; a larger area under the curve means better overall discrimination.

5. P-values, Confidence Intervals and Error Types

• A p-value is the probability of observing these data (or more extreme) if the null hypothesis were true. p < 0.05 is conventionally significant — it is not the probability the null is true, and not proof of clinical importance.

• 95% confidence interval: the range within which the true effect plausibly lies with 95% certainty. If the CI for a ratio (RR/OR/HR) crosses 1, or for a difference in means crosses 0, the result is not statistically significant — regardless of the p-value quoted.

• Larger samples → narrower confidence intervals → greater power.

Type I and Type II Errors

• Type I error (α): rejecting a true null hypothesis — a false positive. Conventionally set at 5%. Analogy: convicting an innocent person.

• Type II error (β): failing to reject a false null — a false negative.

• Power = 1 − β (usually targeted at 80–90%): the ability to detect a real difference. Underpowered trials produce inconclusive (not negative) results.

• Multiple testing: the more comparisons you run, the more likely a spurious significant result; corrections such as Bonferroni adjustment raise the evidence threshold.

6. Survival Analysis

• Kaplan–Meier curves estimate survival over time, handling unequal follow-up through censoring (patients exit — alive — at their last observation).

• The log-rank test compares two Kaplan–Meier curves.

• Cox proportional hazards regression produces the hazard ratio (HR): the relative rate of events at any point in time. HR 0.72 means a 28% relative reduction in the hazard of the event — provided the 95% CI excludes 1.

Classic stem: HR 0.72 (95% CI 0.55–0.94, p = 0.01). Correct interpretation: a statistically significant 28% relative risk reduction. If the CI had been 0.48–1.08, the identical point estimate would be non-significant.

7. Meta-analysis and the Forest Plot

Learn to read a forest plot in five seconds:

• Each square is a study; square size reflects weight (larger = more influential).

• Horizontal lines are each study's confidence interval.

• The vertical line at 1 is the line of no effect (for ratios).

• The diamond is the pooled result — significant if it does not touch the line of no effect.

• I² statistic quantifies heterogeneity: roughly < 25% low, 25–50% moderate, > 50% high. High heterogeneity usually favours a random-effects model and demands caution.

• An asymmetrical funnel plot suggests publication bias.

8. Screening: The Bias Trilogy

Screening questions recur in both papers. A screening programme is justified when the Wilson–Jungner criteria are met: serious condition, early detectable latent phase, safe acceptable test, effective early treatment, and benefits outweighing harm.

The three biases:

1. Lead-time bias — survival appears longer only because disease was detected earlier; death is not delayed.

2. Length-time bias — screening preferentially detects slow-growing, indolent disease.

3. Overdiagnosis — detecting disease that would never have caused symptoms or harm (the classic PSA screening debate).

9. Exam-Style Rapid Practice

Q1. A trial: stroke in 12% of controls vs 8% of treated patients. NNT?
A1. ARR = 4% (0.04); NNT = 25.

Q2. Researchers compare median pain scores across three analgesic groups (data skewed). Correct test?
A2. Kruskal–Wallis (non-parametric, > 2 groups).

Q3. A cancer drug trial reports HR 0.80 (95% CI 0.62–1.02). Interpret.
A3. A point estimate suggesting 20% reduction in hazard, but not statistically significant — the CI crosses 1.

Q4. A screening test with 99% sensitivity and 95% specificity is applied to a population with disease prevalence 0.1%. Most positive results will be false positives — why?
A4. In low-prevalence settings, PPV collapses because false positives (from the large healthy majority) vastly outnumber true positives.

Q5. A screened cancer cohort shows longer survival but unchanged age-adjusted mortality. This illustrates?
A5. Lead-time bias (with a strong hint of overdiagnosis).

10. One-Page Summary for Final-Week Revision

• Ratio CI crosses 1 (or difference CI crosses 0) → not significant

• Case–control → odds ratio only; cohort → risk and RR

• Median/skewed/ordinal → non-parametric test

• NNT = 1/ARR; RRR flatters, ARR informs

• Sens/spec are test properties; PPV/NPV depend on prevalence

• Power = 1 − β; Type I = false positive

• HR with CI excluding 1 → significant reduction in event rate

• Funnel plot → publication bias; I² → heterogeneity

• Screening: lead-time, length-time, overdiagnosis

A Preparation Strategy That Works

1. Make a formula flashcard deck (sens, spec, PPV, NPV, LRs, ARR, RRR, NNT) and drill it with spaced repetition until recall is automatic.

2. Read one abstract per day from a major journal — practise naming the design, the test, and the effect measure before reading the methods.

3. Do EBM-tagged question bank items in blocks, and keep an error log: most repeated mistakes are CI interpretation and test selection.

4. Teach the 2×2 table to a study partner — if you can derive PPV from scratch in thirty seconds, no stem can destabilise you.

Final Word from the Examiner's Chair

EBM questions are rare in medicine: they reward preparation with near-certainty. You cannot bluff your way through a likelihood ratio, but you can absolutely master it in a fortnight of focused work. Candidates consistently underestimate this topic — which means the marks are sitting there, unclaimed, waiting for the candidate who decided that hazard ratio would never intimidate them again. Be that candidate.