Reading a statistical result — branching viva

Opening (candidate)

I would treat this as a structured reading of the effect size and its confidence interval before I read the p-value. First I would classify the data and check that the right test was used, then I would read the confidence interval against the null, then interpret the p-value in context, and finally judge clinical importance and applicability to the child. A p-value on its own never carries the weight of those judgements. [9] [4]

Branch A — Confidence interval before the p-value

Examiner: The abstract quotes p equals 0.04, but the confidence interval crosses zero. Which do you believe? [9]

Candidate: I believe the confidence interval. A 95 percent interval that crosses zero for a difference is compatible with both a benefit and a harm, so the result is not statistically significant at the 0.05 level, however striking its point estimate. The interval carries the magnitude and the precision the p-value omits, and a wide, null-crossing interval is the fingerprint of a small, underpowered study. I would report the result as inconclusive, not as positive. [9] [3]

Branch B — Errors and power

Examiner: Define the two errors and explain why this study might have missed a real effect. [3]

Candidate: A Type I error is a false positive, rejecting a true null, governed by alpha. A Type II error is a false negative, failing to reject a false null, governed by beta. Statistical power is one minus beta, the chance of detecting a true effect, and it rises with the sample size and the size of the effect. This small study likely had low power, so its non-significant result is evidence of insufficient evidence, not proof of no effect. [3]

Branch C — The hazard ratio

Examiner: The oncology abstract reports a hazard ratio of 0.6 with an interval of 0.45 to 0.80. Interpret it. [7]

Candidate: A hazard ratio from a Cox proportional hazards model represents the relative rate of the event between groups, so a hazard ratio of 0.6 indicates about a 40 percent lower rate of the event with the intervention. Because the 95 percent confidence interval excludes one, the result is statistically significant at the 0.05 level. A hazard ratio is a relative rate of events over time, not a ratio of median survival times, which is a cruder summary that carries its own considerable uncertainty. [7]

Branch D — Choosing a test

Examiner: How would you choose the right test for a new paediatric dataset? [3]

Candidate: I would classify the outcome as categorical or numerical, ask whether the comparison is paired or independent, and count the groups. For normally distributed numerical data I would use the paired t-test, independent t-test, or one-way analysis of variance depending on the design; for skewed or ordinal data I would use the rank-based non-parametric equivalents; for proportions I would use the chi-square test or Fisher's exact test. I would check normality and equal variance before I trusted a parametric test. [3] [9]

Branch E — Multiple comparisons

Examiner: The logistic regression ran twenty subgroup analyses. Your concern? [2]

Candidate: Each extra test adds another opportunity for a false positive, so with twenty tests the chance of at least one chance finding is far above the nominal alpha. A lone significant subgroup among twenty is more likely a data-dredged artifact than a real effect. I would apply the Bonferroni correction, which divides alpha by the number of tests, and treat any surviving subgroup as hypothesis-generating until an independent dataset confirms it. [2] [4]

Close

Confirm understanding with teach-back, leave a written summary of the effect size with its confidence interval and its honest uncertainty, name the plan to revisit the decision as larger evidence appears, and document the conclusion with its precision. [9] [4]