Biostatistics, Research Methodology and Critical Appraisal
Study design
Randomisation with allocation concealment removes selection bias; blinding removes measurement bias; intention-to-treat analysis preserves randomisation and is mandatory for non-inferiority trials. Per-protocol analysis answers a different question and can be biased. Phase I establishes the maximum tolerated dose and dose-limiting toxicity, phase II screens for activity, phase III compares against a standard.
Measures of effect
Relative risk for cohort studies, odds ratio for case-control studies, hazard ratio for time-to-event data, absolute risk reduction and number needed to treat (1 divided by ARR). Confidence intervals convey precision better than p-values alone.
Hypothesis testing
Type I error (alpha, false positive) and type II error (beta, false negative); power equals 1 minus beta. A p-value is the probability of data at least as extreme as observed assuming the null hypothesis is true; it is not the probability that the null is true. Multiplicity inflates false positives, hence pre-specified primary endpoints and statistical correction.
Survival analysis
Kaplan-Meier estimates the survival function with censoring; the log-rank test compares curves; Cox proportional hazards yields adjusted hazard ratios. Median survival is the time at which half the cohort has died. Beware crossing curves, which violate proportional hazards.
Bias terms
Lead-time bias, length-time bias, selection bias, recall bias, publication bias and attrition bias; confounding; overdiagnosis.
Critical appraisal
Check whether the question was pre-specified, randomisation was adequate, follow-up was complete, all outcomes were reported, and the effect size is clinically meaningful.
Rule of thumb
A p-value of 0.04 in a subgroup of 40 patients after 15 comparisons is noise until replicated.