Study designs and screening

Study designs

• Phases of drug studies: I safety and kinetics; II efficacy and dose; III large RCT before licensing; IV post-marketing surveillance.
• Superiority versus non-inferiority: a non-inferiority trial aims only to show the new treatment is not worse than the standard by more than a pre-set margin, because it has another advantage (for example a heat-stable uterotonic for hot regions without a cold chain). Failing to show superiority is not the same as showing non-inferiority.
• Intention-to-treat analyses everyone as randomised, preserving randomisation.
• Allocation concealment prevents selection bias at recruitment; blinding prevents observer and placebo effects after allocation.
• Survival analysis: when people are followed for different lengths of time, Kaplan–Meier curves show survival over time, and Cox proportional hazards regression gives a hazard ratio (for example 0.70, 95% CI 0.55–0.89: a 30% lower instantaneous risk of recurrence).

Bias and confounding

• Recall (memory) bias: mothers of malformed babies remember medicines more thoroughly than mothers of healthy babies; use records made before the outcome. Typical of case–control studies.
• Selection bias: convenience samples, hospital controls (Berkson bias), healthy screening attenders.
• Publication bias: positive studies are more likely to be published. A funnel plot (effect size against study size) is asymmetrical when small negative studies are missing.
• Hawthorne effect: people change behaviour because they are studied.
• Screening biases: lead-time (diagnosed earlier, death unchanged), length-time (slow tumours detected more), overdiagnosis.
• Confounder: linked to both exposure and outcome but not on the causal pathway (birth order and Down syndrome, confounded by maternal age). Control by randomisation, restriction, matching, stratification or adjustment.

The two-by-two table

Every test measure comes from these four cells:

• Sensitivity + false-negative rate = 100%; specificity + false-positive rate = 100%.
• SnNOut: a sensitive test, when negative, rules out. SpPIn: a specific test, when positive, rules in. Screen with sensitive tests, confirm with specific ones.
• Prevalence moves predictive values: as a disease becomes more common, PPV rises and NPV falls; sensitivity and specificity do not change. Predictive values cannot be carried unchanged to another population.
• Likelihood ratios: LR+ = sensitivity ÷ (1 − specificity); LR− = (1 − sensitivity) ÷ specificity. Independent of prevalence; pre-test odds × LR = post-test odds.
• Cut-off: lowering the positive threshold raises sensitivity and lowers specificity. The ROC curve plots sensitivity against 1 − specificity; area 0.5 is chance, 1.0 perfect.
• A test can show high accuracy simply because most people tested are healthy.

Worked examples

Measures of frequency

• Incidence = new cases ÷ population at risk over a period. Exclude people who already have the disease: if 96 of 1,200 women with a normal booking test develop gestational diabetes, cumulative incidence is 8%.
• Prevalence = all existing (old + new) cases ÷ population. Prevalence ≈ incidence × duration: incidence 10 per 1,000 per year lasting 5 years gives a prevalence of about 50 per 1,000. A drug that prolongs life without cure raises prevalence.

Measures of risk and effect

• An RR or OR of 1 means no association. The OR approximates the RR when the disease is rare.
• A large RRR can hide a tiny absolute benefit when the outcome is rare; advertisements quote RRR because it sounds bigger.
• AR per cent applies to the exposed group; population attributable risk also depends on how common the exposure is.

Screening

• Main objective: detect disease at an early, treatable stage to reduce death or morbidity (secondary prevention).
• Wilson–Jungner criteria: important problem; recognisable latent or early stage; natural history understood; acceptable, valid test; effective treatment for early disease; cost-effective. Cervical cytology fits; CA-125 for ovarian cancer does not.
• Breast screening: mammography for women from about 50 years; under 35, ultrasound is the first test for a lump because dense breasts hide lesions.
• Mass screening tests everyone; selective targets high-risk groups (higher yield); opportunistic happens when patients attend. In antenatal serum screening, oestriol is part of the triple test.

Chance

• p value: probability of a result at least this extreme if the null hypothesis is true; below 0.05 is called significant. It does not measure effect size.
• Type I error (alpha): false positive. Type II error (beta): missing a real effect; power = 1 − beta, raised by larger samples.
• 95% confidence interval: the likely range of the true value. For ratios an interval including 1 is not significant; for differences, one including 0.
• Standard deviation describes spread of values; standard error the precision of a mean. Chi-square tests association between categories.

Remember Prevalence moves the predictive values; it never moves sensitivity or specificity.

Exam trap The odds ratio comes from case–control studies, the relative risk from cohort studies; do not swap them.

Exam trap For LR+, divide sensitivity by the false-positive rate, not by specificity.

How FCPS asks about this

• Mothers of malformed babies recall drugs better: recall bias.
• PPV from a table: true positives ÷ all positives.
• Disease becomes more common: PPV rises, NPV falls.
• Prevalence from incidence: incidence × average duration.
• Funnel plot: detects publication bias.
• Cox regression result: hazard ratio.
• New drug shown "not worse by more than a margin": non-inferiority trial.