Course 2 · Week 4 — Measurement, change, survival, reporting

Cheatsheet — biostats_courses

Author

R. Heller

Don’t dichotomise

Cutting a continuous variable at the median loses ≈ 37% of the information for a linear association. Keep predictors continuous; model non-linearity with a spline if needed.

Change scores vs ANCOVA

Change scores suffer from regression to the mean.
ANCOVA (adjust for baseline) is the efficient analysis in an RCT.
In observational studies with unequal baselines, both approaches answer subtly different questions — state which.

Agreement & reliability

Statistic	Use	R
Cohen’s κ	two raters, categorical	`irr::kappa2(cbind(r1, r2))`
Weighted κ	ordinal categorical	`irr::kappa2(..., weight = "squared")`
ICC(2,1) / ICC(3,1)	continuous	`psych::ICC(df)`
Bland-Altman	two methods, continuous	plot \(\bar{xy}\) vs \(y - x\)

mean_diff <- mean(y - x)
loa <- mean_diff + c(-1.96, 1.96) * sd(y - x)
ggplot(df, aes((x + y)/2, y - x)) + geom_point() +
  geom_hline(yintercept = c(mean_diff, loa), linetype = 2)

Survival primer

Concept	R
Kaplan-Meier	`survfit(Surv(time, event) ~ g)` + `ggsurvfit::ggsurvfit`
Log-rank test	`survdiff(Surv(time, event) ~ g)`
Cox PH model	`coxph(Surv(time, event) ~ x1 + x2)`
PH check	`cox.zph(fit)`

Report a hazard ratio with 95% CI.
PH violated? → time-varying coefficient or stratify.
Interpret HR only after checking for non-informative censoring.

Decision curves, NRI, IDI

Decision curve: net benefit vs threshold probability — dominates “treat all” and “treat none” when useful.
NRI: how many events move up / non-events move down in risk.
IDI: mean change in predicted probability by class.
Report decision curve first; NRI/IDI as secondary.

Explanation vs prediction (Shmueli)

Explanatory	Predictive
Goal: inference about β	Goal: minimise out-of-sample loss
Tools: ANOVA, diagnostics, intervals	Tools: CV, regularisation, ensembles
Metric: interval coverage, bias	Metric: RMSE, AUC, Brier on hold-out

Reporting guidelines

Design	Guideline
Randomised trial	CONSORT
Observational	STROBE
Diagnostic-accuracy	STARD
Prediction model (incl. AI)	TRIPOD / TRIPOD-AI
Systematic review	PRISMA
Animal research	ARRIVE

Decision rule for Week 4

Continuous predictor + continuous outcome → no median splits.
Two-group with repeated measurement → ANCOVA.
New rater / device → Bland-Altman + ICC.
Time-to-event → KM + Cox, always check PH.
Prediction model → TRIPOD checklist at submission.

Common pitfalls

Reporting κ on near-constant outcomes (high % agreement, low κ).
Quoting median survival that is never reached.
Forgetting to mention PH assumption checking.
Claiming a “prediction model” built on the same data used for evaluation.