Statistical analyses
What each model estimates, and what it assumes to do it.
Logistic regression or Cox regression?
Distinguish an outcome at a fixed horizon from the timing of that outcome.
4 min read
Mixed effects or GEE?
Match the interpretation of a correlated-outcome model to the question.
4 min read
Propensity matching or weighting?
Choose the target population before choosing how to balance observed covariates.
4 min read
Interrupted time series and segmented regression
Estimate a prespecified change in an outcome trajectory after an intervention.
4 min read
Complex survey analysis
Estimate population quantities while respecting sampling weights, strata, clusters and subgroup membership.
4 min read
Prediction model performance and validation
Evaluate calibration, discrimination and usefulness with a validation strategy that protects against leakage and optimism.
4 min read
Linear regression
Estimates how a continuous outcome changes, on average, with one or more predictors.
3 min read
Logistic regression
Models a binary outcome and estimates how predictors relate to the odds of that outcome.
3 min read
Cox proportional hazards regression
Models time until an event, handling censoring and estimating adjusted hazard ratios.
4 min read
Poisson and negative binomial regression
Models counts or event rates, usually with an offset for differing observation time.
3 min read
Mixed-effects models
Combines population-level effects with random effects for people, clinics, or other clusters.
4 min read
Generalized estimating equations
Estimates population-average associations while accounting for within-cluster correlation.
3 min read
Propensity-score methods
Uses the estimated probability of treatment to improve balance between comparison groups.
4 min read
Multiple imputation
Fills each missing value several times so uncertainty from missingness reaches the result.
3 min read
Sensitivity analyses
Tests whether the conclusion survives plausible changes to the study's assumptions.
3 min read
Competing risks
When another event can prevent the outcome from ever happening, a Kaplan-Meier curve overstates how often the outcome occurs.
4 min read
Mediation analysis
Asks how much of an effect travels through a particular intermediate step, which needs stronger assumptions than the total effect did.
4 min read
Kaplan-Meier and the log-rank test
Describes and compares survival over time without a model, which is its strength and also its limit.
3 min read
Cluster-robust inference and few clusters
Corrects standard errors for grouping, but the ordinary correction is unreliable exactly when clusters are few.
4 min read
Marginal standardization and g-computation
Turns a fitted model into the absolute risks and risk differences a clinician can act on.
4 min read
t-tests and ANOVA
Compares average values between two or more groups, assuming roughly normal data and independent observations.
3 min read
Nonparametric and rank-based tests
Compares distributions using ranks rather than means, when the data are ordinal or badly skewed.
3 min read
Chi-square and Fisher's exact test
Tests whether two categorical variables are associated, in a contingency table.
3 min read
Correlation
Summarises how strongly two continuous variables move together, on a scale from -1 to 1.
3 min read
Baseline adjustment and ANCOVA
Analyses a follow-up outcome while adjusting for its own baseline value, which is usually better than analysing change scores.
3 min read
Repeated measures and longitudinal data
Handles several measurements on the same person, which are correlated and cannot be treated as independent.
4 min read
ROC curves and diagnostic thresholds
Describes how well a continuous measurement separates people with and without a condition, across every possible cutoff.
4 min read
Agreement and reliability
Measures whether two methods or two observers give the same answer, which is a different question from whether they correlate.
4 min read
Ordinal regression
Models an outcome with ordered categories without collapsing it to yes-or-no or pretending it is a number.
3 min read
Recurrent events
Analyses every occurrence of an event that can happen to the same patient more than once, not only the first.
4 min read
Descriptive statistics and Table 1
Describes who is in the study, overall and by group, so a reader can judge who the results apply to.
4 min read
Multinomial logistic regression
Models an outcome with three or more categories that have no natural order, without collapsing them to two.
3 min read
Quantile regression
Estimates how an exposure shifts the median, or another chosen quantile, of a continuous outcome rather than its mean.
3 min read
Nonlinearity and splines
Lets a continuous predictor bend, instead of forcing a straight line or chopping it into categories.
3 min read
Variable selection
Deciding which variables to include, which is a question of purpose before it is a question of algorithm.
4 min read
Penalized regression and shrinkage
Deliberately pulls coefficients toward zero to trade a little bias for a large gain in stability.
3 min read
The bootstrap
Estimates uncertainty by resampling the data itself, when a formula for the standard error is unavailable or untrustworthy.
3 min read
Interim analyses and stopping rules
Looking at accumulating trial data before the end, in a way that does not destroy the error rates.
4 min read
Noninferiority and equivalence
Asks whether a new treatment is not meaningfully worse, which needs a margin fixed before the study.
4 min read
Intention-to-treat and per-protocol
Analysing people by the group they were assigned to, rather than the treatment they actually received.
3 min read
Subgroup analyses
Asks whether the effect differs between kinds of patient, which is the question most likely to produce a false finding.
4 min read
Time-varying exposures and treatment
Handles treatment that starts, stops and changes over time, where standard adjustment can make bias worse rather than better.
4 min read
Matching
Builds comparison groups that look alike on measured variables before any outcome is examined.
3 min read
Beyond the hazard ratio
What to do about time-to-event data when proportional hazards does not hold, or when a hazard ratio is not what a patient needs to know.
4 min read
Bayesian analysis
Reports the probability that a treatment works, given the data and what was believed beforehand.
4 min read
Dose-response and trend
Asks whether more exposure means more effect, using the ordering rather than testing each level separately.
3 min read
Meta-analysis models and heterogeneity
Choose between a common-effect and a random-effects model for what the summary must mean, and read heterogeneity for what it is.
4 min read
Run charts and statistical process control
Plots a measure over time to tell ordinary variation from a signal that the process has changed.
4 min read
Network meta-analysis
Compares several treatments at once by combining direct and indirect evidence, which works only when the trials are similar enough to connect.
4 min read
Meta-analysis of diagnostic test accuracy
Combines sensitivity and specificity across studies of a test, modelling them together because they trade off against each other.
4 min read
Individual participant data meta-analysis
Synthesises the raw participant-level data from each study rather than published summaries, which opens analyses aggregate data cannot support.
4 min read