Biostatistical Modelling with R Training Course

Participants can prepare analysis-ready datasets and identify distributional, quality and design issues.

• tidyverse pipelines for filtering, recoding, joining and reshaping health records
• Derive exposure, follow-up, age-band and outcome variables transparently
• Duplicate records, impossible values and inconsistent identifiers investigated systematically
• Frequency tables stratified by site, sex, age and study arm
• Distribution plots selected for continuous, binary, count and categorical variables
• Cross-tabulations reveal sparse cells, separation risks and coding anomalies

Participants can translate a health research question into a defensible modelling strategy.

• Estimands defined before selecting regression families or adjustment variables
• Outcome type, study design and data hierarchy guide model selection
• Confounders, mediators and colliders distinguished using causal reasoning
• Directed acyclic graphs identify minimum sufficient adjustment sets
• Exposure windows, person-time and censoring rules documented explicitly
• Pre-specify primary models, sensitivity analyses and subgroup analyses

Participants can fit, diagnose and interpret linear regression models for continuous health outcomes.

• Ordinary least squares models for blood pressure and quality-of-life scores
• Regression coefficients interpreted as adjusted mean differences with confidence intervals
• Residual-versus-fitted plots assess linearity and constant variance assumptions
• Q-Q plots assess residual normality and influential distributional departures
• Variance inflation factors identify problematic multicollinearity among predictors
• Transformations and spline terms address nonlinear predictor relationships

Participants can build and interpret logistic regression models for binary clinical and public-health outcomes.

• Binary logistic models for vaccination status, mortality and treatment completion
• Odds ratios converted into adjusted probabilities and risk differences
• Likelihood-ratio tests compare nested logistic regression specifications
• Hosmer-Lemeshow limitations and calibration plots assess model fit
• Receiver operating characteristic curves evaluate discriminative performance
• Complete separation detected and managed with penalised logistic regression

Participants can model rates and counts while accounting for exposure time and overdispersion.

• Poisson regression models incidence counts using log person-time offsets
• Incidence rate ratios interpreted for surveillance and service-utilisation outcomes
• Overdispersion assessed using residual deviance and Pearson statistics
• Negative binomial models address extra-Poisson variation in health counts
• Zero-inflated models considered for excess structural zero observations
• Predicted rates compared across exposure groups and covariate profiles

Participants can evaluate whether fitted models are credible, stable and fit for reporting.

• Residual diagnostics selected according to linear and generalised model families
• Cook's distance identifies observations with disproportionate fitted-model influence
• DFBETAs reveal coefficient sensitivity to individual study participants
• Calibration slopes and intercepts assess prediction model agreement
• Bootstrap resampling estimates optimism in apparent model performance
• Diagnostic findings translated into documented model revision decisions

Participants can communicate model outputs as meaningful public-health effect estimates and predictions.

• Marginal effects estimate adjusted outcome differences between policy-relevant groups
• Estimated marginal means compare covariate-adjusted group outcomes clearly
• Confidence intervals distinguished from prediction intervals and credible intervals
• Forest plots display effect estimates across subgroups and outcomes
• Coefficient tables formatted with denominators, reference groups and units
• Plain-language findings separate association, prediction and causal claims

Participants can fit mixed-effects models for data clustered within facilities, districts or individuals.

• Random intercept models account for facility, household and district clustering
• Intraclass correlation coefficients quantify within-cluster outcome similarity
• Random slope models represent varying intervention effects across clusters
• Fixed and random effects distinguished for health-system interpretation
• Likelihood-ratio tests compare nested mixed-effects model structures
• Singular fits and convergence warnings investigated before reporting

Participants can analyse repeated outcome measurements using mixed models and generalised estimating equations.

• Repeated measures data reshaped into long format with measurement timing
• Linear mixed models estimate individual trajectories over follow-up periods
• Time-by-treatment interactions test differential intervention response patterns
• Generalised estimating equations estimate population-average effects under clustering
• Working correlation structures compared using QIC and clinical plausibility
• Subject-specific and population-average coefficients interpreted distinctly

Participants can describe and compare time-to-event patterns under censoring.

• Time origins, event definitions and censoring rules specified precisely
• Kaplan-Meier curves estimate survival probabilities over observed follow-up
• Log-rank tests compare unadjusted survival experience between groups
• Number-at-risk tables accompany survival plots for transparent interpretation
• Median survival estimated only when supported by observed events
• Censoring mechanisms assessed for potential informative follow-up loss

Participants can fit and assess Cox proportional hazards models and competing-risk analyses.

• Cox proportional hazards models estimate adjusted hazard ratios
• Schoenfeld residuals test proportional hazards assumptions by covariate
• Time-varying coefficients address non-proportional exposure effects appropriately
• Restricted cubic splines model nonlinear continuous risk factors
• Competing risks distinguish death from discharge or treatment failure
• Cumulative incidence functions report event probabilities with competing events

Participants can assess missingness mechanisms and implement multiple imputation for regression analysis.

• Missing-data patterns summarised by variable, timepoint and participant characteristics
• MCAR, MAR and MNAR assumptions distinguished for analytical decisions
• Complete-case analysis limitations assessed against plausible missingness mechanisms
• Multiple imputation by chained equations implemented using mice
• Imputation models include outcomes, predictors and auxiliary variables
• Rubin's rules pool estimates and standard errors across imputations

Participants can specify causal analyses that distinguish confounding control from prediction modelling.

• Target trial principles clarify eligibility, treatment and follow-up definitions
• Directed acyclic graphs identify confounders without adjusting for colliders
• Propensity scores estimated using logistic regression and flexible specifications
• Matching, weighting and stratification balance measured baseline covariates
• Standardised mean differences assess covariate balance after adjustment
• Sensitivity analyses state unmeasured confounding assumptions explicitly

Participants can fit and interpret basic Bayesian regression models using prior information appropriately.

• Bayesian inference combines prior distributions with observed health data
• Weakly informative priors regularise coefficients in sparse datasets
• Posterior distributions summarised using means, intervals and probabilities
• Posterior predictive checks assess whether models reproduce observed patterns
• Markov chain convergence assessed with trace plots and R-hat
• Bayesian and frequentist interval interpretations compared carefully

Participants can develop and internally validate prediction models for individual-level health risk.

• Prediction questions distinguished from causal effect estimation questions
• Candidate predictors selected using availability, timing and clinical relevance
• Penalised regression reduces overfitting using lasso and ridge penalties
• Cross-validation estimates out-of-sample discrimination and calibration performance
• Area under the curve interpreted with prevalence and decision context
• Risk thresholds evaluated using sensitivity, specificity and predictive values

Participants can analyse survey data using weights, strata and primary sampling units.

• Survey designs documented through weights, strata and sampling units
• Design-based estimates account for unequal selection probabilities correctly
• Weighted regression models estimate population associations and standard errors
• Taylor linearisation estimates variance under complex sample designs
• Domain analysis preserves design information for population subgroups
• Unweighted and weighted estimates compared for interpretation differences

Participants can assess spatial patterning and introduce area-level random effects where appropriate.

• Geographic identifiers checked for boundaries, aggregation and confidentiality constraints
• Choropleth maps display rates using defensible classification methods
• Standardised mortality ratios compare observed and expected area counts
• Empirical Bayes smoothing stabilises estimates for small populations
• Spatial autocorrelation assessed using Moran's I statistic
• Area-level random effects distinguish signal from sampling variation

Participants can produce a review-ready statistical report with transparent code, tables and limitations.

• Quarto reports integrate narrative, R code, tables and graphics
• STROBE reporting items applied to observational analysis write-ups
• Methods sections document estimands, models, diagnostics and missing-data handling
• Results tables report effect estimates, intervals and analytical denominators
• Figures follow accessible labelling, colour and annotation principles
• Version-controlled scripts support independent review and analytical updates

Participants can build, review and present a complete reproducible R analysis answering a defined public-health question.

• Frame a decision question, estimand and analysis population explicitly
• Prepare documented analysis data from raw public-health source files
• Select regression, survival or longitudinal methods matching study structure
• Fit primary models and conduct planned diagnostic and sensitivity analyses
• Review assumptions, limitations and interpretation with peer technical feedback
• Present a Quarto report containing code, results and recommendations