Epi Analysis

GAMs and GAMMs (mgcv)

• Use smooth terms (s()) for non-linear relationships; parametric terms for categorical variables
• bs = "re" for random intercepts; bs = "fs" for factor-smooth interactions
• bs = "tp" (default thin plate) for spatial/temporal smooths
• Use bam() with discrete = TRUE for large datasets; fREML for fast estimation
• Always check basis dimension: mgcv::k.check()
• Link functions: log link for multiplicative effects on positive outcomes; identity when additive

Adjusted vs unadjusted

• Always fit both unadjusted (single predictor) and adjusted (full) models
• Compare estimates to demonstrate direction and magnitude of confounding
• Do not interpret coefficients of adjustment variables as causal effects (Table 2 fallacy)

Model diagnostics

Run and report these diagnostics:

1. gratia::appraise() for GAM diagnostic plots
2. mgcv::k.check() for basis dimension adequacy
3. Residual vs fitted, QQ plot, scale-location plot
4. For random effects: gammit::extract_ranef() to inspect estimated effects and CIs
5. Check for influential observations if sample size permits

Report what diagnostics show, including imperfections. Do not hide diagnostic issues.

Forecast evaluation

• Use proper scoring rules: WIS (weighted interval score), CRPS, log score
• State which scoring rule and why
• scoringutils pipeline: as_forecast_quantile() then score()
• Log-transformed evaluation: transform BOTH forecasts and observations before scoring
• Pairwise comparisons vs pooled scoring: justify choice
• Report calibration (PIT histograms) and sharpness separately from overall score

Interpreting and reporting results

• Random effects: report as partial effects with 95% CI
• Compare adjusted vs unadjusted to show what confounding adjustment achieves
• Acknowledge residual variation explicitly
• Never claim "no effect" from non-significance; use "no clear evidence of systematic difference"
• Report effect sizes with confidence intervals, not just p-values
• For multiplicative models (log link): back-transform to ratio scale for interpretation

Code conventions

• Use tidyverse with native pipe (|>)
• Document functions with a purpose comment at top
• Separate data processing, model fitting, and output extraction into distinct scripts
• Analysis scripts should be source-able from Quarto/Rmd
• Use here() for all file paths
• Write CSV for intermediate data, RDS for R objects