Report an analysis

What a manuscript's methods section (or supplement) should state about a longmi analysis. Everything below is recorded in the result objects — no reconstruction from memory.

Checklist

  • [ ] Number and pattern of missing outcomes (monotone vs intermittent; retention by wave) — from LongitudinalData / your mask summary.
  • [ ] Assumed missingness mechanism, stated as an assumption (result.validity_report()).
  • [ ] Imputation model: backend, variables, wave terms, interactions, auxiliaries, dependence structure — completed.metadata["model_specification"].
  • [ ] Number of imputations M and how parameters were drawn (exact conjugate vs large-sample approximation — the backend's declaration notes say which).
  • [ ] Diagnostics: optimizer/curvature outcome (NB), chain autocorrelation (Gaussian), and the per-parameter fmi.
  • [ ] Analysis model and covariance estimator (e.g. Poisson GEE, exchangeable working correlation, robust sandwich).
  • [ ] Pooling rule and degrees-of-freedom method (Rubin's rules; Barnard–Rubin, mice-compatible).
  • [ ] Delta values examined for MNAR sensitivity and their substantive justification.
  • [ ] Software version and random seed — completed.metadata["random_state"].

Paragraph template

Missing longitudinal outcomes were multiply imputed using longmi version {X} (seed {S}). The imputation model was a {backend: e.g. negative-binomial generalized linear mixed model with a participant random intercept}, including {variables}, categorical wave effects, and {exposure}-by-wave interactions, so the substantive model's {exposure}-by-time term was represented in the imputation model. We generated M = {M} completed datasets under the missing-at-random assumption; parameter uncertainty was propagated by {exact conjugate data augmentation / draws from the large-sample normal approximation to the posterior}. Each completed dataset was analyzed with {analysis: e.g. a population-averaged Poisson GEE with exchangeable working correlation and robust standard errors, clustered by participant}, and estimates were combined using Rubin's rules with Barnard–Rubin degrees of freedom. The fraction of missing information for the primary contrast was {fmi}. Because missing-at-random is untestable, we repeated the analysis with delta-adjusted imputations shifting imputed {means} by {factors}; conclusions {were / were not} materially altered.

One honesty rule

Do not describe the analysis as having "verified" MAR or congeniality — the validity report deliberately tags these [declared]. State them as assumptions and point to the sensitivity analysis.