cmdpredict_bayesian

Bayesian prediction models for cardiometabolic multimorbidity (CMD) using data from the SCAPIS cohort. The project fits Bayesian logistic regression with a LASSO prior (implemented in Stan) and evaluates prediction performance via 10-fold cross-validation.

---

Project Overview

Outcomes modelled:

• cmd_any — Binary: any cardiometabolic disease
• cmd_multi — Ordinal/multinomial: number of cardiometabolic diseases

Analysis variants (reflected in file names):

Tag	Meaning
logit(any)	Binary logistic regression for any CMD
logit(multi)	Ordinal/multinomial logistic regression for CMD count
cov	Complete-case analysis (covariates only, no imputation)
mice	Multiple imputation (MICE)
lasso	Bayesian LASSO prior on regression coefficients
bart	Bayesian Additive Regression Trees

---

Analysis Pipeline (Steps 1–5)

The core workflow runs sequentially across 5 steps. Each step is a numbered R script. Run them in order.

Step	Script prefix	What it does
1	cmdpredict_1_*	Pre-processing: load SCAPIS data, define predictors, filter complete cases, create 10-fold CV indices, save processed data
2	cmdpredict_2_*	Model fitting: run Bayesian Stan model in 10-fold CV loop, extract fold-level predictions and posterior coefficient summaries
3	cmdpredict_3_*	Results compilation: AUC, Brier score, calibration slope/intercept, Youden's J threshold, ROC and calibration plots
4	cmdpredict_4_*	Full-sample fitting: refit the best model identified in Step #3 on the entire dataset, extract individual-level predicted probabilities with 95% credible intervals
5	cmdpredict_5_*	Subgroup analysis: stratify predictions by age, sex, smoking, hypertension, HDL, waist, and birthplace; compute group-level risk ratios

---

Folder Structure

cmdpredict_bayesian/
├── Main analysis/     # Steps 1–5 for all model variants (primary results)
├── Stepwise/          # Nested predictor models (m1–m8) for model building
├── Interaction/       # Interaction term models for 5 variable pairs
└── Stan/              # Stan model definition (.stan) and compiled object (.rds)

See the README in each subfolder for details.

---

File Naming Convention

cmdpredict_{step}_{outcome}_{imputation}_{method}_v{version}.R

Examples:

• cmdpredict_2_logit(any)_cov_lasso_v06.R — Step 2, binary outcome, complete-case, LASSO
• cmdpredict_3_logit(multi)_mice_lasso_v06.R — Step 3, ordinal outcome, imputed, LASSO

---

Key Predictors

Domain	Variables
Demographic	age1, sex1_2, born1_2, location1_2, site1_*
Socioeconomic	education2_2, expense1_2, fh_cmd_2
Lifestyle	smoking1_2, alcohol2_2, natrium, fibrer, pa_riskvpa, mental_stress2_3, sleep_dur1_2
Clinical	sbp_mean, dbp_mean, waist, hdl, nonhdl, drug_hypertension_2, drug_lipid_2

---

Dependencies

• R packages: haven, dplyr, caret, rstan, pROC, ggplot2, sjmisc, tidyr, tibble
• Stan: requires RStan and a C++ toolchain
• Input data: 05. Result/cmdpredict_python_v05.dta (relative to SCAPIS root)
• Intermediate outputs: saved to 08. Temp/03. Paper 3/
• Final outputs: 05. Result/, 04. Excel/03. Paper 3/, 07. Graph/03. Paper 3/