SGLT2-GLP1

Collection of functions and scripts to investigate how clinical features can be used for prediction.

Final structure for analysis: (Model 11.5)

1. Fit a BART propensity score model (function bartMachine::bartMachine) with all variable available.
2. Perform BART variable selection (function bartMachine::var_selection_by_permute) to choose variables for the PS model.
3. Re-fit a BART propensity score model (function bartMachine::bartMachine) with the selected variables.
4. Fit a SparseBCF model (function SparseBCF::SparseBCF) on complete data and perform variable selection:
   • Only include variable with ~ 20% missingness (discard higher amounts)
   • Select variables with an inclusion proportion above 1/n (n = number of variables used)
5. Fit a BCF model (function bcf::bcf) on complete data with the selected variables.
   • Fit two versions of the model:
     • With propensity scores included in the "control" or mu(x) (use include_pi = "control")
     • Without propensity scores included in the model (use include_pi = "none")
   • Compare individual predictions from both models in order to decide on the use of propensity scores.
6. Check model fit for the outcomes:
   • Plot standardised results to check for any structure in the residuals.
7. Check model fit for the treatment effects: (model fitted in observational data)
   • Plot predicted CATE vs ATE for several ntiles of predicted treatment effect:
     • Propensity score matching 1:1 (check whether matched individuals are well balanced)
     • Propensity score matching 1:1 whilst adjusting for all variables used in the BCF model.
     • Adjust for all variables used in the BCF model.

---

(Developed in CPRD: GOLD download)

Files:

• 0.0: Functions

  • .1: Functions for plotting and some calculations.
  • .2: Functions for bartMachine tree analysis. Slightly modified bartMan R package (not kept up to date) (package requirements were modified).

• 1.0: Detailed explanation of steps taken in the selection of patients for our cohorts.

• 2.0: Descriptive analysis of data

  • .1: Collection of plots demonstrating specific details/quirks of the dataset.
  • .2: Table description of Development and Validation data
  • .3: Treatment Effects/Variable description table/plot (Model 4.5)

• 3.0: R packages to model causal treatment effect.

  • .1: Fitting of a causal model using grf R package. This includes an evaluation of model fit.
  • .2: Fitting of a causal model using bcf R package. This includes an evaluation of model fit.

• 4.0: bartMachine models for treatment heterogeneity.

  • .1: Fitting a BART model with variable selection for propensity score and outcome model. This includes an evaluation of model fit.
  • .2: Fitting a BART model with routine variables in propensity score model and biomarkers in outcome model. This includes an evaluation of model fit.
  • .3: Fitting a BART model with variable selection for propensity score and variable selection using BART + grf for the outcome model. This includes an evaluation of model fit.
  • .4: Fitting a BART model with variable selection using BART + grf for the outcome model. This includes an evaluation of model fit.
  • .5: Fitting a BART model with variable selection for using BART + grf for the outcome model. This includes an evaluation of model fit. (Change from 4.4 - instead of 'score', we use 'score.excl.mi')
  • .6: Fitting a BART propensity score model, variable selection, matching individuals, BART model with all variables.
  • .7: Fitting a BART propensity score model, variable selection, refit propensity score model, BART HbA1c model + propensity score as covariate, variable selection, refit BART HbA1c model.

• 5.0: bartMachine models using no methodical procedure.

  • .1: Fitting a collection of naive Bart models for HbA1c outcome using routine clinical variables / all variables / propensity scores, alternating between them.

• 6.0: Comparing models.

  • .1: Collection of plots comparing naive BART models in 5.0.
  • .2: Collection of plots comparing bcf and bartMachine with the same variables. Head-to-head comparisons of treatment effect for 3.2. vs 5.1. model 1 (Complete/Routine)
  • .3: Collection of plots comparing 4.1-4.4 models.
  • .4: Differential treatment effect.

• 7.0: Sensitivity analysis.

  • .1: Exclusion of GLP1 patients before 2013.
  • .2: Variable importance model 4.4.
  • .3: Modelling predicted treatment effect against model variables.

• 8.0: Presentations/Slides

  • .1: MRC: 29th September London
  • .2: SGLT2 vs GLP1 paper for publish

• 9.0: Shiny App

  • .1: Model 4.4 - probability of achieving target HbA1c.

• 10.0: Modelling accompanying data

  • .1: Weight reduction
  • .2: Discontinuation

Files:

(Developed in CPRD: Aurum download)

• 11.00: Aurum download modelling
  • .01: Functions used specifically for this portion.
  • .02: Detailed explanation of the selection of cohorts.
  • .03: Descriptive analysis of datasets.
  • .04: Propensity score model.
  • .05: Model heterogeneity.
  • .06: Risks/Benefits: hba1c change, eight change, eGFR change, discontinuation, CVD/HF/CKD outcomes, microvascular complications.
  • .07: Differential treatment effects.
  • .08: Paper plots.
    • .1: Main plots of paper
    • .2: Supplementary plots of paper
    • .3: Plots for DUK.
  • .09: Comparison of SGLT2vsGLP1 BCF model to SGLT2vsDPP4 linear model (John Dennis).
  • .10: Validation of treatment effects splitting by ethnicity.
  • .11: Validation of the excluded individuals that were prescribed semaglutide.
  • .12: Validation of treatment effects in those insulin treated.
  • .13: Validation of treatment effects in those with/without baseline CVD.