Strategies for Analyzing Summary Variables in the Presence of Partially Missing Longitudinal Data

First presented at JSM 2018, Vancouver. Code for creating that presentation, including all visuals, is in the JSM2018/ directory. Slides (HTML) are at bit.ly/jlt-jsm2018; a PDF version is here.

Scripts

Data Preparation/Creation

• seed_prep.R: Deidentify/subset data from motivating clinical examples
• simulate_data.R: Create datasets used for simulations. Datasets saved in analysisdata/ (not on GitHub).
• simulate_outcome.R: Given datasets created in simulate_data.R, simulate the outcome of interest given the actual exposure values and specified relationships. Datasets saved in analysisdata/ (not on GitHub).

Simulation Functions

• introduce_missing.R: Functions which introduce missingness to the mental status variable (status) in a dataset, given a type of missingness (MCAR vs MAR/MNAR) and, if MAR/MNAR, a specified relationship between missingness and daily severity of illness.
• summary_strategies.R: Functions which summarize the exposure in various ways, preparing summarized datasets for model fitting. Strategies:
  1. "Ignore": Ignore missing records; total exposure = sum(all observed exposure)
  2. "Worst": Assume all missing records have the exposure; total exposure = sum(all observed + all missing exposure)
  3. "Delete": Any subject with any missing records gets a value of NA, which will be accounted for using multiple imputation at the time of modeling.
  4. "Impute": Impute daily exposure status using daily covariate, then summarize exposure for each imputation. Use those imputed summary datasets in model-based imputation.
• fit_models.R: Two functions to fit models which 1) do not and 2) require mice::mids() objects.
• miss_sum_fit.R: For a given simulated dataset, introduce all kinds of missingness; summarize the exposure in four ways; fit and extract needed info from models
• simulation.R: Run entire simulation on 1000 datasets, in batches of 100

realworld.R applies these strategies to two deidentified actual datasets, saving the results to be plotted.

Sketch of workflow for a single dataset is here.