The original paper used logistic regression using backwards selection on a number of binary variables to attempt to predict the risk of childhood mortality. It created a simple binary risk score to aid in calculating a child's probability of death.
My work attempts to extend this analysis by using machine learning analyses to validate this approach and compare it to other analytical methods. In particular, I aim to produce some decision trees to aid in decision-making.
01_clean_data.do:
- Formats all variables the same way
- Decide what to do with “9” values
- Exclude variables that don’t matter
- Create variable lists: signs, symptoms, treatment, diagnoses, test results, outcomes
- Turn all variables into binary indicators
- May need Herbie’s guidance on indicators not included in the original analysis
02_prep_data.R
- Apply different methods of imputation or observation-dropping
- Output table of missing variables for all variables of interest
- Descriptive table of primary patient characteristics
- Any other descriptive stats
analysis_functions.R
- Specify one function for each method — assume same data structure, and take in arguments for formula etc.
- Logistic regression with backwards selection
- Define variable importance cutoffs for selection (or default)
- Decision trees
- Define ideal tree breakdown — pruning characteristics etc.
- Random forests
- Source up-sampling/down-sampling methods
- Define sampling parameters, number of cv runs, etc.
- Output predictions, graphical representations, ROC analyses.
- Save graphs, predictions, and ROC analyses to flat files (how to toggle by source?)
03_apply_analysis.R
- Source all analysis functions
- Separate data into train and test data
- Apply analysis functions and get/plot results