Comparative Analysis of High Dimensional Propensity Scoring and Dimensionality Reduction in Cohort Studies

Background: Propensity score (PS) models are commonly used to balance non-equivalent exposure groups in observational studies. High-dimensional propensity scores (hdPS) can incorporate information from a large set of covariates into the PS estimation model, but they can be limited by issues arising from multicollinearity and overfitting.

Objectives: To discuss the application and compare the balancing properties and performance of a set of PS estimation and dimensionality reduction (DR) techniques: a priori specified PS, hdPS, principal components analysis (PCA), logistic PCA (LPCA), partial least squares (PLS) regression and autoencoder, in a case study evaluating the association between dialysis and mortality in advanced chronic kidney disease patients.

Methods: We used Optum CDM data with 400 empirical covariates to define a study cohort by identifying individuals aged ≥70 years with a diagnosis of heart failure and ≥1 eGFR value ≤20 recorded 1 year prior to index. A prevalent new-user design was used to compare patients who had undergone chronic dialysis and those that had not yet initiated chronic dialysis. They were matched in a 1:5 ratio on PS estimated using each of the following approaches: a priori specified PS (age, gender, race, eGFR and index date), hdPS, PS based on all empirical covariates, and 4 DR methods (PCA, LPCA, PLS and autoencoder). Balancing properties of each PS estimation method were evaluated using standardised mean differences (SMD) and c-statistics.

Results: We had 485 patients initiating chronic dialysis matched to 2443 patients who did not, with 400 empirical variables. For each method, the following number of imbalanced covariates (SMD >0.1) were observed: a priori specified PS: 83 covariates; all 400 empirical variables: 42 covariates; hdPS: 37 covariates; PLS: 28 covariates; LPCA: 25 covariates; PCA: 20 covariates; autoencoder: 15 covariates. The number of imbalanced covariates with SMD>0.2 were similar across methods. The C-statistic for the various methods were as follows, a priori specified PS: 0.787, PS based on all empirical covariates: 0.944, hdPS: 0.888, with the best implementations of PLS, PCA, LPCA and autoencoder: 0.929, 0.907, 0.882, 0.911respectively.

Conclusions: This example showed the clear benefit of going beyond a priori specified PS and that it may be better to use dimension reduction methods than to use all covariates or even hdPS. PS estimation using DR techniques is feasible and can be useful especially in smaller cohorts with relatively few exposures. Further work related to use of gradient boosting, among others, is needed to avoid any covariate remaining imbalanced as well as impact on outcome analysis and results.