8  Step 5: Covariates

We select 2 types of covariates for the next step (to analyze using propensity score or other alternative approaches):

8.1 Ideal number of prioritised covariates

Based on calculated \(log-absolute-bias\), we select top k recurrence covariates to be used in the hdPS analyses later. Below is a plot of all of the absolute log of the Bias Multiplier:

We used \(k = 100\) covariates selected by the hdPS algorithm (we call them ‘hdPS covariates’). What should be the cutpoint?

Absolute log of the Bias Multiplier has a null value of 0. Anything above 0 is an indication of confounding bias adjusted by the adjustment of the associated recurrent covariate. For large proxy data sources, \(k = 500\) is suggested (Schneeweiss et al. 2009).

8.2 Investigator-specified covariates

\(25\) investigator-specified covariates are selected based on variables in the DAG that are available in the data set.

We should also add necessary interactions of these investigator-specified covariates, or add other useful model-specifications (e.g., polynomials).

Hypothesized Directed acyclic graph drawn based on analyst’s best understanding of the literature

• 14 demographic, behavioral, health history related variables
  • Mostly categorical
• 11 lab variables
  • Mostly continuous

8.3 hdPS model

C = investigator-specified covariates and EC = hdPS covariates (Schneeweiss et al. 2009)

Then the hdPS can be used as matching, weighting, stratifying variables, or as covariates (usuallly in deciles) in outcome model.

(Wyss et al. 2022)

Schneeweiss, Sebastian, Jeremy A Rassen, Robert J Glynn, Jerry Avorn, Helen Mogun, and M Alan Brookhart. 2009. “High-Dimensional Propensity Score Adjustment in Studies of Treatment Effects Using Health Care Claims Data.” Epidemiology (Cambridge, Mass.) 20 (4): 512.

Wyss, Richard, Chen Yanover, Tal El-Hay, Dimitri Bennett, Robert W Platt, Andrew R Zullo, Grammati Sari, et al. 2022. “Machine Learning for Improving High-Dimensional Proxy Confounder Adjustment in Healthcare Database Studies: An Overview of the Current Literature.” Pharmacoepidemiology and Drug Safety 31 (9): 932–43.