Inverse Probability of Treatment Weighting (Propensity Score) using the Military Health System Data Repository and National Death Index

In the past, we've used the military healthcare database to help determine which patients would most benefit from statin treatment to help reduce their chance of a future heart attack. Our protocol offers a roadmap to any medical provider who is interested in using big data to answer their own clinical questions and can be applied to a far ranging and nearly limitless number of topics. After attaining approval from the institutional review board, complete a data sharing agreement application, taking care to specify the data fields and files being requested on the DRT MDR extractions worksheet.

Specify whether the team is requesting a data analyst, supply of the raw data, or if the team will access the MDR directly and whether the request is for a one-time data pull or if regular pulls are requested daily, monthly, or yearly. To access the MDR directly, complete the MDR authorization request in the MDR-CS-2875 forms. When accessing the MDR directly, follow the instructions for accessing and using the MDR, including the software requirements and example SAS programs that are available in the MDR users guide and the MDR functional guide.

To obtain additional MDR beneficiary data for your cohort, acquire the VM6 files from the server and merge with your cohort file. Use different variable names in the VM6 data draw and cohort files for patient names and birth dates to help with subsequent checking for errors following the merge. As database entries are never completely free of error, use the code to preform error checks after each major step, in addition to checking the program log and output for any potential concerns.

When comparing names from the cohort file with the VM6 file, only match the first three letters to improve efficiency and reduce any false errors that may arise with differences in spelling or spacing between files. Then review the error data file, performing a manual review of the health record as needed, to check for other errors of concern. For relevant data extraction, obtain the race and sex data from the VM6 beneficiary files, merge this date with the cohort file, and check for errors as just demonstrated.

Obtain death data from the death master file, merge the data with the cohort file and check for errors. Then obtain additional data files as needed for the analysis. For data merging and summative file construction, extract baseline comorbidities using ICD-9-CM or ICD-10-CM codes from the period before the index date.

Taking care that the patients had eligibility for the military healthcare system during the baseline period as verified in the VM6 beneficiary file. Search the baseline diagnosis codes in the outpatient and/or inpatient files to establish the baseline comorbidities during the baseline 12-month period prior to the index date. If using Elixhauser comorbidities, use the available software from the healthcare cost and utilization project database, making sure to modify the names of the diagnosis variables and files as needed.

To set a study end date for all of the patients as a cut off for followup for patients who have no demonstrated the outcome of interest, search the VM6 beneficiary file to ensure eligibility for healthcare through the study end date. If it is important to limit the study to active users of the healthcare system, independent of eligibility, then determine the last healthcare contact within the data files and censor the patients at that date. When all of the necessary information has been acquired, generate a random patient identifier for each patient using Ranuni.

To model the probability of treatment use logistic progression to calculate the weights from the predicted probability. Then stabilize the propensity score by dividing the score by the mean weight. To verify the balancing after applying the inverse probability of treatment weighting the standard differentiation macro simplifies the computing standardized mean differences for co-variances before and after weighting in SAS.

To generate a cumulative incidence function plot in PROC PHREG, reference a covariate file to specify covariate values to be used when generating the plot in the PROC PHREG syntax, use the weight statement to specify the standardized propensity score variable and use the baseline statement to specify the values for baseline covariates in order to plot the cumulative incidence function. Then specify the strata to use for the plot using ROWID. Here, an example of appropriate balancing in a large cohort of 10000 participants using the absolute standardized difference plot macro as shown.

After application of the propensity score, the absolute standardized differences were reduced significantly. Here, the unsuccessful results of attempting to balance covariates in a cohort of 100 participants can be observed. In this example, PROC PHREG with standardized propensity score weights were used to generate a cumulative incidence function plot, revealing that in this analysis, the untreated control group demonstrated a larger number of events and exhibited a comparatively worse survival than did the treated group.

The MDR includes beneficiaries of the department of defense, including retirees and family members. So it's not restricted to just patients over 65, and it is not restricted to just veterans. This protocol provides a starting point for using the MDR.

And the statistical technique overview can help eliminating biases that can be intrinsic in this type of research.