INTRODUCTION
Researchers have been mimicking randomized trials in nonexperimental
data for decades. One persistent challenge has been allocating time
experienced by individuals while different treatment regimens cannot be
distinguished from one another. Randomized trials can assign patients to
different durations of treatment (e.g., up to six vs up to twelve
chemotherapy infusions)1 after an index date. In
observational data, however, such regimens are often impossible to
distinguish from one another until long after treatment assignment has
occurred. Traditional analyses of these regimens in an observational
context generally change the estimated quantity,2require strong assumptions to estimate that quantity,3or fail to align assessing eligibility, starting treatment, and the
beginning follow-up.4, 5
One analytic solution involves cloning participants into
regimens of interest, censoring clones once incompatible with
their treatment regimens, and weighting uncensored clones based
on their inverse probability of remaining uncensored (IPCW) (hereafter
referred to as CCW for clone-censor-weighting).6, 7CCW can emulate trials of treatment regimens indistinguishable at
baseline without randomization or outcome modeling using the
G-formula8, 9 and compare outcomes when individuals
receive the same treatment at different times (e.g., surveillance
colonoscopy every 3 vs. 7 years)10 or initiate
treatment within specific windows (e.g., statin initiation within 6
months vs no treatment11 or starting hormone therapy
at 3 months vs at disease progression).12
Unfortunately, CCW analyses can be difficult to implement. Moreover, the
causal contrast underlying a CCW analysis comparing regimens like
“start treatment within 30 days” to “start treatment from 30 to 90
days” is difficult to articulate and can vary substantially across
study populations, making information on the study population’s
initiation timings essential context for effect
estimates.13 Here, we walk through using CCW to
estimate the effect of initiating aspirin or clopidogrel treatment
within 30 days of first myocardial infarction (MI) vs within 90 days vs
within 30-to-90 days in the Medicare Claims Synthetic Public Use Files
(SynPUF) and then visualize the hypothetical interventions under study.
The source cohort and commented SAS and R code are posted in a public
GitHub repository at
https://github.com/mawcpharmdphd/CCW_tutorial.