DISCUSSION
We guided readers through a CCW analysis comparing initiation windows
for aspirin and clopidogrel in patients after MI in publicly available
synthetic Medicare claims data. We discussed how to describe the complex
hypothetical interventions under study, whether to clone those
incompatible with their regimens at time 0, and why it is sometimes
necessary to include only recent initiators in creating IPCW at the end
of the treatment window. Finally, we included code visualizing
differences in exposure patterns between regimens to provide helpful
context to the underlying exposure patterns as well as an example of how
to examine censoring weight performance.
Of course, CCW is not a silver bullet. As mentioned previously, the
interventions underlying “start by day X” and “start between day X
and Y” regimens vary depending on the initiation patterns within the
study population. As a result, interpreting treatment effect estimates
or comparing findings across different studies studying the same
regimens requires information on when cloned populations initiate
(similar to studies of other “natural”
interventions).13 Additionally, if predictors of
treatment initiation associated with the outcome are unmeasured or
omitted from IPCW, estimates will be biased. Researchers must also be
concerned about exposure, covariate, and outcome measurement
error,25 differential surveillance between
individuals,26 model
misspecification,27 missing data,26and other common sources of bias in pharmacoepidemiologic
studies.28
This paper focused on providing readers with the most important tools
for performing, understanding, and replicating relatively
straightforward CCW studies contrasting initiation windows. Additional
elements such as requiring individuals to remain on treatment after
initiation or comparing the effects of initiating different treatment
options introduces additional analytic complexity, particularly with
respect to assigning weights at the end of the grace
period.29 We also contrasted effects of different
windows rather than comparing windows to never initiating (though code
to analyze a “never initiate” regimen, which essentially expands the
code covering the first 30 days of the “30-90 day” regimen to the full
180-day period, is also on GitHub). Models for the probability of
remaining uncensored were also fit within clones, rather than within the
base cohort; under additional parametric assumptions, fitting models
within the base cohort can increase precision.16