With my Part D work, I am concerned and trying to protect our ability to do a strong causal inference study when I am worried about the power of the data structure that we have to do such a study.
We are trying to estimate the effect that pharmaceutical access expansions have on long-term care use.
We assume that for the treatment population ages 65-69, the population ages 60-64 offers a good counterfactual for change in long-term care use. The population 65-69 has a shock, exogenous to long-term care use trends, that causes a portion of its uninsured population to switch to an insured population, and this uninsured-to-insured population is a good representation of [##ask-control or treated##] population of interest.
The policy question that Aparna is looking to address is
The empirical intention is to “estimate the impact of prescription drug insurance on elderly individuals’ utilization of formal and informal long-term care” – but impact for whom? the uninsured is assumed
The second empirical intention is to examine how changes in long-term care use affects informal caregivers, but is written as “how changes in LTC use affected labor market and mental health outcomes of informal caregivers.” I’ll need clarification here.
Furthermore, we will do “heterogeneity tests”
The treated group is all Medicare-eligibles. Because they had a shift in drug access caused by Part D. But grouping the three major treated categories (uninsured->uninsured, uninsured->insured, insured->insured) together will dilute the ability to best test our hypothesis about the effect of increased access upon LTC.
The whole endeavor seems ripe for SEM.
But I am also coming around to using
Future topics to cover:
– power analyses
– ATET or ATC estimand – or can we develop a weighted
When you selected the IV method in your 2018 proposal, did you choose not to do propensity score matching/weighting or synthetic control methods for any particular reason? Do you view the IV method as equivalent to using propensity score weights, or as fundamentally different? (I’m thinking that using uninsured-hat changes the estimand from average treatment effect on the treated to ATE on the treated & uninsured.) Are we generally flexible about the estimand? i.e., do we want to estimate average treatment effect of Part D on the treated (which is an overlap group – it includes pre-treatment uninsured and insured), and/or on the treated & pre-treatment uninsured?
If we’re flexible, perhaps we could try to use our study to extend the estimated effect over to today’s elderly population, weighting our study’s estimated ATE based on the demographics & insurance characteristics in today’s post-Part D Medicare-eligibles.
I continue to study ways we can strengthen our causal
inference, which we would need to settle prior to specifying a power analysis.
I am concerned about the applicability of an IV method using demographics
because I think, in theory, that demographics were relevant to changes in LTC
use during the studied period. I am worried about the pooling of effects of the
Part D treatment across multiple groups. I.e. they would not satisfy exogeneity
to the DV except for the ways they relate to Rx insurance. There may be a set of
demographic variables that we could carefully select as instruments that in
theory relate only to Rx insurance, and we could test their exogeneity in the
data. I also am studying if we could generate an estimand of ATE for