ACA Medicaid Expansion: Causal Effects and an HEOR Translation
Did the staggered ACA Medicaid expansions causally reduce cost-related
barriers to care and improve self-reported health among low-income adults? And in the
language a health-economics consultancy speaks, what does that imply for cost per QALY
and budget impact?
How to read this. The causal estimate comes first and stands
on its own. The HEOR overlay (QALYs, ICER, budget impact) sits on top of it: it converts
the causal estimate into cost-effectiveness terms and stress-tests that conversion against
its assumptions. Where the causal evidence is weak, as it is for self-rated health here,
the overlay reports that it cannot demonstrate cost-effectiveness rather than producing a
favourable number anyway.
1. Identification
Treatment is a state adopting Medicaid expansion; timing is staggered from 2014
(KFF tracker).
The estimand is the average treatment effect on the treated (ATT) among
policy-eligible adults (19–64, ≤138% FPL). Estimating on all adults dilutes the
effect, which we show in the robustness suite. Because classic two-way fixed-effects DiD
is biased under staggered timing with heterogeneous effects, the headline estimator is
Callaway–Sant'Anna group-time ATT with not-yet-treated controls; TWFE and a
Goodman-Bacon decomposition are shown only to demonstrate that bias. Inference uses a
state-clustered wild-cluster bootstrap (the correct fix for ~50 clusters).
Falsification. Non-flat event-study pre-trends would sink the
identification; we test them formally. If the coverage/cost effects vanished under CS,
the hypothesis would be falsified. A self-rated-health effect that is null, or that fails
the pre-trend test, yields no credible QALY gain, and is reported as such.
2. Headline causal results
Pre-trend test is a joint Wald test on the event-study leads (event time
< −1). "Flat" (p > 0.05) means the parallel-trends assumption is not rejected.
3. Why not naive TWFE
4. HEOR overlay: QALYs, ICER, budget impact
The QALY gain is derived from the causal change in self-rated health, mapped to
utilities via a sourced crosswalk (base case
4.1 Probabilistic sensitivity analysis
5. Robustness & refutation
6. Limitations
- Self-reported outcomes. BRFSS measures perceived access and health, not
clinical outcomes or mortality.
- QALYs are modelled, not measured. The utility mapping is an assumption layer;
the ICER is an illustrative translation, sensitivity-tested, not a primary cost-utility
measurement. A null health effect yields no QALY gain.
- MEPS is national. The cost side is parameterised from national data, so the
budget impact is a stylised model.
- Few clusters & co-treatments. ~50 states limit precision; concurrent ACA
marketplace changes co-occur, so the ATT is identified off adoption-timing
differences, not a universal expansion effect.
- Not advocacy. The analysis estimates effects and cost-effectiveness under
stated assumptions; it argues neither for nor against expansion.