Work · Analysis
Demonstration · Public data · Not client work
Did Medicaid expansion reduce uninsurance among low-income adults?
Question
When a state expanded Medicaid to adults with incomes under 138% of the poverty line, did fewer of those adults go without health insurance — and by how much, compared with what would have happened anyway?
Data
The uninsured rate among adults aged 19–64 with incomes under 138% of the poverty line, for every state and DC, from the Census Bureau's American Community Survey microdata, 2009–2019 (with 2021–2024 for a secondary look; 2020 had no standard 1-year release). I built the rate from person records rather than the published table because the table changes its age band in 2017; the two agree closely where they overlap. Expansion dates come from KFF; a state counts as treated in the year its expansion took effect if that was on or before July 1.
Design
I compared how the uninsured rate changed in states that expanded with how it changed, over the same years, in states that had not yet expanded.
Result
Expansion lowered the uninsured rate among low-income adults by 7.2 percentage points (95% interval 4.6 to 9.7).
View data table
| year_relative_to_expansion | estimate_pp | ci95_low | ci95_high |
|---|---|---|---|
| -5 | 0.6 | -0.88 | 2.08 |
| -4 | 0.18 | -0.93 | 1.29 |
| -3 | -0.24 | -1.25 | 0.76 |
| -2 | 0.34 | -0.45 | 1.14 |
| -1 | 0.0 | 0.0 | 0.0 |
| 0 | -4.09 | -5.84 | -2.34 |
| 1 | -6.88 | -9.82 | -3.93 |
| 2 | -7.46 | -10.66 | -4.25 |
| 3 | -8.39 | -11.56 | -5.21 |
| 4 | -8.02 | -11.56 | -4.48 |
| 5 | -8.62 | -12.11 | -5.12 |
The comparison group matters more than the method. Between 2013 and the 2016–2019 average, the uninsured rate in the 26 states that expanded in 2014 fell by 18.8 points (95% interval 15.3 to 21.3). But it fell by 10.6 points in states that did not expand, because the rest of the Affordable Care Act arrived everywhere at once. Net of that, the change is −8.2 points (−11.2 to −4.3).
View data table
| comparison | decline_pp | change_pp | ci95_low | ci95_high |
|---|---|---|---|---|
| No comparison group | 18.8 | -18.79 | -21.31 | -15.34 |
| Comparison states | 10.6 | -10.64 | -12.36 | -9.1 |
| Net of comparison | 8.2 | -8.15 | -11.17 | -4.33 |
The simple comparison gives −8.2 points because it looks only at the 2014 states two to five years in, when the effect had grown; averaged over every expanding state and every year after expansion, including the first, smaller year (−4.1 points), the effect is −7.2 points — the headline.
A simpler regression (two-way fixed effects) gives almost the same answer, which is expected here: 26 of the 34 states that expanded by 2019 did so together in 2014, so there are few staggered comparisons for it to get wrong.
| Estimate (percentage points) | Population-weighted | Equal state weights |
|---|---|---|
| Callaway–Sant'Anna | −7.2 (−9.7 to −4.6) | −6.4 (−8.6 to −4.2) |
| Two-way fixed effects | −6.9 (−9.7 to −4.1) | −6.7 (−8.9 to −4.6) |
Checks. The result holds when I change the sample:
| Check | Estimate | 95% interval |
|---|---|---|
| Drop DE, DC, MA, NY, VT | −8.5 | −10.6 to −6.3 |
| Drop DC only | −7.2 | −9.9 to −4.6 |
| Utah coded first treated 2019 | −7.1 | −9.8 to −4.3 |
| Utah dropped | −7.2 | −9.9 to −4.4 |
| Wisconsin dropped | −7.2 | −10.0 to −4.4 |
| New Hampshire dropped | −7.2 | −9.8 to −4.5 |
| All adults 19-64 | −2.3 | −4.4 to −0.2 |
| Equal state weights | −6.4 | −8.6 to −4.2 |
| Drop the 2015 cohort (IN, NH, PA) (added after seeing results) | −7.2 | −10.3 to −4.1 |
| Drop the 2014 adopters (added after seeing results) | −7.4 | −9.4 to −5.3 |
Dropping the five states that already covered this group before 2014 makes the effect larger; using all adults 19–64, most of whom were never eligible, makes it much smaller. Two rows were added after I saw the results, and are labeled so.
Before expansion. The pre-specified test that the four pre-expansion estimates are jointly zero does not reject (p = 0.24). The estimator's built-in test, which stacks every pre-period cell, returns p = 0, but it is unusable here: its covariance matrix has a condition number of about 4.7 × 10⁷ because the 2015, 2016 and 2019 groups have two or three states each. Looked at one group at a time, the 2014 states are flat before expansion, but the 2015 states (Indiana, New Hampshire, Pennsylvania) were already drifting (e = -6: −4.3 points (z = −3.2); e = -5: −3.3 points (z = −2.2)). Dropping them leaves the result unchanged.
How much a pre-trend could matter. Moderate robustness — the average effect stays clear of zero unless the post-expansion violation exceeds 1.5 times the largest pre-period deviation.
Against published estimates. Two published studies of the second year of expansion found −8.2 points (Miller & Wherry (2017), NEJM 376(10):947-956; National Health Interview Survey, U.S. citizens 19–64 under 138% of poverty) and about −6.0 points (Wehby & Lyu (2018), Health Services Research 53(2):1248-1271; ACS, low-educated adults 19–64). My second-year estimate is −6.9 points. Both figures come from the studies' abstracts and use different populations, so the comparison is approximate.
Over the full 2009–2024 period the estimate is −6.2 points, but those later years are confounded by the 2020–2023 rule that kept Medicaid enrollees covered in every state and by the 2023–24 unwinding, so I treat them as context, not as the result.
What this does not show
- Who gained coverage. This is a state-level rate; it does not say which people gained coverage or whether their health changed.
- Clean timing everywhere. Several states covered some low-income adults before 2014, and some expanded partway through a year; the robustness rows bound these choices but cannot remove them.
- Recent years. After 2019 the comparison is confounded by pandemic-era continuous enrollment, its unwinding, and Georgia's limited Pathways program (from July 2023).
Reproduce it
The code, the analysis plan (committed before any estimate was run), the decision log and the results files are at https://github.com/ts2427/ThresholdDataSciences/tree/main/demos/medicaid_did; the run instructions are in the demos README.
If this were your data
The same design answers "did our change work?" for a price, a policy or a program rolled out to some places or groups before others: it separates the change you made from everything else that moved at the same time, and says how sure you can be.