There is great excitement about the potential for AI to reshape science, but so far very little empirical evidence about how that is (or is not??) happening in real time. I'm excited to share a new working paper with Carolyn Stein about the impact of AlphaFold on science ->

(Addendum: Another way of looking at this is that the WH envisions a “minimum premium” requirement that goes far beyond just requiring all enrollees to pay a de minimis premium. Either way, the incipient WH proposal is much stingier than meets the eye.)

But, before even getting there, it’s important to recognize that what’s been called a WH proposal to extend the enhanced subsidies actually falls well short of a full extension. /end

To gauge the proposal’s overall impact, one would also need to consider the proposal’s cuts to the underlying subsidy structure (i.e., the new $5 minimum premium & ending silver loading), the new HSA proposal, and the various regulatory and other changes.

While there’s been a lot of attention to the fact that the White House was going to propose not extending the enhanced credits above 700% of the FPL, these changes below 200% of the FPL would be far more consequential, both in coverage and fiscal impact.

In 2025, people below 133% of the FPL accounted for about a quarter of enrollment (> 6 million people), and the 133-200% of FPL group accounted for close to two-fifths of enrollment (~9 million people). All told, about 2/3 of Marketplace enrollees have incomes below 200% of FPL.

That change would wipe out almost all of the benefit of the enhanced credits for people with incomes below 133% of the FPL and much of the benefit for people in the 133-200% of FPL range (about two-thirds right at 133% of the FPL, falling gradually to zero at 200% of FPL).

Press accounts suggest that, under the incipient WH proposal, all of these applicable percentages would go up to at least 2%. (The phrasing here is a little bit unclear, but this seems like the most sensible interpretation of what’s written.)

apnews.com/article/trum...

What Congress did in the ARP/IRA was lower these applicable percentages (and newly create one for people at or above 400% of the FPL). This is what makes the “enhanced” subsidies “enhanced.” See the enhanced & underlying schedules below.

Some background: the generosity of the premium tax credit is controlled by “applicable percentages” set in law, which determine the percentage of income that a subsidy recipient is expected to contribute to a benchmark plan. They vary based on income.

There’s a feature of the abortive White House ACA subsidy proposal that I think has been underappreciated. Namely, it seems to only partially extend the enhanced subsidies for people with incomes below 200% of the FPL (~$31k for a single person). That’s about 2/3 of enrollees.

Addendum: Ack. This tweet had crucial typo. The 2nd sentence should be "Enrollees who shift from TM to MA when MA grows are likely *cheaper* than the average TM enrollee but *costlier* than the average MA enrollee, so *both* groups likely get more costly as MA grows."

bsky.app/profile/matt...

Ack. This tweet had a crucial typo. The second sentence should read "Enrollees who shift from TM to MA when MA grows are likely *cheaper* than the average TM enrollee but *costlier* than the average MA enrollee, so *both* groups likely get more costly as MA grows."

To be clear, the MA payment system’s *existing* accuracy problems may (and, in my view, do) offer a strong rationale for reform. But the dynamics we consider here seem unlikely to do much, if anything, to bolster that case. /end

A second is that rising MA penetration is unlikely to change selection patterns in ways that seriously reduce the accuracy of the MA payment system and necessitate reforms that would break the link between MA payments and TM costs.

If that’s right, it has a couple of implications. One is that TM is likely at little risk of entering a “death spiral” in which higher MA penetration leads to greater favorable selection that induces still further increases in MA penetration, and so on.

Our approach has limitations, including that it cannot address potential confounding from county differences that vary over time. But these results suggest that further growth in MA will have little effect on the degree of favorable selection into the program.

For example, the results imply that if TM’s market share fell by 50%, then the effect on the TM-MA difference in risk-adjusted costs would lie somewhere between a negligible change and a decline of around 0.6 percentage points.

Details are in the paper, but this figure shows the main results: across a wide range of assumptions about who “switchers” are (reflected in the different values of theta), changes in MA penetration have little effect on the degree of favorable selection into MA.

The panel data allow us to control for persistent cross-county differences, while the model structure allows us to explicitly account for the fact that stayer-switcher cost differences may not coincide with differences in average costs between TM and MA enrollees.

To address these issues, we use county-year panel data on MA penetration and stayer-switcher differences to estimate an empirical version of the theoretical model sketched above.

Under this assumption, the model sketched above suggests that stayer-switcher cost differences may shrink as MA grows even if the difference in average costs between TM and MA enrollees is stable or growing. See, in particular, panels A and B below.

This matters because MA penetration may not affect the two differences in the same way. To see why, suppose we make the (arguably fairly plausible) assumption that TM-to-MA “switchers” correspond to the enrollees on the margin between TM and MA.

The second issue is more subtle. TM-to-MA “switchers” are likely not representative of MA enrollees as a whole, so the cost difference between stayers and switchers may not measure what we’re actually interested in: the difference in the *average* cost of TM and MA enrollees.

Indeed, it’s notable that *changes* in MA penetration are associated with modest declines in stayer-switcher differences, consistent with the concern that cross-sectional relationships are confounded to some degree.

The first is the potential for confounding. Counties with higher MA penetration may differ in other ways that affect stayer-switcher differences, masking the true causal relationship between MA penetration and stayer-switcher differences.

But for a couple of reasons, this may not be a good guide to the causal effect of MA penetration on the degree of favorable selection into MA. There are two main issues.

Prior work has examined the cross-sectional relationship between MA penetration and stayer-switcher differences in risk-adjusted costs, finding little relationship. We replicate that finding:

If “switchers” have lower prior year spending than “stayers” (after risk adjustment), that’s indicative of favorable selection into MA. Larger stayer-switcher differences suggest more intense selection.

Thus, we tackle this question empirically. To do so, we first construct a measure of favorable selection at the county-year level. Following prior work, our measure is the difference in the prior-year spending between TM-to-MA “switchers” and TM “stayers.”