• Plans integrated with providers charge higher premiums

    Here’s the headline finding of my new paper in Health Services Research, coauthored with Steve Pizer and Roger Feldman: Controlling for quality, Medicare Advantage (MA) plans that are fully integrated with providers (owned or controlled by the same entity) charge higher premiums, relative to plans that are not integrated in this way. We also found that such plans have higher star quality ratings, which is why it was important to control for them. Lastly, we found no evidence that integration is associated with more generous benefits.

    Our work is in the context of the incentives of the Affordable Care Act (ACA), which promotes such integration.

    Through bonus payments for quality improvement and cost reduction, the ACA encourages the formation of accountable care organizations (ACOs), networks of providers responsible for the care of a defined group of Medicare patients (Frakt and Mayes 2012). ACOs give providers an incentive to consolidate the spectrum of care under one management because bonus payments will be tied to performance on quality measures and a spending target based on the difference between a benchmark and all Medicare spending attributed to beneficiaries associated with the ACO—even when incurred for services provided outside the ACO. In addition, some ACO contracts put providers at financial risk if their costs are above a benchmark. Consequently, ACOs with risk management capabilities will be better positioned to succeed (Fuchs and Schaeffer 2012). Providers can develop these capabilities internally or acquire them by merging with an insurer. Finally, the ACA offers quality bonus payments to MA plans (Jacobson et al. 2011). To the extent that higher quality can be achieved through plan–provider integration, this is another incentive to integrate. [Hyperlinks added.]

    The study was based on 910 MA plans operating in 2009 in U.S. states and D.C., excluding special need plans, private fee for service (PFFS) plans, plans not offering drug benefits, and other plan types with very low enrollment (like regional PPOs and medical savings accounts). (A sensitivity analysis focused on non-drug MA plans obtained similar results.) Supplementing public data from the Centers of Medicare and Medicaid Service (CMS), we constructed a unique database of fully integrated MA plans by reviewing plans’ websites and governing documents to determine which plan-offering firms had vertically integrated with a hospital or provider group.

    We acknowledge that we examined an extreme form of integration, which can occur, for example, if a provider offers its own plan. It is a special case of the more general concept of vertical restraints or vertical behavior — contracts of various types between plans and providers. A recent example of vertical behavior is the Bon Secours-Aetna ACO deal.

    Our models of plan premiums and benefits control for quality, market structure, cost, and demand factors (all standard in the literature in which our work is placed). Our model of plan quality also controls for market structure, cost, and demand factors.

    We found that integration is associated with an increase of $28 per month in premiums and a monetized increase of just under $8 per month in quality. Consequently, about 70 percent of the total premium difference between integrated and nonintegrated plans is not attributable to quality. Some or all of this premium increase could be associated with benefit enhancements by plans, but we did not observe a statistically significant increase in benefit generosity with integration among the subset of benefit variables we examined. An alternative possibility is that higher premiums for integrated plans are related to higher market power commanded by those plans (whether due to integration or a causal factor of it). Because we did not examine all possible benefits, we cannot completely distinguish between these two possibilities, and that is a natural focus for future investigation.

    Because ours was necessarily an observational study, one must make additional assumptions to infer the extent to which integration caused higher premiums and quality. First, prior theoretical and (limited) empirical literature suggests that our estimate of the extent to which integration causes higher quality is biased upwards. Most likely, integration causes less of a quality increase than we predict.

    Second, we confirmed that our integration variable is historical, i.e., that plans that integrated with providers did so well before 2009. If one assumes that such lagged integration is not correlated with any unobserved factors that also affect premiums in 2009, then one can say that integration causes the premium increase we observed.


    • Isn’t this controlling for an intermediate outcomes? If ownership affects both quality and price, directly, and quality, itself, affects price, then a regression of price on ownership type, controlling for quality, gives biased results for the direct effect. A nice presentation in sections 9.7 and 9.8 of http://www.stat.columbia.edu/~gelman/arm/chap9.pdf

      • You can leave quality out of the premium equation and you get a (without quality) integration coefficient that is exactly the sum of the (with quality) value and the quality coefficient. I would expect an economist to then say, “You’ve got omitted variable bias. Integrated plan-providers are of higher quality and charge higher premiums. That integration leads to better quality is precisely the justification for it. So, you have to control for it.”

        Choose your poison.

        • These are both good points. By including quality we ARE controlling for an intermediate outcome, so our estimate is too low if we were asking the policy question, “what unconditional effect would integration have on premiums?” Instead, we are asking “what effect would integration have on premiums after accounting for induced higher quality?”

          Given the quality-based rationale for integration, we thought this was the key policy question, but PLW is right to point out that it’s often a mistake to include intermediate outcomes on the right hand side of a regression model.

        • Those economists would be wrong. It’s not a missing control; it’s a bad control, and putting it in leaves you with something that is neither the total effect of ownership on price nor the direct effect of ownership on price.

          Think about it in the randomized trial context. Some drug might increase weight by 50% but have no direct effect blood pressure, independent of changes in weight. If that’s true, then if we wanted to find the effect of the drug on blood pressure in a trial, we wouldn’t want to condition on (post-treatment) weight, because a 300 lb person on the drug and a 300 lb person on the placebo are probably quite different in other determinants of blood pressure, and we would (incorrectly) find that the drug helped blood pressure because the treated 300 guy had lower bp than the control 300 pound guy.

          To go back to the health plan example, assume a plan being integrated has no direct effect on price, but instead just being affiliated with a provider makes customers on those plans rate them higher for purely psychological reasons (Warm glow or confidence or something), so that plans that are relatively bad deals for consumers (high prices relative to benefits) get higher satisfaction ratings/fewer complains/fewer customers dropping when offered by integrated providers. Since higher marks on these things yield higher stars, we would find that integrated providers charge higher prices, conditional on stars, even though integration had no real effect on prices at all (and simply regressing prices on integration would find nothing).

          Is something like that happening? I have no idea. But I’d really like to see the unconditional regression just to be sure.

          (Aside.. I’m really “hyper-sensitive” about this because I just blew this same sort of thing in a paper I wrote/submitted and felt like an idiot when a referee pointed it out).

          • We did do an unconditional regression and, as I said, the integration coefficient soaked up the quality one. Another approach is a joint model among premium, quality, and integration (and/or instrumenting for one or several of those). Given the data, we couldn’t take that approach in this paper. But, with more data collection, someone certainly could.

            • I’m sorry. I misunderstood the claim “You can leave quality out ..” I thought you meant you could always do that and get the result you describe. In fact, I think you are saying that you actually did it in this instance and got that result. I’m with you now.