The Profitability Paradox: Why Platinum Plans Are Both the Best and Worst Bet in the ACA
An excursus from NBPP analysis
We’re going to take a detour from NBPP analysis to dive deep into something we uncovered while working on our NBPP analysis.
In my last post, I shared a chart I almost didn’t publish because I thought the results might not be correct, namely, the net margin by segment. The post was focused on the margin economics of CSR, but it also showed something interesting: platinum plans have the highest margin per member, per year, of any plan.
This goes against a common theme that I encounter among health plan executives, actuaries, and others who’ve been around the ACA for a while: everyone says platinum is a disaster, an adverse selection magnet, and that carriers shouldn’t offer them. I’ve been skeptical of this framing, because I’ve seen (and helped lead) health plans be successful with platinum offerings, but I’ve kept my mouth shut in recent years because it’s always seemed like such a settled question that I just resigned myself to the fact that I must be wrong.
And so when I saw this result, I went through the analysis that produced it with a fine-toothed comb, and found why I think both things could be true: platinum is both the most profitable metal level, but also an adverse selection magnet that is too risky for many (perhaps even most) carriers to consider offering.
The Method to the Madness
Before I get there, though, I want to explain in more detail how I calculated these estimates. The basic formula for gross margin (before admin expenses) in the ACA is:
This kind of formula isn’t the sort you’d run at the individual member level because the nature of insurance is pooled risks, but looking at it using different levels of aggregation can provide lots of meaningful insight into trends.
Evensun has the EDGE Limited Data Set, a de-identified dataset containing claims from all ACA members for years 2022 and 2023, which we have used to develop comment letters to proposed regulations. The EDGE LDS has a lot of useful information about the claims (including cost), so the medical claims plus pharmacy claims portion of the formula above we have. But for obvious privacy reasons, the information about the individuals associated with the claims is quite limited. The only information available in the data set about the members is the metal level, the length of their enrollment, CSR variant, gender, age, their market (individual or small group), and (beginning in 2023) a flag indicating whether or not they received APTC.1 We don’t get information about how much premium they were charged, how much APTC they got, what kind of plan they were on (PPO, HMO, etc.), what dates they were enrolled, geography, whether they’re in a state that expanded Medicaid, whether their state used HealthCare.gov or ran its own exchange, etc.
So we’re missing premium and risk adjustment (which are enrollment-dependent), as well as pharmacy rebates for our profit formula. We can’t get any of these perfectly, but we can build some proxies. We’ll start with premium, as that’s going to be the largest part. We had to make some pretty big assumptions, and those assumptions are going to limit the kinds of conclusions we can reasonably make from this data. We decided the closest we could get with the information we had was a nationwide average premium at the metal level, adjusted as best we can. So we used the Unified Rate Review Templates for plan year 2025, which has 2023 experience period data, to come up with a weighted average nationwide premium by metal level, and the URRTs for 2024 for 2022 data.2 There’s some limitations here: a carrier who offered plans in 2023 but left in 2024 or 2025 won’t be in this data set (since they won’t have a 2025 rate filing), and it also assumes the carriers filed the data accurately. The impact of exiting carriers was immaterial for 2023, but for 2022 is a larger factor because Bright and Friday both had substantial market exits in 2023. We considered using the OEP Public Use Files, but found that the cross-tabs of age and premium in state-based exchanges couldn’t get us what we needed. We compared the total member months from the URR files to the total member months in EDGE LDS, and got comfortable that it was directionally correct.
We’re not done, though, because using the same premium across members would be rather inaccurate. In most states, the ACA’s age rating curve applies, and a 64-year-old pays three times what a 21-year-old does. So we calibrated a base rate using the federal age curve, and then multiplied each member’s base rate by their individual age factor. Of course, not every state uses the federal age curve, so that’s another limitation, but most do, and outside of New York, the alternate age curves aren’t too far off, so for this purpose it was close enough.
Next we have risk adjustment. We’ve got everything we need to calculate the risk score, so we built a Python script that managed to calculate all 30 million risk scores in the span of about 25 minutes. We then compared it to a nationwide weighted average risk score from the CMS risk adjustment reports. We were a bit lower, but that makes sense: the EDGE LDS data redacts all claims related to substance use disorder or behavioral health except to researchers with a specific use case related to behavioral health, which isn’t us; thus, we’d expect our calculated risk score to be lower without those conditions in our data.
So now we’ve got risk scores, but that’s not enough to calculate the actual dollars involved. The risk adjustment formula doesn’t just use a member’s risk score. It also uses an induced demand factor (derived from their metal level; we’ve got that), an actuarial value (also derived from their metal level; we’ve got that), an allowable rating factor (an age-based factor that aligns with the federal age curve; we’ll use the same proxy as we did for premium), and a geographic cost factor (GCF). We don’t have the GCF, but the GCF appears on both sides of the transfer formula, and the way it’s calculated in reality, it should be 1 on a weighted average basis, so we just set it to 1 on both sides and forget about it.
Then, the formula calls for statewide averages for all these factors, plus average premium.3 Here, we calculated the nationwide average of our whole dataset for each factor and used that for everyone; we don’t know anything about the geography of the members, so using the nationwide average is the best we can do. Putting it all together and sticking it into the risk adjustment formula, we get an estimated risk transfer at the member level. We double-checked that the whole thing summed up to $0 across our whole data set, since risk adjustment is zero-sum.
Lastly, there’s one other piece to risk adjustment we included, which is something called the high cost risk pool: this is a program where members whose claims exceeded $1M in a year have their costs pooled nationwide at a rate of 60% for every dollar above $1,000,000. So if a person had $2M in claims in one year, the health plan gets $600,000 back from the high cost risk pool (HCRP), and that $600,000 gets paid for as a percentage of premium by all the other health plans in the country. So we accounted for both the cost and the reimbursement side of HCRP.
One thing we’re not capturing: several states operate reinsurance programs under Section 1332 waivers — Colorado, Maryland, New Jersey, Oregon, Pennsylvania, Georgia, and others — where the state subsidizes high-cost claims to reduce premiums. These programs reduce issuer-paid claims (and premium along with it) which means issuers in those states would show higher margins than our national averages. Since we don’t know which state any given member is in, we can’t model this, and we just have to live with this limitation.
We’ve now got the main pieces of the formula: premium, risk adjustment, and claims. There’s one last piece: pharmacy rebates. This one’s a bit harder — a robust analysis would probably calculate average rebates at the NDC level, but that’s not a data set that will be easy to come by (if any PBMs want to give me that data for free, drop me a line; I’ll promise to not say mean things about PBMs for a whole month as a thank you). Without a data set like that, we took a simpler approach: just apply a flat percentage by getting the total pharmacy rebate amounts reported on carrier MLR filings and dividing that by total pharmacy spend reported on the same report. Then we applied that to the overall pharmacy spend by member. That way it’s at least correlated with pharmacy spend, which should generally be true. The limitation is that rebates are not a uniform function of all pharmacy spend; some drugs (especially specialty drugs) get giant rebates that can be more than 50% of their cost, and others may get none (generics have no rebates, but also some orphan drugs that are the only kind in their class get no rebates, since the insurance company basically has no choice but to cover them, so there’s no reason for the manufacturer to pay the rebate). So we have to be cautious depending on the kind of analysis we’re doing, but overall it should be enough to give us some overall insights. We did a quick compare to some NAIC datasets and found everything was reasonable, and the fun begins.
What We Found
What we found was interesting: first, all metal levels, in aggregate, had positive margin in 2022 and 2023. And platinum had the highest gross margin; I checked and double-checked it, and I don’t think we made any material methodological error (although some of the limitations and assumptions discussed in this post could be material to the conclusions with platinum since it’s such a small segment). But I also discovered what I think the reason is that carriers tend to fear platinum plans.
If we break the profitability down by deciles (grouping members into 10 roughly equal buckets ordered by margin, weighted by enrollment duration), we find an interesting pattern:4
The most profitable 10% of platinum members are roughly twice as profitable as any other metal level’s top decile, but the least profitable 10% generate double the losses of the least profitable gold members and more than 4× the losses of the least profitable bronze. In other words, platinum is high risk, high reward.
The thing we don’t have in this data is how these members were distributed. If every health plan got an equal share of all 10 deciles, they’re probably happy. But of course, that isn’t reality. If you offered a platinum plan and got a disproportionate share of the bottom decile (the catastrophic tail) it was a financial disaster.
And with only about 400,000 member-years nationally in platinum (compared to over 10 million in bronze), any one issuer’s platinum book may be small enough that a handful of catastrophic cases can flip the entire metal from profitable to deeply underwater.5
This is where the conventional wisdom is right: platinum is an adverse selection magnet, and platinum is risky. What it gets wrong is concluding that platinum is therefore unprofitable, or that it should be avoided in all circumstances. For the pool overall, the math works: the risk transfer system actually over-compensates for the chronically ill members that platinum attracts, and the margin from those members more than offsets the catastrophic tail. The problem is how variable the outcomes are.
Bronze is boring and reliable: 90% of enrollment is profitable, the worst decile loses $14K/MY, and there are 10 million members to smooth it out. Platinum is volatile: 70% of enrollment is profitable, the worst decile loses $59K/MY, and it’s a small pool: outside of California, you rarely have any competition if you want to offer a platinum plan, and higher risk individuals gravitate to that option.
There’s one more thing I’ll share that I think brings clarity to the adverse selection story a bit. We looked at what percentage of members in each profitability decile have at least one HCC (hierarchical condition category: the chronic and acute diagnoses that drive risk adjustment). In the top profitability decile, 97% of members have at least one HCC across every metal level. The most profitable members in the ACA aren’t the healthy ones; rather, they’re the chronically ill ones whose risk scores generate transfer inflows that exceed their actual claims. This fact is something every ACA actuary knows, and every health plan executive should know.
Platinum amplifies this effect. The top decile is HCC-heavy for every metal level, but with platinum its top three deciles are 60–97% HCC-positive, while the other metal levels all drop to around 10% by the third decile. The adverse selection that everyone warns about (sick people choosing platinum) is real, but it’s also the profit engine. What kills platinum isn’t the chronically ill members; it’s the catastrophically ill members whose costs exceed what even generous risk transfer can cover. Those members show up in the bottom decile, and importantly, the least profitable members on platinum also tend to have HCCs: around 67% have HCCs on platinum, compared to only 30% on bronze. Bronze members who are very unprofitable had an acute, unexpected event happen, while platinum members were already sick and just got sicker, and a catastrophe on top of a chronic illness on average costs more than a catastrophe without one.6
If you’re a plan who isn’t good at managing members with complex, co-morbid conditions, but you offer a platinum plan, you could end up attracting an outsized share of complex case you’ve got neither the financial nor operational capacity to manage well. That’s not a good outcome for anyone, least of all a consumer who picked your plan because they thought platinum would be the key to meeting their health care needs.
But if you design a plan that is targeting people with conditions you know can manage well and you manage to figure out how to get the attention of the people who would benefit the most from the unique aspects of that plan, then the opposite happens: the member is taken care of, which should lead to better outcomes: both clinical and financial. And that member is likely to stick around, too: platinum members are stickier – 75% of the most profitable decile at the platinum metal level in 2023 were on that same plan in 2022. Platinum is a long-term play.
We’ll have more to share as we continue digging into this data. If you’re an issuer who wants to understand your own book-level economics with this kind of granularity, not proxy estimates, but actual member-level profitability using your own claims, premiums, and risk scores — that’s what Evensun does. Reach out to joe@evensun.com (I’m writing this post on the eve of my paternity leave… technically, 40 minutes into my leave and will get in trouble with my wife if I write any more, but the team is on it).
Be sure to subscribe if you’re not already, and if you think we got it wrong, or just want to start a discussion, leave a comment.
We abandoned work that tried to use this meaningfully as we found a lot of discrepancies - e.g. catastrophic members with this flag.
Specifically, we divided the Experience Period Total Premium by Experience Period Member Months for each metal level to get the actual average collected premium per member per month.
For the risk adjustment formula’s statewide factor denominators (FIRS and FERS), we computed these from the EDGE data itself rather than using CMS-published values. The CMS-published values are higher because they include SUD-related claims that are redacted from the LDS. Using CMS denominators with our EDGE-derived numerators would create a systematic bias; computing both from the same dataset ensures the risk transfer pool nets to zero, as it should.
Note the metal level averages are slightly different than the chart from the top of this post / from the prior post; this is because this chart includes the impact of HCRP while the other did not.
This is the key limitation and where this analysis may not generalize: In 2023, California made up 60% of all platinum membership; New York, Florida, Massachusetts, and Hawaii combined make up 24%, and no other state makes up more than 3% of the remainder. We did try the very thing I complained about CMS doing in our comment letter: excluding members with derived claims (which would theoretically remove a lot of members from California, since capitation is commonplace there). It was a huge amount of membership (almost half), but the conclusion wasn’t really meaningfully different, so we abandoned that approach. Excluding members with derived claims has the problem that by definition, anyone who had no claims isn’t getting excluded, so you’re only excluding the higher acuity folks who may be under capitation arrangements, which skews results in problematic ways. Not a big deal for a Substack post; however, if you, for example, created a risk adjustment model for a market serving over 20 million people but excluded all members with derived claims in calibrating the model, you might end up creating some very problematic incentives. Hopefully, no one has done that!
I’ll also note: the high cost risk pool barely matters here. We modeled the HCRP and the effect on the bottom decile was less than $2,000 per member-year even for platinum. The catastrophic tail that drives platinum volatility isn’t the >$1M tail; it’s the $50K–$500K tail, where members are expensive enough to generate six-figure losses but not expensive enough to trigger the HCRP. State based reinsurance (that we couldn’t model from this data set) fixes that problem to some degree and in many cases creates a clear arbitrage scenario for health plans, but surprisingly there hasn’t been a meaningful increase in health plans offering platinum plans in states with reinsurance offerings. Based on a cursory look at the data, from what I can tell, of the 1332 reinsurance states, only Maine and Georgia have issuers offering platinum plans.