Half the menu is a trap.

At roughly half of the American companies that offer employees a choice of health plans, the menu contains an option that arithmetic alone can disqualify. The plan charges more in certain premium than it could ever return in its best possible year, so no state of your health, perfect or catastrophic, can rescue the purchase. Economists call these options dominated. In the best data anyone has collected, 61 percent of employees picked one, the lowest-paid picked one most often, and the money they overpaid did not leave the building.

I will concede up front what this sounds like. It sounds like a lecture about reading the fine print, delivered to people with better things to do, by an industry that wrote the fine print. And it sounds like the standard pitch for herding everyone into high-deductible plans, which carries real baggage. It is neither. The finding is narrower and stranger: the menu itself is often broken, in a precise and checkable way, and the error it invites runs in one direction only. People pay extra to feel covered. At one step of the menu we will examine, the feeling cost $528 a year and could return at most $250. This paper walks through where that happens, how often, to whom, and where the difference goes.

A plan that cannot win

Start with the cleanest natural experiment anyone has run on this question. In 2010, a large US firm let its employees assemble their own coverage from a menu of 48 combinations of deductible, out-of-pocket maximum, coinsurance, and copay. Every combination carried the same insurer, the same doctors, the same covered services. Only the cost sharing and the price varied. Bhargava, Loewenstein, and Sydnor obtained the records for the 23,894 employees choosing single coverage, and the paper they published, Choose to Lose, is now the reference point for everyone who studies benefits menus.

The menu had a flaw you could find with a pencil. Cutting your deductible from $1,000 to $750, holding everything else fixed, cost an average of $528 a year in extra premium. Run the only scenarios that exist. If you stay healthy, the lower deductible returns nothing, and you are out $528. If catastrophe hits and you blow through the deductible, the move returns its absolute maximum, $250, and you are still out $278. There is no third scenario. A plan priced this way cannot win, for the sick or for the healthy, and that is the entire definition of dominated. Not a judgment call. Subtraction. Of the 36 lower-deductible plans on this menu, 35 failed the subtraction.

Figure 2 is the whole mechanism in one transaction. The premium is certain and paid in advance; the deductible relief is contingent and capped. Whenever the certain payment exceeds the contingent cap, the trade loses in every state of the world. Hold on to that handle, because the rest of this paper hangs on it: the total cost of a plan is its premium plus your out-of-pocket spending, and the premium is the only part you know for certain in January. At this firm, the premium alone settled the contest. 61 percent of employees chose a dominated plan. The strictest reading adjusts for the tax break on premiums, which were paid pre-tax; dominance survives the adjustment, at 55 percent. By either count, the losing plan was the popular plan.

The cost of the error was not decorative. The average employee in a dominated plan could have saved $372 a year by switching to the otherwise identical high-deductible plan: 24 percent of the premium they chose to pay, half the premium of the plan they should have chosen, and about 2 percent of annual salary. The authors re-ran the result under risk adjustment, assuming employees might value protection against bad years, and the savings stayed positive at every level of risk aversion they tested, because a dominated plan's bad years are worse too.

Not one strange company

For years the polite reading of Choose to Lose was that one firm had built one weird menu. Then Liu and Sydnor audited the country. They took the national KFF survey of employer health benefits, pulled the 331 firm menus from 2015 that paired a high-deductible plan with a lower-deductible alternative, and ran the same subtraction on every pair. The first result reverses the standard intuition about safety: at 62 percent of those firms, the high-deductible plan has the lower maximum possible cost. The worst year on the "risky" plan is cheaper than the worst year on the "safe" one, once premiums and employer account contributions are counted, and that share held nearly constant in every survey year from 2011 through 2016.

Then the dominance counts. At 37 percent of firms, the high-deductible plan strictly dominates: cheaper at every possible level of medical spending, not only the extremes. Add the menus where it wins for any chooser who dislikes risk, under any of the spending distributions the authors tested, and the share reaches 53 percent. Roughly half. The money at stake is not loose change, either. Across all 331 paired menus, the expected savings from taking the high-deductible option averaged $569 a year, with a quarter of firms above $884. Where the dominance was strict, the average reached $1,121. For scale: the average worker now pays $6,850 a year toward a $26,993 family premium, so $569 is one month of that paycheck deduction, recoverable by reading a table correctly.

And the table goes unread. The KFF survey reports enrollment shares, and even at the firms where the high-deductible plan strictly dominates, only 47 percent of employees on average enroll in it. The majority sits in the plan that cannot win, which is exactly what the single-firm studies predicted. Handel documented the same pattern at another large employer; Choose to Lose measured it head by head. Three independent datasets, one answer.

Nobody had to design this on purpose, which is the unsettling part. Lower-deductible plans attract the people who expect to use them, so their average cost runs high, and employers tend to pass each plan's average cost through to the paycheck. Liu and Sydnor found the pricing patterns consistent with exactly that: total premiums differ by more than the coverage gap can explain, while firms contribute roughly the same dollars toward both plan types. The selection spread keeps climbing with medical costs; the deductible gap it buys is fixed in the plan documents. At some point the certain payment crosses the contingent cap, and a normal-looking menu quietly becomes a trap. By the only national audit on record, at half of firms it has crossed.

The error runs down the pay scale

Back inside the firm with the 48 plans, the losses were not spread evenly. Among employees earning under $40,000, two thirds of the workforce, 63 percent chose a dominated plan. Among their better-paid colleagues, 38 percent did. The gradient survives controls: holding age, sex, tenure, and chronic conditions fixed, employees in the lowest three salary bands were 24 to 30 percentage points more likely to err than colleagues earning over $100,000. Read that with the money attached. The workers least able to absorb a $372 mistake made it most often, and for them the mistake was about 2 percent of salary. Two percent of salary is the raise you would notice, surrendered every year, for misreading a table no one ever taught you to read.

Time did not heal it. The next enrollment season, 23 percent of employees switched plans, the efficiency gains were modest, and the switching itself was regressive: lower-paid employees were less likely to switch at all, and less likely to land on the high-deductible winner when they did. The same gradient reappeared when the authors rebuilt the choice as an online experiment with thousands of subjects: 75 percent of participants earning under $30,000 chose a dominated plan, against 59 percent of those earning over $120,000. Wherever this menu shows up, the people with the least money pay the most for misreading it.

It was never the size of the menu

The reflex explanation is that 48 plans is an absurd menu, and it is. The data killed the explanation anyway. The authors rebuilt the decision as a clean table: four plans, identical in everything but deductible and premium, both stated in plain annual dollars, prices mirroring the firm's. Faced with that table, 66 percent of experimental subjects still chose a dominated plan, a rate slightly worse than the employees who faced all 48. Shrinking the menu fixed nothing, which means menu size was never the disease. Every vendor selling "simplified enrollment" as a checkbox should have to reckon with that number.

What worked was translation. In a second experiment, half the subjects saw the same four plans with one addition: each plan's total annual cost, spelled out for a healthy year and for a bad year. Dominated choices fell from 48 percent to 18 percent. And measured comprehension explained who needed the help. Among subjects who passed all three of the study's insurance-literacy checks, 1.4 percent chose a dominated plan. Among those who failed all three, 86.6 percent. The choice was never a preference. It was a reading test, and the table was written in the wrong language.

The national reading scores say the experiment generalizes. In KFF's 2014 survey, 4 percent of American adults answered all ten basic insurance questions correctly. About seven in ten could define a deductible, but only half could compute the bill for a four-day hospital stay from a deductible and a copay, 33 percent knew what a formulary is, and 16 percent could work the coinsurance arithmetic on an out-of-network lab test. Loewenstein and colleagues, surveying only people who already had insurance, found that 14 percent understood all four of deductible, copay, coinsurance, and out-of-pocket maximum, and that 11 percent could price a four-day hospital stay under their own plan's rules. The comparison table, the one instrument nearly everyone uses to choose insurance, is written in a language roughly one reader in seven can parse, and the cost of illiteracy is 24 percent of your premium.

The overpayment stays in the building

Here is the question almost nobody asks: where does the $372 go? It does not vanish. Premium contributions flow into the plan, and a dominated choice has a precise accounting signature: the extra money paid in exceeds, by construction, the most the plan could ever pay back out. The residue stays with whoever bears the plan's costs. For the 67 percent of covered workers in self-funded plans, that is the employer, so the overpayment from the shop floor quietly offsets the sponsor's own health spend. In fully insured plans it accrues to the carrier. Either way it is a transfer running up the pay scale, from the workers erring at 63 percent to the institution that printed the menu. If a company deducted 2 percent from the salaries of its lowest-paid workers and called it a fee, lawyers would notice. Routed through a benefits menu, the same money moves with no line item, no disclosure, and no agency counting it, and the worker who paid it believes they bought protection.

This does not require a villain, and I want to be careful not to invent one. Carriers price selection because they have to. Employers pass average costs through because that is what the renewal spreadsheet says; the QJE authors found the firm's prices consistent with ordinary average-cost pricing, not a harvest. But notice that nobody in the chain is paid to run the subtraction, and one party is arguably paid not to: broker commissions on fully insured plans run a median of $178 per enrollee and rise with the premium of the plan sold, so the one advisor in the transaction earns more when the expensive plan wins. Every error in this market points the same direction, toward overpaying for the feeling of coverage. The margin does not need a plan. It only needs a misunderstanding that always leans one way.

Three honest objections, and what would change our mind

The first objection is peace of mind. A low deductible is what risk aversion looks like, and buying calm at a negative expected value is the definition of insurance, not a mistake. For most insurance I agree. Dominance is the one case where the defense fails on its own terms, because the worst case is worse too. On the menu in Figure 1, the catastrophic year on the "safest" plan costs $2,300; on the plan everyone avoided, $1,800. You are not buying protection against the bad outcome. You are paying for a feeling the arithmetic contradicts, and Liu and Sydnor's risk-adjusted runs show the same thing nationally: at the strict-dominance firms, more risk aversion makes the high-deductible plan more attractive, not less.

The honest core of the objection is cash flow, not risk. A deductible arrives as a lump, often in January, while a premium leaves as a paycheck trickle, and for a household with no buffer the trickle can be survivable where the lump is not. I take that seriously, and the endnote takes it further. But look at the trade on the Figure 1 menu: $1,150 a year in extra premium, paid for certain, to escape the possibility of an extra $650 lump. As liquidity insurance, that is the most expensive product in the building. And the evidence says the buyers were not making that calculation: dominated choices tracked measured insurance illiteracy, not stated preferences about deductibles, and they collapsed the moment the costs were spelled out.

The second objection comes from economics itself, and it cuts against the obvious fix. Suppose everyone suddenly chose correctly. Handel studied a firm where switching frictions kept employees parked in stale plans, behaving as if re-choosing cost them $2,032 a year, and found that removing the friction made individual choices better and the market worse: healthy people re-sorted into cheap plans, adverse selection deepened, and the welfare loss roughly doubled. Inertia was accidentally holding the risk pool together. So "educate everyone and let them sort" is not free. What that argues for is fixing the menu rather than only the chooser. A dominated plan improves no risk pool; its buyers overpay in every state of the world, sick and healthy alike. Deleting it, or pricing it honestly, upgrades every employee's choice set without re-sorting a single sick person.

The third objection is that people will learn. The best longitudinal test on record says they do not: in Medicare Part D, Abaluck and Gruber found foregone savings grew rather than shrank across five years of experience, and the QJE firm's own second-year switching barely dented the error. The Part D welfare framing is contested, so we lean only on its direction. Here is what would change our mind: a national re-audit of paired menus showing dominance receded after the 2011 to 2016 data Liu and Sydnor used, and we know of no such audit; evidence that liquidity-constrained low-deductible choosers do better in realized outcomes than the arithmetic predicts; or a field result where experience alone, without translation, eliminated dominated choice. Until one arrives, the finding stands. The most common error in American benefits is buying insurance against a risk the cheaper plan already covers better.