The Plateau Is in the Tail, Not the Mean

In 2010 Daniel Kahneman and Angus Deaton published the most quotable regression in the history of happiness research. Working from Gallup’s daily tracking poll of Americans — the same survey file analyzed in this article — they reported that day-to-day emotional well-being rose with household income and then stopped improving at roughly $75,000 a year. The number escaped the journal instantly. It became a personal-finance proverb, a salary-negotiation talking point, and the rare economics result that fits on a bumper sticker: past seventy-five grand, money can’t buy happiness.

It also started a decade-long fight. Betsey Stevenson and Justin Wolfers, surveying every dataset they could find, reported no satiation point anywhere in evaluative well-being — people’s judgments of their lives kept climbing with the logarithm of income, with no kink at $75,000 or anywhere else. In 2021 Matthew Killingsworth, using smartphone experience sampling, found no plateau even in moment-to-moment feelings. And in 2023 Killingsworth, Kahneman, and Barbara Mellers published an unusual adversarial collaboration that ended the argument by dissolving it: both sides were describing the same data. The flattening is real, but it belongs only to the unhappiest fraction of people. For everyone else, well-being keeps rising with income — for the happiest, it rises faster at the top. The famous mean curve everyone had been arguing about was an average of a bending bottom and a rising top.

That resolution was reached on an experience-sampling panel of some thirty thousand people. The original battlefield — Gallup’s daily poll — allows the same question to be asked of more than fifty times as many, with both kinds of well-being measured in the same interviews. This article re-stages the resolution there: 2,587,175 interviews from 2009 through 2017 with a usable household-income answer, of which 1,808,814 rated their lives on the 0–10 ladder. Instead of plotting the average of well-being at each income, we plot its distribution. (One caution applies throughout: this extract of the Gallup file carries no survey weights, so every number here is an unweighted sample estimate — good for shapes and contrasts, not for precise population levels.)

Two different questions about a life

The fight was always partly about constructs, so let us be explicit. Gallup asks both of the questions that the literature keeps conflating. The first is evaluative: the Cantril ladder, “on which step of a 0-to-10 ladder do you stand, where 10 is the best possible life for you?” It invites a judgment, an audit of one’s life from the outside. The second is experiential: did you experience enjoyment, happiness, stress, worry, sadness “during a lot of the day yesterday” — yes or no. It samples feelings from the inside. Kahneman and Deaton’s $75,000 plateau was a claim about the second construct; their own paper found that the first, the ladder, rose with log income without satiating. Much of the subsequent decade of argument consisted of people checking one construct against headlines written about the other.

Household income comes in six harmonized brackets, from under $24,000 to $90,000-and-up. Because the bracket frame changed between 2008 and 2009, we pool the years 2009–2017. To put brackets on a dollar axis we use midpoints, with the open top bracket assigned $135,000; the x-axis of every chart below is logarithmic, so “slope” always means change per doubling of income.

The curve everyone fights about

Start where the argument always starts: the mean. Average ladder rises from 6.278 in the bottom bracket to 7.602 in the top one — 1.324 rungs across roughly three and a half doublings of income, a gradient of 0.3881 rungs per doubling. And it is strikingly straight in the logarithm. Fit the slope over the bottom four brackets and you get 0.3798 rungs per doubling; the final step, from the $60–90k bracket to $90k+, runs at 0.359 rungs per doubling — about 95 percent as steep. On this construct, at these incomes, there is no plateau in the mean. Stevenson and Wolfers would recognize the picture immediately.

But the mean is a summary of 1.8 million people who do not have to resemble each other. Press the button under the chart and the average splits into the curves it was averaging.

The fan is the argument. At the 10th percentile of well-being — the unhappiest tenth at each income — the ladder climbs from 3.123 in the bottom bracket to 5.605 at the top: 0.7103 rungs per doubling, nearly twice the gradient of the mean. At the 25th percentile the slope is 0.6307; at the median, 0.3819; at the 75th, 0.1283. And at the 90th percentile the curve is simply pinned: 9.522 at the bottom bracket, 9.451 at the top, slope −0.0243 — statistically a flat line lying against the ceiling of the scale. The happiest tenth of poor Americans already rate their lives 9 or above; there is nowhere on a 0–10 ladder for money to take them.

So the income gradient is not a property of “happiness” in general. It is a property of the lower tail. Income predicts how bad the bad lives are at each rung of the economy far better than it predicts how good the good ones are. The mean curve everyone fights about is real, but it is an artifact of aggregation: a steep bottom and a flat top, averaged.

Counting heads instead of slicing percentiles

Percentiles of an 11-point scale are chunky objects — the raw 90th percentile is literally the integer 9 in five of the six brackets — which is why the curves above use a continuity correction and why the cleaner way to see the tails is to count heads on either side of fixed thresholds. Call a rating of 4 or below a suffering ladder score and 8 or above a thriving one, and ask what share of each income bracket sits past each threshold.

The two curves tell opposite stories. The share at 4 or below collapses from 19.43 percent in the bottom bracket to 3.388 percent at $90k+ — misery falls by more than four-fifths across the income range. But the collapse is front-loaded. Over the first step it falls 5.632 points per doubling; by the final step, from $60–90k to $90k+, the pace has slowed to 1.850 points per doubling — 32 percent of the bottom-half gradient. By the $48–60k bracket the share is already down to 6.775 percent: roughly two-thirds of the misery visible in the bottom bracket is gone before the famous $75,000 is reached. Part of this deceleration is arithmetic necessity — a share approaching zero has to flatten — but that is precisely the substantive point: by upper-middle incomes there is very little measured suffering left for money to remove.

The thriving share does the opposite. It rises from 34.39 percent to 59.07 percent, and it rises fastest at the top: 3.214 points per doubling over the first step, 10.82 over the last — the largest single jump in the whole curve is the final one, from 49.90 percent at $60–90k to 59.07 percent at $90k+. The per-doubling gradient at the top is 2.133 times the bottom-half slope. Whatever is satiating in these data, it is not the propensity of richer Americans to call their lives excellent.

This is the adversarial collaboration’s resolution, re-drawn with 1.8 million respondents: a flattening that lives in the lower tail of the well-being distribution, and an acceleration that lives in the upper tail. Killingsworth, Kahneman, and Mellers located the unhappy minority’s plateau above roughly $100,000; our top bracket starts at $90,000 and is open-ended, so the data here cannot resolve the exact altitude of the bend. What they show, on the original survey, is the shape the resolution requires: by the upper brackets, a doubling of income buys far less relief at the bottom of the well-being distribution than it did lower down — and more abundance at the top than it ever did.

The plateau that replicates: feelings

Kahneman and Deaton’s headline claim was never about the ladder — it was about emotional experience. The same interviews carry five binary yesterday-affect items, so the experiential construct can be checked directly, against the evaluative fan, in the same people.

Here the satiation is unmistakable. Worry is the cleanest case: 40.73 percent of the bottom bracket worried a lot of yesterday, against 25.49 percent at the top — but the first income step removes 9.054 points of worry and the last step removes 0.2965, barely a thirtieth as much. Per doubling, the top-step gradient is 5 percent of the bottom-half gradient. Enjoyment climbs from 76.71 to 89.91 percent but does almost all of that climbing early; its final-step slope is 23 percent of its bottom-half slope, and happiness-yesterday (80.67 to 92.45 percent) is at 19 percent. Sadness keeps falling the longest — 28.92 down to 10.83 percent — yet its top step too runs at less than a quarter of its early pace. And stress does something the others don’t: it falls from 43.82 percent to a minimum of 35.30 percent in the $48–60k bracket, then rises again to 37.77 percent at $90k+. Money buys off worry and sadness; it does not buy off stress, which the highest earners report more often than the middle class does.

So the $75,000 legend was not wrong about its own construct. In these data, collected by the same poll that produced it, day-to-day feeling really does stop improving much past the middle of the income distribution — the experiential curves are concave to the point of flatness exactly where Kahneman and Deaton said they were. The error was in the retelling: a fact about yesterday’s emotions, true mainly of the unhappier end of the distribution, was promoted into a fact about how people judge their entire lives. The judgment curve never flattened at all.

Is it just composition?

A fan like Fig. 1 invites a deflationary reading: maybe the steep bottom tail is not about money but about who is poor — the bottom bracket holds more of the old without degrees, the young, the retired, groups with their own well-being profiles. The check is to re-draw the threshold curves inside demographic cells, holding age band and education fixed.

The pattern survives everywhere we can cut it. Twenty-two age × education cells are large enough to demand at least 1,000 ladder ratings in every income bracket; the six largest, shown above, range from 112,996 to 176,187 respondents. In all 22 cells the suffering share falls with income and the thriving share rises. In 21 of 22, the suffering curve flattens at the top exactly as it does in aggregate (top-step gradient smaller than the bottom-half gradient); in 18 of 22, the thriving curve is steeper at the top than at the bottom. College-educated 55-to-65-year-olds — the single biggest cell — lose 6.608 points of suffering share per doubling of income overall, with a top step running at 22 percent of the early pace, while their thriving share rises at 10.45 points per doubling. The scissors are not a composition artifact; they open inside every large demographic cell in the file.

What the average was hiding

Three findings, then, from returning to the original battlefield with the distribution in view. First, the mean evaluative curve — the object of a decade of argument — has no plateau in these data: 0.38 rungs per doubling in the bottom half of the income range, 0.36 in the last step. Second, the income gradient is wildly unequal across the well-being distribution: 0.71 rungs per doubling for the unhappiest decile, nothing for the happiest, whose ratings sit against the top of the scale at every income. The flattening that exists is concentrated where the unhappy are — the suffering share’s decline slows to a third of its early pace by the top brackets, while the thriving share accelerates. Third, the construct matters as much as the percentile: the experiential items satiate dramatically — worry’s gradient is 95 percent exhausted past $60,000 — even as the evaluative top tail keeps growing.

The usual caveats are load-bearing and worth stating plainly. Income here is a six-bracket household measure, not equivalized for household size, with an open top bracket; dollar positions are midpoint conventions. The data are a repeated cross-section — these are gradients across people, not the causal effect of giving anyone a raise, and the well-off differ from the poor in countless unmeasured ways. The ladder is an 11-point discrete scale, which is why tails are framed as threshold shares and why a flat 90th percentile partly reflects the scale’s ceiling, just as the suffering share’s late flatness partly reflects its floor. And nothing is weighted, because the file ships no weights: shapes and contrasts, not levels, are the deliverable.

Within those limits, the lesson of the adversarial collaboration scales up cleanly from a thirty-thousand-person smartphone study to nearly two million interviews. “Does money buy happiness?” was never one question. It is a different question at every percentile of well-being and for every construct of it — and the famous answer, the one on the bumper sticker, was true mainly of the people having the worst time, feeling their way through yesterday. For everyone else, the ladder just keeps going up.