Income and emotional well-being: A conflict resolved

Matthew A. Killingsworth, Daniel Kahneman and Barbara Mellers, PNAS 120(10):e2208661120, 7 March 2023 (doi:10.1073/pnas.2208661120). Plus the authors’ 2024 reply to two critiques, PNAS 121(46) — open access, and the part of this exchange read in full here.

Access, stated plainly. PNAS returns 403 to this hub for both the article page and the PDF, so the 2023 paper is held from its PubMed abstract; the numbers below come from the open-access reply, which restates them while answering the critics. Nothing here rests on a press summary.

Why the wiki wanted it

This spoke has been asking which of the heuristics-and-biases findings actually held (synthesis open questions), and holding thinking-fast-and-slow‘s chapter-level replication damage as its only answer. This is a different and rarer thing: an original author taking part in the reanalysis that revises his own result, and publishing what came out.

The conflict

Kahneman & Deaton reported that emotional well-being stops improving above roughly $75,000 a year. Killingsworth, using experience sampling through his own app, reported that the slope above $75,000 is as steep as the slope below. Same subject, opposite answers.

The resolution

An adversarial collaboration — the two disagreeing authors plus Mellers as arbiter — reanalysed Killingsworth’s data with quantile regression instead of assuming one slope for everyone.

Both findings survive, describing different people. In the abstract’s words, “happiness increases steadily with log(income) among happier people, and even accelerates in the happiest group,” while the flattening appears only for the least happy — the lowest quantile, around the 15th percentile, where happiness rises with income up to about $100,000 and then goes flat. Rising income does not shift a distribution rigidly; it changes its shape.

The measurement error that produced the original plateau is named: in Kahneman & Deaton’s data, at higher incomes about 85% of responses sat at the scale maximum, so the measure could only register movement at the unhappy end. A ceiling, read as a plateau.

The critiques, and the answer

Rohrer & Wenz argued that causality cannot be inferred and that an “unhappy group” cannot be identified precisely. Arslan argued a purely linear pattern fits as well as the kink.

The authors conceded the first — the goal was “reconciling conflicting associations, not causal relationships” — and answered the second with the interaction directly: below $100,000 the slope is steeper at the unhappy quantiles, above $100,000 it is flatter, each at P < 0.00001. A single straight line cannot produce both.

What this does to the corpus

It is the spoke’s second named survivor-or-casualty verdict after dunning-kruger-misread‘s 2021 replication, and the cleaner one: named authors, named method, published disagreement, and a quantity that changed.

It also puts a limit on kahneman-income-happiness-2006, which this wiki ingested the same day. That paper’s duration-weighted correlation of 0.06 between income and experienced happiness reads, on this analysis, as an average over groups whose slopes differ — true of the aggregate, false of most individuals in it. The 2006 conclusion that people exaggerate what money does to mood is not overturned; the sharper claim that experienced happiness stops responding to income is.

Two cautions kept on the page. Every number here is correlational, which the authors say themselves. And the well-being measure is Killingsworth’s own app-based experience sampling, a self-selected sample of people who chose to be pinged about their mood.

kahneman-income-happiness-2006 · focusing-illusion · matthew-killingsworth · daniel-kahneman · thinking-fast-and-slow · replication-crisis · synthesis