# Pre-registered plan: does guaranteed income improve recipients' mental health, and does it reduce work?

## Question

Since 2017, more than a hundred American cities and counties, and governments in Finland, Germany, and Ireland, have tested guaranteed income: regular cash with no conditions, paid for a year or more to people with low incomes. Its advocates expect it to relieve stress and improve mental health without discouraging work; its critics expect less work and little lasting gain. Many of these pilots were randomized. Most report alone, in reports their evaluators publish outside journals, and most are small.

The largest, OpenResearch's trial in Illinois and Texas (1,000 people paid $1,000 a month for three years), found that recipients worked somewhat less and that improvements in stress and mental health were large in the first year and gone after it (NBER working papers 32719 and 32711, 2024). A Campbell systematic review (Campbell Systematic Reviews 2024, doi:10.1002/cl2.1414; searched to December 2022) found 10 experiments, most of them older or with payments that fell as earnings rose, and pooled self-rated mental distress across a few of them. A working paper from the American Enterprise Institute (Corinth and Mayhew, February 2026) averaged the employment effects of 30 randomized American pilots without weighting them, giving a mean of +0.8 percentage points and −3.2 among the four largest, with no pooled model, interval, or mental-health outcomes. No synthesis has pooled the trials' mental-health results, weighted their employment results by their precision, or asked whether the mental-health benefit lasts while the payments go on.

**Questions.**

1. **P1.** While payments are being made, about a year in, does unconditional cash improve the mental health of the people who receive it?
2. **P2.** Does it change whether they work?
3. **P3.** Does the mental-health effect fade while payments continue?
4. **Secondary.** How the mental-health effect changes over the time payments last and after they end; perceived stress, life satisfaction, full-time work, and hours; one-time lump sums; whether larger payments do more; whether smaller trials report larger effects; and how many registered trials never reported.

## What was known before this plan

Before writing this plan the author read the abstracts of the Campbell review and of the two OpenResearch papers named above (so OpenResearch's direction of effect on work and mental health is known), the AEI working paper with its table of each pilot's employment effect (so the direction of most American pilots' employment effects is known), and, to see what the Center for Guaranteed Income Research's reports print, the appendix table of impacts of one of them (Cambridge, Massachusetts, whose difference in psychological distress at 12 months was small and not significant). To code designs, the design coders read whole reports, results included; the author has read their design codes and notes, which mention results only in passing (for example that one trial's brief found no apparent effect on employment). No number has been extracted for this review. This plan fixes, before extraction: which trials and comparisons count, which estimate each contributes, how estimates are combined, which sensitivity analyses run, and how each result will be read.

## Data

**Reports.** PubMed (E-utilities) and Europe PMC (titles and abstracts, which adds preprints), searched on 2026-10-17 for guaranteed or basic income, unconditional cash, cash transfers, and similar terms, with mental health, well-being, work, or health, and randomization (queries in `data/search.json`): 529 PubMed records and 504 Europe PMC records, 582 distinct. OpenAlex, searched the same day for the same cash and design terms with places in high-income countries, published from 2017, when the current wave of flat guaranteed-income trials began, which adds working papers, reports, and books the medical databases miss: 866 records, 736 not already found.

**Registrations.** ClinicalTrials.gov, searched the same day for interventional studies of the same cash terms (104), and the AEA RCT Registry's full export (12,932 registrations), from which 337 name any of the cash terms and a high-income country or no country: 441 registrations in all.

**Grey literature.** The Center for Guaranteed Income Research's publications page (22 reports in English), the Guaranteed Income Pilots Dashboard's pilot pages (one report beyond the Center's), the documents the AEI working paper cites for its randomized pilots (18 retrieved), the Baby's First Years trial's publications page (its own copies of six papers), and the Stanford Basic Income Lab's database of basic income experiments as its public map served it (280 rows, 91 marked as randomized controlled trials): for each randomized pilot not already covered, 41 in all, one or two web searches for an evaluation report found documents for 18. For each of the 17 eligible registrations due to have reported with no report found, the same kind of search for its results found them for 10 (`data/screening.json`).

## Screening

Two coders, separate instances of Claude Opus 5.5 (the model that wrote this study), each working alone from `screening_codebook.md` with only titles and abstracts, coded every record (`trial_report`, `protocol`, `review`, `exclude`) and every registration (`eligible`, `exclude`). On the 582 PubMed and Europe PMC records they agreed on keeping 578 (κ = 0.96), on the 736 OpenAlex records 720 (κ = 0.91), and on the 441 registrations 439 (κ = 0.98). A third instance, given the codebook, the item, and both decisions, decided the 37 items on which the coders' codes differed. 53 PubMed and Europe PMC records and 99 OpenAlex records were kept as trial reports or protocols, and 68 registrations as eligible (`data/records.json`, `data/screening.json`). Each review kept was read for trials the searches missed.

## Trials and their designs

Reports, registrations, and grey documents were grouped into trials by registration number, trial name, place, and links in their text (`data/groups.json`, each grouping with its reason): 107 trials. Nine kept records belong to no trial (a natural experiment, commentaries, a design proposal, books, and a pre-analysis plan for pooling three trials' data, with reasons). The included studies of the reviews found (the Campbell review and four reviews of income support and health) were checked against this list; they named no randomized trial of flat unconditional cash missing from it (`data/reviews_checked.json`). Of the 107 trials, 25 are known only by a registration.

Two further coders, working alone from `design_codebook.md`, coded the design of each of the other 82 trials from all its documents: eligibility (randomized; unconditional; flat, not falling as earnings rise; a control arm with no cash or a token amount; a high-income country; mental health or employment among its outcomes), the arms with their amounts and sizes, each comparison of a cash arm with a control arm (recurring payments or a lump sum), when payments began and how long they lasted, and each wave of follow-up with its months since the first payment and whether payments were still being made. Nine trials whose documents grew after a later search (results of registered trials) were coded again from all of them. The coders agreed on every field the analysis uses for 50 trials; a third reader decided the other 32 against the documents (most differences were interim waves one coder listed and the other didn't, or arm sizes), with reasons (`coding/design/`).

Sixty-two trials are eligible, with 69 comparisons (59 of recurring payments, 10 of lump sums); 53 of them have reports with results, 48 from the United States and one each from Finland, Germany, Ireland, Canada, and Italy, and 9 have no results yet. Twenty aren't eligible: no control arm (5: recipients chosen at random but no one followed as a control), payments that fell with income (5: the negative income tax experiments, Mincome, Ontario's pilot, B-MINCOME, the Dutch social-assistance experiments, and PHLHousing+'s rental assistance), not randomized (4), no outcome of mental health or employment reported or planned (5), and payments tied to a spending plan (1). `data/trials.json` holds each trial's decided design.

Where a trial's cash arm also got something its control didn't (TAYportunity's optional job training), the comparison stands as the trial made it, and the paper says so. Where two cash arms share one control (Compton's twice-monthly and quarterly payments, Denver's two schedules, In Her Hands' two schedules, Newark's recurring and lump-sum payments), each is a comparison of its own.

**Precision.** At the wave closest to 12 months, up to 43 comparisons of recurring payments in 40 trials report mental health, with about 36,000 people randomized (not counting the Finnish experiment's control group, drawn from registers), and up to 41 comparisons in 38 trials report employment. If half the people randomized are analysed and the effects are homogeneous, a pooled standardized mean difference of 0.05 would be detected with power above 90% at two-sided α = 0.05; heterogeneity and Knapp-Hartung intervals lower that.

## Integrity

For each eligible trial, whether it was registered before its payments began (by the month). Of the 53 eligible trials with results, 31 have no registration this review found; of the 22 registered, 11 were registered before payments began, 9 after, and for 2 the reports don't say when payments began. A sensitivity analysis of each primary keeps only trials registered before payments began. No retraction or expression of concern is linked to any included report.

## Outcomes

Numbers are extracted after this plan is registered, by two coders working alone from `extraction_codebook.md`, copying what each source prints in the form printed; a third coder decides every entry on which they differ, against the source. The code does every conversion.

- **Mental health**: one measure per comparison and wave, the first of these classes the trial reports: psychological distress (Kessler 6 or 10), then depressive symptoms (PHQ, CES-D, and the like), then anxiety symptoms (GAD and the like), then a general mental-health scale or index (MHI-5, SF-12 or SF-36 mental component, WHO-5, WEMWBS, or an index the trial built). A comparison keeps the class it reports first at its waves during payments; at a wave without that class, the next one there is used. A scale's score comes before a share above a cut-off on it.
- **Employment**: the share of people in any paid work.
- **Secondary**: perceived stress, life satisfaction, the share employed full-time, and hours worked per week.

## Effect measures

- **Scales** become standardized mean differences (Hedges' g), positive when the cash arm is better off. A difference in a scale's points is divided by the scale's standard deviation: the arms' standard deviations at that wave where printed (pooled), else the baseline standard deviation printed by the trial, else the median of the standard deviations the same trial prints for the same measure at other waves, else the median of those all included trials print for the same instrument, else, as a last resort, the standard deviation implied by the difference's standard error and the wave's sample sizes (Cochrane Handbook 6.5.2.3). A difference's standard error comes from the standard error printed, else its confidence interval, else its exact p value; else from the sample sizes. A share above a cut-off becomes a log odds ratio and then a standardized difference by Chinn's conversion (multiplied by √3/π). An estimate the trial prints in standard-deviation units is used as printed.
- **Employment** becomes a difference in percentage points, cash minus control, positive when more of the cash arm worked, with its standard error from the one printed, else the interval, else the exact p value, else the arms' shares and sizes.
- **Hours** become a difference in hours per week (a month's hours times 12/52).
- **Which estimate.** For each comparison, outcome, and wave: the whole sample before any subgroup; then an estimate adjusted for covariates before an unadjusted one, before one printed in standard-deviation units, before an odds ratio, before one computed from the arms' means or shares; then the trial's first-listed results report.
- **Shared control.** Where two comparisons of one trial share its control arm in one analysis, each comparison's variance is recomputed as if the control's participants were split equally between them (Cochrane Handbook 23.3.4).
- **Waves.** A wave is "during payments" if payments were still being made or had ended less than a month before. For each comparison of recurring payments, P1 and P2 use its wave during payments closest to 12 months after the first payment (the later of two as close) that has an estimate of that outcome: most trials paid for 12 to 24 months, so this compares trials at one horizon. A lump sum has no "during": its latest wave at most 12 months after the payment is used.

## Estimates

All pooled effects are random-effects estimates by restricted maximum likelihood with Knapp–Hartung confidence intervals and tests (`metafor::rma`, `method = "REML"`, `test = "knha"`), with τ², I², and the 95% prediction interval.

- **P1:** the pooled standardized mean difference in mental health at the wave during payments closest to 12 months, over the comparisons of recurring payments.
- **P2:** the pooled difference in the share employed, in percentage points, at the wave during payments closest to 12 months, over the comparisons of recurring payments.
- **P3:** over comparisons of recurring payments with mental-health estimates at two waves during payments at least 6 months apart, the pooled change in the standardized mean difference from the earliest such wave to the latest. The two waves' estimates are treated as independent, which overstates the change's variance when they are positively correlated, as repeated measures of one sample are.
- **Multiplicity:** Holm's correction over the three primary p values.

**Sensitivity analyses** for P1 and P2: each comparison's latest wave during payments instead; a common-effect estimate; a Bayesian random-effects estimate with a half-normal prior on τ (scale 0.2 for standardized differences, 5 percentage points for employment) and a flat prior on the mean; each comparison left out in turn; only comparisons whose wave kept at least 60% of those randomized with arms differing in retention by at most 10 points; only trials registered before payments began; only comparisons with at least 1,000 people analysed; and, for P1, only estimates standardized by a standard deviation the trial itself printed. Egger's regression (the estimate on its standard error) tests whether smaller comparisons report larger effects. For P2, also with the share employed full-time where a trial prints no share employed at all.

## How results will be read

- **P1** says cash improved mental health if its Holm-corrected p < 0.05 with a positive estimate (worsened it, if negative). It is **robust** if, in addition, the common-effect estimate, every leave-one-out estimate, and the estimate restricted by retention each have p < 0.05.
- **P2** says cash changed employment if its Holm-corrected p < 0.05, read the same way; otherwise the paper states the largest reduction and the largest increase its 95% confidence interval allows.
- **P3** says the mental-health effect faded while payments went on if the pooled change is negative with Holm-corrected p < 0.05.
- Every pooled result is reported with its interval whatever its p value, and the paper reports every analysis below, planned or not, and says which is which.

## Secondary analyses

- **Time during payments**: every wave during payments of every recurring comparison, in a model with a random effect for each comparison and for each wave within it, giving the slope of the mental-health effect per year since the first payment; and the pooled effect at waves within the first 6 months, 7 to 12, 13 to 24, and after 24 months (each comparison's latest wave in each band).
- **After payments ended**: the mental-health effect and the employment effect at the first wave after payments ended, at most 12 months after the last payment.
- **Other outcomes during payments**: perceived stress, life satisfaction, the share employed full-time, hours worked per week.
- **Lump sums**: the mental-health and employment effects of one-time payments, pooled on their own.
- **All cash**: recurring payments and lump sums together.
- **Amount**: among recurring payments in US dollars, the meta-regression slope of each primary effect on log₂ of the monthly amount.
- **The file drawer**: of the registrations the screeners found eligible (the third reader deciding where they differed), less those of trials the design coders found ineligible, whose primary completion (or end) date was at least 12 months before the search and that weren't withdrawn or still recruiting, how many reported results this review found and how many didn't.

## Software

R 4.6.1 with metafor 5.2.1 and jsonlite 2.0.0, in the container `env/Dockerfile` builds. `code/analyze.R`, registered with this plan, reads `data/` and writes every result; it was run end to end on made-up data before registration. Once the numbers are extracted, it is first run with `--blind`, which flips the sign of each comparison's effects at random (seed in `data/parameters.json`), to check that every estimate is found and every model fits before the results mean anything. Any change to the code after registration is listed in `deviations.json`.