1 of 7
Each mark is one study
The map sorts studies by HR practice (rows) and by what was measured (columns).
Pay and performance management has the most evidence, with studies. Bundles of HR practices, which many HR strategies rely on, have only .
What recent studies say about HR practices and business performance
Studies published from 1 January 2025 to 18 October 2026 that tested an HR practice with an experiment or a quasi-experiment.
Studies that measured business output
These are counts of studies, not sizes of effects. Each study tests a different practice in a different setting, and published studies tend to lean positive. How to read this
The findings, step by step
Scroll to move through seven steps. The map stays in view and changes with each step. Click any mark to open that study.
2 of 7
Only studies measured business output
The first column holds studies that measured productivity, sales, quality, innovation or safety.
The rest measured turnover or absence (), rated or self-reported performance (), or settings further from everyday HR, such as AI tools, public services and short lab tasks ().
Lower turnover is not proof of higher profit, so the columns stay separate.
3 of 7
Business output: what the studies found
A small gain and a large gain each count once. Only of the studies report an effect size we could extract.
4 of 7
Some sensible-sounding practices made things worse
- Public performance rankings lowered productivity in a sweater factory. Private rankings had no average effect.
- Surprise bonuses lowered output in short data-entry jobs.
- Adding explicit “potential” ratings to promotion decisions led to weaker promoted employees.
5 of 7
How a practice is run decides the result
- Mentoring raised sales when it was mandatory. A voluntary version did not, because the people who needed it most did not sign up.
- An AI rollout raised sales when it was fitted to how salespeople work, and lowered them when it was not.
- Letting store managers hire raised average sales, but some stores lost.
6 of 7
How solid is each result?
7 of 7
Could your organization do this?
Most results come from one firm, one country or one period. Before you rely on one, check that its setting resembles yours, or test the practice on a small scale first.
Search all studies ↓The table below the walkthrough gives the same counts as text.
Show the map as a table
All studies
Filter by what was measured, the practice or the result. Each entry gives the finding, its main limitation and a link to the paper.
No studies match these filters
Remove a filter or try a related word. No match here does not mean no such research exists; this list covers selected journals only.
Method and limits
How to read this
- Counts are not effects. Each study counts once, whatever its size or quality. There is no pooled “HR effect”, because the studies measure different things.
- The result codes are a first draft. An AI assistant assigned them from each study's reported finding, mostly taken from abstracts, and no person has reviewed them yet. Every study note shows the finding the code rests on, and the reasoning where the call was not obvious.
- Most checks were light.
- Published results lean positive. Studies that find nothing are less likely to be written up and published.
- This is not the whole literature. It covers 49 selected journals, screened by a single reviewer. It is a scoping review, not a systematic one.
Checked through 18 October 2026. Download all data (JSON)
How the studies were found
We pulled 9,600 metadata records from Crossref for 49 selected journals in industrial-organizational psychology, HR, management and economics, published from 1 January 2025. After removing non-research items, 9,229 remained. Title filters and targeted searches for causal designs narrowed these to 107 candidates, each assessed individually; 68 were kept.
The first online publication date decides whether a study falls in the window. Where only an issue date was available, the study note says so. One further candidate is listed separately because its original publication date could not be confirmed.
We did not use licensed databases such as PsycINFO, Scopus, Web of Science or EconLit.
What the outcome columns mean
- Business output
- Measured productivity, sales, innovation, quality, safety or time use. Broader than profit: only a few studies measure profit directly.
- Turnover, absence, hiring
- Retention, absence, recruitment, labor supply and workforce mix. Useful, but not by themselves proof of higher productivity or profit.
- Rated or self-reported performance
- Supervisor ratings, self-reports, creativity and learning measures.
- Related settings
- AI tools at work, public-sector staff, entrepreneur training and short lab or game tasks. Relevant, but further from everyday HR in private firms.
How the result codes were assigned
Each study gets one code for the direction of its main business outcome, using the study's own reported finding:
- Improved
- At least one tested version improved the outcome, and none worsened it.
- Mixed or depends
- Both gains and losses were reported: across versions, groups of people, settings, or between two business outcomes.
- Worsened
- At least one tested version worsened the outcome, and none improved it.
- No detectable effect
- No tested version changed the outcome detectably. This does not prove the effect is zero.
- Unclear
- The source we checked does not report a direction.
The codes were assigned by an AI assistant and have not yet been reviewed by a person. The full rules and every judgment call are in the codebook.
What each design can and cannot show
Randomized experiments show the effect of assigning the practice in that setting. They do not show that the effect carries over to other organizations.
Natural and quasi-experiments (difference-in-differences, thresholds, staggered rollouts) need the comparison group to have followed the same trend as the treated group would have without the practice.
Experiments whose randomization we could not confirm are labelled as such. We did not assume random assignment from the word “experiment”.
Observational studies with adjustment (fixed effects, matching, instruments, selection models) depend on assumptions that the data cannot fully test.
Journals covered
Chosen for prominence and relevance; no impact-factor cut-off. A journal with no study listed may still have published relevant work.
Excluded and borderline studies
Near misses, date exclusions and candidates that need a closer look at their methods. “Methods check needed” does not mean the study lacks a causal design.
Known gaps
Not yet verified: DOIs and online dates have not been checked against Crossref since the original collection; a one-off check run on GitHub will do this. Pre-registration has not been checked in trial registries: the 3 studies that state it in their design are marked, and the rest are marked “not checked”, not “no”.
A single edition: this map covers studies first published online from 1 January 2025 to 18 October 2026 and will not be updated. Studies published later are not included.
What would make it more complete: human review of the result codes, a second independent coder, full-text checks of methods for every study, licensed database searches, citation chasing and a formal risk-of-bias assessment.