Training (employees)
Agreement with “I receive the training I need to do my job well.” This is perceived training adequacy: how trained people feel, not tested skill. In 2024, 67% answered positively (n = 666,259).
Does getting the training you need keep people from leaving?
Organizations spend heavily on training but rarely know whether it keeps people. Using 15 years of the U.S. Federal Employee Viewpoint Survey, I asked whether employees who feel well trained are less likely to plan to leave, and whether agencies that improve on development see less leaving afterward.
- Data
- OPM FEVS public files, 2010–2024
- Scale
- 1.65M responses · 29-agency panel
- My role
- Designed and directed; AI-assisted analysis
- Tools
- Python · DuckDB · statsmodels · matplotlib
72% vs 23%considering leaving: strongly disagree vs strongly agree they get needed training (2024)
~5 ptsadjusted gap in intent to leave, stable across 2022–2024
0.0change in next-year leaving after an agency’s development score rose (p = 0.81)
29agencies tracked for 15 years, 387 agency-year pairs
01
The question and the data
Training is usually judged by how people feel right after it, and research shows those reactions say little about learning. Data that links training to outcomes is almost always internal to companies. FEVS is the largest public alternative: several hundred thousand federal employees a year, asked about training, development, their supervisors and whether they plan to leave.
Development (agencies)
The training item starts in 2022, too late for a 15-year panel. Agencies are tracked on “I am given a real opportunity to improve my skills,” asked every year. The two move together across agencies (r = 0.71–0.85).
Outcome
“Are you considering leaving your organization within the next year?” Any “yes” counts. This is intent, not actual departures, and every result is an association, not proof of cause.
02
Feeling well trained goes with staying, but less than it first looks
The raw pattern is steep. But people who feel good about their job tend to rate everything higher, so I compared employees in the same agency who also rated their supervisor’s support, recognition, workload and pay alike. The gap shrinks from about 23 points to about 5, and I headline the cautious number.
Share considering leaving, by answer to “I receive the training I need”
Each step up the scale goes with fewer people thinking about leaving. 2022 and 2023 differ from 2024 by at most 2.3 points at any answer.
2024, survey-weighted shares, no controls
The gap, measured two ways (percentage points)
Same agency onlyFully adjusted
2022
2023
2024
Fully adjusted for supervisor support, recognition, workload, pay, tenure, supervisory status and age: −5.2 points in 2024 (95% CI −5.8 to −4.7). The true link likely sits between 5 and 23 points; the defensible claim is the lower one.
Survey-weighted logistic regression, agency fixed effects, average marginal effects; SEs clustered by agency
Who it matters for: adjusted gap by years of service (2024)
Any reason for leavingLeaving for another job
10 or fewer
11–20 years
Over 20
The link is concentrated among staff in their first 20 years. For veterans it nearly vanishes, partly because many are leaving to retire; counting only job-switching, their gap reappears (−3.8).
Where people would go
The gap is largest for moving to another federal job, which suggests internal mobility, not only exit, is at stake. The result barely moves without survey weights, with complete demographics only, or for employees under 40.
Adjusted gap in percentage points, 2024, each reason vs not considering leaving
03
Agencies that improved on development didn’t see less leaving
Agencies with better development climates do have lower leaving (r = −0.45 across 15-year averages). But that could just mean some agencies are better places to work for years at a time. So I compared each agency with itself and asked whether a rise in its score came before a fall in leaving.
Government-wide, 2010–2024
Development score“Training I need” (from 2022)Considering leaving
Both development scores and intent to leave rose over the period. National swings (pay freezes, shutdowns, the pandemic) are why the agency tests remove each year’s government-wide change.
246,846 to 686,499 respondents a year, survey-weighted
Four agency tests, per typical difference in the development score
Next year’s intent to leaveNext year’s Global Satisfaction
Score this year → outcome next year
−1.7+2.9
Same, holding next year’s score constant
−0.3+0.6
Change in score → next year’s change
0.0+0.6
Same, plus each agency’s own trend
0.0+0.4
−2−10+1+2+3
A rise in an agency’s development score was not followed by less leaving (−0.02 per point, 95% CI −0.15 to 0.11, bootstrap p = 0.81), a precise null: at most about 0.3 points either way. It was followed by slightly higher satisfaction (bootstrap p = 0.004), a smaller, same-survey finding.
29 agencies, agency and year fixed effects; bars are 95% intervals; wild cluster bootstrap (9,999 draws)
Reverse direction
High leaving this year predicts a lower development score next year just as strongly (−0.58 vs −0.48), so the year-ahead pattern can’t show which comes first.
Horse race
Once next year’s score is known, this year’s score adds nothing detectable to predicting next year’s leaving (−0.08, CI −0.22 to 0.07).
Changes on changes
An agency’s rise in score doesn’t predict a fall in leaving the next year, with or without each agency’s own long-run trend, and without small agencies.
04
What this means for HR leaders
Employees’ sense of being well trained is a useful warning signal for retention risk, especially among newer staff. On this evidence, raising that score is not a lever that lowers turnover intent on its own.
01
A signal, not a lever
Use perceived training adequacy to flag where retention risk is building, not as a promise that a new program will cut attrition.
02
Target the first 20 years
The link is strongest for newer and mid-career staff, and for people weighing another job in the same sector: internal mobility is at stake.
03
Don’t promise turnover cuts from a survey score
Agencies whose development scores rose saw slightly higher satisfaction, not lower intent to leave.
04
Measure what training builds
Test skill before, right after and 3–6 months after training, roll programs out in stages so later groups act as a comparison, and link training records to real separations, moves and promotions.
My role
- Chose the question, wrote the research proposal and made the calls that shape what the report claims: which survey questions to use, how to treat the 2020 COVID survey year, headlining the conservative estimate, and leading with the null result.
- Directed Claude Code (an AI coding agent), which wrote and ran the data pipeline, the models and the first draft of the report; I reviewed and approved each of four milestones, with every decision and its reason logged.
- The project is fully reproducible: six numbered scripts rebuild everything from OPM’s public files, the report is filled from a template so no number is typed by hand, and the pipeline reproduces OPM’s published 2024 figures for 4 agencies to within 0.5 points.
Methods: Survey-weighted logistic regression · average marginal effects · fixed-effects panels · first differences · wild cluster bootstrap · robustness and reverse-causality checks
Full report with appendixLimits: the predictor is perceived, not measured, training; training and intent come from the same survey and person, so results are associations; intent to leave is not turnover; no person is followed over time; the agency panel uses a development item rather than the training item; these are federal employees, and private-sector patterns may differ. Data: U.S. Office of Personnel Management, Federal Employee Viewpoint Survey public release files, 2010–2024 (2026 re-releases).