Kevin Michael Frick, Jonas Gathen
10 min
This paper by Kevin Michael Frick and Jonas Gathen tackles a tough problem: how do you measure the skills of government workers when their output is hard to observe (e.g., bureaucracy) and wages are often rigid, not reflecting true productivity? Traditional methods fail here—output isn't measurable like in private firms, and wages don't signal skills due to fixed pay scales. The authors propose a clever workaround using private sector wages as benchmarks for comparable jobs, combined with machine learning to predict skills from observables like education, cognitive tests, and personality traits.
Applied to Indonesia's Indonesian Family and Life Survey (IFLS) panel data from 1988-2014, the method reveals striking insights into government hiring and pay. Government workers were initially positively selected on skills but have seen a continuous relative skill decline as top talent shifts to the private sector. Despite this, the government pays a 43% wage premium conditional on skills, highlighting inefficiencies in wage structures.
The core idea: private sector jobs with similar tasks (e.g., social services) have wages that, while noisy, signal productivity due to market competition. The authors extract 'skill fixed effects'—individual-specific skill measures—from these wages, residualizing them to remove other factors like experience. This treats private wages as noisy indicators of the skills needed for government-equivalent roles.
For example, a government clerk's skills might mirror a private administrative role. By benchmarking against these, they get a skill signal untainted by government rigidities.
With skill fixed effects in hand, they use ML algorithms (e.g., regressions flexible enough for nonlinearities) to link them to rich observables from IFLS: education, cognitive tests, personality, preferences. ML shines here by disentangling true skill signals from noise—using techniques like covariance shrinkage (inspired by Best et al., 2023) to separate variance from estimation error.
This model then predicts out-of-sample skills for government workers, addressing endogeneity and selection bias. It's robust: even instrumenting with parental public employment doesn't change core results.
Over 25 years, government skills declined relative to private sector. Why? Skilled workers select into private jobs as the economy develops, depleting the top of the skill distribution. Government hiring 'waves'—large cohorts post-reforms—force hiring lower down, diluting average quality.
Post-2000 reforms (democratization, decentralization) barely improved skill-based selection. Government wages are less predictive of skills than private ones but align when controlling for experience's predictable progression.
Strikingly, Indonesia's government pays 43% more than private sector equivalents conditional on skills, despite rigid structures. This premium doesn't attract top talent effectively, suggesting misaligned incentives.
The method generalizes: estimate 'potential skills' in any job using observables, useful beyond government (e.g., counterfactuals in mismatched labor markets). It builds intuition that skills are portable across similar tasks, but selection and institutions shape where talent flows.
We propose a new approach to estimate government worker skills, a setting where output is hard to observe and wages may be uninformative about skills. The approach uses wages in comparable jobs in the private sector and machine learning tools to link skills to skill-related observables. We apply the approach to rich Indonesian household-level panel data from 1988-2014, showing two main applications. First, government skills have continuously declined relative to the private sector, driven by the most skilled workers ending up in the private sector. Second, the Indonesian government pays a wage premium of 43% conditional on skills.
Alex: Right—like the private data teaches the connection between traits and skills, letting them guess skills for folks without private wage info. Do they worry the private workers aren't similar enough?
Sam: They do address that by reweighting the private sample to match government workers' basics like age, gender, job type, and location—ensuring the comparison is fair. This out-of-sample prediction holds up because the trait-to-skill link stays steady over time, letting them spot how government skills lagged private ones in Indonesia.
Alex: So those hiring booms pulled in lower talent as the best went private... but what does the data actually look like for Indonesian workers? Like, how common are government jobs, and do they match up well with private ones?
Sam: The main data comes from a long-running survey called the Indonesian Family Life Survey, or IFLS, which tracks thousands of people from 1988 to 2014 with details on jobs, wages, and tests of things like memory or math ability. Government jobs make up about 6% of employment—mostly in social services like teaching and health—while private comparable roles are in similar office or service work. They reweight private workers by how closely their age, gender, job type, and location match government ones, creating a fair comparison group, much like balancing teams in a game to test real differences.
Alex: Right, so the data covers a broad mix but focuses on similar roles. And once they clean the private wages for market shifts and experience, how do they turn those leftovers into skill predictions?
Sam: After netting out time-varying skill prices—estimated from late-career wage changes where pay flattens because skills aren't growing much anymore—they fit a simple curved line for how pay rises then levels with years on the job. That leaves a residual mainly signaling core skill, which they feed into machine learning models trained to connect it to traits like education or test scores. Among options, the gradient boosted model performs best on holdout data.
Alex: Huh—so the experience curve lets the model reliably map traits to skills across groups. That out-of-sample bridge is what uncovers the relative decline.
Sam: Precisely. The evidence points to government skills rising overall but falling behind private ones steadily, as top talent shifted and hiring surges tapped deeper into the pool. This holds across checks like different price methods or job subsets.
Alex: Okay, so the machine learning bridges those cleaned wage leftovers to traits reliably enough for out-of-sample predictions. But how well does it actually capture the skill signal?
Sam: The study tests several machine learning approaches on private workers' data, holding some out to check prediction accuracy. The gradient boosted model performs best, explaining a notable share of the leftover wage variation. Most of that gain comes from adding more traits like test scores and personality measures; education still dominates as the top predictor.
Alex: Right, so even if wages are noisy from luck or job perks, the model squeezes out a meaningful skill signal without fooling itself on training data alone. And they stress checking on unseen data to avoid overfitting?
Sam: Yes, out-of-sample checks are crucial here; in-sample fits can look great by memorizing noise. With these skill estimates, they track changes over time: government workers start ahead of similar private ones in 1989, with skills rising across the board from better schooling, but government lags relatively as private pulls even or ahead.
Alex: So absolute skills improve everywhere, but government's edge erodes. What explains that gap widening?
Sam: Looking by birth year, newer cohorts from school-building booms boost private skills more, as top talent skips government for those jobs. They also map a selection pattern: government favors mid-to-high skilled workers over low ones, but skips the very top—think picking strong players but missing the stars who go elsewhere. Hiring surges mainly force pulling from lower down as elites deplete.
Alex: Got it—like during boom hiring, with top talent already taken, averages drop even if selection stays fair. Does the data back that cohort dilution clearly?
Sam: Yes, cohorts facing big government hiring waves show lower relative skills, confirming fluctuations hurt quality by exhausting the pool. The method's checks, like matching job types, reinforce this.
Alex: So those cohort regressions confirm hiring waves lower relative skills... but with limited cohorts, the results must be noisy as the paper notes. Beyond selection, what about government wages themselves?
Sam: Yes—they predict wages using estimated skills for government and matched private workers. Government wages explain less skill variation at first, but close the gap once accounting for rigid experience schedules. Re-estimating skills directly from government wages correlates highly overall, though it flattens for top talent, meaning pay there hides skill differences.
Alex: Huh—so wages signal skills okay in government, but less sharply, especially at the high end. And do they earn more than similars?
Sam: Controlling for skills and jobs, government workers get a wage premium over comparable private roles. Much of that stems from no gender pay gap in government, unlike the private sector.
Alex: Got it—a premium partly from fairer pay across genders, despite rigid scales. That might explain why skills haven't dropped more.
Sam: Exactly. Checks like splitting data pre- and post-2000 show skill estimates stable over time, supporting the mapping holds as general skills.
Alex: Those stability checks across years sound solid. But to wrap this up, what about worries that skills might be job-specific, or hidden traits not captured?
Sam: They test that by retraining the model just on private workers in one job type at a time, then comparing predictions to the full baseline. For bigger groups, predictions line up closely, suggesting the estimates mostly grab general skills useful across jobs.
Alex: Right, so occupation-specific training doesn't shift things much for main roles. And for unmeasured traits driving job choices?
Sam: A key check uses parents' private jobs as a tool to model sorting into those roles based on hidden factors. Correcting via a standard adjustment barely budges estimates, leaving rankings and time trends intact.
Alex: Huh—so even accounting for unobservables, the core story holds.
Sam: Yes, it assumes those observables link steadily to skills over time and sectors. That could understate gaps if hidden factors sort strongly, though checks suggest stability. Practically, it quantifies how uneven hiring dilutes quality by exhausting top applicants, pointing to smoothing hires for better state capacity.
Alex: That's a meaningful step for understanding talent in government. Thanks, Sam—this method sheds clear light on a tough puzzle.
Sam: My pleasure, Alex. It equips sharper analysis of public sector skills worldwide. Thanks for listening to ResearchPod.