Most of the analysis on this site is built on the Census Bureau's Post-Secondary Employment Outcomes (PSEO) dataset. PSEO links institutional records to federal wage data through the Longitudinal Employer-Household Dynamics (LEHD) system and reports earnings at 1, 5, and 10 years after completion, with distributions at the 25th, 50th, and 75th percentiles (P25, P50, P75). For trajectories, distributional spread, and cross-state mobility, nothing else comes close.
But a dataset good enough to anchor serious analysis is not a dataset without gaps. The gaps in PSEO are specific, and they do not fall evenly across programs, fields, or institutions. The five layers below describe what PSEO misses, why, and what that means for interpretation.
Layer 1: Unemployment insurance wage records miss entire categories of work
PSEO's earnings come from state unemployment insurance (UI) wage records, a system designed to administer unemployment benefits, not to measure earnings comprehensively. That design choice carries through:
Self-employment is excluded. Unincorporated self-employed graduates do not appear in UI wage records at all.
Several categories adjacent to standard W-2 employment are missing. Some federal civilian employees, railroad workers, and informal or family-based employment arrangements fall outside the system.
These gaps are not uniform across fields. They hit hardest in the skilled trades, cosmetology, independent healthcare roles, and other technical fields common in short-term certificate programs, precisely the fields where self-employment and independent contracting are most prevalent. A field where a meaningful share of graduates work as independent contractors will look lower-earning in PSEO than it actually is.
Layer 2: Small-cell suppression removes programs entirely
PSEO applies differential privacy protections that suppress earnings cells when the underlying cohort is too small to protect individual identity. Suppressed cells are not rounded, flagged, or imputed. They are removed from the published data.
This is a necessary privacy safeguard. But it has a systematic analytical consequence: programs with small numbers of graduates produce no reportable earnings at any percentile or time horizon. They are simply absent from any analysis that takes the dataset at face value.
The threshold binds harder at longer horizons. A program that clears the cell-size threshold at Year 1 can fall below it by Year 5 or Year 10, because the underlying employment sample thins as graduates move across state lines, out of UI-covered work, or out of the labor force. Longer-horizon data is therefore available for a narrower and narrower slice of programs.
Layer 3: The suppression is not random
The cell-size threshold is correlated with cohort size, and cohort size is correlated with almost everything that matters for program type. Which means the programs that disappear from the visible data share a profile:
Small programs. Short-term certificates (less than one year) are often small by design, tied to specific occupations and local labor markets. They are disproportionately suppressed.
Rural and specialized institutions. Colleges serving smaller regions, or offering specialized training like wildland fire, agricultural technology, or mine safety, are systematically underrepresented in the reportable data.
Entire program segments. Across the national PSEO data, a substantial share of short-term certificate program-institution combinations have no reportable earnings. In some states, more than half of all such programs are fully suppressed.
This is the layer that matters most for interpretation. If the programs that disappear are a random sample, the visible programs are representative of the whole. If they disappear in patterns, as they do here, the visible programs are a biased sample, skewed toward larger institutions, higher-enrollment programs, and urban areas. Any "national" conclusion drawn from the visible programs alone carries that selection into the result.
Layer 4: Other datasets each have their own missingness
There is no dataset that closes PSEO's gaps. The commonly used alternatives have their own:
The College Scorecard provides national coverage but its post-enrollment earnings are built from Treasury tax records matched only to recipients of federal student aid under Title IV of the Higher Education Act, meaning students who filed the Free Application for Federal Student Aid (FAFSA) and received federal grants or loans. Students who paid without federal aid are absent from the earnings measures even when their institution reports into the Integrated Postsecondary Education Data System (IPEDS). Title IV eligibility is also a program-level attribute: undergraduate certificate programs below 600 clock hours or 15 weeks have historically been ineligible, so none of their completers appeared in Scorecard earnings regardless of a student's own aid history. The Workforce Pell Grant, authorized in July 2025 and implemented starting January 2026, extends Title IV eligibility to many 150 to 600 clock-hour programs, and coverage in forthcoming Scorecard releases should improve accordingly. Historical earnings windows that closed before 2026 still reflect the pre-Workforce-Pell exclusion.
State longitudinal data systems can provide detailed program-level outcomes, but they are bounded by state borders and subject to state-specific methodology. Some state systems exclude lower-earning individuals from reported distributions. The Colorado Department of Higher Education, for example, drops graduates earning below a minimum annualized threshold from its program earnings reports, which truncates the lower tail of the distribution and biases program-value estimates upward.
Alumni-based datasets like Lightcast provide occupational and career pathway context, but coverage depends on observable online activity and produces uneven representation across fields, demographics, and regions. Results also depend heavily on how queries are tailored: adjusting a few filter parameters, such as geography, occupation codes, the profile date window, or degree-match logic, can shift headline figures substantially, so two reports drawn from the same underlying data can tell noticeably different stories.
Each source has its own pattern of missingness. Triangulating across them is possible, but there is no assembled dataset that provides a complete picture of postsecondary labor market outcomes.
Layer 5: Analytical implications
PSEO is a significant improvement over the evidence base that existed a decade ago. But four constraints should inform any conclusion built on it, particularly aggregate conclusions drawn across colleges, states, or program types:
First, not all forms of employment are captured. Estimates for fields with high self-employment rates should be read as lower bounds on actual earnings.
Second, small-cell suppression removes a non-random subset of programs from the observable data. The remaining programs skew toward scale.
Third, the suppression profile changes with the time horizon. Year 10 findings draw on a much thinner, and differently selected, slice of programs than Year 1 findings.
Fourth, PSEO is best interpreted alongside complementary sources, not as a standalone measure of program value. When a substantial share of programs in a given state or sector are suppressed, any conclusion drawn from the visible programs alone reflects the outcomes of scale, not the outcomes of the full workforce training system.
In workforce analysis, what is not observed is often as important as what is.
Data: U.S. Census Bureau, Post-Secondary Employment Outcomes (PSEO), fourth-quarter 2025 release. Earnings in 2023 dollars, adjusted for inflation using the Consumer Price Index for All Urban Consumers (CPI-U). Explore the data in the interactive earnings explorer.