The Earnings Decline That May Not Be There
This series examines how earnings profiles differ across the working life for people with different college majors, focusing on the apparent decline of engineering and computer-science earnings relative to other fields. The opening question is deceptively simple: when older technical graduates appear to have a smaller earnings advantage than younger ones, does that mean their skills or careers have deteriorated faster? The useful judgment is more careful: before treating an age curve as a record of career growth, separate an average major premium from a lifecycle profile and ask who is being compared.
An average return is not a career path
A college major can be associated with higher or lower earnings in at least two different senses. An average return compares the typical earnings of people with one major against the typical earnings of people with another, often after adjusting for observable characteristics. That comparison can be important for understanding the economic stakes of education. But it compresses people at different ages, from different cohorts, and with different work histories into one number.
A major-specific age–earnings profile asks a different question. Holding the comparison framework fixed, how does the earnings gap between major groups vary at different ages? A profile might show that engineering and computer-science graduates earn more when they first enter the labor market but have a smaller advantage by age 50. That is a statement about the shape of an estimated relationship across age. It is not automatically a time-lapse record of a typical engineer or computer scientist, nor does it by itself identify what caused the gap to change.
In the paper, these profiles are estimated for age groups rather than read directly from a single person’s pay history. The coefficients describe how a major’s earnings gap relative to the reference group changes at points such as ages 29–30, 39–40, and 49–50, with other measured characteristics and calendar years accounted for in the model. That construction is useful because it puts fields on a common scale. It is also why the result must be read as a conditional comparison: the line is shaped by who appears in each age group, which earnings measure is used, and which differences the model leaves unmeasured.
A profile can therefore be informative without being a forecast. Consider two people with the same degree field, one observed near the beginning of a career and one observed decades later. The difference between their earnings may reflect experience, job changes, further education, employment intensity, the labor market they entered, or the characteristics that led them to choose and complete that major. A cross-sectional profile compresses those possibilities into an age-related pattern. The research task is to determine which interpretation the available evidence can sustain.
The distinction matters because people make decisions over a long horizon. A student deciding whether to borrow for college, a university evaluating fields of study, or a policymaker assessing the payoff to education needs more than an entry salary. They need to know whether an early premium persists, grows, or narrows—and whether the evidence actually measures that trajectory. The paper “College Majors and Earnings Growth” begins from this problem: average returns can conceal important differences in earnings growth over the lifecycle.
The pattern that needs explaining
The motivating comparison comes from a short repeated cross-section of the American Community Survey. In that pattern, engineering, computer science, and business majors start with a substantial earnings advantage over a broad group of other majors, but the advantage becomes smaller as workers age. In the replicated ACS specification, the engineering-and-computer-science estimate falls by 0.167 log points between age 23 and ages 49–50 relative to the other-major group. That is an estimated relative profile, not a claim that every person in those fields loses a particular percentage of pay.
The reference group also matters. “Other majors” is not a synonym for humanities. In the study’s main grouping it includes humanities, education, health, vocational fields, and related degrees; it is a large and diverse residual category. The comparisons also include life and physical science and social science majors. If a technical field’s profile is described as declining, it is declining relative to this defined comparison group under a particular model. Changing the reference group can change the meaning of the statement.
That relative framing matters for decisions about education. Two majors could have the same average lifetime return while producing different timing: one might pay more soon after graduation, while another catches up later. For someone facing borrowing constraints, the early part of the profile can affect the ability to repay, even if the long-run average is similar. Conversely, a high entry premium can be mistaken for a permanently superior path if the comparison does not account for how cohorts and work patterns differ. The point is not to turn a profile into a prescription, but to recognize why its shape has consequences.
The earlier cross-sectional interpretation also came with a plausible substantive story. Technical graduates often enter occupations in which skill requirements change quickly, so their initial knowledge might appear to lose value faster. That story connects an observed narrowing earnings gap to human-capital depreciation. It is a hypothesis about a mechanism, however, not something the age curve proves by itself. A different population composition can produce the same visual pattern, which is why the paper treats the measurement question as prior to the mechanism question.
The paper’s longer-panel result challenges the most intuitive interpretation of the cross-sectional pattern. By linking major information from the ACS to earnings histories in the Longitudinal Employer–Household Dynamics data, the authors find that engineering and computer-science majors experience earnings growth comparable to or faster than that of the broad other-major group. In the baseline worker-fixed-effects specification, the apparent technical penalty is no longer a large negative profile; after additional controls, the estimates are consistent with technical wage growth being on par with the reference group. The business profile changes for a different reason, involving the composition of annual earnings and labor supply, which will require separate attention.
This is not simply a dispute over which number is correct. The two designs ask questions with different observational leverage. One sees many people once, at different ages. The other uses the earnings histories of the same linked workers across many years. To understand why those approaches can diverge, we first need to see why an age comparison may contain more than age.
When age is carrying other information
Imagine a survey conducted in the same calendar year. The 25-year-olds and 50-year-olds are not the same people observed at different moments. They generally belong to different birth cohorts. They entered college and the labor market under different technologies, institutions, and labor-market conditions. If the economic value of a technical major has risen for newer cohorts, the younger group may carry a larger entry premium for reasons that have nothing to do with what happens as any one worker grows older.
That creates a misleading visual story. Suppose older engineering graduates began their careers with a modest premium, while newer engineering graduates began with a much larger one. A short cross-section will tend to place the newer, high-premium group at younger ages and the older, lower-premium group at older ages. The line drawn across ages slopes downward. Yet each cohort could maintain its initial advantage, or even gain relative to other majors, as its members acquire experience. The downward line would describe a changing mix of cohorts rather than a verified process of technical skill decay.
This is the difference between a level shift and a slope change. A newer cohort may sit higher on the earnings scale at every comparable age because its members were selected differently or acquired skills with greater market value. That shifts the profile’s level across cohorts. A slope change would mean that the newer cohort’s earnings rise or fall differently as its members age. Short cross-sections have difficulty distinguishing those statements because the newer cohort is mostly visible at younger ages. A long record can place more cohorts at more stages of their working lives and make the distinction testable.
The problem is not confined to a single survey year. The available ACS observations are concentrated in a relatively short recent period, so the age variation is doing much of the work of representing career progression. The LEHD history reaches back to 1985 for covered states, allowing young workers to be observed in earlier calendar years and older workers in later ones. That does not remove every selection or measurement issue, but it changes the balance between “different people at different ages” and “the same people over time.”
Calendar year adds another dimension. A worker’s age plus their birth cohort determines the survey year in which they are observed, so age, period, and cohort are mechanically related. A fully saturated analysis cannot freely estimate every combination without running into that collinearity. Researchers therefore have to decide which variation to use and which assumptions to make. The paper compares targeted approaches rather than pretending that one specification can reveal every age, cohort, and time effect at once.
The practical lesson is not that cross-sectional evidence is useless. A short survey can describe how earnings are distributed among people observed at different ages, and it can reveal a real association worth investigating. But the interpretation must stop at what the design can support. A cross-sectional age profile can show that older technical graduates have a smaller relative advantage in the observed sample. It cannot, on its own, establish that a typical technical graduate’s skills depreciate faster than those of a graduate in another field.
The same caution applies in reverse. Finding that a panel estimate is flat or positive does not prove that technology never makes skills obsolete. Occupations differ, workers change jobs, and the content of a major changes over time. The central claim is narrower: the observed cross-sectional convergence is not enough evidence for a universal lifecycle-decay story, because the people represented at each age may have entered the labor market with different major premiums.
The question worth carrying forward
The title’s apparent decline is therefore a real feature of one comparison, but not yet a settled description of technical careers. It tells us what the short cross-section looks like; it does not tell us which part comes from aging, which part comes from cohort change, or which part comes from how earnings are measured.
That distinction gives this series its central rule. Whenever an age–earnings curve appears, ask five questions before drawing a lifecycle conclusion: Who is observed? Over what years? What is the comparison group? Does the outcome combine wages with employment or education? And are the same workers followed, or are different cohorts being substituted for one another?
The next question follows naturally: when a dataset moves from snapshots to long earnings histories, what new comparison does it make possible—and which part of the apparent decline changes first?