The Hidden Mix Inside an Average Wage
This chapter follows an occupational average after its comparison frame has been made explicit. The question is why the average wage of an occupation can move differently from the occupation’s underlying wage premium when people enter, leave, or accumulate experience. The goal is to separate the price attached to the work from the changing mix of workers observed doing it, while keeping the remaining selection problem visible.
The distinction is central to interpreting occupational wage growth. A wage premium is about how the labor market prices an occupation relative to other work. An average wage is about the people and circumstances represented in the data at a particular time. Those quantities may move together, but there is no reason they must.
The price of work is not the average of its occupants
Imagine an occupation whose employers suddenly need more of its output. The additional demand may raise the wage premium: employers are willing to pay more for the work, relative to the alternatives available to workers. But the average wage observed in the occupation is not the premium alone. It also reflects the characteristics of the workers who occupy the jobs and the experience they bring with them.
This is the difference between a price and a composition. If a premium rises while the same workers remain in place, the average wage may show much of that movement. If the premium rises and the occupation attracts many new workers, the average can rise less, remain nearly flat, or even move in an unexpected direction, depending on the entrants’ wages and experience. The occupation can become more expensive to hire for while its observed workforce becomes less selected than before.
The reverse can happen in a shrinking occupation. If workers with lower wages or less favorable job matches leave first, the remaining workers can have a higher average wage even while demand for the occupation is weakening. A simple comparison of averages would then suggest improvement where the underlying demand signal is negative. The average is not lying; it is answering a different question from the one the analyst may have intended.
The IZA paper makes this distinction operational. Its wage equation treats log wages as containing a time-varying occupation premium, a worker–occupation component, skill returns, an occupation-specific experience profile, and an idiosyncratic shock. In plain language, a worker’s wage reflects both what the occupation commands and what the worker brings to it, including how experience is rewarded. When researchers observe an average change without separating these pieces, the result can be called occupational wage growth even though it is partly a change in who is being counted.
The paper therefore focuses on workers who stay in the same occupation across adjacent years. Comparing stayers removes time-invariant worker–occupation characteristics through first differences in the paper’s framework. A person’s persistent comparative advantage or stable match with an occupation is not allowed to masquerade as a new premium simply because the person appears in both periods. The design also allows the researchers to examine year-to-year wage changes rather than comparing only two broad cross-sections.
This does not mean that stayers are a perfect stand-in for the entire occupation. They are a deliberately selected group. Their wage growth may differ from the wage growth of entrants, leavers, or workers who switch occupations. The value of the design is narrower and more defensible: under its assumptions, the occupation-specific component of the change among stayers is closer to a change in the occupation’s premium than the raw change in the mean for everyone classified into it.
The distinction also matters for how we talk about skills. The study backs out an “implied” change in average skills by comparing observed average wage changes with estimated premium changes. That residual is useful for showing whether composition counteracts or reinforces premium growth. It is not a complete structural measurement of each worker’s productivity, nor is it a direct test score. It is a component of the paper’s decomposition, defined by what the model has separated from the observed wage.
A reader can therefore ask two different questions about the same occupation. “Did the occupation’s premium rise?” asks how the labor market’s price for the work changed after accounting for the model’s worker and experience terms. “Did the average worker in the occupation earn more?” asks what happened to the observed group. Both questions matter, but they do not have the same answer by construction.
The difference becomes especially important when comparing occupations with different starting wages. If the occupations that already paid more also receive larger premium increases, the premium component widens the gap between them. But if workers with lower measured skills move into the growing occupations, the composition component can push the observed average back toward the middle. An analyst who looks only at the final averages may see a modest change and miss a larger change in the occupation-specific price that was partly offset by sorting.
The reverse combination is possible as well. A premium can fall while selective departures leave behind a group whose average wage is relatively high. The average then contains a favorable composition effect that hides a weaker price signal. This is why a decomposition is not an ornamental refinement. It determines whether the observed change should be read as a change in demand for the work, a change in the people performing it, or a mixture of both.
For workers, employers, and policymakers, the distinction changes the decision that follows. A firm deciding whether a role has become harder to fill wants information about the price of the work and the available labor pool, not merely the current mean wage. A worker comparing an occupation wants to know whether a high average reflects a premium accessible to newcomers or the accumulated experience of incumbents. A policymaker interpreting inequality wants to know whether occupations changed their prices, whether employment moved between them, or whether the people inside them changed. The same statistic cannot settle all three questions.
How sorting changes the signal
The mechanism becomes clearest when a demand shock changes who wants to work in an occupation. The paper uses computer programmers as an example. Suppose a positive demand shock raises the price paid for programming output. The higher occupational premium attracts workers from other occupations. Some entrants may be less productive at programming than the incumbents they join, not because they lack ability in every setting but because they have had less time to accumulate occupation-specific skills or because they were previously selected for different work.
The sequence is then easy to miss:
If the entrants’ wages are lower than incumbents’ wages, they can dilute the average even as the premium rises. A researcher who sees only a modest increase in the mean might conclude that demand had little effect. The stayer comparison points to a different possibility: the occupation’s price moved substantially, but the inflow of workers changed the composition at the same time.
A declining occupation creates the mirror image. Workers with better alternatives may leave, but workers with weaker matches may leave first as jobs disappear or employers contract. Depending on who exits, the remaining average can decline sharply, remain stable, or rise. The direction of the average therefore cannot identify the direction of the premium without information about sorting and selection.
This is why the paper’s use of occupation stayers is more than a sample convenience. First differences among stayers remove time-invariant worker and occupation-match terms, while a polynomial in potential experience accounts for the way wage growth changes over the life cycle. The main specification estimates separate regressions for adjacent years and uses occupation-specific fixed effects to capture the premium-growth component under the paper’s assumptions. The study analyzes Swedish employees aged 18–64, uses the Swedish Wage Structure Statistics and related administrative data, and works with 101 occupations at the three-digit SSYK96 level.
The design also clarifies what a short panel can and cannot say. It follows adjacent-year changes and accumulates the estimated occupation effects across the 1996–2013 period; it does not observe every worker’s entire career or directly watch every transition between occupations. That is a trade-off. Short panels make it possible to compare many occupation-year changes while controlling for stable worker characteristics, but they leave more work for the experience model and the assumptions about who remains. The resulting premium estimate is therefore more informative than a raw cross-section for the paper’s question, yet narrower than a complete account of occupational mobility.
There is an important measurement limit inside that design. The researchers cannot reconstruct each worker’s complete history of actual occupation-specific experience from the relatively short panel and sampling structure. They therefore use potential overall labor-market experience, based on age and education. That measure is useful for describing how wages change with experience in the observed occupation, but it is not identical to years spent mastering that exact occupation. A worker with ten years in the labor market may have ten years of relevant experience, or a much more mixed history.
The experience issue means that even the stayer estimate is not simply “what the worker became worth.” A worker may earn more because the occupation rewards accumulated experience, because portable skills became more valuable, because the occupation’s common premium rose, or because the particular worker remained after others left. The research design separates some of these channels, but the claim remains conditional on its assumptions.
Selection on idiosyncratic shocks is one remaining problem. If an occupation’s premium grows quickly, it may retain more workers who happen to receive a bad wage shock; if its premium falls, it may retain only workers with unusually favorable shocks. The stayer sample can then make premium-growth variation look smaller than it really is. This is a different issue from time-invariant selection: even after stable differences are removed, who remains can depend on transitory outcomes.
The distinction between composition and capability also protects against a common overclaim. If an expanding occupation attracts workers with lower measured skills than incumbents, the average skill component may fall while the occupation’s premium rises. That does not mean the entrants are simply “worse workers” in an absolute sense. It means the composition of the occupation changed relative to the incumbents and relative to the model’s skill measure. They may be highly capable in another occupation, have transferable skills that the data do not capture, or be acquiring the occupation-specific skills that the demand shock now rewards.
The same caution applies to a shrinking occupation whose implied average skills rise. Selective exit can improve the composition of those who remain without making the occupation more valuable as a destination for new workers. The observed average describes the surviving group; it does not automatically reveal the direction of the occupational market.
Once price and composition are separated, the interpretation becomes more disciplined. A higher average wage can be evidence consistent with a higher premium, but it is not enough by itself. A stronger premium estimate can suggest a change in demand, but it is still an estimate under a design. A residual change in implied skills can show that sorting counteracted the premium, but it is not a complete account of worker productivity.
The practical questions now follow from the mechanism: what exactly was measured, who was present in the comparison, and which part of the movement could be ordinary experience growth? Keeping those questions together prevents a wage statistic from becoming a premature verdict about workers. It also prepares the ground for a compact diagnostic that can be applied before accepting any explanation of occupational wage growth.