What the Evidence Does—and Does Not—Let Us Claim
This chapter tests how far the occupational wage-growth interpretation can travel. The Swedish results are strongest inside their own design, so the relevant questions are whether they survive plausible specifications, how changing skill returns affect them, what the supplementary U.S. analysis adds, and how current BLS wage comparisons fit alongside the research evidence. The practical judgment is not to choose a single “best” source, but to match each claim to the population, period, measure, and precision that support it.
A result can be robust without being assumption-free. A comparison can be useful without being causal. A second country can make a mechanism worth investigating without confirming the first country’s magnitudes. Keeping those distinctions explicit is what turns evidence about occupational wage growth into calibrated judgment rather than a collection of impressive-looking numbers.
Robustness changes confidence, not scope
The IZA paper tests how sensitive the premium and inequality results are to the choices that make premium growth identifiable. The flat-spot location is the most visible choice. The baseline centers potential experience at 30 years, while alternatives use 25 and 35 years. The decomposition is sensitive to this choice, especially in the covariance between premium growth and initial wages. That sensitivity matters because the covariance is a large part of the direct inequality effect.
The paper also estimates occupation-specific flat spots under an additional assumption that experience profiles are strictly concave except for possible flat regions. Those results are similar to the baseline. By contrast, setting the flat spot at zero—effectively treating the wage growth of entrants as the reference for premium growth—produces unreasonable results. The comparison does not show that one number is metaphysically true. It shows that re-centering is necessary and that reasonable alternatives can be checked rather than hidden.
Other checks change the flexibility of the experience profile, adjust for endogenous mobility, allow premium growth at regional and industry levels, pool data to impose more stable profiles, and divide the data by gender or by the availability of enlistment measures. In most cases, the no-sorting counterfactual is similar to or larger than the baseline. That pattern supports the main conclusion that worker composition masks part of occupation-level premium movement in the raw data.
But robustness has a precise meaning. It tells us that the result is not driven entirely by one convenient polynomial order, one mobility adjustment, or one sample restriction among the alternatives the authors examine. It does not establish that all unmeasured selection is harmless, that every labor market follows the same process, or that the estimate would survive a fundamentally different data-generating model. The study itself notes that forward-looking occupational choice, amenities, search frictions, and long-term contracts are outside its scope.
Selection on idiosyncratic shocks is a particularly important remaining boundary. Even among occupation stayers, who remains can depend on temporary wage realizations. A fast-growing occupation may retain workers who happened to receive a bad shock, while a declining occupation may retain only those with favorable shocks.