The Flat Spot: How to Separate Premium Growth from Experience

This chapter explains the IZA study’s central method for estimating occupational premium growth when experience also raises wages. The specific problem is that a worker can earn more because the occupation’s relative price increased, because the occupation rewards another year of experience, or because both happened together. Understanding the method lets us read its premium estimates as conditional measurements rather than as raw wage growth or a direct score of worker value.

The method begins with a simple warning: “we followed workers who stayed” is not enough. Stayers make selection more manageable, but they still accumulate experience. If experience profiles differ across occupations or change over time, the wage growth of stayers can be mistaken for a shift in the occupation’s premium. The paper’s answer is to model those profiles and re-center them around a flat spot.

What remains entangled after comparing stayers

The paper represents a worker’s log wage as a combination of an occupation-specific premium, a worker–occupation component, returns to observed skills, an occupation-specific experience profile, and an idiosyncratic shock. Each component answers a different question. The premium is the relative price attached to the occupation at a given time. The worker–occupation term captures persistent differences in the worker’s match or comparative advantage. Skill returns describe how measured skills are rewarded. The experience profile describes how wages change as potential experience accumulates. The shock captures a transitory component that the model does not predict.

This decomposition is not a claim that a wage record arrives with these labels attached. It is the structure the researchers use to interpret the record. The observed wage is the sum of components that must be separated through assumptions, comparisons, and estimation.

Now compare the same occupation stayer in two adjacent years. The worker–occupation component is differenced out because it is stable in the paper’s framework. That is a meaningful gain: a persistent comparative advantage no longer looks like new premium growth merely because the same person appears twice. But the change still contains the occupation’s premium movement and the wage growth associated with experience. In the paper’s notation, the constant term in the differenced regression combines the change in the premium with the linear part of the experience profile.

That is the identification problem. A profile that rises quickly for inexperienced workers can create large year-to-year wage changes even if the occupation’s common premium is unchanged. An occupation with slower early-career growth can show less wage movement even if its premium rises by the same amount. If a model simply adds experience terms without deciding where the profile is centered, it may attribute the wrong part of the curve to the occupation.

The paper addresses the problem with separate regressions for adjacent years. Its main specification uses workers who stayed in the same occupation, a fourth-order polynomial in potential experience, and occupation-specific fixed effects. Potential experience is re-centered at a chosen flat spot. Under the paper’s assumptions, the occupation effect in each adjacent-year regression estimates premium growth at that point, and the estimates are accumulated across the 1996–2013 period.

The design has a useful division of labor. Stayers address time-invariant worker–occupation selection.