Approaches to learn about employer learning

The empirical literature on employer learning assumes that employers learn about unobserved ability differences across workers as they spend time in the labour market. This article describes testable implications arise from this basic hypothesis and how have been used to quantify contribution of job market signalling human capital measured returns education. While basis is still thin, results suggest contributes at most 25% observed Approches de l'apprentissage par l'employeur. La littérature empirique sur des employeurs part du principe que les apprennent différences compétences non observées entre travailleurs au fur et à mesure qu'ils passent temps le marché travail. Le présent décrit tenants aboutissants vérifiables qui découlent cette hypothèse base la façon dont ils ont été utilisés pour quantifier indicateurs l'emploi humain dans l'incidence mesurée l'éducation. Bien soit encore mince, résultats laissent entendre contribuent plus environ observée What believe skills bound determine their outcomes. How schooling affects these beliefs thus crucial for determining an additional year schooling. Thus, might act a signal function distort decisions acquire social optimum (Spence 1973). Job difficult test because it objects are inherently hard observe: latent costs vary with skills. As such, measure much private education diverge each other due signalling. viewpoint one approach large gap be, builds specification proposed common model by Farber Gibbons (1996) further developed Altonji Pierret (2001). Central assumptions that: (i) individual productivity has persistent, time-invariant component, (ii) compensation equals expected (iii) persistent component To empirically operationalize assumptions, (or public) grew out (2001) explicitly formulates both potential what information underlying researchers access to. These leveraged derive predictions earnings proxies life cycle. In paper, we use simplified version illustrate important is. Two papers form discussion. Lange (2007) follows researcher correlate do not wage setting.1That paper: shows functional can be summarize process single parameter, “speed learning,” derives conditions under which identify parameter proceeds estimate it. goes impose structure decision economic years second Aryal al. (2022), takes different but related approach. Again, analysis starts circumstances interpret instrumental variable estimates causal effect representing either or education.2The main insight interprets IV depends whether know instrument induces variation population. extend variation, represent not, returns. discuss approaches signalling, also reference work implements approaches. two online appendices, provide evidence (appendix A) review existing B). Our view field suggests make up However, find conclusion narrow more work, using highlighted above, necessary firm our confidence assessment. Ideally, new will following compared estimates. We begin first introducing section 2 setting rely throughout remainder article. It based less general than analyzed (1996), (2007). then show 3 hidden correlate, employers, enables its parameters partial correlation briefly reviews discusses proposes section, 4, explores regressions learning. need observe, notably induced instrument. Instrumental variables advantage rather schooling, allow, suitable point-identify and/or Section 5 concludes. A i ˜ $$ \tilde{A_i} stands components caused Crucially, restrict c o v ( , S ) = 0 \mathit{\operatorname{cov}}\left(\tilde{A_i},{S}_i\right)=0 so correlated coefficient δ | {\delta}^{A\mid S} represents return captures increase independent accrues worker. Further, assume ε t \left(\tilde{A_i},{S}_i,{\varepsilon}_{it}\right) jointly normally distributed { } \left\{{\varepsilon}_{it}\right\} iid. Throughout, maintain W {W}_{it} experience conditional level ℰ y τ < . {\mathcal{E}}_{it}=\left\{{S}_i,{\left\{{y}_{i,\tau}\right\}}_{\tau <t},t\right\}. set grows t, implies hypothesizes revealed lim → ∞ − E [ e x p ] {\lim}_{t\to \infty}\left({W}_{it}-E\Big[\mathit{\exp}\left({y}_{it}\right)|{\mathcal{E}}_{it}\Big]\right)=0. becomes useful differentiating between substantive assumptions. brief remarks order. First, sense any shared among sufficient number competing ensure productivity. rule long-term contracts linking over longer periods. Common strong assumption, models asymmetric entail complex strategic interactions firms them cycle earnings.3 Second, straightforward incorporate observable controls analysis. Importantly, all authors footsteps allow observe correlates Q {Q}_i standard Mincer equations nor estimated always biased {S}_i. paper abstracts simplifies expressions communicate arguments without these. Here r σ \mathit{\operatorname{var}}\left(\tilde{A_i}\right)={\sigma}_0^2 \mathit{\operatorname{var}}\left({\varepsilon}_{it}\right)={\sigma}_{\varepsilon}^2 κ \kappa signal-to-noise ratio process, ie variance object (ie ξ + {\xi}_{it}=\tilde{A_i}+{\varepsilon}_{it} wages far forms discussion follows. invoke stronger strictly required, sometimes significantly so, uncovering fairly transparent way reasoning behind literature. foundational literature, (2001), combine wage-setting similar assumption Z direct refer correlate. such regression coefficients log evolve experience.4A reason popularity conform simple intuitions models. evolution primary interest especially (1997), Each associated coefficients. (in analogous specification) β ^ w \left\{{\hat{\beta}}_{wZ,t}\right\} (equation (10)) if {Z}_i positively \tilde{A_i}. when projecting {S}_i projection exceeds {\hat{\beta}}_{wS,t} 1 θ / \left(1-\theta \left(t,\kappa \right)\right){\hat{\beta}}_{AZ}\mathit{\operatorname{cov}}\left(Z,S\right)/\mathit{\operatorname{var}}(S). 5Thus, experience, emerging alone screen Third, exploits restrictions implied normal–normal \theta \right) (see equation (7)) \left\{{\hat{\beta}}_{wS,t},{\hat{\beta}}_{wZ,t}\right\} contrast (9) defined (6) directly deliver even controlling long Empirically, relied heavily NLSY 1979, data 1979 panel individuals during early careers good well cognitive skill serve correlate.6This measure, Armed Forces Qualification Test (AFQT) score battery tests administered respondents 1979. started collecting US residents born 1957 1964. From 1997 on, collection cohort 1980 1984 commenced. known 1997. Figure AFQT data. scatter point (measured deviations) 17.7 NOTES: scatters display standardized level. line predicted table 1. estimation described appendix A2. For sets, obtain those decrease expect model. Visual inspection above fits quite well. fitted profiles obtained estimating normal (2007), weighted averages t=0 limit value converges t\to \infty three only (the initial values ), fitting non-linear least squares.8 Data generated processes would generate features shown figure usefulness delivers particularly fit (even though does), allows summarizing manner whose summarized Table speed sets {S}_i,{Z}_i\Big). Columns (1) (2) reproduce stemming period document identical findings (column (2)) expectation errors average decline after around After years, error declined 50%. nine market, fallen 75%, 27 90%. therefore, worker's mostly within few her career. columns (3) (4) hand relatively slow entering 2000s. Specifically, column 11% worker spends eight 50%, 23 75% full 70 finding slowed down 1980s (5) report sample Norwegian males IQ scores collected military conscription procedures.9The bracket cohorts. Those imply very rapid learning, while Nevertheless, lower indicate substantial does take place 10 individual's career.10 Unfortunately, countries settings abundant desirable. And, unfortunately, comparing synthesizing environments. B. step uses total gains schooling.11 aspects bear particular mentioning. available (which abstracted above) researchers, arrives upper he relies cost here tradition views primarily investment determined largely opportunity discount rate future earnings. His discounting used. Both central provides. rates 3% 7% his preferred 26 =0.26 finds overall varies 26%. 95% interval 14 =0.14. If slow, 6% 47%. discussed single, estimable rapidly ability. produce required arrive strong. (2022) having Rather, willing stand inform discussions Equation (13) interpreting b I V {b}_{IV,t} comes making L {L}_i enters into {\mathcal{E}}_{it} held employers. focus polar cases. consider “transparent instruments” {\mathcal{E}}_{it}=\left\{{S}_i,{\xi}_i^t,{L}_i\right\} “hidden set. these, ∉ {L}_i\notin {\mathcal{E}}_{it}. instruments, applying LIE (12) {b}_{IV,t}={\delta}^{A\mid S}. Intuitively, aware instrument, infer à {\overset{\widetilde }{A}}_i. By imposed orthogonal endogenous IVs itself: