“…And, unlike restricted maximum likelihood estimators, Bayesian methods directly estimate 4 Unlike Combes, Duranton and Gobillon (2008), our data do not follow workers or dwellings over time, preventing us from including individual effects in the first-stage regressions. Donovan et al (2021) find that including such effects reduces estimates of agglomeration economies in production by 1.1% whereas Ahlfeldt and Pietrostefani (2019) conclude controlling for "sorting" reduces estimates by 2-4%. 5 For workers, we control for gender, age (polynomial by gender), qualification (n = 10, except in 1976 when n = 4), two-digit industry sector (n = 54), ethnicity (n = 15), religion (n = 11), and birthplace (n = 12).…”