Como medir o invis\'ivel? Guerras, pizzarias do Pent\'agono e o uso de vari\'aveis proxy em econometria

Abstract

Many economically relevant variables (risk, confidence, uncertainty) are latent and therefore not directly observable, which creates identification challenges in applied regressions. This text formalizes how omitting latent factors generates omitted-variable bias and discusses when including a proxy variable can mitigate it. We distinguish the case of a perfect proxy, which can eliminate the bias, from the more realistic case of an imperfect proxy, where residual bias remains and the estimated effect is attenuated. We propose a practical evaluation protocol based on four properties: relevance, conditional sufficiency, exogeneity, and stability. As an illustration, we use micromobility data from Arlington together with the U.S. Geopolitical Risk Index, estimating cointegration and a bivariate VEC model to interpret local activity as a high-frequency signal of the latent component of geopolitical tension.