Instrument-free inference under confined regressor endogeneity and mild regularity

The instrument-free approach adopts flexible bounds on the correlation between regressors and disturbances, instead of exploiting instruments presupposing their asymptotic uncorrelatedness with the model errors. Earlier findings on such instrument-free inference methods assumed the observations to be mesokurtic and independent and identically distributed. Adopting substantially weaker regularity, this alternative to Two-Stage Least-Squares (TSLS) is developed and simulated for general linear regression models, permitting time-dependent regressors with heterogeneous excess kurtosis. Replicating three prominent empirical studies TSLS is shown to be based on untenable exclusion restrictions, whereas instrument-free inference can arguably be more credible, while potentially producing narrower confidence intervals than (weak-instrument robust) TSLS.