Comment on ‘The case against JIVE’

摘要: While Davidson and MacKinnon (2006) (DM) present an interesting analysis of recently advocated estimators for overidentified instrumental variables models, we disagree with their conclusion that the LIML estimator should almost always be prefered to JIVE Phillips Hale (1977) (PH), Angrist et al. (1999) (AIK) Blomquist Dahlberg (BD). Our disagreement stems from three general points. First, show many advantages over DM illustrate are significantly reduced when one uses 'improved' proposed by Ackerberg Devereux (2003) (AD). Second, argue resulting in terms median bias dispersion small except cases where instruments so weak meaningful or relevant inference is likely not possible any estimator. Lastly, discuss some important estimator?in particular, citing new results AD regarding robustness heteroskedasticity. suggest two improve sample properties original PH, AIK BD, studied DM. The first these, IJIVE ((Improved) JIVE) estimator, removes partialling out exogenous explanatory (including constant term) before running procedure. this eliminates a term proportional number such variables. second UIJIVE ((Unbiased) IJIVE) additional term. Monte Carlo also UIJIVE, particularly appear more precise than JIVE. panels Figure 1 replicate middle DM's A, series UIJIVE. panel indicates region R2 > 0.2 (we values < moment), all shows improves on 9 decile range about 20%, although still considerably larger LIML. Interestingly, does do as well bias, levels equal (except opposite direction). However, adjustment designed eliminate mean bias. This verified 6, which looks at trimmed estimators.1 very small, most