Econometric analysis of dynamic panel-data models using stata精品.pdfVIP

Econometric analysis of dynamic panel-data models using Stata David M. Drukker StataCorp Summer North American Stata Users Group meeting July 24-25, 2008 1 / 32 1 Dynamic panel-data models 2 The Arellano-Bond estimator 3 The Arellano-Bover/Blundell-Bond estimator 2 / 32 Dynamic panel-data models Why dynamic panel-data models require special estimators Introduction We are interested in estimating the parameters of models of the form y = y γ + x β + u + ǫ it it−1 it i it for i = {1, . . . , N} and t = {1, . . . , T} using datasets with large N and fixed T By construction, yit−1 is correlated with the unobserved individual-level effect ui Removing ui by the within transform (removing the panel-level means) produces an inconsistent estimator with T fixed First difference both sides and look for instrumental-variables (IV) and generalized method-of-moments (GMM) estimators 3 / 32 The Arellano-Bond estimator The Arellano-Bond estimator I First differencing the model equation yields ∆y = ∆y γ + ∆x β + ∆ǫ it it−1 it it The ui are gone, but the yit−1 in ∆yit−1 is a function of the ǫit−1 which is also in ∆ǫit So ∆yit−1 is correlated with ∆ǫit by construction [Anderson and Hsiao(1981)] suggested a 2SLS estimator based on further lags of ∆yit as instruments for ∆yit−1 For instance, if ǫit is IID over i and t, ∆yit−2 would be a valid instrument for ∆yit−1 [Anderson and Hsiao(1981)] also suggested suggested a 2SLS