A Solution to the Problem of Too Many Instruments in Dynamic Panel Data GMM
By Jens Mehrhoff. Social Science Research Network, 2009
127 citations recorded by the source corpus.
The well-known problem of too many instruments in dynamic panel data GMM is dealt with in detail in Roodman (2009, Oxford Bull. Econ. Statist.). The present paper goes one step further by providing a solution to this problem: factorisation of the standard instrument set is shown to be a valid transformation for ensuring consistency of GMM. Monte Carlo simulations show that this new estimation technique outperforms other possible transformations by having a lower bias and RMSE as well as greater robustness of overidentifying restrictions. The researcher's choice of a particular transformation can be replaced by a data-driven statistical decision.
Subfields: Economics and Econometrics
View Paper: https://www.semanticscholar.org/paper/16741f4f9ea99e1f6aea1e56b6925376c578d665
Citations in the corpus
- Estimation of linear dynamic panel data models with time‐invariant regressors (2019, unknown): “Remark 3: As an alternative to the strong Assumption 2 that requires some regressors to be uncorrelated with the unobserved unit-specific effects αi in model (1), we can consider a correlated random effects assumption in the spirit of Mundlak (1978), E[αi|Xi,Fi] = b+x̄iπ, or Chamberlain (1982), E[αi”
- Counting Chickens when they Hatch: Timing and the Effects of Aid on Growth* (2012, unknown): “…2010) while the Anderson–Hsiao estimator does; the system-GMM estimator is now known to generate biased coefficients in many applied settings (Mehrhoff, 2009; Roodman, 2009); and the Anderson–Hsiao estimator does substantially eliminate dynamic panel bias in expectation even if it is less…”
- An Empirical Examination of the Antecedents and Consequences of Contribution Patterns in Crowd-Funded Markets (2013, unknown): “To alleviate any concerns about a high instrument count in the GMM estimations, we also use a collapsed instrument set comprised only of low order lags, as this has been shown to produce more reliable results in scenarios where the instrument count is on the higher side (Mehrhoff 2009).”