Dear Statalist community,
I use system GMM to estimate the following model:
y_gr: labour productivity growth, dependent variable
y: labour productivity, levels
L.y: lag of level of labour productivity, levels
L.y_gr: lag of
x: independent variable
x2: independent variable squared
c_gr: capital intensity growth
yr11-yr19: year dummies
Ind2-ind42: industry dummies
pftdum2-pftdum5: business outlook indicator
There are a couple of issues that are unclear to me:
1) I would like to include the levels of the dependent variable because with higher levels, it is likely more difficult to achieve growth. Am I mis-specifying the model if I do not also include lags of the dependent variable (L.y_gr)?
2) I understood form Roodman (2009) that generally speaking, all regressors should either enter gmmstyle() or ivstyle(). How can I implement a ‘classical’ instrumentation approach where some variables only enter the first stage? (These are not part of the above model.
3) The models that I have specified have far more groups than instruments (4000+ groups, max 500 instruments). Yet, the Hansen test is often significant. What other options, than reducing the instrument count, do I have to improve the Hansen statistic? I understood from Sebastian Kripfganz that I should avoid omitted variables in the Stata output as this can impact the Hansen statistic. Hence, I have manually specified the dummies to be included.
4) The data spans the period 2009-2019 (yearly), but I have rather short panels (unbalanced, on average, 3.8 observations), so including higher-order lags as regressors strongly reduces the sample size. What other options do I have to improve the AR(2) statistic?
Many thanks in advance,
Best regards,
Vladimir Sobota
I use system GMM to estimate the following model:
Code:
xtabond2 y_gr L.y x x2 c_gr contro1 control2 pftdum2-pftdum5 yrdum11-yrdum19 /// control3 ind2-ind43, ivstyle(c_gr x x2 contro1 contro2 yrdum11-yrdum19 control3 ind2-ind43) /// gmmstyle(L.y pftdum2-pftdum5) robust twostep
y: labour productivity, levels
L.y: lag of level of labour productivity, levels
L.y_gr: lag of
x: independent variable
x2: independent variable squared
c_gr: capital intensity growth
yr11-yr19: year dummies
Ind2-ind42: industry dummies
pftdum2-pftdum5: business outlook indicator
There are a couple of issues that are unclear to me:
1) I would like to include the levels of the dependent variable because with higher levels, it is likely more difficult to achieve growth. Am I mis-specifying the model if I do not also include lags of the dependent variable (L.y_gr)?
2) I understood form Roodman (2009) that generally speaking, all regressors should either enter gmmstyle() or ivstyle(). How can I implement a ‘classical’ instrumentation approach where some variables only enter the first stage? (These are not part of the above model.
3) The models that I have specified have far more groups than instruments (4000+ groups, max 500 instruments). Yet, the Hansen test is often significant. What other options, than reducing the instrument count, do I have to improve the Hansen statistic? I understood from Sebastian Kripfganz that I should avoid omitted variables in the Stata output as this can impact the Hansen statistic. Hence, I have manually specified the dummies to be included.
4) The data spans the period 2009-2019 (yearly), but I have rather short panels (unbalanced, on average, 3.8 observations), so including higher-order lags as regressors strongly reduces the sample size. What other options do I have to improve the AR(2) statistic?
Many thanks in advance,
Best regards,
Vladimir Sobota
