Hi all,
First post here! Thanks in advance for your kind attention and assistance.
I'm trying to use the -xtabond2- command to investigate the determinants of loan growth by banks. Having read the excellent material provided by Roodman himself, I am aware of the potential danger of having too many instruments. However, in my currently estimated equation I am running into the problem of constantly obtaining a high number of instruments. As can be seen below, I have tried options such as -lags()- and -collapse- to restrict the number of instruments, but I am not sure what the right approach would be.
The idea is to explain loan growth by means of various bank balance sheet characteristics and risk indicators. I have unbalanced panel-data of about 650 banks (=n) over 10 years which nets me 4033 observations (=N). I have classified each bank either as a retail, wholesale or investment bank (this is what the bm-dummy stands for) and this interacts with the balance sheet characteristics and risk indiactors. The idea is that a retail bank might react differently to an increase in its customer deposits than an investment bank.
I have added gdp growth (gdpgr), return on equity (roae) and the change in interest rate (intdif) as control variables. The bank-lending channel literature states that there is a potential endogeneity problem here. Roae and gdpgr can affect loangrowth, but the causality can also flow the other way around.
What I feed into Stata is as follows:
xtabond2 loangrowth l.loangrowth i.year intdif gdpgr l.roae (i.bm)#(c.l.custdep c.l.sec c.l.stliab c.l.ltliab c.l.derliab c.l.lev c.l.liq c.l.caprwa c.l.resimploans c.l.nii) if loangrowth<=1 | l.loangrowth<=1, gmmstyle(l2.loangrowth l2.roae l.gdpgr) ivstyle(i.year i.bm#(c.l.custdep c.l.sec c.l.stliab c.l.ltliab c.l.derliab c.l.lev c.l.liq c.l.resimploans c.l.caprwa c.l.nii)) robust nodiffsargan orthog
This yields 128 instruments and the following Sargan, Hansen and Aerllano-Bond output:
Sargan test of overid. restrictions: chi2(85) = 183.07 Prob > chi2 = 0.000
(Not robust, but not weakened by many instruments.)
Hansen test of overid. restrictions: chi2(85) = 185.22 Prob > chi2 = 0.000
(Robust, but weakened by many instruments.)
Arellano-Bond test for AR(1) in first differences: z = -8.40 Pr > z = 0.000
Arellano-Bond test for AR(2) in first differences: z = 1.06 Pr > z = 0.289
I am uncertain about how to proceed from here. The variables in -gmmstyle- are specified as recommended by Roodman if one suspects that those variables are endogenous. Yet, I obtain so many instruments.
When I write -gmmstyle(l2.loangrowth l2.roae l.gdpgr), lag(1 1) collapse)- I get 45 instruments and the problem seems to go away (see below), but I wonder whether this is a valid way to treat the potential endogeneity. Indeed, the Arellano-Bond test for AR(2) now becomes a problem!
Sargan test of overid. restrictions: chi2(2) = 5.67 Prob > chi2 = 0.059
(Not robust, but not weakened by many instruments.)
Hansen test of overid. restrictions: chi2(2) = 4.92 Prob > chi2 = 0.085
(Robust, but weakened by many instruments.)
Arellano-Bond test for AR(1) in first differences: z = -5.71 Pr > z = 0.000
Arellano-Bond test for AR(2) in first differences: z = -3.13 Pr > z = 0.002
I guess my question boils down to the following: how to specify -gmmstyle- correctly such that the instruments seem valid while still maintaining valid regression estimation?
First post here! Thanks in advance for your kind attention and assistance.
I'm trying to use the -xtabond2- command to investigate the determinants of loan growth by banks. Having read the excellent material provided by Roodman himself, I am aware of the potential danger of having too many instruments. However, in my currently estimated equation I am running into the problem of constantly obtaining a high number of instruments. As can be seen below, I have tried options such as -lags()- and -collapse- to restrict the number of instruments, but I am not sure what the right approach would be.
The idea is to explain loan growth by means of various bank balance sheet characteristics and risk indicators. I have unbalanced panel-data of about 650 banks (=n) over 10 years which nets me 4033 observations (=N). I have classified each bank either as a retail, wholesale or investment bank (this is what the bm-dummy stands for) and this interacts with the balance sheet characteristics and risk indiactors. The idea is that a retail bank might react differently to an increase in its customer deposits than an investment bank.
I have added gdp growth (gdpgr), return on equity (roae) and the change in interest rate (intdif) as control variables. The bank-lending channel literature states that there is a potential endogeneity problem here. Roae and gdpgr can affect loangrowth, but the causality can also flow the other way around.
What I feed into Stata is as follows:
xtabond2 loangrowth l.loangrowth i.year intdif gdpgr l.roae (i.bm)#(c.l.custdep c.l.sec c.l.stliab c.l.ltliab c.l.derliab c.l.lev c.l.liq c.l.caprwa c.l.resimploans c.l.nii) if loangrowth<=1 | l.loangrowth<=1, gmmstyle(l2.loangrowth l2.roae l.gdpgr) ivstyle(i.year i.bm#(c.l.custdep c.l.sec c.l.stliab c.l.ltliab c.l.derliab c.l.lev c.l.liq c.l.resimploans c.l.caprwa c.l.nii)) robust nodiffsargan orthog
This yields 128 instruments and the following Sargan, Hansen and Aerllano-Bond output:
Sargan test of overid. restrictions: chi2(85) = 183.07 Prob > chi2 = 0.000
(Not robust, but not weakened by many instruments.)
Hansen test of overid. restrictions: chi2(85) = 185.22 Prob > chi2 = 0.000
(Robust, but weakened by many instruments.)
Arellano-Bond test for AR(1) in first differences: z = -8.40 Pr > z = 0.000
Arellano-Bond test for AR(2) in first differences: z = 1.06 Pr > z = 0.289
I am uncertain about how to proceed from here. The variables in -gmmstyle- are specified as recommended by Roodman if one suspects that those variables are endogenous. Yet, I obtain so many instruments.
When I write -gmmstyle(l2.loangrowth l2.roae l.gdpgr), lag(1 1) collapse)- I get 45 instruments and the problem seems to go away (see below), but I wonder whether this is a valid way to treat the potential endogeneity. Indeed, the Arellano-Bond test for AR(2) now becomes a problem!
Sargan test of overid. restrictions: chi2(2) = 5.67 Prob > chi2 = 0.059
(Not robust, but not weakened by many instruments.)
Hansen test of overid. restrictions: chi2(2) = 4.92 Prob > chi2 = 0.085
(Robust, but weakened by many instruments.)
Arellano-Bond test for AR(1) in first differences: z = -5.71 Pr > z = 0.000
Arellano-Bond test for AR(2) in first differences: z = -3.13 Pr > z = 0.002
I guess my question boils down to the following: how to specify -gmmstyle- correctly such that the instruments seem valid while still maintaining valid regression estimation?
