Hi,
I am currently investigating the impact of macroprudential policies on credit growth, with Stata 13.0. The idea is to reproduce the model used by Cerutti and al (2015) : ''The Use and Effectiveness of Macroprudential Policies : New Evidence'', but with another set of data.
In their paper, they used the '' xtdpd '' command in order to obtain the Arellano-Bond GMM estimator. They used a dynamic panel data because past credit growth affect the actual ones.
So I tried this command, but because GMM is very new to me, I tried Xtabond2 command by Roodman (2007), which is less complex. In order to replicate the model used by Cerutti and al (2015), I choose the difference GMM estimator, and considered that all variables are endogeneous.
My understanding is that :
- The number of instruments has to be lower than the number of groups (of observations). In fact, I have N = 21 and T = 56 (14 years, 2000 to 2014, with a quarterly frequency) and several missing values. I know that xtabond2 is not very recommanded when there is a small N and a big T, but with the ''collapse'' option and only lags(2 3) for each instruments, I obtain a number of intruments equals to 18, so maybe it should work. In comparison, Cerutti and al (2015) have N=119 and T = 13.
- Because all my variables are endogeneous, i can't use the first lag of a variable as an instrument.
- Twostep method is preferred over the onestep method.
In fact, there is 5 macroprudential measures in the model, which are coded as follows : if a measure is tightened (loosened), the associed value is ''1'' (''-1''). If there is no change, the value is 0. If a measure is tightened (loosened) twice over the years, then the value will be ''2'' (''-2'') after the second tightening (loosening). These are the variables with the "c_"
In addition to macroprudential variables, the other independent variables are the lagged dependent variable, the real GDP growh in the previous quarter, a dummy variable which capture the presence of a banking crisis during the previous quarter, and a variable which capture the impact of the interest rate in the previous quarter.
So all independent variables have a lag.
This is the model: CreditGrowthi,t= CreditGrowthi,t-1 α + C_macroprudentialmeasuresi,t-1 β + GDPGrowthi,t-1 + BankingCrisisi,t-1 δ + InterestRatei,t-1θ + CountryFixedEffect + ei
And this is an example of what I tried:
And this is the post-estimation results (Postestimation → Manage estimation results → Table of estimation results):
If I'm not mistaken, the model passes both the AR test and the Hansen test (more appropriate than the Sargan test because of the "twostep" option) but for the last one I'm not sure because of this :''(Robust, but weakened by many instruments.)''
The problem is that there is some of these variables, like crisis_banking, which are supposed to have a huge negative impact on the credit growth.
Because these macroprudential measures are very new, I have a lot of ''0'' in my data, and so Stata dropped many of them. So I had to agregate several variables together in order to solve this issue, and this is why I have only 5 macroprudential variables.
In addition, one issue is that some coefficients have the wrong sign, and their coefficients are volatile (when I delete one variable which is not significant, some coefficients change completly, both in terms of sign and impact (see below)).
I read that this issue may come from multicollinearity. So I did a multicollinearity diagnostic, but because the VIF is not available after xtabond2, I used ''regress'' instead of ''xtabond2'', and here are the results:
1) Is this the good way to perform statistics tests when we use GMM ? I have the same remark for the heteroskedasticity test.
2) Apparently, there is no multicollinearity (the VIF is very low for each variable), what I'm supposed to do in order to obtain strong coefficients ?
3) In case several regression with different lags successfully pass the Sargan/Hansen test and the AR test, what critererion is used to choose the best ? The one with the higher number of instruments ?
4) with fixed effects models, some variables are highly significant, as when I use xtabond2 without the ''robust '' option. When I add this option, no variables are significant. My understanding is that the''robust'' option allows us to work with heteroskedasticity. In this particular case, heteroskedasticity is an assumption or I have to test it ? (as for my first question, I can't test it after xtabond2).
5) Finally, I'am used to go to Statistics → Postestimation → Manage estimation results → Table of estimation results, in order to obtain the significance of coefficients with stars. Is it possible to do that directly in the regression results ?
I'm a little lost with the choice of options, especially with the option ''orthogonal'', I don't know if I can use it.
Here is a sample of my data:
So, it's a very long post with a lot of questions, and I think I forgot some, so the title is not very accurate. Sorry for that, but I did a lot of research, especially on this forum.
Thank you in advance for your answers.
References :
Cerutti, E., S. Claessens and L. Laeven. 2015. “The Use and Effectiveness of Macroprudential Policies: New Evidence.” IMF Working Paper WP/15/61.
Cerutti, E., R. Correa, E. Fiorentino, and E. Segalla. 2017. “Changes in Prudential Policy Instruments—A New Cross-Country Database.” International Journal of Central Banking 13 (S1).
Roodman, D. (2009) : How to do xtabond2: An introduction to difference and system GMM in Stata. Stata Journal 9(1): 86-136
Roodman, D. (2009), “A note on the theme of too many instruments,” Oxford Bulletin of Economics and Statistics, 71, 135-158
Mileva, E. (2007) "Using Arellano – Bond Dynamic Panel GMM Estimators in Stata, Tutorial with Examples using Stata 9.0 (xtabond and xtabond2)," Economics Department, Fordham University July 9, 2007
I am currently investigating the impact of macroprudential policies on credit growth, with Stata 13.0. The idea is to reproduce the model used by Cerutti and al (2015) : ''The Use and Effectiveness of Macroprudential Policies : New Evidence'', but with another set of data.
In their paper, they used the '' xtdpd '' command in order to obtain the Arellano-Bond GMM estimator. They used a dynamic panel data because past credit growth affect the actual ones.
So I tried this command, but because GMM is very new to me, I tried Xtabond2 command by Roodman (2007), which is less complex. In order to replicate the model used by Cerutti and al (2015), I choose the difference GMM estimator, and considered that all variables are endogeneous.
My understanding is that :
- The number of instruments has to be lower than the number of groups (of observations). In fact, I have N = 21 and T = 56 (14 years, 2000 to 2014, with a quarterly frequency) and several missing values. I know that xtabond2 is not very recommanded when there is a small N and a big T, but with the ''collapse'' option and only lags(2 3) for each instruments, I obtain a number of intruments equals to 18, so maybe it should work. In comparison, Cerutti and al (2015) have N=119 and T = 13.
- Because all my variables are endogeneous, i can't use the first lag of a variable as an instrument.
- Twostep method is preferred over the onestep method.
In fact, there is 5 macroprudential measures in the model, which are coded as follows : if a measure is tightened (loosened), the associed value is ''1'' (''-1''). If there is no change, the value is 0. If a measure is tightened (loosened) twice over the years, then the value will be ''2'' (''-2'') after the second tightening (loosening). These are the variables with the "c_"
In addition to macroprudential variables, the other independent variables are the lagged dependent variable, the real GDP growh in the previous quarter, a dummy variable which capture the presence of a banking crisis during the previous quarter, and a variable which capture the impact of the interest rate in the previous quarter.
So all independent variables have a lag.
This is the model: CreditGrowthi,t= CreditGrowthi,t-1 α + C_macroprudentialmeasuresi,t-1 β + GDPGrowthi,t-1 + BankingCrisisi,t-1 δ + InterestRatei,t-1θ + CountryFixedEffect + ei
And this is an example of what I tried:
- first attempt
Code:
xtabond2 credit_growth L1.credit_growth L1.c_sscb L1.c_cap_req L1.c_ltv_cap L1.c_rr L1.c_exposition > L1.interest_rate L1.GDP_growth L1.banking_crisis, gmmstyle(L1.credit_growth L1.c_sscb L1.c_cap_req > L1.c_ltv_cap L1.c_rr L1.c_exposition L1.interest_rate L1.GDP_growth L1.banking_crisis, lag(2 3) > collapse) noleveleq twostep Favoring space over speed. To switch, type or click on mata: mata set matafavor speed, perm. Dynamic panel-data estimation, two-step difference GMM ------------------------------------------------------------------------------ Group variable: num Number of obs = 1007 Time variable : q_date Number of groups = 21 Number of instruments = 18 Obs per group: min = 37 Wald chi2(9) = 74.81 avg = 47.95 Prob > chi2 = 0.000 max = 58 ----------------------------------------------------------------------------------------- credit_growth | Coef. Std. Err. z P>|z| [95% Conf. Interval] ------------------------+---------------------------------------------------------------- credit_growth | L1. | -.0390548 .1130852 -0.35 0.730 -.2606977 .1825881 | c_sscb | L1. | -4.62242 .9516904 -4.86 0.000 -6.487699 -2.757141 | c_cap_req | L1. | -1.407323 1.02413 -1.37 0.169 -3.41458 .5999351 | c_ltv_cap | L1. | 4.428849 3.137135 1.41 0.158 -1.719824 10.57752 | c_rr | L1. | -1.116012 1.679616 -0.66 0.506 -4.407999 2.175975 | c_exposition | L1. | .0300822 1.116648 0.03 0.979 -2.158508 2.218673 | interest_rate | L1. | .0787794 .2102005 0.37 0.708 -.333206 .4907648 | GDP_growth | L1. | -.2289244 .2287425 -1.00 0.317 -.6772514 .2194026 | banking_crisis | L1. | 1.426424 1.563614 0.91 0.362 -1.638203 4.491051 ----------------------------------------------------------------------------------------- Warning: Uncorrected two-step standard errors are unreliable. Instruments for first differences equation GMM-type (missing=0, separate instruments for each period unless collapsed) L(2/3).(L.credit_growth L.interest_rate L.c_sscb L.c_cap_req L.c_ltv_cap L.c_rr L.c_exposition L.GDP_growth L.banking_crisis) collapsed ------------------------------------------------------------------------------ Arellano-Bond test for AR(1) in first differences: z = -2.34 Pr > z = 0.019 Arellano-Bond test for AR(2) in first differences: z = 0.29 Pr > z = 0.775 ------------------------------------------------------------------------------ Sargan test of overid. restrictions: chi2(9) = 26.54 Prob > chi2 = 0.002 (Not robust, but not weakened by many instruments.) Hansen test of overid. restrictions: chi2(9) = 11.35 Prob > chi2 = 0.253 (Robust, but weakened by many instruments.) . end of do-file
Code:
Variable active
credit_gro~h
L1. -.03905478
c_sscb
L1. -4.6224202***
c_cap_req
L1. -1.4073225
c_ltv_cap
L1. 4.428849
c_rr
L1. -1.1160123
c_exposition
L1. .03008222
interest_r~e
L1. .07877938
GDP_growth
L1. -.2289244
banking_cr~s
L1. 1.4264237
legend: * p<.1; ** p<.05; *** p<.01
The problem is that there is some of these variables, like crisis_banking, which are supposed to have a huge negative impact on the credit growth.
- second attempt
Code:
Variable active
credit_gro~h
L1. -.03905478
c_sscb
L1. -4.6224202
c_cap_req
L1. -1.4073225
c_ltv_cap
L1. 4.428849
c_rr
L1. -1.1160123
c_exposition
L1. .03008222
interest_r~e
L1. .07877938
GDP_growth
L1. -.2289244
banking_cr~s
L1. 1.4264237
legend: * p<.1; ** p<.05; *** p<.01
Because these macroprudential measures are very new, I have a lot of ''0'' in my data, and so Stata dropped many of them. So I had to agregate several variables together in order to solve this issue, and this is why I have only 5 macroprudential variables.
In addition, one issue is that some coefficients have the wrong sign, and their coefficients are volatile (when I delete one variable which is not significant, some coefficients change completly, both in terms of sign and impact (see below)).
- third attempt:
Code:
xtabond2 credit_growth L1.credit_growth L1.c_sscb L1.c_cap_req L1.c_ltv_cap L1.c_rr L1.c_exposition
> L1.interest_rate L1.GDP_growth, gmmstyle(L1.credit_growth L1.c_sscb L1.c_cap_req L1.c_ltv_cap L1.c_rr
> L1.c_exposition L1.interest_rate L1.GDP_growth, lag(2 3) collapse) noleveleq twostep
Favoring space over speed. To switch, type or click on mata: mata set matafavor speed, perm.
Dynamic panel-data estimation, two-step difference GMM
------------------------------------------------------------------------------
Group variable: num Number of obs = 1180
Time variable : q_date Number of groups = 21
Number of instruments = 16 Obs per group: min = 41
Wald chi2(8) = 8.45 avg = 56.19
Prob > chi2 = 0.391 max = 58
-------------------------------------------------------------------------------
credit_growth | Coef. Std. Err. z P>|z| [95% Conf. Interval]
--------------+----------------------------------------------------------------
credit_growth |
L1. | -.2786686 .2411136 -1.16 0.248 -.7512425 .1939053
|
c_sscb |
L1. | -2.977535 3.035509 -0.98 0.327 -8.927022 2.971952
|
c_cap_req |
L1. | -.2339573 1.101806 -0.21 0.832 -2.393458 1.925544
|
c_ltv_cap |
L1. | 5.151175 7.673965 0.67 0.502 -9.88952 20.19187
|
c_rr |
L1. | -1.594947 2.977368 -0.54 0.592 -7.430481 4.240588
|
c_exposition |
L1. | -.824738 3.071016 -0.27 0.788 -6.843818 5.194342
|
interest_rate |
L1. | .1985294 .3771818 0.53 0.599 -.5407333 .9377921
|
GDP_growth |
L1. | -.0554081 .3740414 -0.15 0.882 -.7885158 .6776997
-------------------------------------------------------------------------------
Warning: Uncorrected two-step standard errors are unreliable.
Instruments for first differences equation
GMM-type (missing=0, separate instruments for each period unless collapsed)
L(2/3).(L.credit_growth L.c_sscb L.c_cap_req L.c_ltv_cap L.c_rr
L.c_exposition L.interest_rate L.GDP_growth) collapsed
------------------------------------------------------------------------------
Arellano-Bond test for AR(1) in first differences: z = -0.77 Pr > z = 0.440
Arellano-Bond test for AR(2) in first differences: z = -1.13 Pr > z = 0.257
------------------------------------------------------------------------------
Sargan test of overid. restrictions: chi2(8) = 18.29 Prob > chi2 = 0.019
(Not robust, but not weakened by many instruments.)
Hansen test of overid. restrictions: chi2(8) = 10.57 Prob > chi2 = 0.227
(Robust, but weakened by many instruments.)
.
end of do-file
. estimates table, star(.1 .05 .01) style(oneline)
------------------------------
Variable | active
-------------+----------------
credit_gro~h |
L1. | -.27866862
|
c_sscb |
L1. | -2.977535
|
c_cap_req |
L1. | -.23395735
|
c_ltv_cap |
L1. | 5.151175
|
c_rr |
L1. | -1.5949469
|
c_exposition |
L1. | -.82473802
|
interest_r~e |
L1. | .19852941
|
GDP_growth |
L1. | -.05540809
------------------------------
legend: * p<.1; ** p<.05; *** p<.01
Code:
regress credit_growth L1.c_sscb L1.c_cap_req L1.c_ltv_cap L1.c_rr L1.c_exposition L1.interest_rate
> L1.GDP_growth L1.banking_crisis
estat vif
Variable VIF 1/VIF
banking_crisis
L1. 1.25 0.803007
c_cap_req
L1. 1.24 0.807731
c_exposition
L1. 1.20 0.831029
c_sscb
L1. 1.18 0.849848
interest_rate
L1. 1.14 0.879999
c_rr
L1. 1.14 0.881049
GDP_growth
L1. 1.11 0.903970
c_ltv_cap
L1. 1.02 0.976797
Mean VIF 1.16
2) Apparently, there is no multicollinearity (the VIF is very low for each variable), what I'm supposed to do in order to obtain strong coefficients ?
3) In case several regression with different lags successfully pass the Sargan/Hansen test and the AR test, what critererion is used to choose the best ? The one with the higher number of instruments ?
4) with fixed effects models, some variables are highly significant, as when I use xtabond2 without the ''robust '' option. When I add this option, no variables are significant. My understanding is that the''robust'' option allows us to work with heteroskedasticity. In this particular case, heteroskedasticity is an assumption or I have to test it ? (as for my first question, I can't test it after xtabond2).
5) Finally, I'am used to go to Statistics → Postestimation → Manage estimation results → Table of estimation results, in order to obtain the significance of coefficients with stars. Is it possible to do that directly in the regression results ?
I'm a little lost with the choice of options, especially with the option ''orthogonal'', I don't know if I can use it.
Here is a sample of my data:
Code:
* Example generated by -dataex-. To install: ssc install dataex
clear
credit_growth c_sscb c_cap_req c_ltv_cap c_rr c_exposition interest_rate GDP_growth banking_crisis
"2000q1" 0 0 0 0 -1 1 3.5423 .93956 0
"2000q2" 1.322812 0 0 0 -1 1 4.263 1.013441 0
"2000q3" .6646443 0 0 0 -1 1 4.7376 -.141768 0
"2000q4" 1.7269744 0 0 0 -1 1 5.024167 .076452 0
"2001q1" .4631569 0 0 0 -1 1 4.745033 1.634062 0
"2001q2" .224378 0 0 0 -1 1 4.590766 .085889 0
"2001q3" 1.0374751 0 0 0 -1 1 4.267833 -.289637 0
"2001q4" 1.0920715 0 0 0 -1 1 3.4435 .118342 0
"2002q1" .24551354 0 0 0 -1 1 3.362233 -.308947 0
"2002q2" .8305011 0 0 0 -1 1 3.446 .258706 0
"2002q3" 1.6563503 0 0 0 -1 1 3.357333 .473059 0
"2002q4" .241526 0 0 0 -1 1 3.1088 -.214017 0
"2003q1" 1.4132304 0 0 0 -1 1 2.6831 -1.211775 0
"2003q2" 1.1803162 0 0 0 -1 1 2.3619 .021715 0
"2003q3" -.19664878 0 0 0 -1 1 2.139233 .499232 0
"2003q4" .3575432 0 0 0 -1 1 2.149633 .367172 0
"2004q1" .7381726 0 0 0 -1 1 2.062967 -.024217 0
"2004q2" -.3454666 0 0 0 -1 1 2.082467 .344379 0
"2004q3" .4714422 0 0 0 -1 1 2.1163 -.182331 0
"2004q4" .35406515 0 0 0 -1 1 2.1636 .096709 0
end
label var GDP_growth "taux_de_croissance_PIB_reel"
Thank you in advance for your answers.
References :
Cerutti, E., S. Claessens and L. Laeven. 2015. “The Use and Effectiveness of Macroprudential Policies: New Evidence.” IMF Working Paper WP/15/61.
Cerutti, E., R. Correa, E. Fiorentino, and E. Segalla. 2017. “Changes in Prudential Policy Instruments—A New Cross-Country Database.” International Journal of Central Banking 13 (S1).
Roodman, D. (2009) : How to do xtabond2: An introduction to difference and system GMM in Stata. Stata Journal 9(1): 86-136
Roodman, D. (2009), “A note on the theme of too many instruments,” Oxford Bulletin of Economics and Statistics, 71, 135-158
Mileva, E. (2007) "Using Arellano – Bond Dynamic Panel GMM Estimators in Stata, Tutorial with Examples using Stata 9.0 (xtabond and xtabond2)," Economics Department, Fordham University July 9, 2007
Comment