Hello,
I'm working on a university project, investigating the relationship between institutional trust and economic growth.
As a starting point I've tried to replicate a paper studying the same question by Hwang (2017), referenced below.
I got my data from the sources quoted by Hwang (2017). It is panel data with 46 countries and 5 time periods. All variables have been averaged over 4-year periods to avoid the influence of business cycle effects on growth.
Comparing the summary statistics of all variables to the appendix in Hwang (2017), the data look exactly the same. Here is a sample of my "replication data":
The estimator employed is Arellano-Bond Difference GMM, for one specification the paper gives the following Stata command:
"justice" (the trust variable of interest) and the controls "p_inv" (relative price of investment), "pop" (population size) are considered endogenous. The time-dummies are the only exogenous regressors.
I'm using Stata 15.1 on a Mac. My "replication code" and the corresponding results are as follows:
Although the results are similar (signs and significance levels of coefficients is the same as in Hwang (2017)), the coefficients differ in magnitude. Can my way of incorporating the time-fixed effects be a reason for that? Or could a different version of Stata used by Hwang (2017) be the reason for differing results? Unfortunately there is no official replication data or code to compare my work to.
More general econometric questions I have:
(1) In more or less all specifications I've tried, the lagged dependent variable ("growth") has been insignificant. Does that basically show me that it shouldn't be included to the model?
(2) With xtabond2, is it possible to include lags of my independent variable of interest ("justice") as regressors, when I already instrument for that variable with its lags? In Roodman (2009) there are examples where this happens, but I couldn't yet wrap my head around that issue.
(3) When including interaction terms as regressors in xtabond2, do I also need to include the interaction term into the gmmstyle() argument? Or is it enough to have the individual components of the interaction as arguments for gmmstyle()?
(4) Similar to (3), if I include the squared version of a regressor to the model, do I need to include the squared form into gmmstyle(), or just the linear form?
Thank you for any help and advice!
References:
Hwang, In Do, Which Type of Trust Matters?: Interpersonal vs. Institutional vs. Political Trust (May 15, 2017). Bank of Korea WP 2017-15. Available at SSRN: https://ssrn.com/abstract=2967051 or http://dx.doi.org/10.2139/ssrn.2967051
David Roodman, 2009. "How to do xtabond2: An introduction to difference and system GMM in Stata," Stata Journal, StataCorp LP, vol. 9(1), pages 86-136, March.
I'm working on a university project, investigating the relationship between institutional trust and economic growth.
As a starting point I've tried to replicate a paper studying the same question by Hwang (2017), referenced below.
I got my data from the sources quoted by Hwang (2017). It is panel data with 46 countries and 5 time periods. All variables have been averaged over 4-year periods to avoid the influence of business cycle effects on growth.
Comparing the summary statistics of all variables to the appendix in Hwang (2017), the data look exactly the same. Here is a sample of my "replication data":
Code:
* Example generated by -dataex-. To install: ssc install dataex clear input str70 CountryName float period double growth float(pop p_inv) double justice "Argentina" 1 2.58 35.05271 94.7473 1.98 "Argentina" 2 1 36.719788 90.40239 2.4925 "Argentina" 3 -.44 38.19308 51.14445 1.4249999999999998 "Argentina" 4 7.06 39.83163 58.22219 1.845 "Argentina" 5 4.02 41.1284 60.54487 2.8766666666666665 "Australia" 1 2.89 17.890095 89.05154 7.73 "Australia" 2 2.96 18.729885 77.52952 8.120000000000001 "Australia" 3 2.24 19.6477 73.25857 8.6725 "Australia" 4 1.5 20.61969 93.53814 8.592500000000001 "Australia" 5 .84 21.3892 95.87012 8.473333333333334 end
Code:
xtabond2 growth L.growth justice p_inv pop time*, gmm(L.(growth justice p_inv pop)) iv(time*) noleveleq small robust
I'm using Stata 15.1 on a Mac. My "replication code" and the corresponding results are as follows:
Code:
xtabond2 growth L.growth justice p_inv pop i.period, gmm(L.(growth justice p_inv pop)) iv(i.period) noleveleq small robust
Favoring space over speed. To switch, type or click on mata: mata set matafavor speed, perm.
Warning: Two-step estimated covariance matrix of moments is singular.
Using a generalized inverse to calculate robust weighting matrix for Hansen test.
Difference-in-Sargan/Hansen statistics may be negative.
Dynamic panel-data estimation, one-step difference GMM
------------------------------------------------------------------------------
Group variable: CountryName1 Number of obs = 138
Time variable : period Number of groups = 46
Number of instruments = 27 Obs per group: min = 3
F(9, 46) = 10.27 avg = 3.00
Prob > F = 0.000 max = 3
------------------------------------------------------------------------------
| Robust
growth | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
growth |
L1. | .0299311 .1501954 0.20 0.843 -.2723966 .3322588
|
justice | 2.015258 .6446599 3.13 0.003 .7176253 3.312892
p_inv | .0019294 .0456303 0.04 0.966 -.0899196 .0937784
pop | .0457723 .0194937 2.35 0.023 .0065335 .0850111
|
period |
1 | 0 (empty)
2 | 2.928606 .8752231 3.35 0.002 1.166873 4.690338
3 | 1.688182 .8004807 2.11 0.040 .0768983 3.299466
4 | 2.263644 .3493905 6.48 0.000 1.560357 2.966931
5 | 0 (omitted)
------------------------------------------------------------------------------
Instruments for first differences equation
Standard
D.(1b.period 2.period 3.period 4.period 5.period)
GMM-type (missing=0, separate instruments for each period unless collapsed)
L(1/4).(L.growth L.justice L.p_inv L.pop)
------------------------------------------------------------------------------
Arellano-Bond test for AR(1) in first differences: z = -1.49 Pr > z = 0.137
Arellano-Bond test for AR(2) in first differences: z = -0.15 Pr > z = 0.881
------------------------------------------------------------------------------
Sargan test of overid. restrictions: chi2(18) = 46.53 Prob > chi2 = 0.000
(Not robust, but not weakened by many instruments.)
Hansen test of overid. restrictions: chi2(18) = 19.25 Prob > chi2 = 0.376
(Robust, but weakened by many instruments.)
Difference-in-Hansen tests of exogeneity of instrument subsets:
iv(1b.period 2.period 3.period 4.period 5.period)
Hansen test excluding group: chi2(15) = 15.07 Prob > chi2 = 0.447
Difference (null H = exogenous): chi2(3) = 4.19 Prob > chi2 = 0.242
More general econometric questions I have:
(1) In more or less all specifications I've tried, the lagged dependent variable ("growth") has been insignificant. Does that basically show me that it shouldn't be included to the model?
(2) With xtabond2, is it possible to include lags of my independent variable of interest ("justice") as regressors, when I already instrument for that variable with its lags? In Roodman (2009) there are examples where this happens, but I couldn't yet wrap my head around that issue.
(3) When including interaction terms as regressors in xtabond2, do I also need to include the interaction term into the gmmstyle() argument? Or is it enough to have the individual components of the interaction as arguments for gmmstyle()?
(4) Similar to (3), if I include the squared version of a regressor to the model, do I need to include the squared form into gmmstyle(), or just the linear form?
Thank you for any help and advice!
References:
Hwang, In Do, Which Type of Trust Matters?: Interpersonal vs. Institutional vs. Political Trust (May 15, 2017). Bank of Korea WP 2017-15. Available at SSRN: https://ssrn.com/abstract=2967051 or http://dx.doi.org/10.2139/ssrn.2967051
David Roodman, 2009. "How to do xtabond2: An introduction to difference and system GMM in Stata," Stata Journal, StataCorp LP, vol. 9(1), pages 86-136, March.
