Model specification using xtdpdsys and xtabond2

[Kieran Hughes]

Hi all,

I have previously been using a fixed-effects ordered logit model to test the relationship between unemployment and self-assessed health. My results have indicated that unemployment is significantly related to health. However, when I use duration effects unemployment is only significant in short-term, this implies to me that unemployment is the result of workers with declining health being selected into the unemployed category rather than a result of poor health. I have tried many different specifications with GMM, but I fear I don't understand it well enough, because with xtabond2 my equations normally fail all of the tests, and with xtdpdsys my results always fail the Sargan test.

In my equation, I assume individual characteristics: education (dummies), gender, marriage status and age are exogenous. I assume that log income and employment status are endogenous. My employment status dummies are: unemployed, sick, retired, student and out of labour force. Time periods are measured in waves, the waves vary from 54,000 responses in wave 2 to 27,000 in wave 13.

The code I have currently tried is :

xtabond2 healthcon l.healthcon l2.healthcon unemployed OoLF lnincome GCSE OtherDegree Alevel degree age female married, gmm(healthcon unemployed OoLF lnincome, collapse) iv( married degree GCSE OtherDegree Alevel female age) artests(4) robust twostep small

xtdpdsys healthcon, lags(2) maxldep(2) maxlags (2) twostep pre(age, lagstruct(0,2)) pre(education, lagstruct(0,2)) pre(female, lagstruct(0,2)) endog(unemployed, lagstruct(0,2)) endog(lnincome, lagstruct(0,2)) endog(OOLF, lagstruct(0,2)) endog(student, lagstruct(0,2)) endog(sick, lagstruct(0,2)) endog(retired, lagstruct(0,2)) artests(4)

Initially I tried xtivreg2 and the results said my estimates were efficient for homoskedasticity only. I've tried reading Arellano and Bond (1991) and Roodman (2009) to get a better understanding, but I am worried I still don't understand well enough to improve the specification.


Can anyone help me improve the specification of this model?

Any help will be much appreciated,
Thank you.

[Kieran Hughes]

Solution:

I solved this problem by reducing the number of waves to the first 7 waves of the panel and regressing male and females separately. I also greatly restricted the number of endogenous variables in my gmm equation. By treating unemployment and the dependent variable as the only endogenous regressors passed all of the tests apart form AR(1). The Sargan statistic was 0 when goodhealth is included as the only endogenous variable and 0.3 when unemployment is included, this leads me to reject the null that unemployment is exogenous to health.

My code was:
. xtabond2 goodhealth l.goodhealth l2.goodhealth unemployed sick retired OOLF student lnincome Alevel GCSE degree OtherDegree children married age nonwhite wave2 wave3 wave4 wave5 wave6 wave7 if female==1 & wave<8, gmm(goodhealth unemployed, lag(2 3) collapse) iv(married GCSE Alevel degree student unemployed OOLF retired sick OtherDegree lnincome nonwhite age wave2 wave3 wave4 wave5 wave6 wave7) artests(2) small two robust