Dear Statalist community,
I am currently struggling with several questions regarding my time-series poisson regression (Stata 15.1) and was hoping to get your feedback on this.
My time-series covers only 17 years, with the dependent variable being a count (99 policy announcements of one country) and the independent variables being a set of continuously measured factors of economic change (e.g., changes in the country’s productivity level). All independent variables are lagged by 1 year and standardized to have a mean of 0 and sd of 1.
From what I read in the forum (1) (2) (3) (4), glm …, family(poisson) link(log) vce(hac nwest n) t(year) might be suitable to account for the time-series characteristics of the dataset. Further, with using Newey-West standard errors from Hardin “vce(hac nwest)” it seems to be possible to account for autocorrelation and heteroskedasticity problems.
Code:
Result:
Following Hilbe (2011: 177) as well as this forums’ suggestion, I chose a poisson instead of a negative binomial, as negative binomial models cannot properly estimate under dispersed models ((1/df) Pearson = .3978726).
Reading on small samples and short time-series, I tried to include the bootstrap option into the glm model to figure out if my sample suffers from the standard error bias and also to tackle the underdispersion. However, bootstrapping seems not possible for time-series or I might have done something wrong.
Hence, my questions:
1. How to handle the underdispersion in my model?
2. Is there any goodness-of-fit test for this short time-series?
3. Any other suggestions on how to improve this short time-series?
I would very much appreciate getting your insights on this (and also on what else you see that might be wrong with this model).
Thank you very much – I highly appreciate your responses.
Fredrik
Reference:
Hilbe, J.M. (2011), Negative Binomial Regression, 2nd ed., Cambridge University Press, New York.
I am currently struggling with several questions regarding my time-series poisson regression (Stata 15.1) and was hoping to get your feedback on this.
My time-series covers only 17 years, with the dependent variable being a count (99 policy announcements of one country) and the independent variables being a set of continuously measured factors of economic change (e.g., changes in the country’s productivity level). All independent variables are lagged by 1 year and standardized to have a mean of 0 and sd of 1.
From what I read in the forum (1) (2) (3) (4), glm …, family(poisson) link(log) vce(hac nwest n) t(year) might be suitable to account for the time-series characteristics of the dataset. Further, with using Newey-West standard errors from Hardin “vce(hac nwest)” it seems to be possible to account for autocorrelation and heteroskedasticity problems.
Code:
Code:
glm Policies l.z1Pol l.z1Ind1 l.z1Ind2 l.z1Nat1 l.z1Nat2 l.z1Nat3, family(poisson) link(log) vce(hac nwest 14) vfactor(1.5454545) t(year)
Following Hilbe (2011: 177) as well as this forums’ suggestion, I chose a poisson instead of a negative binomial, as negative binomial models cannot properly estimate under dispersed models ((1/df) Pearson = .3978726).
Reading on small samples and short time-series, I tried to include the bootstrap option into the glm model to figure out if my sample suffers from the standard error bias and also to tackle the underdispersion. However, bootstrapping seems not possible for time-series or I might have done something wrong.
Hence, my questions:
1. How to handle the underdispersion in my model?
2. Is there any goodness-of-fit test for this short time-series?
3. Any other suggestions on how to improve this short time-series?
I would very much appreciate getting your insights on this (and also on what else you see that might be wrong with this model).
Thank you very much – I highly appreciate your responses.
Fredrik
Reference:
Hilbe, J.M. (2011), Negative Binomial Regression, 2nd ed., Cambridge University Press, New York.
Comment