An error occurred when executed xtnbreg

[Geralt Ji]

Hi Stata members,

I'm currently running -xtnbreg- to explore the association between No_admission (Number of hospital admissions) and phi (private health insurance status) with a rich set of covariates (interaction term, secondary care, expenditure and so on) in Stata/SE 16 . As a dependent variable, No_admission is an integer variable, ranging from 0 to 120. It is not continuous and takes only certain values. And phi is a binary variable where 0 refers to 'Without insurance' 1 refers to 'Insured'. This panel data involves 3 waves of survey and information on around 1000 observations and 500 groups/

When I run the following codes, the process takes forever until I click the 'Break' button. Then Stata will pop up a warning: an error occurred when jackknife/bootstrap executed xtnbreg. I have also tried to report the type of std. err. by default or by vce(oim) option, the Stata showed continuous messages such as 'Iteration 1..: log likelihood = -44xx.xxxx (not concave)'

xtnbreg No_admission hos_status interaction_phi secondary_care all_expenditure $xvars_fe, re vce(boot) i(waveid)
xtnbreg No_admission hos_status interaction_phi secondary_care all_expenditure $xvars_fe, fe vce(jacknife) i(waveid)

I'm wondering what happened to my dataset? Does the warnings suggest that I'm using wrong method to predict the impacts of phi on the number of hospital admissions? Because I have read previous post Model for dependent variable with integer values of certain magnitude - Statalist and am currently think about trying other model to explore the association.

Warm regards,
Geralt

[Carlo Lazzaro]

Geralt:
with MLE the risk of no convergence cannot be ruled out.
The usual recipe is to start out with a more parsimonious regression model adding one predictor at a time and diagnose when Stata starts to gasp.

[Geralt Ji]

Geralt:
with MLE the risk of no convergence cannot be ruled out.
The usual recipe is to start out with a more parsimonious regression model adding one predictor at a time and diagnose when Stata starts to gasp.

Hi Carlo,

Thanks for your suggestion! I deleted 'secondary care' (a variable refers to who used hospital care) and 'interaction_phi' (the product of secondary_care and phi status) and then the revised codes succeeded for vce(oim), vce(boot) and vce(jackknife).

But I still have questions about this error:
1. In view of these two deleted variables, does they suggest that this error is related to collinearity?
2. Some red 'x' and black dots were displayed when I chose to report std. error by Bootstrap. Existing posts show that red 'x' is displayed if command returns an error (svybootstrap_options.pdf (stata.com)) and it could be a convergence failure or prefect prediction. (Re: st: Question concerning Bootstrapping in Stata), but Stata still can produce results successfully. I'm wondering whether these red 'x' matter?

. xtnbreg No_admission hos_status all_expenditure $xvars_fe, re vce(boot) i(waveid)
(running xtnbreg on estimation sample)

Bootstrap replications (50)
----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5 
xxxx.xxxxxxxxxxxx.xxxxx.xxxx.xxxxxxx.xxxxxxx.xx.xx    50

[Carlo Lazzaro]

Geralt:
1) all these issues seem to be related to convergence problems;
2) the cautionary tale is to limit your non-default standard errors to vce(jacknife).

[Geralt Ji]

Geralt:
1) all these issues seem to be related to convergence problems;
2) the cautionary tale is to limit your non-default standard errors to vce(jacknife).

Thanks for your patience and explanation, Carlo!