Interrupted Time series (ITS) - Convergence not achieved

[Anseh Danesharasteh]

Dear STATA members,


I am running an ITS on a panel data using the start date of a testing program (as my interruption date). When I run xtitsa command using just my dependent variable, I get the error of "convergence not achieved".

Can anyone help me with that, please?

Best,
Anseh

gen date1= date(date, "YMD")
gen surv_date= date(surv_policy_go_live_date, "YMD")

gen weeks_from_surv_start= round((date1 - surv_date)/14)

egen pannel_var= group(related_organization poc_test_type reason_for_poc_test test_type account_role__c test_result)

gen weeks_from_surv_start_pos = weeks_from_surv_start + 21

xtset pannel_var weeks_from_surv_start

xtitsa positivity_rate if weeks_from_surv_start < 12 & weeks_from_surv_start > -8, single trperiod(0) posttrend figure replace





. xtset pannel_var weeks_from_surv_start

Panel variable: pannel_var (unbalanced)
 Time variable: weeks_from_surv_start, -21 to 22, but with gaps
         Delta: 1 unit

.
end of do-file


. xtitsa positivity_rate if weeks_from_surv_start < 12 & weeks_from_surv_start > -8, single trperiod(0) posttrend figure replace



Panel variable: pannel_var (unbalanced)
 Time variable: weeks_from_surv_start, -21 to 22, but with gaps
         Delta: 1 unit

(resetting alpha to -0.0582)
Iteration 1: tolerance = .00329983
(resetting alpha to -0.0582)
Iteration 2: tolerance = 0

GEE population-averaged model                       Number of obs    =  33,245
Group variable: pannel_var                          Number of groups =   8,657
Family: Gaussian                                    Obs per group:  
Link:   Identity                                                 min =       1
Correlation: exchangeable                                        avg =     3.8
                                                                 max =      18
                                                    Wald chi2(3)     = 3699.33
Scale parameter = .0011869                          Prob > chi2      =  0.0000

------------------------------------------------------------------------------
_positivit~e | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
          _t |  -.0005946   .0000525   -11.33   0.000    -.0006974   -.0004918
         _x0 |  -.0169303   .0007456   -22.71   0.000    -.0183917   -.0154689
       _x_t0 |   .0038683   .0000667    58.01   0.000     .0037377     .003999
       _cons |   .0357383   .0006089    58.70   0.000     .0345449    .0369316
------------------------------------------------------------------------------
convergence not achieved
r(430);

end of do-file

r(430);

[Jared Greathouse]

No idea. Likely has to do with the unbalanced panel- your data has a panel with one observation. Why? And, 8657 panels at that. Are you sure ITS is the right design here?

EDIT: Also, your method uses maximum likelihood. So, maybe due to the issues with the panel structure (maybe it needs a balanced panel), ti cannot estimate the log-likelihood

[Anseh Danesharasteh]

Thank you for your response, Jared.

So the data includes testing information from late August 2021 to late June 2022 in a biweekly period. Not all test centers have information in all dates. I think that is the reason of unbalanced panel.

I am interested to see the changes (in the slope, if any) after the intervention and that is why I choose ITSA. I tried eventdd but was not that much informative.

For the panel with one observation, I have no idea! Any suggestions on how can I check that?

And any suggestions on how I can fix this problem?