Inverse Proability Weighted Regression Adjustment - Models do not converge

[Kevin Teopista Akoyi]

I run several IPWRA models using the teffects ipwra command in stata with 4 outcome variables; Coffee production, Coffee income, Household income and Headcount poverty, against a large set of household characteristics, farm characteristics and village characteristics as independent variable. Some of the models with Coffee income and household income however, do not converge, so the standard errors are missing or cannot be calculated. How can I solve this problem of non-convergence? My aim is to estimate the impact of coffee certification on household income and poverty

[Clyde Schechter]

There are a number of reasons models fail to converge. You need to show the exact commands you gave and the exact response you got from Stata. There are often clues lurking in the details of the output. But without seeing that, it's hard to advise. Copy (do not retype) the command(s) and Stata response directly from the Results window or your log file directly into a code block in the forum editor. (See FAQ for how to create a code block.) Perhaps with more information somebody will be able to help.

[Kevin Teopista Akoyi]

Dear Clyde,

Thanks for the advice. See the model I run and the response I got:

. teffects ra (HH_povC educ_head fem_head age_hhead age_hhead2 HHsize_adults HHsize_children Total_area_cult Total_area_cult2 livestock_unit value_agequipment hmstd_
> forest_km hmstd_river_km hmstd_road_km, logit) (cert), aequations

Iteration 0: EE criterion = 3.510e-19
Iteration 1: EE criterion = 2.620e-32
convergence not achieved
The Gauss-Newton stopping criterion has been met but missing standard errors indicate some of the parameters are not identified.

. teffects ipwra (Tot_coffee_harvest educ_head fem_head age_hhead age_hhead2 HHsize_adults HHsize_children Area_coffee Area_coffee2 livestock_unit value_agequipment
> hmstd_forest_km hmstd_river_km hmstd_road_km, linear) (cert BCU_yrs educ_head age_hhead Total_area_cult Total_area_cult2), aequations iterate(20)

Iteration 0: EE criterion = 2.411e-17
Iteration 1: EE criterion = 5.766e-25
convergence not achieved
The Gauss-Newton stopping criterion has been met but missing standard errors indicate some of the parameters are not identified.

. teffects ipwra (INC_coffee_z educ_head fem_head age_hhead age_hhead2 HHsize_adults HHsize_children Area_coffee Area_coffee2 livestock_unit value_agequipment hmstd_
> forest_km hmstd_river_km hmstd_road_km, linear) (cert BCU_yrs educ_head age_hhead Total_area_cult Total_area_cult2), aequations iterate(20)

Iteration 0: EE criterion = 1.295e-15 (not concave)
Iteration 1: EE criterion = 1.083e-15 (not concave)
Iteration 2: EE criterion = 9.945e-16 (not concave)
Iteration 3: EE criterion = 9.365e-16 (not concave)
Iteration 4: EE criterion = 8.973e-16 (not concave)
Iteration 5: EE criterion = 8.697e-16 (not concave)
Iteration 6: EE criterion = 8.402e-16 (not concave)
Iteration 7: EE criterion = 8.402e-16 (not concave)
Iteration 8: EE criterion = 8.402e-16 (not concave)
Iteration 9: EE criterion = 8.402e-16 (not concave)
Iteration 10: EE criterion = 8.402e-16 (not concave)
Iteration 11: EE criterion = 8.402e-16 (not concave)
Iteration 12: EE criterion = 8.402e-16 (not concave)
Iteration 13: EE criterion = 8.402e-16 (not concave)
Iteration 14: EE criterion = 8.402e-16 (not concave)
Iteration 15: EE criterion = 8.402e-16 (not concave)
Iteration 16: EE criterion = 8.402e-16 (not concave)
Iteration 17: EE criterion = 8.402e-16 (not concave)
Iteration 18: EE criterion = 8.402e-16 (not concave)
Iteration 19: EE criterion = 8.402e-16 (not concave)
Iteration 20: EE criterion = 8.402e-16 (not concave)
convergence not achieved
convergence not achieved
The Gauss-Newton stopping criterion has been met but missing standard errors indicate some of the parameters are not identified.

Treatment-effects estimation Number of obs = 600
Estimator : IPW regression adjustment
Outcome model : linear
Treatment model: (multinomial) logit

[Kevin Teopista Akoyi]

Dear Statalisters,

Am having problems with non convergence when I used the teffects ipwra command, as indicated below: Is there anyone who can suggest a solution?

. teffects ra (HH_povC educ_head fem_head age_hhead age_hhead2 HHsize_adults HHsize_children Total_area_cult Total_area_cult2 livestock_unit value_agequipment hmstd_
> forest_km hmstd_river_km hmstd_road_km, logit) (cert), aequations

Iteration 0: EE criterion = 3.510e-19
Iteration 1: EE criterion = 2.620e-32
convergence not achieved
The Gauss-Newton stopping criterion has been met but missing standard errors indicate some of the parameters are not identified.

. teffects ipwra (Tot_coffee_harvest educ_head fem_head age_hhead age_hhead2 HHsize_adults HHsize_children Area_coffee Area_coffee2 livestock_unit value_agequipment
> hmstd_forest_km hmstd_river_km hmstd_road_km, linear) (cert BCU_yrs educ_head age_hhead Total_area_cult Total_area_cult2), aequations iterate(20)

Iteration 0: EE criterion = 2.411e-17
Iteration 1: EE criterion = 5.766e-25
convergence not achieved
The Gauss-Newton stopping criterion has been met but missing standard errors indicate some of the parameters are not identified.

. teffects ipwra (INC_coffee_z educ_head fem_head age_hhead age_hhead2 HHsize_adults HHsize_children Area_coffee Area_coffee2 livestock_unit value_agequipment hmstd_
> forest_km hmstd_river_km hmstd_road_km, linear) (cert BCU_yrs educ_head age_hhead Total_area_cult Total_area_cult2), aequations iterate(20)

Iteration 0: EE criterion = 1.295e-15 (not concave)
Iteration 1: EE criterion = 1.083e-15 (not concave)
Iteration 2: EE criterion = 9.945e-16 (not concave)
Iteration 3: EE criterion = 9.365e-16 (not concave)
Iteration 4: EE criterion = 8.973e-16 (not concave)
Iteration 5: EE criterion = 8.697e-16 (not concave)
Iteration 6: EE criterion = 8.402e-16 (not concave)
Iteration 7: EE criterion = 8.402e-16 (not concave)
Iteration 8: EE criterion = 8.402e-16 (not concave)
Iteration 9: EE criterion = 8.402e-16 (not concave)
Iteration 10: EE criterion = 8.402e-16 (not concave)
Iteration 11: EE criterion = 8.402e-16 (not concave)
Iteration 12: EE criterion = 8.402e-16 (not concave)
Iteration 13: EE criterion = 8.402e-16 (not concave)
Iteration 14: EE criterion = 8.402e-16 (not concave)
Iteration 15: EE criterion = 8.402e-16 (not concave)
Iteration 16: EE criterion = 8.402e-16 (not concave)
Iteration 17: EE criterion = 8.402e-16 (not concave)
Iteration 18: EE criterion = 8.402e-16 (not concave)
Iteration 19: EE criterion = 8.402e-16 (not concave)
Iteration 20: EE criterion = 8.402e-16 (not concave)
convergence not achieved
convergence not achieved
The Gauss-Newton stopping criterion has been met but missing standard errors indicate some of the parameters are not identified.

Treatment-effects estimation Number of obs = 600
Estimator : IPW regression adjustment
Outcome model : linear
Treatment model: (multinomial) logit

[Clyde Schechter]

One possibility is that there is near collinearity among some of your predictor variables, so that their separate effects are not identified.

My advice here would be to start over with a much simpler model incorporating just one predictor, and then add predictor variables one at a time until you encounter difficulties. That will help you identify the "culprit."

You don't show us the regression output beyond the iterations. If not all of the standard errors are missing, the variables causing problems will almost certainly be among those whose standard errors are missing. So you can probably start out retaining those variables for which you got standard errors and then adding in from there to find out where things fall apart. One caveat to that: another possibility for non-convergence of a logit model is that some of your variables come close to separating the response groups. You would recognize variables like that in the regression coefficient table by an unreasonably large magnitude regression coefficient--any such variable is highly suspect.

[Arthur Langlois-Droal]

Clyde Schechter I am encountering a similar issue, and have isolated the problem to a single binary variable. From your explanation, it would seem that the non-convergence is occurring because the variable comes "close to separating the response groups", as its coefficient is unusually large and does not seem to be collinear with my target variable. Would you please mind clarifying what you mean by separating the response groups, if the independent is not strongly correlated with the target? Thanks in advance!

[Clyde Schechter]

I don't know what you mean by "target variable." If you mean the key independent variable in the model, estimating the effect of which is a key research goal, then the correlation between that and the offending variable with an outlandishly large coefficient is irrelevant.

If you mean the outcome (dependent) variable of the model, it cannot be the case that your offending variable is close to separating the response groups and yet it is not correlated with the outcome. In that case, it may be that the outlandishly large coefficient arises from some extreme outlier values or some other problem. With no code or actual results to examine, nor any data example that reproduces your problem, nothing more can be said.