Logistic Regression - Hessian is not negative semidefinite

[Nicola Shelley Williams]

I am conducting assumption tests for four logistic regression models with different binary outcomes and the same independent variables. I have adjusted independent vars in the models; I have logged one of the continuous variables (GDP per capita) and turned another into a categorical variable. I am using the Box-Tidwell test to verify the assumption of a linear relationship between explanatory variables and the logit of the response variable. For three models, the test is satisfied. For one of the logistic regression models, I am not achieving convergence.

Code
. boxtid logistic outcome i.preorgmil i.preorg_polisoc i.splinter i.merger i.ethnic i.rebhigh i.rebels1 i.govs_1 i.polity2_democracy i.intensity_level2 lgdp_percap totalduration_cat

Iteration 0: Deviance = 214.2661
Iteration 1: Deviance = 211.4495 (change = -2.816559)
Iteration 2: Deviance = 211.4274 (change = -.0221194)
Iteration 3: Deviance = 211.3873 (change = -.0401122)
Iteration 4: Deviance = 211.2426 (change = -.1447378)
Iteration 5: Deviance = 210.8297 (change = -.4128671)
Hessian is not negative semidefinite
r(430);

end of do-file

Question
What else can I do to achieve convergence, given that all independent continuous variables are already transformed?

[Carlo Lazzaro]

Nicola:
the best recipe is to start with amnore parsimonious model, add one more predictor at a time and see when Stata starts complaining.

[Nicola Shelley Williams]

Thank you Carlo! It is the remaining continuous variable, log(GDP_percap), that is the problem. When I remove it, all the assumptions are satisfied for this particular logistic regression model. Would removing it be a reasonable next step? Or, is there another option I should explore? To note, there is only a very minor change (0.01) in the coefficient and no difference in the statistical significance of relationships between IVs and the DV when this variable is removed from the model.