Convergence problems

[Andrea Parisi]

Hi,

I am trying to calculate RR using logistic regression but there are apparently some convergence problems:

. glm resistance_new i.sa4code_analysis, family (binomial) link(log) eform

Iteration 0: log likelihood = -4999.9042 (not concave)
Iteration 1: log likelihood = -3998.9689 (not concave)
Iteration 2: log likelihood = -3961.6864 (not concave)
Iteration 3: log likelihood = -3954.0891 (not concave)
Iteration 4: log likelihood = -3953.9001 (not concave)
Iteration 5: log likelihood = -3953.8996 (not concave)
Iteration 6: log likelihood = -3953.8996 (not concave)
Iteration 7: log likelihood = -3953.8996 (not concave)
Iteration 8: log likelihood = -3953.8996 (not concave)
Iteration 9: log likelihood = -3953.8996 (not concave)
Iteration 10: log likelihood = -3953.8996 (not concave)
Iteration 11: log likelihood = -3953.8996 (not concave)
Iteration 12: log likelihood = -3953.8996 (not concave)
Iteration 13: log likelihood = -3953.8995 (not concave)
Iteration 14: log likelihood = -3953.8995 (not concave)
Iteration 15: log likelihood = -3953.8995 (not concave)
Iteration 16: log likelihood = -3953.8995 (not concave)
Iteration 17: log likelihood = -3953.8995 (not concave)
Iteration 18: log likelihood = -3953.8995 (not concave)


When I specify number of iterations at 20, I get this:

. glm resistance_new i.sa4code, family (binomial) link(log) eform iter(20)

Iteration 0: log likelihood = -4999.9042 (not concave)
Iteration 1: log likelihood = -3998.9689 (not concave)
Iteration 2: log likelihood = -3961.6864 (not concave)
Iteration 3: log likelihood = -3954.0891 (not concave)
Iteration 4: log likelihood = -3953.9001 (not concave)
Iteration 5: log likelihood = -3953.8996 (not concave)
Iteration 6: log likelihood = -3953.8996 (not concave)
Iteration 7: log likelihood = -3953.8996 (not concave)
Iteration 8: log likelihood = -3953.8996 (not concave)
Iteration 9: log likelihood = -3953.8996 (not concave)
Iteration 10: log likelihood = -3953.8996 (not concave)
Iteration 11: log likelihood = -3953.8996 (not concave)
Iteration 12: log likelihood = -3953.8996 (not concave)
Iteration 13: log likelihood = -3953.8995 (not concave)
Iteration 14: log likelihood = -3953.8995 (not concave)
Iteration 15: log likelihood = -3953.8995 (not concave)
Iteration 16: log likelihood = -3953.8995 (not concave)
Iteration 17: log likelihood = -3953.8995 (not concave)
Iteration 18: log likelihood = -3953.8995 (not concave)
Iteration 19: log likelihood = -3953.8995 (not concave)
Iteration 20: log likelihood = -3953.8995 (not concave)
convergence not achieved

Generalized linear models No. of obs = 7,651
Optimization : ML Residual df = 7,633
Scale parameter = 1
Deviance = 7907.799058 (1/df) Deviance = 1.036001
Pearson = 6400006511 (1/df) Pearson = 838465.4

Variance function: V(u) = u*(1-u) [Bernoulli]
Link function : g(u) = ln(u) [Log]

AIC = 1.038269
Log likelihood = -3953.899529 BIC = -60351

--------------------------------------------------------------------------------
| OIM
resistance_new | Risk Ratio Std. Err. z P>|z| [95% Conf. Interval]
---------------+----------------------------------------------------------------
sa4code |
302 | .3186184 .0543363 -6.71 0.000 .2280916 .4450742
303 | .27928 .0513765 -6.93 0.000 .1947387 .4005229
304 | .4046953 .1011731 -3.62 0.000 .2479301 .6605824
305 | .2603081 .0624054 -5.61 0.000 .1627134 .4164395
306 | .231827 .0414789 -8.17 0.000 .1632547 .3292019
307 | .4443705 .1560037 -2.31 0.021 .2233144 .8842472
308 | .5627156 .0598641 -5.40 0.000 .4568089 .6931758
309 | .2273888 .0338319 -9.95 0.000 .1698727 .304379
310 | .5681479 .0649963 -4.94 0.000 .4540292 .71095
311 | .3718394 .0459352 -8.01 0.000 .2918784 .4737058
312 | 2.470708 .0673254 33.19 0.000 2.342214 2.60625
313 | .2835836 .0781954 -4.57 0.000 .165185 .4868458
314 | .327395 .059736 -6.12 0.000 .2289618 .4681458
315 | .6347662 .0340739 -8.47 0.000 .5713758 .7051894
316 | 2.231104 .0917053 19.52 0.000 2.058415 2.418282
317 | 1.011726 .2652771 0.04 0.965 .6051682 1.691415
318 | 3.322281 . . . . .
319 | .59012 .0511128 -6.09 0.000 .4979826 .6993048
|
_cons | .3088778 3.86e-10 -9.4e+08 0.000 .3088778 .3088778
--------------------------------------------------------------------------------
Note: _cons estimates baseline risk.
Warning: parameter estimates produce inadmissible mean estimates in one or
more observations.
Warning: convergence not achieved


Apparently, there is a problem with sa4code=318. Is there any way I could correct this? Option of converting this categorical variable to continuous is not possible in this case. I tried to change the baseline but didn't work either.

Thanks in advance.

Andrea

[Richard Williams]

Do you want the link to be log or logit?

Output would be much easier to read if you used code tags. See pt #12 in the FAQ.

[Rich Goldstein]

for RR, there are well-known problems using -glm-; I recommend using -poisson- to get IRR (essentially the same thing), just remember to use "vce(robust)" as an option; the main problem with using -poisson- for this is inefficiency; thus, if your sample size is small, you have no good alternatives