Using the attached data, I am working on the effects of poverty on farmers' adaptive capacities (N=1211). Multivariate analysis is used to isolate interactions between different determinants. My dependent variable (nb_success) measures the number of changes in agricultural practises. The variable is bounded (0-9) and a binomial distribution with a logit link function fits the data well.

However, I have an issue with the poverty variable (DEP_TOT). When running the regression with other determinants, the variable has zero degrees of freedom and SAS displays a value of zero for both the parameter estimate and its standard error

I used the following SAS command where nb_trials is =5:

proc genmod data= data1 ; model nb_success/nb_trials = m002 Dist_marche SS_nb_s SS_part_travaux_agri SS_CeRPA m104 com dep_tot m15 i705 PART_ASSO part_ceremonies i102 i101 Educ p_sp p_temp/ dist=bin link=logit ;run;

My understanding is that SAS finds (DEP_TOT) to be linearly dependent, or aliased, with columns corresponding to parameters preceding it in the model. However, when I run the model with this variable as a unique predicator, the problem persists. Another variable (number of childrens in the households) is correlated with this variable. When I drop the number of children, the issue remains. The model works well when DEP_TOT is log transformed or stratified in deciles and even quartiles.

Any idea where the issue may lies? Would log-transform an independant variable in a GLM be an issue?

Many thanks for your help, I am confused

  • $\begingroup$ Is it possible to disclose your data? $\endgroup$ – Zhanxiong Jul 2 '15 at 17:32
  • $\begingroup$ It sounds like a problem with the data. Does the predictor actually take on distinct values? $\endgroup$ – shadowtalker Jul 2 '15 at 17:32
  • $\begingroup$ Thanks, I attached a link with the problematic predictor $\endgroup$ – Frédéric K. Jul 3 '15 at 6:06
  • $\begingroup$ Would you be so kind as to write what you regress on what? I am not sure I understand which/what model, you are trying to estimate from your question. $\endgroup$ – Repmat Jul 3 '15 at 19:17

NB_trials is always 5, you can't use that in a regression where you also include a intercept. I have no (computational at least) problems with regressing: nb_succes on a constant and dep_tot

  • $\begingroup$ Thanks. The binomial GLM i m run is regressing nb.succes / nb trials. If we do not include the nb of trials and only regress nb success, would it still be a valid GLM binomial model? $\endgroup$ – Frédéric K. Jul 3 '15 at 16:38
  • $\begingroup$ @FrédéricK. post your code in the question as well $\endgroup$ – shadowtalker Jul 3 '15 at 18:14
  • $\begingroup$ @Repmat. Which code did you use? dropping the constant nb_trials from the command results in a SAS error. $\endgroup$ – Frédéric K. Jul 5 '15 at 16:48
  • $\begingroup$ I am not sure we are running the same model then. Would be so kind as to update your question, with a mathematical stated model? $\endgroup$ – Repmat Jul 5 '15 at 18:56
  • $\begingroup$ One thing I do not fully understand is whether it is valid or not to use an independent log varible in a GLM with logit link? $\endgroup$ – Frédéric K. Jul 6 '15 at 15:36

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.