Errors

[oic...@gmail.com]

What is the meaning of these errors?

glm.fit: fitted probabilities numerically 0 or 1 occurred
glm.fit: algorithm did not converge

Thanks

[Isaac Subirana]

If I'm not wrong, these messages comes from when estimating the Odds Ratio (OR) of a continuous variable. Probably, they are not errors but warnings and they may indicate that this continuous variable separates the two groups (cases/controls) totally, and provide non-finite OR estimates...

Regards,

Isaac.

[oic...@gmail.com]

I am not estimating the Odds Ratio.
I used the following commands:
dmtab<-compareGroups(dmtab1s, y = dmtab1$Treatment)
and I obtained 8 error messages
Mensajes de aviso perdidos
1: glm.fit: algorithm did not converge
2: glm.fit: fitted probabilities numerically 0 or 1 occurred
3: glm.fit: fitted probabilities numerically 0 or 1 occurred
4: glm.fit: fitted probabilities numerically 0 or 1 occurred
5: glm.fit: fitted probabilities numerically 0 or 1 occurred
6: glm.fit: fitted probabilities numerically 0 or 1 occurred
7: glm.fit: algorithm did not converge
8: glm.fit: fitted probabilities numerically 0 or 1 occurred

Also, the results for p.overall resulted were different from the ones obtained with the t.test function.
Can you help me please?
Thank you very much in advance,
Rocío Zamanillo

[Isaac Subirana]

I suspect that these warnings arise when fitting a logistic model, whether the odds ratio are displayed or not.

In general, the p-values obtained from a logistic model (where the treatment is the response) are different from the ones obtained by a t-test. However, these two p-values should not differ very much.

The p.overall column displayed in the bivariate table computed with compareGroups package corresponds to a t-test (t.test function in R) if the variable is continuous and considered normal-distributed. So, I do not understand why the p.overall values you have are different from the ones obtained with a t-test. Please, could you provide the data to the compareGroups package mantainer, in order to better answer to you question? Thanks.

Isaac. 

[Isaac Subirana]

The warnings messages appeared because there are extreme (not outliers) values in the markers (variables) that makes the fitted probabilities in the logistic model be 0 or 1. But in your case, it does not affect your results, since you report means and SD per group of treatment and p.overall which is computed by a t-test and not a logistic model. So, you can ignore these warnings.

Regarding the p.overall values, the compareGroups performs a t.test without assumiing equal variances (default). In fact, under statistical point of view, it is suggested to compute the Welch t-test directly (default option of t.test function) instead of testing previously whether the variances are equal with a var.test, So, you can check that calling t.test setting var.equal=FALSE provides exactly the same p-value than the p.overall column obtained with compareGroups. 

Isaac.