[R] Warning message: In eval(expr, envir, enclos) : non-integer #successes in a binomial glm!

[David Winsemius]

On Dec 10, 2012, at 5:40 AM, Simon Kiss wrote:

> Hi there
> I'm trying to fit a logistic regression model to data that looks  
> very similar to the data in the sample below.   I don't understand  
> why I'm getting this error; none of the data are proportional and  
> the weights are numeric values.  Should I be concerned about the  
> warning about non-integer successes in my binomial glm? If I should  
> be, how do I go about addressing it?
> I'm pretty sure the weights in the data frame are sampling weights.
>
> What follows is the result of str() on my data, the series of  
> commands I'm using to fit the model, the responses I'm getting and  
> then some code to reproduce the data and go through the same steps  
> with that code.  One last (minor) question.  When calling svyglm on  
> the sample data, I actually get some information about the model  
> fitting results as well as the error about non-integer successes.   
> In my real data, you only get the warning. Calling summary(mod1) on  
> the real data does return information about the coefficients and the  
> model fitting.
>
> I'm grateful for any help. I'm aware that the topic of non-integer  
> successes has been addressed before, but I could not find my answer  
> to this question.
>
> Yours, Simon Kiss
>
> ######str() on original data
> str(mat1)
> 'data.frame':	1001 obs. of  5 variables:
> $ prov  : Factor w/ 4 levels "Ontario","PQ",..: 2 2 2 2 2 2 2 2 2  
> 2 ...
> $ edu   : Factor w/ 2 levels "secondary","post-secondary": 2 2 2 1 1  
> 2 2 2 1 1 ...
> $ gender: Factor w/ 2 levels "Male","Female": 1 1 2 2 2 2 1 1 2 2 ...
> $ weight: num  1.145 1.436 0.954 0.765 0.776 ...
> $ trust : Factor w/ 2 levels "no trust","trust": 2 1 1 1 1 2 1 2 1  
> 2 ...
>
> #######Set up survey design
> des.1<-svydesign(~0, weights=~weight, data=mat1)
>
> #######model and response to svyglm
> mod1<-svyglm(trust ~ gender+edu+prov, design=des.1,  
> family='binomial').

>
> Warning message:
> In eval(expr, envir, enclos) : non-integer #successes in a binomial  
> glm!

I also got:

> 2: glm.fit: algorithm did not converge


And the gender coefficient was NA.

Your problem is not reproducible because no seed was set. Does this  
suggest any issues?

with(mat.test, table(trust, gender)  )
          gender
trust     Female Male
   notrust    300  100
   trust        0  200

-- 
David.

>
> ########Model Summary
> summary(mod1)
>
> Call:
> svyglm(formula = trust ~ gender + edu + prov, design = des.1,
>    family = "binomial")
>
> Survey design:
> svydesign(~0, weights = ~weight, data = mat1)
>
> Coefficients:
>                   Estimate Std. Error t value Pr(>|t|)
> (Intercept)       -0.625909   0.156560  -3.998 6.87e-05 ***
> genderFemale       0.013519   0.140574   0.096    0.923
> edupost-secondary -0.011569   0.141528  -0.082    0.935
> provPQ            -0.006614   0.172105  -0.038    0.969
> provatl            0.335166   0.297860   1.125    0.261
> provwest          -0.053862   0.174826  -0.308    0.758
> ---
> Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
>
> (Dispersion parameter for binomial family taken to be 1.002254)
>
> Number of Fisher Scoring iterations: 4
>
>
>
>
> #########Attempt To Reproduce The Problem
> ########Data
> mat.test<-data.frame(edu=c(rep('secondary', 300), rep('post- 
> secondary', 300)), prov=c(rep('ON', 200), rep('PQ', 200), rep('AB',  
> 200)), trust=c(rep('trust',200), rep('notrust',400)),  
> gender=c(rep('Male', 300), rep('Female', 300)), weight=rnorm(600,  
> mean=1, sd=0.3))
> #######Survey Design object
> test<-svydesign(~0, weights=~weight, data=mat.test)
>
> #####Call To svyglm
> svyglm(trust ~ edu+prov+gender, design=test, family='binomial')
>
> #Reults
> Independent Sampling design (with replacement)
> svydesign(~0, weights = ~weight, data = mat.test)
>
> Call:  svyglm(formula = trust ~ edu + prov + gender, design = test,
>    family = "binomial")
>
> Coefficients:
> (Intercept)  edusecondary        provON        provPQ    genderMale
>  -2.658e+01    -8.454e-04     5.317e+01    -1.408e-02            NA
>
> Degrees of Freedom: 599 Total (i.e. Null);  596 Residual
> Null Deviance:	    759.6
> Residual Deviance: 3.406e-09 	AIC: 8
> Warning messages:
> 1: In eval(expr, envir, enclos) :
>  non-integer #successes in a binomial glm!
> 2: glm.fit: algorithm did not converge

-- 

David Winsemius, MD
Alameda, CA, USA