R help - Apr 2010 - Weights in binomial glm

I have some questions about the use of weights in binomial glm as I am
not getting the results I would expect. In my case the weights I have
can be seen as 'replicate weights'; one respondent i in my dataset
corresponds to w[i] persons in the population. From the documentation
of the glm method, I understand that the weights can indeed be used
for this: "For a binomial GLM prior weights are used to give the
number of trials when the response is the proportion of
successes."
>From "Modern applied statistics with S-Plus 3rd ed." I understand
the
same.

However, I am getting some strange results. I generated an example:

Generate some data which is simular to my dataset
> Z <- rbinom(1000, 1, 0.1)
> W <- round(rnorm(1000, 100, 40))
> W[W < 1] <- 1
Probability of success can either be estimated using:
> sum(Z*W)/sum(W)
[1] 0.09642109

Or using glm:
> model <- glm(Z ~ 1, weights=W, family=binomial())
Warning message:
In glm.fit(x = X, y = Y, weights = weights, start = start, etastart etastart,  :
  fitted probabilities numerically 0 or 1 occurred
> predict(model, type="response")[1]
           1
2.220446e-16

These two results are obviously not the same. The strange thing is
that when I scale the weights, such that the total equals one, the
probability is correctly estimated:
> model <- glm(Z ~ 1, weights=W/sum(W), family=binomial())
Warning message:
In eval(expr, envir, enclos) : non-integer #successes in a binomial
glm!
> predict(model, type="response")[1]
         1
0.09642109


However scaling of the weights should, as far as I am aware, not have
an effect on the estimated parameters. I also tried some other
scalings. And, for example scaling the weights by 20 also gives me the
correct result.
> model <- glm(Z ~ 1, weights=W/20, family=binomial())
Warning message:
In eval(expr, envir, enclos) : non-integer #successes in a binomial
glm!
> predict(model, type="response")[1]
         1
0.09642109


Am I misinterpreting the weights? Could this be a numerical problem?

Regards,

Jan

Jan,

It looks like you did not understand the line "For a binomial GLM prior
weights are used to give the number of trials when the response is the
proportion of successes."

Weights must be a number of trials (hence integer). Not a proportion of
a population. Here is an example that clarifies the use of weights.

library(boot)
library(reshape)
dataset <- data.frame(Person = c(rep("A", 20), rep("B",
10)), Success c(rbinom(20, 1, 0.25), rbinom(10, 1, 0.75)))
Aggregated <- cast(Person ~ ., data = dataset, value = "Success",
fun list(mean, length))

m0 <- glm(Success ~ 1, data = dataset, family = binomial)
m1 <- glm(mean ~ 1, data = Aggregated, family = binomial, weights length)

inv.logit(coef(m0))
inv.logit(coef(m1))

Have a look at the survey package is you want to analyse stratified
data.

Thierry

------------------------------------------------------------------------
----
ir. Thierry Onkelinx
Instituut voor natuur- en bosonderzoek
team Biometrie & Kwaliteitszorg
Gaverstraat 4
9500 Geraardsbergen
Belgium

Research Institute for Nature and Forest
team Biometrics & Quality Assurance
Gaverstraat 4
9500 Geraardsbergen
Belgium

tel. + 32 54/436 185
Thierry.Onkelinx at inbo.be
www.inbo.be

To call in the statistician after the experiment is done may be no more
than asking him to perform a post-mortem examination: he may be able to
say what the experiment died of.
~ Sir Ronald Aylmer Fisher

The plural of anecdote is not data.
~ Roger Brinner

The combination of some data and an aching desire for an answer does not
ensure that a reasonable answer can be extracted from a given body of
data.
~ John Tukey
  
> -----Oorspronkelijk bericht-----
> Van: r-help-bounces at r-project.org 
> [mailto:r-help-bounces at r-project.org] Namens Jan van der Laan
> Verzonden: vrijdag 16 april 2010 14:11
> Aan: r-help at r-project.org
> Onderwerp: [R] Weights in binomial glm
> 
> I have some questions about the use of weights in binomial 
> glm as I am not getting the results I would expect. In my 
> case the weights I have can be seen as 'replicate weights'; 
> one respondent i in my dataset corresponds to w[i] persons in 
> the population. From the documentation of the glm method, I 
> understand that the weights can indeed be used for this: "For 
> a binomial GLM prior weights are used to give the number of 
> trials when the response is the proportion of successes."
> >From "Modern applied statistics with S-Plus 3rd ed." I
understand the
> same.
> 
Druk dit bericht a.u.b. niet onnodig af.
Please do not print this message unnecessarily.

Dit bericht en eventuele bijlagen geven enkel de visie van de schrijver weer 
en binden het INBO onder geen enkel beding, zolang dit bericht niet bevestigd is
door een geldig ondertekend document. The views expressed in  this message 
and any annex are purely those of the writer and may not be regarded as stating 
an official position of INBO, as long as the message is not confirmed by a duly 
signed document.

Jan,

Thierry is correct in saying that you are misusing glm(), but there is also a
numerical problem.

You are misusing glm() because your model specification claims to have
Binomial(n,p) observations with w in the vicinity of 100, where there is a
single common p but the observed binomial proportion is either 1 or 0, never
anything in between.  These data are a very poor fit to a binomial model.

The correct specification if you have what you call replicate weights and I call
frequency weights is to produce a single data record for each covariate pattern
that has both the 1 and 0 observations. This can either be two columns for
successes and failures, or one column of proportions and one column of weights. 
As your quote from MASS says "weights are used to give the number of trials
when the response is the proportion of successes." In your data the
response is *not* the proportion of successes.


However, the MLE should still be equal to the weighted mean even with this
misuse.  The reason it is not is because of the starting values.  R has to find
some starting values for the iterative maximization of the likelihood, and for
binomial data with y successes out of n it uses  starting values for the fitted
means of  (y+0.5)/(n+1).  Starting the iteration at the data in this way usually
makes the Fisher scoring algorithm very reliable -- it is correctly scaled to
the data, in some sense.   Unfortunately, if you separate out the successes and
failures, you have some points starting with values very close to 0.  When I
used your code the starting value for the point with the largest weight was
0.5/199.   At iteration 2, the estimated mean ends up very small for all
observations, and then the iteration diverges.  However, if you provide a
starting value then the fitting works, even if you start the iteration at, say
beta=1, corresponding to a fitted mean of over 70%.

So, the result is wrong in the sense that it is not the mle, because of a
failure of convergence, which happens because specifying the weights the way you
did rather than the documented way leads to bad default starting values for the
iteration.  You need either to specify the data as recommended or supply
starting values.

     =thomas


On Fri, 16 Apr 2010, Jan van der Laan wrote:
> I have some questions about the use of weights in binomial glm as I am
> not getting the results I would expect. In my case the weights I have
> can be seen as 'replicate weights'; one respondent i in my dataset
> corresponds to w[i] persons in the population. From the documentation
> of the glm method, I understand that the weights can indeed be used
> for this: "For a binomial GLM prior weights are used to give the
> number of trials when the response is the proportion of successes."
>> From "Modern applied statistics with S-Plus 3rd ed." I
understand the
> same.
>
> However, I am getting some strange results. I generated an example:
>
> Generate some data which is simular to my dataset
>> Z <- rbinom(1000, 1, 0.1)
>> W <- round(rnorm(1000, 100, 40))
>> W[W < 1] <- 1
>
> Probability of success can either be estimated using:
>> sum(Z*W)/sum(W)
> [1] 0.09642109
>
> Or using glm:
>> model <- glm(Z ~ 1, weights=W, family=binomial())
> Warning message:
> In glm.fit(x = X, y = Y, weights = weights, start = start, etastart >
etastart,  :
>  fitted probabilities numerically 0 or 1 occurred
>> predict(model, type="response")[1]
>           1
> 2.220446e-16
>
> These two results are obviously not the same. The strange thing is
> that when I scale the weights, such that the total equals one, the
> probability is correctly estimated:
>
>> model <- glm(Z ~ 1, weights=W/sum(W), family=binomial())
> Warning message:
> In eval(expr, envir, enclos) : non-integer #successes in a binomial glm!
>> predict(model, type="response")[1]
>         1
> 0.09642109
>
>
> However scaling of the weights should, as far as I am aware, not have
> an effect on the estimated parameters. I also tried some other
> scalings. And, for example scaling the weights by 20 also gives me the
> correct result.
>
>> model <- glm(Z ~ 1, weights=W/20, family=binomial())
> Warning message:
> In eval(expr, envir, enclos) : non-integer #successes in a binomial glm!
>> predict(model, type="response")[1]
>         1
> 0.09642109
>
>
> Am I misinterpreting the weights? Could this be a numerical problem?
>
> Regards,
>
> Jan
>
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>
Thomas Lumley			Assoc. Professor, Biostatistics
tlumley at u.washington.edu	University of Washington, Seattle