Dear Gavin,
glm reported exactly what it noticed, giving a warning that some very
small fitted probabilities have been found.
However, your data are **not** quasi-separated. The maximum likelihood
estimates are really those reported by glm.
A first elementary way is to change the tolerance and maximum number
of iterations in glm and see if you get the same result:
#
> mod1
Call: glm(formula = analogs ~ Dij, family = binomial, data = dat,
control = glm.control(epsilon = 1e-16, maxit = 1000))
Coefficients:
(Intercept) Dij
4.191 -29.388
Degrees of Freedom: 4033 Total (i.e. Null); 4032 Residual
Null Deviance: 1929
Residual Deviance: 613.5 AIC: 617.5
#
This is exactly the same fit as the one you have. If separation
occured the effects ususally diverge as we allow more iterations to
glm and at some point.
**************
Secondly an inspection of the estimated asymptotic standard errors,
reveals nothing to worry for.
#
> summary(mod1)
Call:
glm(formula = analogs ~ Dij, family = binomial, data = dat, control =
glm.control(epsilon = 1e-16,
maxit = 1000))
Deviance Residuals:
Min 1Q Median 3Q Max
-1.676e+00 -1.319e-02 -1.250e-04 -1.958e-06 4.104e+00
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 4.1912 0.3248 12.90 <2e-16 ***
Dij -29.3875 1.9345 -15.19 <2e-16 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
(Dispersion parameter for binomial family taken to be 1)
Null deviance: 1928.62 on 4033 degrees of freedom
Residual deviance: 613.53 on 4032 degrees of freedom
AIC: 617.53
Number of Fisher Scoring iterations: 11
#
If separation occurred the estimated asymptotic standard errors would
be unnaturally large. This is because, in the case of separation
(quasi or not) glm would calculate the standard errors taking the sqrt
of the diagonal elements of minus the hessian of the log-likelihood,
in a point where the log-likelihood appears to be flat for the given
tolerance.
**************
To be certain, you could also try fitting with brglm, which is
guaranteed to give finite estimates, that have bias of smaller order
than the MLE and compare the results.
#
> library(brglm)
> mod.br <- brglm(analogs ~ Dij, data = dat, family = binomial)
> mod.br
Call: brglm(formula = analogs ~ Dij, family = binomial, data = dat)
Coefficients:
(Intercept) Dij
4.161 -29.188
Degrees of Freedom: 4033 Total (i.e. Null); 4032 Residual
Deviance: 613.5448
Penalized Deviance: 610.2794 AIC: 617.5448
#
The estimates are similar a bit shrunk towards the origin which is
natural for bias removal. If separation occurred, and given the
previous discussion, the bias-reduced estimates would be considerably
different than the estimates that glm reports.
**************
Lastly, the more certain way to check for separation is to inspect the
profiles of the log-likelihood. Vito suggested this but the chosen
limits for the xval are not appropriate. If separation would occur the
estimate would be -Inf so that the profiling as done in his email
should be done starting from example from -40 rather than -20. This
would reveal that the profile deviance starts increasing again, while
if separation occured there would be an asymptote on the left. Below I
give the correct profiles, as reported by profileModel.
> library(profileModel)
> pp <- profileModel(mod1, quantile = qchisq(0.95, 1), objective =
"ordinaryDeviance")
Preliminary iteration .. Done
Profiling for parameter (Intercept) ... Done
Profiling for parameter Dij ... Done
> plot(pp)
The profiles are quite quadratic. In the case of separation you would
have seen asymptotes on the left or on the right (see
help(profileModel) for an example).
**************
It appears that the fitted logistic curve, while steep still has a
finite gradient, for example, at the LD50 point
> library(MASS)
> dose.p(mod)
Dose SE
p = 0.5: 0.1426167 0.003646903
When separation occurs the LD50 point cannot be identified (computer
software would return something with enormous estimated standard error).
In conclusion, if you get data sets that result in large estimated
effects on the log-odds scale, the above checks can be used to
convince you whether separation occurred or not. If there is
separation (not the case in the current example) then, you could use
an alternative to maximum likelihood for estimation ---such as
penalized maximum likelihood in brglm--- which always return finite
estimates. Though in that case, I suggest you incorporate the
uncertainty on how large the estimated effects are in having
confidence intervals with one infinite endpoint, for example
confidence intervals as in help(profile.brglm).
Hope this helps,
Best wishes,
Ioannis
On 15 Dec 2008, at 18:03, Gavin Simpson wrote:
> Dear List,
>
> Apologies for this off-topic post but it is R-related in the sense
> that
> I am trying to understand what R is telling me with the data to hand.
>
> ROC curves have recently been used to determine a dissimilarity
> threshold for identifying whether two samples are from the same "type"
> or not. Given the bashing that ROC curves get whenever anyone asks
> about
> them on this list (and having implemented the ROC methodology in my
> analogue package) I wanted to try directly modelling the probability
> that two sites are analogues for one another for given dissimilarity
> using glm().
>
> The data I have then are a logical vector ('analogs') indicating
> whether
> the two sites come from the same vegetation and a vector of the
> dissimilarity between the two sites ('Dij'). These are in a csv file
> currently in my university web space. Each 'row' in this file
> corresponds to single comparison between 2 sites.
>
> When I analyse these data using glm() I get the familiar "fitted
> probabilities numerically 0 or 1 occurred" warning. The data do not
> look
> linearly separable when plotted (code for which is below). I have read
> Venables and Ripley's discussion of this in MASS4 and other sources
> that
> discuss this warning and R (Faraway's Extending the Linear Model
> with R
> and John Fox's new Applied Regression, Generalized Linear Models, and
> Related Methods, 2nd Ed) as well as some of the literature on Firth's
> bias reduction method. But I am still somewhat unsure what
> (quasi-)separation is and if this is the reason for the warnings in
> this
> case.
>
> My question then is, is this a separation issue with my data, or is it
> quasi-separation that I have read a bit about whilst researching this
> problem? Or is this something completely different?
>
> Code to reproduce my problem with the actual data is given below. I'd
> appreciate any comments or thoughts on this.
>
> #### Begin code snippet
> ################################################
>
> ## note data file is ~93Kb in size
> dat <- read.csv(url("http://www.homepages.ucl.ac.uk/~ucfagls/
> dat.csv"))
> head(dat)
> ## fit model --- produces warning
> mod <- glm(analogs ~ Dij, data = dat, family = binomial)
> ## plot the data
> plot(analogs ~ Dij, data = dat)
> fit.mod <- fitted(mod)
> ord <- with(dat, order(Dij))
> with(dat, lines(Dij[ord], fit.mod[ord], col = "red", lwd = 2))
>
> #### End code snippet
> ##################################################
>
> Thanks in advance
>
> Gavin
> --
> %~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%
> Dr. Gavin Simpson [t] +44 (0)20 7679 0522
> ECRC, UCL Geography, [f] +44 (0)20 7679 0565
> Pearson Building, [e] gavin.simpsonATNOSPAMucl.ac.uk
> Gower Street, London [w] http://www.ucl.ac.uk/~ucfagls/
> UK. WC1E 6BT. [w] http://www.freshwaters.org.uk
> %~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%~%
>
> ______________________________________________
> R-help@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.
------------------------------------------
Ioannis Kosmidis
D0.05, Dept. of Statistics,
University of Warwick,
Coventry, CV4 7AL, UK
Webpage: http://go.warwick.ac.uk/kosmidis
Voice: +44(0)2476 150778
Fax: +44(0)2476 524532
------------------------------------------
[[alternative HTML version deleted]]