On Mon, 31 Jan 2011, Dale.W.Steele at gmail.com wrote:
I'm interested in modeling the probability of appendicitis in patients with abdominal pain. The R binary data file 'http://www.ped-em.org/appy.rda' contains the following variables from a pilot study of 138 children with abdominal pain. 'dx' eventual diagnosis: 0=no appendicitis, 1=appendicitis 'gender' Male/Female 'wbc' total white blood cell count 'priorprob' Clinical predicted probability of appendicitis 'doc' doctor who assigned 'priorprob' My initial thought was to fit a multiple logistic regression model: m1 <- glm(dx ~ gender + priorprob + wbc + doc, family=binomial, data=appy) However, it seems likely that each doctor interpreted the probability scale differently. The 23 doctors evaluated from 1 to 17 patients each. I'm not primarily interest in predictions by a specific clinician. Thus, it seems to make sense to fit a generalized linear mixed model.
I think this is a fun kind of dataset. I don't think there is a large amount of "slop" coming from the different clinicians -- I looked at this by a cheap and nasty mixed model of priorprob as a continuous variable, noting dx was not correlated with doc. It's amusing that there is little correlation between wbc and priorprob.
m2 <- glmer(dx ~ priorprob + gender + wbc + (1 | doc), family=binomial, data=appy) m3 <- glmer(dx ~ priorprob + gender + wbc + (priorprob | doc), family=binomial, data=appy)
These seemed OK to me. You might look at a GAM too. Just 2c, David Duffy.
| David Duffy (MBBS PhD) ,-_|\ | email: davidD at qimr.edu.au ph: INT+61+7+3362-0217 fax: -0101 / * | Epidemiology Unit, Queensland Institute of Medical Research \_,-._/ | 300 Herston Rd, Brisbane, Queensland 4029, Australia GPG 4D0B994A v