[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
Re: st: re: bivariate random effects diagnostic meta-analysis
Ben Dwamena wrote:
>(1) How may the same problem be implemented in gllamm?
I'm not sure what assumptions are being made and what countermeasures are
being taken by the other investigators in their use of PROC MIXED for
meta-analysis of reports in the diagnostic medical literature, but to answer
your question naively, to use -gllamm-, I would use the counts from the
fourfold tables directly from the publications, if they're given.
If the counts are not given in the publications, then I would -generate-
counts of positive and negative cases from the sample size and prevalence
covariates that you mentioned having, and convert the specificity and
sensitivity proportions into counts of positive diagnoses from them.
Now you have three variables for each of two rows (records) for each
publication: positive diagnoses, cases (count) and disease status.
Positive diagnoses would be the dependent variable, the cases would be the
number of trials to reference in the denom() option, and disease status is
an indicator-variable covariate analogous to the
diagnostic_accuracy_indicator in the last post; thus, the rows for a Study 1
that reported a specificity of 75%, a sensitivity of 63%, a sample size of
100 and a prevalence of 80% would be:
study_id diagnosis_positive cases disease_status
1 50 80 1
1 5 20 0
And the command would be -gllamm diagnosis_positive ///
disease_status [other covariates], i(study_id) ///
family(binomial) denom(cases) link(logit) adapt . . .-
>(2) How do I obtain the mean logit-transformed sensitivity and
>specificity after xtreg transformation?
I would use either -predict- or -adjust-.
* For searches and help try: