[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

st: Does Stata have an "exact likelihood approach " to estimate the variance of a proportion?

From   "Tiago V. Pereira" <>
Subject   st: Does Stata have an "exact likelihood approach " to estimate the variance of a proportion?
Date   Wed, 18 Mar 2009 11:19:22 -0300 (BRT)

Hello, stalisters!

I would like to explore the method mentioned by Hamza et al (2008). The
binomial distribution of meta-analysis was preferred to model within-study
variability. Journal of Clinical Epidemiology, Volume 61, Issue 1, January
2008, Pages 41-51

The authors compared an exact likelihood approach to the standard method
that approximates the within-study variability of a proportion by a normal

They comment that main packages are able to do that, indluign SAS, R AND
S-plus. Hence, is there a way to perform such analysis using Stata?

If you need more info, please, let me know.

all the best,


"Now the model is a GLMM, and the parameters can be
estimated by standard likelihood procedures. The practical
disadvantage is that software is much more scarce and not
yet available in all statistical packages. We used the
NLMIXED procedure from the SAS package [17]. It is also
possible to use the recently included GLIMMIX procedure
in the SAS package, which is still experimental in SAS
version 9.1. The GLIMMIX procedure allows more random
effects, but it has the disadvantage that it uses an approximation
instead of the true log likelihood."

"In this paper,we compared the use of the approximate normal
within-study likelihood that is used in practice with the
alternative exact binomial likelihood. Calculation of the exact
binomial likelihood involves an approximation of the integral.
In NLMIXED, the method of Gaussian quadrature is
used, with the number of quadrature points to be specified
by the user or automatically by SAS. The larger that number
is chosen, the better the approximation, but at the cost of more
computational time. For example, Carlin et al. [34] have
shownthat for binary outcome longitudinal data, a reasonably
large number of quadrature points (i.e., 20) is required to ensure
convergence on model parameter estimates. In our data
example, to study the impact of the number of quadrature
points we fitted the model for varying number of quadrature
points. It turned out the estimates (SE) of sensitivity and
specificity did not change for a number of quadrature points
greater than or equal to 10 and 15, respectively. We used 20
quadrature points for our simulation study."

*   For searches and help try:

© Copyright 1996–2021 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index