[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

st: RE: longitudinal data with non-detectable measurements

From   "Nick Cox" <>
To   <>
Subject   st: RE: longitudinal data with non-detectable measurements
Date   Fri, 9 May 2008 16:27:22 +0100

I regard zero inflation and non-detectable measurements as different
problems. That is,  a spike of "known to be zero"s and a spike of "less
than threshold"s are not identical. 

I have just a small marginal suggestion. 

One thing that is easy in the GLM context (and indeed in any modelling
context) is to add an indicator variable (censored or not) as an extra
covariate. At best you might find it unnecessary and even at worst if it
was necessary there are further things to try such as adding interaction

Others should have more direct suggestions. 


Leny Mathew

                          I'm trying to model data on change in
certain hormones over a period of time. The data on some of these have
a number of values that are set (currently) at a non-zero minimum
detectable value (MDL). Currently, I'm using a GLM with a gamma family
along with robust estimation of SE to model this data. The model works
reasonably fine on the data but I would like to improve this by
properly accounting for the information on the non-detectables.
I'm interested in the change over time of these hormones and so change
from the MDL to a higher value is useful information. Apart from the
GLM, I've looked at the tobit models and zero-inflated models. From
what I understand, the tobit model assumes that the data that is not
censored follows a normal distribution and this not the case in my
data. Zero-inflated models seem to be used for count rather than
continuous data.
Does anybody in the list have experience on modeling continuous data
with zero-inflation or data with a MDL issue?

*   For searches and help try:

© Copyright 1996–2017 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index