Conditional logistic regression model for paired data

[Abay Burusie]

I am studying "Determinants of Childhood TB Disease Development". I randomly selected health facilities providing TB treatment from the capital city (23 facilities) and two other large towns closer to the capital (3 facilities from each). Then, I randomly selected 256 cases (diagnosed with TB and on TB treatment) from these healthcare facilities and matched them by age with control children (free of TB disease) who were selected at triage sequentially. I then analyzed the data using the conditional logistic regression model. This data is from multi-centers (multiple healthcare facilities). However, there are no suspected healthcare facility-level factors that risk TB disease development (the risk of developing TB is independent of healthcare facilities providing services to the community). Should I use multilevel analysis for such data? If yes, is there a multilevel analysis method for conditional logistic regression (paired data) analysis?

[Clyde Schechter]

Although it does not answer the intent of your question, let me point out that conditional logistic regression is itself a multilevel model: you have individuals nested in pairs.

However, your data has some hierarchical structure that goes beyond that. You have individual nested in pairs. Then there are hospitals nested in towns. I cannot tell from your description whether the individuals in each pair are always from the same hospital or not. If so, then you have a four level design: individuals in pairs in hospitals in towns. If not, it is more complicated.

You state that you do not expect any hospital level risk factors for TBC. I'm skeptical of that: even if the development of actual TBC infection and disease is truly independent of the hospital, different hospitals may have different abilities to make the diagnosis. So I'd be inclined to begin by assuming that there is a hospital-level effect: if subsequently we don't find one in the data, we can always eliminate hospital from the modeling afterwards. And I would certainly expect town level effects on this.

I am also confused about how many hospitals you have. On the one hand you say you have 23. But you also say you selected 3 facilities from each of the capital and two towns nearby, which would be only 9. In this situation it actually makes a difference. 9 hospitals is too few to use hospital as a level in the model, but 23 is enough.

If you have 23 hospitals, I would probably use mixed (or one of the other -me- commands, depending on the outcome you are analyzing). I would set up the model with individuals in pairs in hospitals. As you have only 3 towns, that is not enough to meaningfully add a town-level to the model, though I would certainly include town indicator variables. So something like

melogit outcome i.tbc covariates i.town || hospital: || pair:

Added: If the variance component at the hospital level turns out to be very close to zero, as you suspect it will, you can always rerun the model without that level--in which case using conditional logistic regression (as shown below) is better. /Added.

If you only have 9 hospitals, then I would just do

xtset pair
xtlogit outcome i.tbc covariates, fe

Note that in this approach, since the pairs come (I assume) from the same hospital, and each hospital is located in a single town, the hospital and town-level effects cannot be estimated, but they are automatically adjusted for even without being mentioned in the model. If the members of a matched pair can come from different hospitals, then I would add i.hospital to the -xtlogit- command. If the i.hospital coefficients turn out to be, for practical purposes, zero, as you believe they will, then you can re-do the analysis without i.hospital.