2
$\begingroup$

I need to identify cases of type 2 diabetes in a health care database for a specific population using latent class analysis (LCA).

In addition to doctor's diagnosis of diabetes and the receipt of diabetes medications, I was planning to include glycosuria (sugar in urine; True/False) as one of the manifest (predictor) variables. But I found that:

  • 30% of people in the cohort have data for urine test

  • of those having data for urine test, 5% have glycosuria

Assume that:

  • according to another study, %25 of those with confirmed diagnosis have data for urine test

  • the prevalence of glucosuria (any level) in the community is thought to be also 5%

Is it OK to use this variable in the model if it shows the expected distribution of values?

What important assumptions/conditions/checks that should be fulfilled by manifest variables in LCA?

$\endgroup$
1
+50
$\begingroup$

Are diagnosis/receipt of medication (binary?/and glycosuria the only variables you have? Why do you need to do a LCA? I mean, if there is a diagnosis you should be reasonably certain that the patient has diabetes. That leaves you with four combinations of medication and glycosuria when there is no diagnosis of diabetes. Diabetes or even internal medicine is not my field, but it seems you might be better of just deciding how you should judge each of these combinations, especially the presence of glycosuria when there is no diagnosis or medications:

diabetes dx (regardless of other variables) = diabetes

diabetes medication (regardless of other variables) = diabetes

no dx and no med, but glycosuria = ? (probably undetected diabetes?)

none of the above = no diabetes

But in any case, I think that missing data should't be a problem in LCA. In the documentation to R package poLCA it says that missing data on manifest variables can be handled, so you should be good to go, I think. I think you might risk selection bias if there is an association between the test being taken and higher levels (more often glycosuria), but because glycosuria is most often caused by diabetes anyway (if I recall correctly) this shouldn't be a problem.

Give it a try, using 2 latent classes, and examine the results. Do they seem reasonable?

My guess is that the identified latent classes will be any positive of diagnosis/medication/glucosuria vs. none of them.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ If there is a diagnosis you should be reasonably certain that the patient has diabetes. In my data, I think this would require validation against external reference standard, which is unavailable, and that's why I'm using LCA. By looking at the data, a significant number of patients received diabetes medication without having any related diagnosis (most likely missing data). Those could be false-negatives if I used diagnosis alone. $\endgroup$ – Orion Oct 14 '15 at 10:54
  • $\begingroup$ My understanding is that with LCA the resulting classes will be identified according to how the population tends to cluster based on carefully chosen related manifest variables. I did use poLCA but didn't know that it handles missing data in manifest variables. Thank you. $\endgroup$ – Orion Oct 14 '15 at 11:01
  • $\begingroup$ I mean that the risk of a doctor setting the diabetes diagnosis when the patient doesn't have diabetes seems very small. So in the cases where there is a diagnosis, you can be reasonably certain that the diagnosis is correct, but the difficulty is the cases in which there is no diagnosis but perhaps medication and/or glycosuria. Of course, I don't know how doctors in your country do their work, but I'd say that in Sweden, the risk of a patient getting a diabetes diagnosis when he/she doesn't have diabetes is very very small. $\endgroup$ – JonB Oct 14 '15 at 11:03
  • $\begingroup$ As for missing data in the manifest variables, take a look at the first paragraph of the fifth page in the document you link to. $\endgroup$ – JonB Oct 14 '15 at 11:05
  • $\begingroup$ I'm no expert in using LCA though I have some experience, and I'm no expert on diabetes, though I am a medical doctor in a different field. I don't think I would use LCA here, but I would do as I suggest above and consider all who have a diagnosis and/or medication as having diabetes, then calculate prevalences including or excluding those who have glucosuria. Perhaps you have data on how many in a "normal" population without diabetes diagnosis who are identified as having diabetes with the use of positive glucosuria? $\endgroup$ – JonB Oct 14 '15 at 11:15

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.