1
$\begingroup$

I want to test whether changes in Hemoglobin (Hgb) levels over time can help diagnose Myelodysplastic syndrome (MDS).

I have a cohort of thousands of patients. Each patient has several Hgb measurements over a period of 3 years. Measurements are unevenly spaced and few (anything from two to 4 measurements).

My questions: 1. Is a time series the correct model for me considering I have so few observation over time for each subject ? if not how do I analyze the trend ? 2. If I should use a time series, I thought of tagging each lab measurement by its year so as to make the measurements evenly spaced (this seems sensible enough from a physiological perspective). Is this valid ? Is there a better way to handle the inequality of measurement spacing ? 3. What do I use to compare trends between two groups ?

I would be most obliged if you could refer me to the appropriate further reading

$\endgroup$
  • $\begingroup$ this is a difficult (and recurring) question. I don't have good answers. But (1) time series is not the correct approach given too few measurements (2) you'd be much more successful if you include race, age, WBC, PLT - if they have Hb they had a CBC. (3) are the measurements iid? low Hb leads to recheck...not sure how you deal with that. $\endgroup$ – charles Mar 10 '14 at 14:35
  • $\begingroup$ Thank you. Of course I would add other covariates but the main issue is how do I model the trend of Hgb over the 1-3 years preceding the diagnosis. My hypothesis is that there would be a decreasing trend in the blood counts of patients with MDS even while in the normal range (as the disease develops over a long time). In order to get even time intervals I thought of taking the first or lowest blood count per 6 months. This will solve the problem of confirmatory blood tests but is prone to introduce bias into the data. $\endgroup$ – user2387584 Mar 10 '14 at 15:53
  • $\begingroup$ (1) not sure I have great ideas (2) would consider modeling as change from baseline. Recode readings as percentage change from first reading. This "flattens" out the data and you looks pattern of change but easier and likely fairly good. Just need to be careful how you code time. $\endgroup$ – charles Mar 10 '14 at 17:58

Your Answer

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

Browse other questions tagged or ask your own question.