No announcement yet.
  • Filter
  • Time
  • Show
Clear All
new posts

  • To nest or cluster in multivariable logistic regression

    Hi, I wanted to perform a multivariable logistic regression to check which of the patient, indwelling catheter and insertion factors predispose bacteriuria from 900 patients and 1500 catheters. I have age per 10 year increment, sex, ethnicity, type of catheter, location (4 wards), and inserter (specific to wards) as variables. The data were collected in 5 large hospitals (clusters), each with similar wards (location) structure. I wanted to check how do I go about it, e.g. whether I consider the nesting of the patients to the hospitals, or perform clustering at the hospital-level? Thanks in advance.

    The code I used so far:
    melogit bacteriuria age_10 sex... || _all.R:patient || hospital: , or
    Last edited by Jason Avino; 29 Dec 2023, 08:28.

  • #2
    Well, aren't the patients also nested in wards? I think ward needs to be a level in the model as well, because the ward is a proxy variable for the nursing and other ancillary care given to the patient and his or her catheter. And I don't see why you are crossing patient with hospital. I get that the same patient might be found in multiple hospitals, so that you have a multiple membership model (which would be represented as crossed effects.) But do you really have many of those? If there are only a small number of patients who appear in multiple hospitals, I would randomly select a single hospital for each patient and use the simpler model with patients nested in, rather than crossed with, hospitals. It's computationally less burdensome, and also, as I expect that the multiple membership model is actually very sparse, the crossed model may have trouble converging. Also, if restricting the data to one hospital per patient you end up with only one observation per patient (it isn't clear whether you have longitudinal data within hospitalizations), then the patient level disappears from the model. (Or, strictly speaking, it merges with the residual level.) That simplification will also reduce computational burden considerably.

    Finally, the pathogenesis of urinary infections is very different in men and women, and I think just having sex as a variable in the model is not adequate to capture that. I think it needs to be interacted with most or even all of the other fixed effects.

    In short, my model would look more like:
    melogit bacteriuria ...) || hospital: || ward: || patient:
    with the -patient:- level being omitted if there is only one observation per patient.


    • #3
      Thanks a ton, Clyde. The code has provided me with a more meaningful result. Happy New Year!