3

I want to analyze a table with surgery operations data. For some patients, about 10%, multiple operations have been made, so there may be multiple rows for a patient in the table. My goal is to compare certain two types of operations. A patient may undergo multiple operations of one type or operations of the both types.

What are the standard approaches to the analysis? Considering each operation as a single case? Considering each patient as a single case and considering multiple operations as a confounder for each case?

To be more concrete, let us have two types of operation: type 0 and type 1. Both types can be done on each of the two arteries: the left one and the right one. One person can undergo several operations. About 10% of the patients undergo multiple operations. We register deaths during some period. So we have some longitudinal censored data for mortality. Let also the groups for the two types of operation be adjusted for all the confounders.

How can we compare mortality for the two types of operation?

Viktor
  • 853
  • 1
  • 8
  • 22
  • 1
    Multilevel analysis using mixed models? Surgeries nested within patients (nested within surgeons and/or within clinics)? – Alexis Jun 11 '18 at 05:46
  • @Alexis There is one hospital. Surgeries are nested within patients. The problem is that there are relatively small number of patients who underwent two or three operations. I do not understand how can I apply mixed models in this situation. Could you please explain? – Viktor Jun 11 '18 at 08:09
  • You can first apply a two-way ANOVA to see if the interaction between the two categorical variables (operation type and artery) is significant on the mortality. If you end up with a result saying that the interaction is insignificant, then you can omit the artery variable and apply one-way ANOVA, using only the operation type variable as the input. Otherwise, if you have enough data, you may split your data as left-artery-patients, right-artery patients,both-artery-patients and comment on the results separately. – tyumru Jun 20 '18 at 14:52
  • @tyumru The main problem here is that some patients undergo two or even three operations, while the majority of the patients undergo only one operation. Am I to consider different operations as independent observations even if they are done on one person? – Viktor Jun 20 '18 at 22:41
  • 1
    Well, saying that the interaction of consecutive operations doesn't have any significant influence on the likelihood of mortality is a strong assumption to be made. However, as I said above, if you are able to prove that the interaction is insignificant, you can safely consider different operations as independent observations even if they are done on one person. The questions to be asked before proceeding are: How many patients in total? How many operations in total? How many of the patients have multiple operations? Also, how many different types of operations do you have in hand? – tyumru Jun 21 '18 at 09:19
  • @tyumru There are ~2400 operations, ~75 patients with two operations, one patient with 3 operations. There are two types of operations for primary analysis. It seems unlikely that a test will yield any significance for the interaction of the operations, but I will try to see. – Viktor Jun 21 '18 at 10:54
  • @tyumru The other problem is that, strictly speaking, I am able to prove only that one variable has influence on another one, not that the former one does not have influence on the latter one. I ended at exact matching on the four covariates: the presence of preceding operation of the same/other type on the same/other side. – Viktor Jun 21 '18 at 16:39

0 Answers0