1
$\begingroup$

Short version:

I am fairly new to multilevel modelling and would like to avoid it. My question is basically if I can leave it aside given the nature of my data and the goals of my analysis.

My explanatory variables are all constant within the clusters. However, my binary dependent variable does vary within the clusters (but not within all of them). Do I need a multilevel model here? Or should I decide this based on theoretical considerations on the independence of the outcomes within the clustering?

Additional information:

The data:

The observations are legal acts. My dependent variable is 1/0 indicating one of the two types of legal act I am interested in. These legal acts are clustered through so called "parent acts", i.e. legal acts that multiple (or single) acts refer to. All my explanatory variables are on the level of these "parent acts".

My goal:

So several observations have the same "parent act" and therefore the exact same values for their explanatory variables. However, the observations refering to the same "parent act" can differ in their outcome (1/0). I basically want to use the characteristics of the "parent act" to estimate the likelihood of one outcome over the other.

My question:

Given this data structure, do I need to use a multilevel model that takes into account the clustering in the different "parent acts". I am unsure because, except of the dependent variable, all variables are on this "level 2" (if this is the correct term here). Since I expect that certain characteristics of a "parent act" will make one of the outcomes more likely, the observations referring to the same "parent act" should not be independent. On the other hand, there are quite few "parent acts" to which only one of both types (1/0) refers to.

Although my results are in line with my theoretical expectations, my intuition tells me not to trust them. Please tell me if I'm wrong here. Thank you!

Improve this question
$\endgroup$
3
  • $\begingroup$ It is difficult to address this question without some domain-specific knowledge of the relationship between "parents" and "child" (?) acts. Do you expect that a cluster of observations sharing the same 'parent' act do share some similarities above and beyond the features that are explicitly represented in your other variables/covariates? $\endgroup$
    – matteo
    Commented Apr 30, 2020 at 13:30
  • $\begingroup$ My description is confusing, sorry. Assume the "parent acts" are schools and the observations students accepted to these schools. I have variables describing the schools, only the binary dependent variable varies across observations (the students, let's say male/female student). I want to analyse if the characteristics of the school are correlated to a higher probability of accepting male students. I expect that this only depends on the characteristics of the school. The schools vary in their number of students, and some only accept male or female students. I hope that makes it more clear! $\endgroup$
    – Rowanlegere
    Commented Apr 30, 2020 at 16:51
  • $\begingroup$ OK I see, I think in that case I don't think you need multilevel modelling. essentially in your case each school is a binomial observation (with a certain proportion of 1 and 0 outcomes), so a standard logistic regression would be appropriate. $\endgroup$
    – matteo
    Commented May 1, 2020 at 9:03

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.