Announcement

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

  • Stratifying conditional logit model on unmatched variable

    Hello all- trying my luck again by rewording a question i had posted previously. Essentially, i want to understand how one would handle the stratification of a conditional logistic regression model by a variable we didnt match our cases and control on. Basically, i foresee some matched-sets breaking up if they fall in different strata of this variable. Is it fair to take an approach where we count all our matched controls in the strata of its case irrespective of what they reported for the stratifying variable?

  • #2
    You can emulate stratification by interacting the stratification variable with everything in the model. That will still work with matched data even though matched observations might disagree on the stratifier. So
    Code:
    regression_command outcome_variable i.stratification_variable##(right_hand_side_variables)
    Don't forget to prefix each of the right_hand_side_variables with c. or i. as appropriate. Then you can get the stratum-specific marginal effects of each right hand side variable with
    Code:
    margins stratification_variable, dydx(right_hand_side_variables)

    Comment


    • #3
      Originally posted by Clyde Schechter View Post
      You can emulate stratification by interacting the stratification variable with everything in the model. That will still work with matched data even though matched observations might disagree on the stratifier. So
      Code:
      regression_command outcome_variable i.stratification_variable##(right_hand_side_variables)
      Don't forget to prefix each of the right_hand_side_variables with c. or i. as appropriate. Then you can get the stratum-specific marginal effects of each right hand side variable with
      Code:
      margins stratification_variable, dydx(right_hand_side_variables)
      Thank you for this helpful response. when i run clogit this way, i notice that it only gives me the interaction estimates for exposed categories of both variables..

      Comment


      • #4
        Yes, that is the way interaction models are reported. But you can get the marginal effects for the exposed and unexposed categories from the --margins- command.

        Comment

        Working...
        X