Dear Statalisters,
I have a question on how to interpret predicted probabilities after a discrete time event history analysis with interaction effects.
I use a person-month datafile (dependent variable is event occurrence, the event is getting active on labour market after a period of inactivity). Length of duration of inactivity is recorded in months: individuals are followed from the month of initial inactivity until the month of becoming active (or right-censoring). I employ a logit model for discrete time event history analysis with clustered standard errors for individuals, duration is modelled as a quadratic function.
Using the exponentiated coefficients the interpretation is – but correct me if I’m wrong – rather straightforward: one-unit increase in x leads to an decrease or increase in the odds of making the transition to activity by a certain amount of percentage points given survival up to the end of the previous month.
My confusions starts when it comes to including interaction effects and calculating predicted probabilities: I included an interaction effect between x1 and x2, where x1 is a continuous and x2 a categorical variable. I choose to fix the continuous variable at -2, -1, 0, 1 and 2 and I’m only interested in the predicted probability of the 4th category of x2. Using the default asobserved in margins, for the variables that are not fixed (x3, x4, but also duration) the actual observed values on each person-month are used and, as far as I understand, the predicted probabilities are then averaged across these observations.
My question is how to interpret the predicted probability in light of the person-month data set-up and the fact that predicted probabilities are averaged across observations (so across person-months with different duration times). Can I still refer to monthly transition rates? More specific;
I have a question on how to interpret predicted probabilities after a discrete time event history analysis with interaction effects.
I use a person-month datafile (dependent variable is event occurrence, the event is getting active on labour market after a period of inactivity). Length of duration of inactivity is recorded in months: individuals are followed from the month of initial inactivity until the month of becoming active (or right-censoring). I employ a logit model for discrete time event history analysis with clustered standard errors for individuals, duration is modelled as a quadratic function.
Code:
logit y x1 i.x2 x3 x4 duration duration2, robust cluster(id) or
My confusions starts when it comes to including interaction effects and calculating predicted probabilities: I included an interaction effect between x1 and x2, where x1 is a continuous and x2 a categorical variable. I choose to fix the continuous variable at -2, -1, 0, 1 and 2 and I’m only interested in the predicted probability of the 4th category of x2. Using the default asobserved in margins, for the variables that are not fixed (x3, x4, but also duration) the actual observed values on each person-month are used and, as far as I understand, the predicted probabilities are then averaged across these observations.
Code:
logit y c.x1##i.x2 x3 x4 duration duration2, robust cluster(id) or margins, at(x1=(-2(1)2) x2=(4)) marginsplot
- Can I say that the reported predicted probability is the probability that an individual makes the transition to activity in a certain month given that the respondent was inactive up until that month?
- Let’s say that the predicted probability is 0.10 at x1=-2 (and x2=4) and 0.05 for x1=2 (and x2=4). Can I then say that the predicted probability of making the transition to activity reduces with 5 percentage points in a certain month from going from -2 to 2 on x1?
Comment