Interpeting interaction terms for a variable with a time varying effect

[Bob Sanders]

I am doing a survival analysis to estimate the risk of child abuse (yes or no) for children in foster care as a function of placement setting (group home, foster home, other home). Hazard curves suggests there is a time varying effect of setting on the event. For example, in group homes the risk of abuse is high early in the placement but then starts to drop over time.

I am trying to introduce an interaction term to account for this time-varying effect.

I first set the data:

stset duration_months, failure(abuse==1) id(placement_id)

Then split the data into these time intervals: 0 - 1 months, 2 - 3 months, 4 - 7 months, and 8 months or more.

stsplit interval, at(1,3,8)

I then calculated the lag of duration as follows:

generate lag_exposure = ln(_t - _t0)

I then tried to introduce the interaction like below:

mepoisson _d ib0.interval##ib3.setting i.gender i.race i.age_at_placement, offset(lag_exposure) irr allbaselevels || child_id:

I was trying to make the reference group "Other home in 0 - 1 months," but the interaction term I used assigned the base to *all* settings in 0 - 1 months (0#grouphome = base, 0#fosterhome = base, and 0#otherhome = base). Further, for the subsequent intervals it assigned "Own home" as the base.

Any suggestions on how to change the interaction term?

Thank you.