stcox model fails to converge when adding interaction effects

[Bob Sanders]

I'm doing a cox proportional hazard model to estimate the risk of a patient getting infection (yes or no) as a function of his treatment technique (categorical, 4 levels), controlling for age at hospital stay (categorical, 5 levels), gender (male, female), and race (4 levels). Survival time is measured in days (from hospital start to either infection or censoring). A patient may have multiple hospital stays so the unit of analysis is the hospital stay, and I'm accounting for correlation with patient by adjusting for standard errors using cluster (patient_id).

I have 30,837 observations (i.e., hospital stays) with 532 failures (i.e., infections). Total analysis time at risk and under observation is 6,520,775.

The following model with no interactions works fine:

stcox ib3.technique ib1.gender ib4.race ib5.age_at_stay, efron vce(cluster child_id) allbaselevels

However, if I try to include an interaction term with the log of the survival time (or just survival time - same issue), the model fails to converge around iteration 4 due to "flat region resulting in a missing likelihood."

For example, this model with a term that interacts technique with the log of survival time fails to converge:

stcox ib3.technique ib1.gender ib4.race ib5.age_at_stay i.technique#c.log_survival_time, efron vce(cluster child_id) allbaselevels

This happens with any interaction I try with survival time. However, interactions that don't use survival time pose no problem (i.technique#i.age, i.gender#i.race, etc.).

I've tried changing the tie method to Breslow but the problem persists.

Any idea what could be causing this issue?