MerQuR+Statistics Glossary

Statistics Glossary → Time-dependent covariate and counting-process data

Time-dependent covariate and counting-process dataconcept

A predictor whose value changes during follow-up, and the data layout that splits each person's follow-up into (start, stop] intervals so that such predictors can enter survival models.

In survival analysis some covariates are not fixed: a patient may start treatment midway through follow-up, or blood values may be measured every month. In counting-process form each person is represented by one row for every period during which the covariate is constant: the start time, the stop time, whether the event occurred at the end of that interval and the covariate value for the interval. At each event time the Cox model uses only the rows at risk at that moment and their current covariate values. Robust standard errors clustered by person can account for the dependence between a person's rows.

Time-dependent covariates prevent immortal time bias, which arises when future information is carried back in time: if 'people who received a transplant during the study' are treated as a group from the start, they gain an artificial advantage because they had to survive until the transplant. The same data layout is used to model a covariate whose effect changes over time through an x · g(t) interaction term. Internal covariates affected by the disease process (such as markers that change as the illness progresses) need extra care in causal interpretation.

An everyday example

A heart patient receives a transplant in month 3 of the study and dies in month 10. The data hold two rows for this patient: (0, 3] transplant = 0, event = 0 and (3, 10] transplant = 1, event = 1. The first 3 months are thus analysed together with patients who have not had a transplant and the next 7 months with those who have; the pre-transplant time is not wrongly credited to the transplant group.

In MerQur

In Time-dependent Cox Regression, 'ID, Start, Stop, Event column and Covariates' define the (start, stop] counting-process data, and 'Time interaction (x·g(t)) and g(t) function (log t / t)' adds the x·g(t) term for the selected covariate and splits the data at event times.

Analyses where this concept appears

Time-dependent Cox Regression

Related concepts

Proportional hazards assumptionHazard ratio (HR)