SPSS Statistics

I have a Cox regression problem where the dependent variable is the time to some clinical event. One of my covariates is an intervention that may or may not occur during follow-up, and if it occurs, the time between baseline and the start of the intervention differs from patient to patient. Now my questions are as follows:

- Since Cox regression in SPSS does not support data in the 'long' format (at least so I understood), how can I analyze data where the time varying covariate changes more often than once (e. g., an intervention that has a start and an end, both of which may affect the hazard, or an 'no/off' series of successive interventions of the same type)?

- As I have written above, most of my patients receive the intervention coded in a time varying covariate, but not all. My understanding of the output of Cox regression with time varying covariates is that the hazard ratio for this covariate applies to the time until the intervention occurs. How do I interpret the coefficient in case of patients in whom the intervention does not occur at all?

------------------------------
Frank Furter
------------------------------

#SPSSStatistics

From one of our statisticians:

----

COXREG indeed does not handle data in a format allowing for multiple cases per subject. Handling of a time-dependent covariate that changes at arbitrary times for different cases (what's sometimes referred to as a segmented time-dependent covariate) makes use of a computation using Boolean operators to select values from different variables at different time points in order to create a single variable that contains the correct value for each case at each event time. This is discussed in the Technote at https://www.ibm.com/support/pages/coxreg-segmented-time-dependent-covariates.

 

The CSCOXREG procedure in the SPSS Statistics Complex Samples module does allow for data input with multiple cases. The second of the Case Studies for CSCOXREG at https://www.ibm.com/docs/en/spss-statistics/28.0.0?topic=edition-complex-samples-cox-regression shows how to use multiple-case input, as well as creation of an analysis plan file specifying a simple random sample, as would be needed for analyzing data typically addressed in COXREG.

 

It's not entirely clear whether the data under discussion involve simply cases where an intervention that can be applied only once per subject may occur at a different time for different subjects, or whether interventions may be applied for periods of time and reapplied to the same subject. Interpretation of the coefficient for a segmented time-dependent covariate may have subtle differences in these different scenarios, but in the simplest case, a binary 0-1 covariate coefficient is interpreted as the relative log-hazard ratio for those experiencing the intervention compared to those not experiencing it.

 

----

------------------------------
Rick Marcantonio
Quality Assurance
IBM
------------------------------

Original Message:
Sent: Thu June 23, 2022 01:23 AM
From: Frank Furter
Subject: Cox regression with time varying covariate

I have a Cox regression problem where the dependent variable is the time to some clinical event. One of my covariates is an intervention that may or may not occur during follow-up, and if it occurs, the time between baseline and the start of the intervention differs from patient to patient. Now my questions are as follows:

- Since Cox regression in SPSS does not support data in the 'long' format (at least so I understood), how can I analyze data where the time varying covariate changes more often than once (e. g., an intervention that has a start and an end, both of which may affect the hazard, or an 'no/off' series of successive interventions of the same type)?

- As I have written above, most of my patients receive the intervention coded in a time varying covariate, but not all. My understanding of the output of Cox regression with time varying covariates is that the hazard ratio for this covariate applies to the time until the intervention occurs. How do I interpret the coefficient in case of patients in whom the intervention does not occur at all?

------------------------------
Frank Furter
------------------------------

#SPSSStatistics