Hi everyone,

I’m currently modelling loss in life expectancy of cervical cancer patients, with a flexible parametric relative survival model. What I want to do is a traditional cohort approach for the years 1989–2009 and a period approach for the years 2010*–2014. However, the period approach is not going as I expected.

Cohort approach (blue lines in Figure 1):

Age at diagnosis and year of diagnosis are both modelled using restricted cubic splines with 3 degrees of freedom. Interactions between age and year are included. All main and interaction effects were allowed to be time-varying. The model had 4 degrees of freedom for the baseline and 2 for time-varying effects.

Code:
* stset the data
stset vitfup_years, failure(vitstat==1) id(patid) exit(time 10)
 
* Model
stpm2 agespl* yearspl* a1y? a2y? a3y?, scale(hazard) df(4) bhazard(rate) tvc(agespl* yearspl* a1y? a2y? a3y?) dftvc(2)
 
* Predict loss in life expectancy
predict ll, lifelost mergeby(_year sex _age) diagage(age) diagyear(yydx) nodes(40) tinf(80) using(popmortNL) survprob(prob) stub(surv) maxyear(2020) ci
This approach gives me no trouble and provides me, for example for patients aged 30, with the following estimates:

yydx ll ll_lci ll_uci
1989 9,390 5,994 12,785
1990 8,712 6,131 11,292
1991 8,096 6,136 10,056
. . . .
. . . .
. . . .
2010 7,508 5,739 9,276
2011 7,244 5,184 9,305
2012 6,956 4,465 9,447
2013 6,650 3,664 9,635
2014 6,345 2,855 9,836

Period approach (pink lines in Figure 1):

For the period approach, age is modelled continuously using restricted cubic splines and year of diagnosis is NOT included in the model (Andersson et al. 2015; supplementary material; doi: 10.1186/s12885-015-1427-2). Age is allowed to be time-varying. The model had 4 degrees of freedom for the baseline and 2 for time-varying effects.

Code:
* stset the data
stset vitdat, origin(incdat) enter(time mdy(1,1,2010)) exit(time mdy(12,31,2014)) fail(vitstat==1) id(patid) scale(365.24)

* Model
stpm2 agespl*, scale(hazard) df(4) bhazard(rate) tvc(agespl*) dftvc(2)
yydx ll ll_lci ll_uci
2010 7,150 5,597 8,704
2011 7,149 5,596 8,702
2012 7,153 5,599 8,707
2013 7,152 5,598 8,706
2014 7,153 5,598 8,707

This approach gives estimates which seem to be linear over time for the period 2010–2014 (pink line). This is contradictory to what I expected. I expected loss in life expectancy to vary over time, like in the periode 1989-2009.

Can anyone tell me what I should do differently?