# Problem with reference category after survSplit in Cox regression

I have two covariates in a survival data, Acet with four factor and Age as a continuous variable. I want category AAA to be the reference in the covariate Acet. Normal Cox regression is fine:

> summary(fit)
Call:
coxph(formula = Surv(Seurantaaikavuosia, Endpoint == 1) ~ Acet +
Age, data = datah)

n= 1262, number of events= 179

coef exp(coef)  se(coef)      z Pr(>|z|)
AcetBBB  0.983612  2.674098  0.467294  2.105 0.035299 *
AcetDDD  0.749951  2.116897  0.506057  1.482 0.138353
AcetOOO  1.555544  4.737661  0.458826  3.390 0.000698 ***
Age     -0.021558  0.978672  0.007824 -2.755 0.005863 **
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

exp(coef) exp(-coef) lower .95 upper .95
AcetBBB    2.6741     0.3740    1.0701    6.6825
AcetDDD    2.1169     0.4724    0.7851    5.7076
AcetOOO    4.7377     0.2111    1.9276   11.6444
Age        0.9787     1.0218    0.9638    0.9938

Concordance= 0.632  (se = 0.021 )
Rsquare= 0.029   (max possible= 0.862 )
Likelihood ratio test= 37.19  on 4 df,   p=2e-07
Wald test            = 34.13  on 4 df,   p=7e-07
Score (logrank) test = 36.46  on 4 df,   p=2e-07


But the covariate Acet violates the proportional hazards assumptions and therefore I use the survSplit function to have time-dependent covariates.

> datah2<-survSplit(Surv(Seurantaaikavuosia,Endpoint==1)~.,data=datah,cut=c(2,7),episode="tgroup",id="id")
> summary(fit2)
Call:
coxph(formula = Surv(tstart, Seurantaaikavuosia, event == 1) ~
Acet:strata(tgroup) + Age, data = datah2)

n= 3596, number of events= 179

coef exp(coef)  se(coef)      z Pr(>|z|)
Age                            -0.021470  0.978759  0.007835 -2.740 0.006136 **
AcetAAA:strata(tgroup)tgroup=1  0.124520  1.132605  0.801794  0.155 0.876583
AcetBBB:strata(tgroup)tgroup=1 -0.495755  0.609111  0.626804 -0.791 0.428988
AcetDDD:strata(tgroup)tgroup=1  0.412182  1.510110  0.626907  0.657 0.510869
AcetOOO:strata(tgroup)tgroup=1        NA        NA  0.000000     NA       NA
AcetAAA:strata(tgroup)tgroup=2 -2.157216  0.115647  0.718084 -3.004 0.002663 **
AcetBBB:strata(tgroup)tgroup=2 -0.883994  0.413130  0.232617 -3.800 0.000145 ***
AcetDDD:strata(tgroup)tgroup=2 -0.844255  0.429878  0.326606 -2.585 0.009740 **
AcetOOO:strata(tgroup)tgroup=2        NA        NA  0.000000     NA       NA
AcetAAA:strata(tgroup)tgroup=3 -1.622617  0.197382  1.019774 -1.591 0.111575
AcetBBB:strata(tgroup)tgroup=3 -0.136532  0.872378  0.271062 -0.504 0.614476
AcetDDD:strata(tgroup)tgroup=3 -1.386624  0.249918  0.607573 -2.282 0.022476 *
AcetOOO:strata(tgroup)tgroup=3        NA        NA  0.000000     NA       NA
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

exp(coef) exp(-coef) lower .95 upper .95
Age                               0.9788     1.0217   0.96384    0.9939
AcetAAA:strata(tgroup)tgroup=1    1.1326     0.8829   0.23528    5.4521
AcetBBB:strata(tgroup)tgroup=1    0.6091     1.6417   0.17830    2.0808
AcetDDD:strata(tgroup)tgroup=1    1.5101     0.6622   0.44196    5.1598
AcetOOO:strata(tgroup)tgroup=1        NA         NA        NA        NA
AcetAAA:strata(tgroup)tgroup=2    0.1156     8.6470   0.02831    0.4725
AcetBBB:strata(tgroup)tgroup=2    0.4131     2.4205   0.26187    0.6518
AcetDDD:strata(tgroup)tgroup=2    0.4299     2.3262   0.22664    0.8154
AcetOOO:strata(tgroup)tgroup=2        NA         NA        NA        NA
AcetAAA:strata(tgroup)tgroup=3    0.1974     5.0663   0.02675    1.4566
AcetBBB:strata(tgroup)tgroup=3    0.8724     1.1463   0.51283    1.4840
AcetDDD:strata(tgroup)tgroup=3    0.2499     4.0013   0.07597    0.8222
AcetOOO:strata(tgroup)tgroup=3        NA         NA        NA        NA

Concordance= 0.647  (se = 0.02 )
Rsquare= 0.014   (max possible= 0.501 )
Likelihood ratio test= 50.39  on 10 df,   p=2e-07
Wald test            = 41.94  on 10 df,   p=8e-06
Score (logrank) test = 47.37  on 10 df,   p=8e-07


Now the OOO is suddenly the reference category. It should still be the same variable type:

\$ Acet                               : Factor w/ 4 levels "AAA","BBB","DDD",..: 4 4 4 4 4 4 4 4 4 4 ...


How can I have the AAA as the reference in the splitted regression?