Unlike [wikipedia](https://en.wikipedia.org/wiki/Proportional_hazards_model), the linear predictor (LP) in `coxph` is made dependent from the baseline hazard function $\lambda$. For example, if a model has two predictors `x1`, `x2` stratified by `sex`, the linear predictor would be like:
$$LP = log (  \lambda _{sex}) + \beta _{1}x _{1}+\beta _{2}x _{2}$$

So we could calculate the LP without group information (sex) by hand using the $\beta _{1} \, \beta _{2}$ provided by `coef(fit)` and a made up constance $\lambda _{unkown}$.

Below is a simulation comparing the results from `predict` and by hand `coef(fit)`.
```
## Toy data
df <- list(time=c(4,3,1,1,2,2,3), 
          status=c(1,1,1,0,1,1,0), 
          x1=c(0,2,1,1,1,0,0), 
          x2 = 1:7,
          sex=c(0,0,0,0,1,1,1)) 

## Fit a stratified model 
fit <- coxph(Surv(time, status) ~ x1 + x2 + strata(sex), df) 
coef(fit)
#         x1          x2 
# 0.79451881 -0.01917633

## Baseline linear predictor
lambda_sex1 <- predict(fit, newdata=list(x1=0, x2=0, sex=0))
# -0.746578
lambda_sex2 <- predict(fit, newdata=list(x1=0, x2=0, sex=1))
# -0.1497816

## by predict function
predict(fit, newdata=list(x1=1, x2=2, sex=0))
# 0.009588167
predict(fit, newdata=list(x1=1, x2=2, sex=1))
# 0.6063845

## by hand
lambda_sex1 + coef(fit)[1]*1 + coef(fit)[2]*2
# 0.009588167
lambda_sex2 + coef(fit)[1]*1 + coef(fit)[2]*2
# 0.6063845
```
Finally, to calculate "a" linear predictor for an unknown `sex`:
```
# prevented by predict function
predict(fit, newdata=list(x1=1, x2=2, sex=2))
#Error in model.frame.default(data = list(x1 = 1, x2 = 2, sex = 2), formula = ~x1 +  : 
# factor strata(sex) has new level sex=2

# by hand, assuming lambda for sex 2 is 0
0 + coef(fit)[1]*1 + coef(fit)[2]*2
# 0.7561661
```