LOS_boot <- function(data1)
{
  ## Authors: Derek Hazard, Jerome Lambert
  ## Returns the duration/length of stay estimates for each bootstrap cohort for use with msboot function in mstate package
  ##
  ##
  ## Args:
  ##    data1: data frame with rows of all possible transitions for which a 
  ##    patient is at risk (i.e. returned data frame from ext_mstate function)
  ##
  ## Returns:
  ##    vector with c(1,2,...): non-ventilation/ventilation duration estimates for patients non-MV at admission
  ##    c(...,3,4,...): non-ventilation/ventilation duration estimates for patients MV at admission
  ##    c(...,5): length of ICU stay estimates for entire cohort
  
# Cox model stratified by transition
coxf <- coxph(Surv(entry, exit, status) ~ strata(trans), data1, method = "breslow")

# msfit calculates baseline hazards
msf_f <- msfit(coxf,trans = tra)

# probtrans calculates transition probabilities, prediction from day 0
pt_f <- probtrans(msf_f, predt = 0)

# ELOS gives expected length of stay for 28 days in the states for patients in the states at day 0
LOS_f <- ELOS(pt_f, 28)
    
# Determine initial distribution of patients at time = 0
# Subset of transtions from time = 0
data1_start1 <- subset(data1, data1$entry ==0)

# Keep relevant variables
data1_start2 <- data.frame(data1_start1$id, data1_start1$from)

# Remove multiple entries
data1_start3 <- unique(data1_start2)

colnames(data1_start3) <- c("id", "from")
    
# Count of patients in each initial state
temp_table <- table(data1_start3$from)

# Calculate inital proportions, no patients are discharged or dead at admission
init_dis <- c(temp_table[1]/nrow(data1_start3), temp_table[2]/nrow(data1_start3), 0, 0)
    
# Multiply LOS_mat_1 with initial distribution to get weighted average of expected lengths of stay for entire cohort
LOS_f_init <- init_dis %*% LOS_f

# Return estimates    
return(c(LOS_f[1,1:2],LOS_f[2,1:2], LOS_f_init[1:2]))
  }