R data.table replace missing value by next non missing value

data[,age.fill := nafill(age,'nocb'),by=horsenum][, 
     rdate.fill:=ifelse(is.na(age),rdate[which.min(age.fill==age)],rdate),by=horsenum][,
      age.fill:=unclass(age.fill - round((rdate.fill-rdate)/365))
      ]



  index      rdate horsenum age ind4 age.fill rdate.fill
1: 14704 2009-03-01     K123  NA   10        3 2015-09-06
2: 14767 2009-03-01     K212  NA    9        2 2015-10-10
3: 39281 2011-10-09     K123  NA   10        6 2015-09-06
4: 39561 2011-10-19     K212  NA    9        5 2015-10-10
5: 74560 2015-04-07     K212  NA    9        8 2015-10-10
6: 77972 2015-09-06     K123  10   NA       10 2015-09-06
7: 79111 2015-10-10     K212   9   NA        9 2015-10-10
8: 84233 2016-03-28     K212  10   NA       10 2016-03-28

Your algorithm systematically underestimates age. For example, horse K212's estimated age on 2015-04-07 (row 5) is 8. However, we know K212's age on 2016-03-28 is 10 (row 8), so K212 must be 9 on 2015-04-07, not 8. Here I address this problem by calculating an estimated birthdate from each non-NA rdate, then calculating the earliest estimated birthdate for each horse.

library(data.table)
data=data.table(index=c(14704L,14767L,39281L,39561L,74560L,77972L,79111L,84233L),rdate=structure(c(14304,14304,15256,15266,16532,16684,16718,16888),class="Date"),horsenum=c("K123","K212","K123","K212","K212","K123","K212","K212"),age=c(NA,NA,NA,NA,NA,10,9,10))

lt = data[!is.na(age),as.POSIXlt(rdate)]
lt$year = lt$year - data[!is.na(age),age]
data[!is.na(age),bday:=as.Date(lt)]
data[,bday:=min(bday,na.rm=T),horsenum]
data[,age.fill:=floor(as.numeric(rdate-bday)/365)]
data[order(index)]

Output:

   index      rdate horsenum age       bday age.fill
1: 14704 2009-03-01     K123  NA 2005-09-06        3
2: 14767 2009-03-01     K212  NA 2006-03-28        2
3: 39281 2011-10-09     K123  NA 2005-09-06        6
4: 39561 2011-10-19     K212  NA 2006-03-28        5
5: 74560 2015-04-07     K212  NA 2006-03-28        9
6: 77972 2015-09-06     K123  10 2005-09-06       10
7: 79111 2015-10-10     K212   9 2006-03-28        9
8: 84233 2016-03-28     K212  10 2006-03-28       10

Note: this algorithm could be improved. Consider K212 is 9 on 2015-10-10 and 10 on 2016-03-28. This means that K212's actual birthday is after 10-10 and before 3-28. Instead of assuming 3-28, we could assume it is halfway between 10-10 and 3-28, or, more specifically, if there is more than one estimated birthdate, calculate both the max and the min possible birthdate for each horse, then find the date that's halfway between max - 1 year and min.

The approach below is slightly different:

It calculates, based on the given ages, the possible 'range' of the birthday from the horse. It then uses this window to calculate the minimum and maximum age a horse can haveon the given rdate.

So, the more infor you have in te horse's age, the smaller the window of possible birthdays, and the bigger the chance that the minimum estimates age equals the maximum estimated age (of they are the same, you know the age of the horse for sure )..

Here we go:

library( data.table )
library( lubridate )  #for the %m+% and %m-% operators
library( intervals )  #to calculate with intervals and find overlaps
library( eeptools )   #for age_calc function; calculating the age, given a date and a birthday (respects leap yaers, etc..)

#read sample data
DT <- fread("
index      rdate horsenum age ind4
14704 2009-03-01     K123  NA   10
14767 2009-03-01     K212  NA    9
39281 2011-10-09     K123  NA   10
39561 2011-10-19     K212  NA    9
74560 2015-04-07     K212  NA    9
77972 2015-09-06     K123  10   NA
79111 2015-10-10     K212   9   NA
84233 2016-03-28     K212  10   NA")

#set dates as IDate
DT[, rdate := as.POSIXct(rdate) ]
#set keys
setkey( DT, horsenum, rdate, age )
#calculate bandwidth date of birth (dob) based on age and date
DT[!is.na( age ), dob_min := as.integer( rdate %m-% lubridate::years(age + 1) %m+% lubridate::days(1) ) ]
DT[!is.na( age ), dob_max := as.integer( rdate %m-% lubridate::years(age) ) ]

#function to get get overlap of birthday-intervals
myfun <- function( y ) {
  all_intervals <- intervals::Intervals( as.matrix( y ), check_valid = TRUE ) 
  int_min <- all_intervals[1]
  for (i in 1:nrow(all_intervals) ) int_min <- interval_intersection( all_intervals[1], all_intervals[i] )
  as.data.table( int_min )
}

#get range of possible date of birth for each horsenum
dob_range <- DT[ !is.na(age), myfun( .SD ), by = .(horsenum), .SDcols = c("dob_min", "dob_max") ]
dob_range <- dob_range[, .(horsenum, dob_from = as.POSIXct(V1, origin = "1970-01-01"), 
                           dob_to = as.POSIXct(V2, origin = "1970-01-01"))]

#use found ranges of birthday to estimate ages
#first join dob-ranges by horsenum
DT[ dob_range, `:=`( dob_from = i.dob_from, dob_to = i.dob_to), on = .(horsenum)]
#now calculate ages (minimum and maximum)
DT[, age_min := floor( eeptools::age_calc( as.Date(dob_to), as.Date(rdate), units= "years" ) )]
DT[, age_max := floor( eeptools::age_calc( as.Date(dob_from), as.Date(rdate), units= "years" ) )]

#remove helper columns
DT[, `:=`( dob_min = NULL, dob_max = NULL, dob_from = NULL, dob_to = NULL)]


#    index      rdate horsenum age ind4 age_min age_max
# 1: 14704 2009-03-01     K123  NA   10       3       4
# 2: 39281 2011-10-09     K123  NA   10       6       7
# 3: 77972 2015-09-06     K123  10   NA      10      10
# 4: 14767 2009-03-01     K212  NA    9       2       3
# 5: 39561 2011-10-19     K212  NA    9       5       6
# 6: 74560 2015-04-07     K212  NA    9       9       9
# 7: 79111 2015-10-10     K212   9   NA       9       9
# 8: 84233 2016-03-28     K212  10   NA      10      10