按时间窗口条件组替换缺失值

Question

我按subject_id分组。

变量 acpa 在 row_type == BIOBANK 所在的行上测量。 我想使用index_date 变量替换其他行上的<NA>，但只能在给定的时间窗口内。

这是我的桌子：

   subject_id index_date      row_type acpa
 1:        155 2010-05-12 BASELINE_info <NA>
 2:        155 2010-05-12 BASELINE_info <NA>
 3:        155 2010-05-12       BIOBANK    N
 4:        155 2010-05-12 PHYSICAL_exam <NA>
 5:        155 2010-09-29     FOLLOW_UP <NA>
 6:        155 2011-11-30     FOLLOW_UP <NA>
 7:        155 2013-06-01     FOLLOW_UP <NA>
 8:        155 2014-06-01     FOLLOW_UP <NA>
 9:        155 2015-06-01     FOLLOW_UP <NA>
10:        155 2016-08-31     FOLLOW_UP <NA>
11:        568 2012-04-07 BASELINE_info <NA>
12:        568 2012-04-07 BASELINE_info <NA>
13:        568 2012-04-07     FOLLOW_UP <NA>
14:        568 2012-04-19       BIOBANK    H
15:        568 2012-04-19 PHYSICAL_exam <NA>
16:        568 2013-06-01     FOLLOW_UP <NA>
17:        568 2013-12-12       BIOBANK    H
18:        568 2013-12-12 PHYSICAL_exam <NA>
19:        568 2014-04-27     FOLLOW_UP <NA>
20:        568 2014-05-22       RA_DIAG <NA>
21:        568 2014-05-28 PHYSICAL_exam <NA>
22:        568 2018-11-06       BIOBANK    L
23:        568 2018-11-06 PHYSICAL_exam <NA>
24:        568 2018-11-27     FOLLOW_UP <NA>
25:        568 2019-06-20     FOLLOW_UP <NA>
    subject_id index_date      row_type acpa

例如，在第 3 行，acpa = "N"。 我想执行诸如应用na.locf() 之类的操作，但仅适用于对应index_date 周围-180 到+365 天之间的行。

library(date.table)
library(zoo)
TABLE[, acpa := na.locf(acpa, na.rm = F), by = subject_id] 

我如何为na.locf() 构建一个rule 参数，这样它只会填补我的时间窗口中的空白？

我想要的输出看起来像：

   subject_id index_date      row_type acpa
 1:        155 2010-05-12 BASELINE_info    N
 2:        155 2010-05-12 BASELINE_info    N
 3:        155 2010-05-12       BIOBANK    N
 4:        155 2010-05-12 PHYSICAL_exam    N
 5:        155 2010-09-29     FOLLOW_UP    N
 6:        155 2011-11-30     FOLLOW_UP <NA>
 7:        155 2013-06-01     FOLLOW_UP <NA>
 8:        155 2014-06-01     FOLLOW_UP <NA>
 9:        155 2015-06-01     FOLLOW_UP <NA>
10:        155 2016-08-31     FOLLOW_UP <NA>
11:        568 2012-04-07 BASELINE_info    H
12:        568 2012-04-07 BASELINE_info    H
13:        568 2012-04-07     FOLLOW_UP    H
14:        568 2012-04-19       BIOBANK    H
15:        568 2012-04-19 PHYSICAL_exam    H
16:        568 2013-06-01     FOLLOW_UP    H
17:        568 2013-12-12       BIOBANK    H
18:        568 2013-12-12 PHYSICAL_exam    H
19:        568 2014-04-27     FOLLOW_UP    H
20:        568 2014-05-22       RA_DIAG    H
21:        568 2014-05-28 PHYSICAL_exam    H
22:        568 2018-11-06       BIOBANK    L
23:        568 2018-11-06 PHYSICAL_exam    L
24:        568 2018-11-27     FOLLOW_UP    L
25:        568 2019-06-20     FOLLOW_UP    L
    subject_id index_date      row_type acpa

如果由于时间窗口重叠而发生冲突，我会保留最后一个（前一个）非 NA 的值。

Answer 1

我的数据经理同事提出了这个建议：

从前瞻预测开始。

    DT[!is.na(acpa), variable_t := index_date] # variable_t copied as the index date of the variable to impute
    DT[, acpa_c := na.locf(acpa, na.rm = F), by = subject_id] # locf for the variable in a new variable
    DT[, variable_t := na.locf(variable_t, na.rm = F), by = subject_id] # locf of the variable_time
    DT[, delay := as.numeric(difftime(index_date, variable_t ,units = "days"))] # calculating on each row the delay 
    DT[delay > 365, acpa_c := NA] # cutting off values above my threshold

那么，向后投影也是如此。

    DT[, acpa_c := na.locf(acpa_c , na.rm = F, fromLast = T),by = subject_id] ### na.lofb, in the corrected variable
    DT[!is.na(acpa), variable_t := index_date] # reset of the temporary variable_time, so we can projet it backward
    DT[, variable_t := na.locf(ratio_t, na.rm = F, fromLast = T),by = subject_id] # locb of the variable_time
    DT[, delay := as.numeric(difftime(index_date, variable_t , units = "days"))] # delay
    DT[delay < -180, acpa_c := NA] # cutting off values above my threshold

如果一切正常，则用新值替换。

DT[, acpa := acpa_c]
DT[, acpa_c := NULL]