← SASnRRetain minimum score upto the current observation

Introduction

A running minimum is the smallest score seen so far within a subject — useful for worst-observed-value summaries, minimum nadir calculations, and similar clinical derivations.
The running minimum resets at the start of each subject so values from one subject never leak into another.
SAS uses RETAIN with min(score, minscore) inside a DATA step, resetting the retained value on first.usubjid.
Tidyverse uses group_by() + cummin(), with an Inf placeholder for NAs so missing values do not break the running calculation.
Base R uses ave() with a helper that replaces NAs with Inf and calls cummin() per subject.
This lesson also fills the rows where the original score was NA with the running minimum, so the score column becomes a monotonically non-increasing series within each subject.

Retain minimum score upto the current observation

SAS (Base SAS)+

*==============================================================================;
* Create the input dataset;
*==============================================================================;
data pmin;
    input usubjid visitnum score;
    datalines;
101 1 20
101 2 .
101 3 30
102 1 90
102 2 10
102 3 .
102 4 91
102 5 .
;
run;

*==============================================================================;
* Sort the dataset by usubjid and visitnum;
*==============================================================================;

proc sort data=pmin;
    by usubjid visitnum;
run;

*==============================================================================;
* Retain previous non-missing low score;
*==============================================================================;

data pmin01;
    set pmin;
    by usubjid visitnum;

    retain minscore;

    * Store the original score;
    orig_score = score;

    if first.usubjid then minscore = .;

    *Store minimum score and assign minscore when score is missing;
    if not missing(score) then minscore = min(score,minscore);
    else score=minscore;
run;

To carry the running minimum score forward within each subject, we combine RETAIN with a conditional min() on each row.
After sorting by subject and visit, we reset minscore at the start of each subject with if first.usubjid then minscore = .; — otherwise the minimum from the previous subject would leak in.
On every non-missing row, we update minscore = min(score, minscore). SAS treats missing minscore (the reset value) as smaller than any number, so the first non-missing score becomes the starting minimum automatically.
On rows where score itself is missing, we substitute the running minscore into score — filling NAs with the running low.
We keep orig_score as a snapshot of the original value so the substitutions are auditable.
After running, we check that minscore never increases within a subject and that filled rows show the running minimum rather than the original missing.

R (tidyverse)+

#==============================================================================;
#Retain previous non-missing low score;
#==============================================================================;

pmin <- tribble(
  ~usubjid,~visitnum,~score,
  101,1,20,
  101,2,NA,
  101,3,30,
  102,1,90,
  102,2,10,
  102,3,NA,
  102,4,91,
  102,5,NA
)

pmin01 <- pmin %>%
  arrange(usubjid,visitnum) %>% 
  group_by(usubjid) %>% 
  mutate(
    orig_score = score,
    temp_score = if_else(is.na(score),Inf,score),
    minscore=cummin(temp_score),
    score = if_else(is.na(score),minscore,score)
  )

In tidyverse we reach for cummin() — the built-in cumulative minimum — and wrap it inside a grouped mutate().
cummin() does not know how to handle NAs the way SAS does. If an NA slips in, everything after it becomes NA too.
We work around that by replacing NAs with Inf in a temporary column (temp_score) before running cummin(). Inf is larger than every real score, so it never wins the minimum — the running low is preserved.
group_by(usubjid) resets the cumulative minimum at each subject boundary, matching the SAS if first.usubjid then minscore = .; pattern.
Finally we put the cumulative minimum back into score only where the original value was NA, keeping real observations untouched.
After running, we verify that within each subject the minscore column is monotonically non-increasing, and NAs in the original score have been replaced with the running low.

Happy learning!

Search lessons

Introduction

Retain minimum score upto the current observation