8.4 Screening Extreme & Missing Values with PROC UNIVARIATE

8.4 Screening Extreme & Missing Values with PROC UNIVARIATE

Quick Answer: The content guide's summary-reporting objective ends with a bullet that names one procedure by name: "Identify extreme and missing values with the UNIVARIATE procedure." PROC UNIVARIATE produces a far richer distributional profile than PROC MEANS, and its Extreme Observations table — the five lowest and five highest values with their observation numbers — is the feature the exam is pointing at.

---

1. The Simplest Invocation

proc univariate data=sashelp.class;
   var Height Weight;
run;

For each analysis variable, SAS prints five tables:

That last table is the reason the procedure exists for this objective. PROC MEANS tells you the minimum is -999; PROC UNIVARIATE tells you the minimum is -999, that it occurs at observation 412, and that the next four smallest values are 50, 51, 52, and 53 — which immediately identifies -999 as a sentinel code rather than a real measurement.

---

2. Reading the Extreme Observations Table

                          Extreme Observations

        -----Lowest-----            -----Highest----

        Value      Obs              Value      Obs

          -999     412                  78       19
            50      88                  79        3
            51     205                  80       11
            52      17                  81       15
            53      94                 999      377

Two sentinel values (-999 and 999) stand out instantly against a body of plausible heights, and the Obs column tells you exactly which rows to inspect.

Controlling How Many Extremes Appear

/* Show the 10 lowest and 10 highest instead of the default 5 */
proc univariate data=work.patients nextrobs=10;
   var SystolicBP;
run;

NEXTROBS=n sets the number of extreme observations listed. A closely related option, NEXTRVAL=n, prints the n extreme values without observation numbers.

Making the Listing Meaningful with ID

Observation numbers are useless once a data set has been sorted or subset. The ID statement replaces them with a real key.

proc univariate data=work.patients nextrobs=10;
   var SystolicBP;
   id PatientID;
run;

The Extreme Observations table now shows PatientID beside each value, so the outlier can be traced back to a record rather than a row position.

---

3. Counting Missing Values

The Moments table reports N and N Miss side by side:

                                Moments

N                          185    Sum Weights              185
Mean                  62.336757   Sum Observations       11532.3
Std Deviation         5.1270985   Variance              26.287139
N Miss                      15

N is the count of non-missing values used in every statistic; N Miss is the count of missing values excluded from them. For a variable meant to be fully populated, a non-zero N Miss is an immediate data-quality finding — and you get it without writing a separate PROC MEANS ... NMISS or PROC FREQ ... MISSING step.

Exam Trap: every statistic in PROC UNIVARIATE — like every statistic in PROC MEANS — is computed on non-missing values only. A mean of 62.3 over N = 185 says nothing about the 15 missing rows. N Miss is what quantifies the gap.

---

4. Subgroup Screening with CLASS and BY

/* One profile per treatment group, no pre-sorting required */
proc univariate data=work.trial;
   class Treatment;
   var Response;
   id SubjectID;
run;

/* BY requires the data to be sorted first */
proc sort data=work.trial out=work.trial_sorted;
   by Site;
run;

proc univariate data=work.trial_sorted;
   by Site;
   var Response;
run;

As in PROC MEANS, CLASS handles unsorted data while BY requires a prior PROC SORT. Screening by subgroup often reveals that an "outlier" is perfectly normal within its own stratum.

---

5. Trimming the Output

The full five-table report is heavy when you only want the extremes. Two mechanisms cut it down:

/* Statement-level: keep only the extremes and the moments */
ods select ExtremeObs Moments;

proc univariate data=work.patients nextrobs=10;
   var SystolicBP;
   id PatientID;
run;

ods select all;   /* restore default behaviour */

/* Capture the extremes into a data set for programmatic follow-up */
ods output ExtremeObs = work.outliers;

proc univariate data=work.patients nextrobs=10;
   var SystolicBP;
run;

The NOPRINT option suppresses all printed output, which is appropriate when the goal is purely to build an output data set.

---

6. UNIVARIATE versus MEANS

The one-line rule: PROC MEANS summarizes the data you believe in; PROC UNIVARIATE tells you whether you should.