*Copyright @ www.mycsg.in;

What is PROC UNIVARIATE

• PROC UNIVARIATE produces a detailed statistical summary of one or more numeric variables
• Compared to PROC MEANS, which focuses on basic summary statistics, PROC UNIVARIATE also reports percentiles, extreme values, tests for normality, and optionally produces histograms and probability plots
• It is especially useful for understanding the distribution of a variable — whether it is symmetric, skewed, or has extreme outliers
• In clinical programming, PROC UNIVARIATE is commonly used for verifying value ranges, identifying outliers, and assessing whether a continuous endpoint meets normality assumptions before statistical testing

Create sample data

• The dataset `lab_results` contains simulated haemoglobin values for a small group of subjects
• A few values are intentionally set at the extremes to make the percentile and extreme-value output more informative

data lab_results; infile datalines truncover; input usubjid $ sex $ age hgb; datalines; 1001 F 45 12.1 1002 M 52 14.8 1003 F 38 11.5 1004 M 61 16.2 1005 F 29 13.0 1006 M 47 15.5 1007 F 55 9.8 1008 M 33 17.1 1009 F 42 12.8 1010 M 58 14.0 ; run; Copy Code View Log

• `hgb` is the haemoglobin value in g/dL — the dataset has ten subjects with a deliberate low value (9.8) and a high value (17.1) to exercise the extreme-values section of the output
• Inspect the dataset and note the range of values before running PROC UNIVARIATE

View Data

Basic PROC UNIVARIATE output

• The default PROC UNIVARIATE output contains several panels for each variable
• The Moments panel shows the mean, standard deviation, variance, skewness, kurtosis, and sample size
• The Basic Statistical Measures panel shows mean, median, mode, standard deviation, variance, and range
• The Tests for Location panel tests whether the mean or median differs significantly from zero
• The Quantiles panel shows percentiles including the minimum, maximum, and quartiles
• The Extreme Observations panel lists the five smallest and five largest values with their observation numbers

proc univariate data=lab_results; var hgb; run; Copy Code View Log

• Review the Quantiles section — the minimum should be 9.8 and the maximum 17.1
• The median (50th percentile) is the middle value when all ten are ranked — verify it is between 12.8 and 13.0
• The Extreme Observations section lists subject 1007 (9.8) as the lowest value and subject 1008 (17.1) as the highest — confirm this matches the source data

Adding an ID variable to extreme observations

• By default, the Extreme Observations panel identifies rows by their observation number
• The `id` statement replaces the observation number with a chosen variable, making it much easier to trace extreme values back to specific subjects
• In clinical data, you would typically use the subject identifier variable as the ID

proc univariate data=lab_results; var hgb; id usubjid; run; Copy Code View Log

• The Extreme Observations panel now shows `usubjid` values instead of row numbers
• Confirm that the five lowest and five highest values are identified by subject ID — this is far more useful than a row number when investigating outliers

Running PROC UNIVARIATE by group

• A `by` statement produces a separate PROC UNIVARIATE output panel for each group
• The data must be sorted by the BY variable before the procedure runs
• This is useful when you need distribution summaries broken out by sex, treatment group, or visit

proc sort data=lab_results; by sex; run; proc univariate data=lab_results; var hgb; id usubjid; by sex; run; Copy Code View Log

• Two separate output sections are produced — one for female subjects and one for male subjects
• Compare the medians and ranges between the two groups — the female group includes subject 1007 with the low value of 9.8 which will pull the female minimum down

Saving key statistics to an output dataset

• The `output` statement saves selected statistics from PROC UNIVARIATE into a new dataset
• Each statistic keyword is followed by `=` and the name to give that variable in the output dataset
• Commonly saved statistics include `mean=`, `median=`, `std=`, `min=`, `max=`, `q1=` (25th percentile), and `q3=` (75th percentile)
• Saving statistics to a dataset is essential when you need to use them programmatically — for example to flag values more than two standard deviations from the mean

proc univariate data=lab_results noprint; var hgb; output out=hgb_stats mean = hgb_mean median = hgb_median std = hgb_std min = hgb_min max = hgb_max q1 = hgb_q1 q3 = hgb_q3 n = hgb_n; run; Copy Code View Log

• The `noprint` option suppresses the usual printed output — the statistics are written only to `hgb_stats`
• Inspect `hgb_stats` — it should contain a single row with one column per requested statistic
• Verify the mean, min, max, and n values against what you observed in the earlier full PROC UNIVARIATE output

View Data

Using saved statistics to flag outliers

• Once statistics are in a dataset, they can be merged back to the source data to derive flags
• A common rule flags values more than two standard deviations above or below the mean as potential outliers
• The step below merges the summary statistics into the subject-level data using a cross-join pattern and then creates the flag

data lab_flagged; if _n_ = 1 then set hgb_stats; set lab_results; lower_limit = hgb_mean - 2 * hgb_std; upper_limit = hgb_mean + 2 * hgb_std; outlier_flag = (hgb lt lower_limit or hgb gt upper_limit); run; Copy Code View Log

• The `if _n_ = 1 then set hgb_stats;` pattern reads the single-row statistics dataset once and retains its values across all subsequent rows — this is a standard SAS idiom for merging a summary row into detail data
• Inspect `lab_flagged` and check the `outlier_flag` column — subjects whose `hgb` falls outside the two-standard-deviation band should have `outlier_flag = 1`
• Also inspect `lower_limit` and `upper_limit` to confirm the band is calculated correctly

View Data

Key points to remember

• PROC UNIVARIATE provides richer distributional detail than PROC MEANS — moments, quantiles, extreme values, and normality tests
• The `id` statement replaces observation numbers with a subject identifier in the Extreme Observations panel
• A `by` statement produces separate output panels per group — data must be pre-sorted
• The `output` statement saves named statistics to a dataset — use `noprint` to suppress the printed output when only the dataset is needed
• Saved statistics can be merged back to detail data using the `if _n_ = 1 then set` pattern to derive per-subject flags or derived values