Summarise and test numeric variables, in the style of SAS PROC UNIVARIATE

Produces the statistics data set SAS PROC UNIVARIATE writes with OUTPUT OUT=: one row per analysis variable (per class level), one column per requested statistic in the order requested, then one column per PCTLPTS= percentile. Beyond proc_means it adds weighted and decimal percentiles, VARDEF=, and the tests for location and normality.

Usage

proc_univariate(
  data,
  vars = NULL,
  class = NULL,
  stats = c("n", "median", "mean", "std", "cv", "min", "max"),
  weights = NULL,
  pctlpts = NULL,
  pctlpre = "p",
  mu0 = 0,
  vardef = c("df", "n", "wdf", "weight")
)

Arguments

data: A data frame, tibble, or similar tabular object.
vars: Character vector of columns to analyse. NULL (default) selects every numeric column not named in class or weights.
class: Character vector of grouping columns, prepended to the result as leading columns. Rows with a missing value in any class variable are dropped, and levels follow ORDER=INTERNAL, as in proc_means. Unlike proc_means(), a level whose every weight is missing has no row, as in SAS.
stats: Character vector of SAS statistic keywords: every keyword proc_means accepts, plus "stdmean", "t", "probt", "msign", "probm", "signrank", "probs", "normal" and "probn".
weights: Character or NULL. Name of a single numeric column of data to use as an observation weight, SAS WEIGHT. Observations whose weight is missing are excluded from every statistic except nobs, which counts them.
pctlpts: Numeric vector of percentile points in [0, 100], decimals allowed, or NULL. SAS PCTLPTS=.
pctlpre: Single string prefixed to the pctlpts column names. SAS PCTLPRE=. The point follows with . replaced by _, so 2.5 gives p2_5.
mu0: Single finite number, the null value for t, msign and signrank. SAS MU0=.
vardef: The variance divisor: "df" (default), "n", "wdf" or "weight". SAS VARDEF=.

Value

A data frame with one row per analysis variable per class combination. Columns are the class variables (when supplied), then variable, label, then one column per stats keyword in the order given, then one column per pctlpts point in the order given. n, nmiss and nobs are integer; all others are numeric.

Details

Weighting. Weights apply as in proc_means, except that the quantiles are weighted: median, q1, q3, qrange, pNN and the pctlpts columns. The weighted percentile sorts the values with their cumulative weights \(S_i\); the 0th percentile is the minimum and the 100th the maximum; otherwise, when \(S_i = pW\) the result is the mean of \(x_i\) and \(x_{i+1}\), else the first \(x_i\) with \(S_i > pW\). Unweighted quantiles use stats::quantile(type = 2), SAS's default PCTLDEF=5. mode stays unweighted, as in SAS. stderr and stdmean divide the standard deviation by the square root of the sum of the weights. A zero or negative weight is an error; SAS instead treats a negative weight as zero and still counts the observation.

VARDEF. The variance divisor is \(n - 1\) ("df"), \(n\) ("n"), \(W - 1\) ("wdf") or \(W\) ("weight"), where \(W\) is the sum of the weights (\(n\) without weights). It changes var, std and cv. Under "n", skewness and kurtosis are the moment forms rather than the adjusted ones.

Statistics SAS does not compute are NA, as SAS writes missing:

msign, probm, signrank, probs, normal, probn when weights is given;
t, probt, stderr and stdmean when vardef is not "df";
skewness and kurtosis when vardef is "wdf" or "weight";
normal and probn above 2000 observations, with one warning per call: SAS switches to a Kolmogorov D test there, which is not ported;
std, var, cv, stderr, stdmean, t, probt at one observation, skewness below three and kurtosis below four;
t, probt, skewness, kurtosis, normal, probn when every value is equal (cv is then 0 unless the mean is zero, where it is NA);
msign, probm, signrank, probs when no value differs from mu0;
mode when no value repeats among two or more values.

Tests. t is Student's t for the mean against mu0; msign the sign statistic; signrank the Wilcoxon signed rank statistic, with an exact p-value for 20 or fewer non-zero differences and SAS's t approximation above that. normal is the Shapiro-Wilk \(W\), from stats::shapiro.test(), which implements Royston's 1995 algorithm; SAS implements Royston's 1992 one, so normal and probn agree with SAS to about \(10^{-8}\), not to machine precision. At two observations both are 1, as SAS reports.

Examples