sasLM for SAS Users
In sasLM: 'SAS' Linear Model

knitr::opts_chunk$set(comment = NA)
options(width = 100)
library(sasLM)

Why sasLM?

Many statisticians move between SAS and R, and they expect the same numbers from both. For unbalanced or complex designs, however, popular R functions often produce sums of squares, standard errors, or least squares means that differ from those of SAS PROC GLM. The differences are not necessarily errors - they come from different conventions (types of sums of squares, coding of singular designs, denominators, quantile definitions, and so on).

sasLM implements the conventions of SAS so that the results match SAS PROC GLM, REG, ANOVA, TTEST, FREQ (2x2 tables), and UNIVARIATE. The package is written in base R with only mvtnorm as an additional dependency, and its results have been validated against SAS outputs and the textbooks listed in ?sasLM.

Procedure-to-function map

| SAS | sasLM | |-----------------------------|----------------------------------------| | PROC GLM | GLM, aov1, aov2, aov3, EMS | | PROC GLM SOLUTION | GLM(..., BETA=TRUE), REG | | PROC GLM LSMEANS | LSM, GLM(..., EMEAN=TRUE) | | PROC GLM LSMEANS / PDIFF | PDIFF, Diffogram | | PROC GLM ESTIMATE/CONTRAST | est, ESTM, CIest, CONTR | | PROC GLM RANDOM / TEST | RanTest, EMS, T3test | | PROC GLM SLICE | SLICE | | PROC REG | REG, lr, lr0, regD, Coll | | PROC ANOVA | aov1 | | PROC TTEST | TTEST, tmtest, mtest, vtest | | PROC UNIVARIATE | UNIV and the descriptive functions | | PROC FREQ (2x2, stratified) | RD, RR, OR, RDmn, RRmn, ORmn, ORcmh |

PROC GLM

The SAS code

PROC GLM DATA=np;
  CLASS block N P K;
  MODEL yield = block N*P*K / SOLUTION;
  LSMEANS N*P / CL;
RUN;

corresponds to:

GLM(yield ~ block + N*P*K, npk)

GLM returns the overall ANOVA, fit statistics, and the Type I, II, and III tables in one call. Note the agreement with SAS for this unbalanced design - this is the core purpose of the package.

Coefficients as with the SOLUTION option:

GLM(yield ~ block + N*P*K, npk, BETA=TRUE)$Parameter

Least squares means with confidence limits:

LSM(weight ~ feed, chickwts, "feed")

Pairwise differences as with LSMEANS / PDIFF, including the Tukey adjustment:

PDIFF(weight ~ feed, chickwts, "feed", adj="tukey")

The Dunnett test uses the first level as the control when ref is not given, as SAS does:

PDIFF(weight ~ feed, chickwts, "feed", adj="dunnett")

Expected mean squares and random effects

EMS produces the expected mean squares table, and RanTest performs the tests with a random factor as the RANDOM / TEST statement of SAS PROC GLM:

EMS(yield ~ block + N*P*K, npk)
RanTest(yield ~ block + N*P*K, npk, Random="block")

PROC REG

REG(mpg ~ wt + hp, mtcars)

Heteroscedasticity-consistent standard errors (HC0, HC3) and White's test:

REG(mpg ~ wt, mtcars, HC=TRUE)

Weighted least squares follows the SAS WEIGHT statement: observations with nonpositive weights are excluded, and Resid=TRUE returns fitted values and residuals in the original scale as OUTPUT P= R= does.

w = mtcars$cyl
head(REG(mpg ~ wt + hp, mtcars, Weights=w, Resid=TRUE)$Fitted)

PROC TTEST

TTEST(mtcars$mpg[mtcars$am==0], mtcars$mpg[mtcars$am==1])

With summarized input (mean, SD, n) only:

tmtest(5.4, 2.2, 30, 4.8, 2.0, 28)

PROC UNIVARIATE

UNIV(mtcars$mpg)

The quartiles and the interquartile range use quantile type 2, the SAS default definition.

2x2 tables

Risk difference, relative risk, and odds ratio with score confidence intervals:

RD(7, 10, 3, 10)
RR(7, 10, 3, 10)
OR(7, 10, 3, 10)

See the vignette Stratified 2x2 Tables for the stratified Miettinen-Nurminen methods and meta-analysis.

Notes

All functions take a formula and a data.frame; factors follow the order of levels, which corresponds to the sorted CLASS levels of SAS.
GLM is fast even for large data sets, since version 1.0.0.
For the philosophy and validation materials, see the references in ?sasLM.