regress([Model] [,print:F or silent:T,pvals:T,coefs:F,marginal:T]) |

regress(Model) performs a least squares fit of the regression model given in the quoted string or CHARACTER variable Model. It prints out the regression coefficients, their standard errors, and t statistics, plus other summary statistics (see below). If option 'pvals' is True, regress() also prints P values for each coefficient based on Student's t distribution. See subtopic 'options:"pvals"'. No ANOVA table is printed by regress(). To see one, type 'anova()' as the next GLM command after regress(). Examples (y, x, x1, x2, and x3 all REAL vectors of length 10): regress("y = x") Simple linear regression of y on x regress("y = x - 1") Linear regression through origin of y on x regress("y = {run(10)}") Simple linear regression of y on vector(1,2,3,4,5,6,7,8,9,10) regress("y = x1 + x2 + x3") 3 variable multiple regression of y on x1, x2 and x3 regress("{sqrt(y)} = x + {x^2}") Quadratic polynomial regression of sqrt(y) on x regress("{sqrt(y)} = P2(x)") Same as preceding. regress(Model,weights:Wts) performs a weighted least squares fit, using REAL vector Wts as case weights. The elements of Wts must not be negative. The results are what you would get by multiplying by sqrt(Wts) the response vector and all independent variables, including the contant vector, and then doing a least squares fit. Model is of the form "Response = Var1 + Var2 + ... + Vark", where Response, Var1, ..., Vark are either variable names or have the form {expr}, where expr is a MacAnova expression. All variables or evaluated expressions must be REAL with the same number of rows. The variables to the right of '=' must be vectors or n by 1 matrices. If any right hand side variable is actually a factor, it is treated as a quantitative variate whose values are the levels of the factor. The associated sum of squares has only 1 degree of freedom regardless of the number of levels of the factor. You specify regression through the origin by including "-1" in the model. See also topic 'models'. regress() or regress(,weights:Wts) with no model specified computes a least squares regression using the same model as was used by the most recent GLM command such as regress(), anova(), or poisson(). See topic 'glm'. Other printed output from regress() includes multiple R-squared, the overall F-statistic for the model excluding the constant term, the mean squared error and the Durbin Watson statistic. regress() computes side effect variables RESIDUALS, HII, SS, DF, DEPVNAME, TERMNAMES, STRMODEL, COEF, and XTXINV. When weights are used, RESIDUALS = response - fit and WTDRESIDUALS = sqrt(Wts)*RESIDUALS is an additional side effect variable. When an independent variable is of the form {expr}, the corresponding element of TERMNAMES is "{expr}". See topic 'glm'. You can retrieve coefficients and/or their standard errors using coefs() or secoefs(). Other Keywords Keyword phrase Default Meaning print:F T Suppress all output except warning and error messages. Side effect variables are set. silent:T F Suppress all output except error messages. Side effect variables are set. pvals:T F Print P values. Default is T when option pvals is T (see subtopic 'options:"pvals"') marginal:T F Specifies that the elements of the side effect variable SS are computed marginally. When none of the X-variables are aliased, the computed SS are equivalent to SAS Type III SS. See topic 'glm' for details. SS will be printed by anova() with no intervening GLM command. Keyword phrase 'coefs:F' is not legal with regress(). See also anova().

Gary Oehlert 2003-01-15