Here is a list of keyword phrases recognized by more than one GLM command: Keyword phrases Commands recognizing --------------- ------------------------------------------- print:F All GLM commands Directs that most of the output to the screen is suppressed, although side effect variables are created. silent:T All GLM commands but screen() Directs that all output except error messages is suppressed; side effect variables are computed when there are no errors, coefs:F All GLM commands but screen(), regress(), fastanova(), ipf(), robust(); Suppresses the computation of coefficients or a generalized inverse to X'X (X'WX when there are weights). Except in the case of balanced ANOVA, coefs() and secoefs() cannot be used to retrieve coefficients later. In addition, some of modelinfo() options are effectively disabled after using 'coefs:F' on a GLM command. 'coefs:F' is not legal with 'marginal:T'. fstats:T regress(), anova(), manova(), robust() Directs that F-statistics and P values are computed and printed. The denominator is the mean square for the next following term whose name is of the form "ERROR1", "ERROR2", ... . For manova(), statistics are given separately for each variable and printing of the SS/SP matrices is suppressed, although they are created as side effect variables. pvals:T All GLM commands except screen() Directs that F or Chi-Squared P values are computed and printed for F-statistics, t-statistics, and deviances. pvals:F suppresses P values when they might otherwise be printed. inc:T poisson(), ipf(), logistic(), glmfit() Specifies that an incremental analysis of deviance table is to be computed and printed. No longer legal on robust(). marginal:T anova(), manova(), robust() Specifies that SS (SS/SP matrices for manova()) are computed marginally. When there are no empty cells, and sometimes when there are, the computed SS or SS/SP are equivalent to SAS Type III quantities. 'marginal:T' is not legal with 'coefs:F'. maxiter:n fastanova(), poisson(), ipf(), logistic(), robust(), glmfit() Specifies the maximum number of iterations allowed in fitting eps:smallVal fastanova(), poisson(), ipf(), logistic(), robust(), glmfit() Specifies the a threshhold in relative change of objective function for determining when convergence has been reached problimit:smallVal glmfit() with 'dist:"binomial"', logistic(), probit() Restricts iteration so that fitted probabilities are between smallVal and 1 - smallVal, where 1e-15 <= smallVal < .0001. wts:vec anova(), manova(), regress() weights:vec Specifies a REAL vector to be used as weights. 'wts' and 'weights' are equivalent. offsets:vec poisson(), logistic(), probit(), glmfit() Causes the model to be fit to link to be 1*vec + Model, where vec is a REAL vector the same length as response y. vec must be in the same units as the link function. Thus for poisson() and ipf(), vec should be units of log(E[response]); for logistic(), vec should be in units of log(p/(1-p)) and for probit() vec should be in units of invnor(p). When vec is a linear combination of X-variables in the model, say b01*x1 + b02*x2 + b03*x3, the coefficients computed for x1, x3 and x5 will b1-b01, b2-b02 and b3-b03, where b1, b2 and b3 are the values that would be computed when offsets:vec is not an argument. The residual deviance is not affected. If inc:T is an argument, the deviance associated with x1, x2, and x3 reflects the departure of b1 from b01, b2 from b02, and b3 from b3. This makes it possible to test a hypothesis that one or several coefficients to have specified values. See logistic(), poisson() and probit() for examples. If neither fstats:T nor fstats:F is an argument, for anova() and fastanova() (but not manova()), the printing of F-statistics is controlled by option 'fstats'. See topic 'options'. If neither pvals:T nor pvals:F is an argument, for all GLM commands except manova(), robust() and screen(), the printing of P values is controlled by option 'pvals', except that if options 'pvals' has value False, P values will be printed if F-statistics are. Sums of squares or deviances are normally computed sequentially. For anova(), manova(), and robust() these are SAS Type I quantities. Thus in the unbalanced case, several analyses may be necessary to compute all the sums of squares or deviances needed. Keyword phrase, 'marginal:T', when it can be used, causes SS or SS/SP to be computed differently. When there are no empty cells in the design and no aliased variates, and sometimes when there are, the SS or SS/SP computed are SAS Type III quantities. In every case, they are numerator SS or SS/SP for a test that all the coeffients of non-aliased X-variables in a term are 0, where aliasing is determined by the original order of the sequential fit. If there is aliasing, the quantities computed may depend on the order in which the terms are fit.

Gary Oehlert 2003-01-15