[ MULTILOG ] [ OPTIONS ] [ OUTPUT_GROUPS ] [ STATEMENTS ]
The MULTILOG procedure extends the modeling capabilities of SUDAAN to include categorical outcomes with more than two categories which may or may not have a natural ordering. These models can be viewed as generalizations of logit models for binary outcomes already available in SUDAAN in the LOGISTIC procedure. MULTILOG analyzes data from sample surveys as well as from randomized experiments and other observational studies involving cluster-correlated or longitudinal responses.
Two models have been implemented in the MULTILOG procedure: the proportional odds model with cumulative logit link for ordinal responses and a generalized multinomial logit model for nominal outcomes. Both models handle continuous as well as discrete explanatory variables. The generalized Multinomial Logit Model produces separate parameter vectors for each of the generalized logit equations of interest; the Proportional Odds Model produces a common slope but separate intercepts for each of the cumulative logit equations of interest.
The MULTILOG procedure estimates model parameters using generalized estimating equations (GEE). For estimating variance of the parameter estimates, MULTILOG implements two robust methods described in Binder (1983) and Zeger and Liang (1986), as well as a model-based (naive) variance estimation method. All three variance estimation methods allow a choice of independent vs exchangeable working correlations for describing the dependence of responses within clusters. By default, the GEE iterative fitting procedure in the exchangeable case uses the one-step approach, although a multistep GEE procedure can also be obtained.
MULTILOG produces estimates of the model parameters and their standard errors, odds ratios and their confidence limits, and tests the null hypothesis that individual regression coefficients associated with each variable in the model are equal to zero. MULTILOG also provides tests for overall model significance, model minus intercept, as well as model main effects and interactions. In addition, you can test linear combinations of the model parameters and output many statistics to an output data set. You can also estimate and test linear combinations of the conditional and predicted marginals (generalizations of adjusted group means to non-linear models).
Procedure Enhancements:
Analysis of multiply imputed data has been implemented for all estimates, variances and tests of hypothesis.
The CLASS statement is available in MULTILOG (providing an alternative to SUBGROUP statement).
Estimates of confidence limits for the model parameters are now produced by default in the BETAS group.
A new continuous_variable=(value(s)) option on the CONDMARG, PREDMARG, COND_EFF And PRED_EFF statements optionally specifies particular values at which the predicted and conditional marginals and their contrasts are to be evaluated with respect to that continuous variable.