Overview of the Generalized Regression Personality
The Generalized Regression personality of the Fit Model platform features regularized, or penalized, regression techniques. Such techniques attempt to fit better models by shrinking the model coefficients toward zero. The resulting estimates are biased. This increase in bias can result in decreased prediction variance, thus lowering overall prediction error compared to unpenalized models. Two of these techniques, the Elastic Net and the Lasso, include variable selection as part of the modeling procedure.
Modeling techniques such as the Elastic Net and the Lasso are particularly useful for large data sets, where collinearity is typically a problem. In addition, modern data sets often include more variables than observations. This situation is sometimes referred to as the p > n problem, where n is the number of observations and p is the number of predictors. Such data sets require variable selection if traditional modeling techniques are to be used.
The Elastic Net and Lasso can also be used for small data sets with little correlation, including designed experiments. They can be used to build predictive models or to select variables for model reduction or for future study.
The personality provides the following classes of modeling techniques:
• Maximum Likelihood
• Step-Based Estimation
• Penalized Regression
The Elastic Net and Lasso are relatively recent techniques (Tibshirani 1996; Zou and Hastie 2005). Both techniques penalize the size of the model coefficients, resulting in a continuous shrinkage. The amount of shrinkage is determined by a tuning parameter. An optimal level of shrinkage is determined by one of several validation methods. Both techniques have the ability to shrink coefficients to zero. In this way, variable selection is built into the modeling procedure. The Elastic Net model subsumes both the Lasso and ridge regression as special cases. See Statistical Details for Estimation Methods.
Details about Generalized Regression Modeling Techniques
• The Maximum Likelihood method is a classical approach. It provides a baseline to which you can compare the other techniques, and it is the most appropriate place for traditional inference techniques such as hypothesis testing.
• Forward Selection is a method of stepwise regression. In forward selection, terms are entered into the model. The most significant terms are added until all of the terms are in the model or there are no degrees of freedom left.
• The Lasso has two shortcomings. When several variables are highly correlated, it tends to select only one variable from that group. When the number of variables, p, exceeds the number of observations, n, the Lasso selects at most n predictors.
• The Elastic Net, on the other hand, tends to select all variables from a correlated group, fitting appropriate coefficients. It can also select more than n predictors when p > n.
• Ridge regression was among the first of the penalized regression methods proposed (Hoerl 1962; Hoerl and Kennard 1970). Ridge regression does not shrink coefficients to zero, so it does not perform variable selection.
• The Double Lasso attempts to separate the selection and shrinkage steps by performing variable selection with an initial Lasso model. The variables selected in the initial model are then used as the input variables for a second Lasso model.
• Two-Stage Forward Selection performs two stages of forward stepwise regression. It performs variable selection on the main effects in the first stage. Then, higher-order effects are allowed to enter the model in the second stage.
The Generalized Regression personality also fits an adaptive version of the Lasso and the Elastic Net. These adaptive versions attempt to penalize variables in the true active set less than variables not contained in the true active set. The true active set refers to the set of terms in a model that have an actual effect on the response. The adaptive versions of the Lasso and Elastic Net were developed to ensure that the oracle property holds. The oracle property guarantees the following: Asymptotically, your estimates are what they would have been had you fit the model to the true active set of predictors. More specifically, your model correctly identifies the predictors that should have zero coefficients. Your estimates converge to those that would have been obtained had you started with only the true active set. See Adaptive Methods.
The Generalized Regression personality enables you to specify a variety of distributions for your response variable. The distributions fit include normal, Cauchy, Student’s t, exponential, gamma, Weibull, lognormal, negative lognormal, beta, binomial, beta binomial, Poisson, negative binomial, zero-inflated binomial, zero-inflated beta binomial, zero-inflated Poisson, zero-inflated negative binomial, and zero-inflated gamma. This flexibility enables you to fit categorical and count responses, as well as continuous responses, and specifically, right-skewed continuous responses. You can also fit quantile regression and Cox proportional hazards models. For some of the distributions, you can fit models to censored data. The personality provides a variety of validation criteria for model selection and supports training, validation, and test columns. See Specify a Distribution.