Main Methods

Click on a button corresponding to a predictive modeling main method. All processes require a wide data set. (See Tall and Wide Data Sets.) For a more thorough introduction to predictive modeling and these processes, see Predictive Modeling.

Refer to the table below for key features and general guidance on these processes. You are encouraged to explore multiple processes and use the individual process links for a more detailed explanation of each.

Tip: When in doubt, there is no harm in trying several predictive modeling methods on your data. The Predictive Modeling Review enables you to standardize model parameters and specifications. Additional tools are also available in the Model Comparisons submenu for this purpose.

Process	Uses SAS PROC(s)	Permits dependent variables of type	Particularly appropriate for data with these characteristics	Classification boundary shape for binary dependent variable	Other classification and process characteristics
Discriminant Analysis	STEPDISC DISCRIM	• Nominal • Binary • Ordinal	• Can be represented by a multivariate normal distribution with known classes • Fewer variables than observations	Linear, parabolic, or S-shaped	• Based on Fisher discriminant analysis
Distance Scoring	DISTANCE	• Nominal • Binary • Ordinal • Continuous		Variable; depends on the distance metric	• Nonparametric discriminant method Tip: Diagonal Linear Discriminant Analysis can be performed via the Euclidean Distance Metric.
General Linear Model Selection	GLMSELECT	• Binary • Ordinal • Continuous		Linear	• Flexible • Many model selection methods available • Many inclusion and stopping criteria available
K Nearest Neighbors	DISCRIM	• Nominal • Binary • Ordinal	• Fewer variables than observations	Any shape	• Nonparametric discriminant method • Predictions based on the set of k training observations that are closest in feature space distance (instance-based learning)
Logistic Regression	LOGISTIC	• Nominal • Binary • Ordinal	• Fewer variables than observations	S-shaped	• Data fit to a logistic curve using a logit link function Caution: This process can take a long time to run, depending on the number of predictor variables and the speed of your machine.
Partial Least Squares	PLS	• Binary • Ordinal • Continuous	• More variables than observations (wide data sets) • Multicollinearity among predictor variables exists	Linear or quadratic	• Linear regression model • Simultaneously models variability in both dependent and predictor variables
Partition Trees	GENESELECT	• Nominal • Binary • Ordinal • Continuous	• Can be represented as a hierarchy of partitions	Step function	• Simple tree-based rule sets from optimal splitting relationships between dependent and predictor variables are used
Quantile Regression Selection	QUANTREG	• Binary • Ordinal • Continuous	• The median or particular quantiles of the dependent variables are better measures of central tendency than the mean	Linear or quadratic	• Flexible • Many model selection methods available • Many inclusion and stopping criteria available • Model robustness; data robustness to outliers
Radial Basis Machine	GLIMMIX	• Binary • Ordinal • Continuous	• More variables than observations (wide data sets)	Any shape	• Dimensions of calculations are based on the number of observations, rather than the number of variables
Ridge Regression	MIXED	• Binary • Ordinal • Continuous	• More variables than observations (wide data sets) • Multicollinearity among predictor variables exists • Continuous dependent variable	Linear or quadratic	• Computes Best Linear Unbiased Predictions (BLUPs) of the responses based on a mixed model • Shrinks (regresses) estimates toward a common mean
Life Regression	LIFEREG	• Binary • Time-to-event • Censor	• Time-to-event data • Data that follows one of the time-to-event distributions (Weibull, for example)		• Fits parametric models to time-to-event data. • Predictor reduction methods can be used to trim a large set of predictors.
Proportional Hazards Regression	PHREG	• Binary • Ordinal • Continuous	• Survival data with time-to-event variable (censor variable optional) • Fewer variables than observations	Exponential family	• Uses a Cox proportional hazards model • Many model selection methods available Caution: This process can be computationally intensive for large data sets.
Genomic BLUP	MIXED HPMIXED	• Binary • Continuous •	• Continuous dependent variable	Linear or quadratic	• Computes Best Linear Unbiased Predictions (BLUPs) of the responses based on a mixed model. • Shrinks (regresses) estimates toward a common mean.

Predictive Modeling Review

Click Predictive Modeling Review to sets up a predictive modeling review that can be used to compare the efficacy of different models, applied to one or more dependent variables, at making predictions under the same conditions and compare the models using cross validation, test sets, or learning curves.

See Predictive Modeling for other subcategories.