polyreg (version 0.8.0)

polyFit: Polynomial Fit

Description

Fit polynomial regression using a linear or logistic model; predict new data.

Usage

polyFit(xy, deg, maxInteractDeg=deg, use = "lm", glmMethod="one", 
     return_xy=FALSE, returnPoly=FALSE, noisy=TRUE)
# S3 method for polyFit
predict(object, newdata, ...)

Arguments

xy

Data frame with response variable in the last column. Latter is numeric, except in the classification case: Categorical variables (> 2 levels) must be passed as factors or character variables if use is 'glm'; an integer vector must be used for the for the 'mvrlm' case, or for the 2-class case (0s and 1s).

deg

The max degree for polynomial terms. A term such as uv, for instance, is considered degree 2.

maxInteractDeg

The max degree of interaction terms.

use

Set to 'lm' for linear regression, 'glm' for logistic regression, or 'mvrlm' for multivariate-response lm.

glmMethod

Defaults to 'one,' meaning the One Versus All Method. Use 'all' for All Versus All.

newdata

Data frame, one row for each "X" to be predicted. Must have the same column names as in xy (without "Y").

object

An item of class 'polyFit' containing output. Can be used with predict().

return_xy

Return data? Default: FALSE

returnPoly

return polyMatrix object? Defaults to FALSE since may be quite large.

noisy

Logical: display messages?

...

Additional arguments for getPoly().

Value

The return value of polyFit() is an polyFit object. The orginal arguments are retained, along with the fitted models and so on.

The prediction function predict.polyFit returns the predicted value(s) for newdata. It also contains probability for each class as an attribute named prob. In the classification case, these will be the predicted class labels, 1,2,3,...

Details

The polyFit function calls getPoly to generate polynomial terms from predictor variables, then fits the generated data to a linear or logistic regression model. (Powers of dummy variables will not be generated, other than degree 1, but interaction terms will calculated.)

When logistic regression for classification is indicated, with more than two classes, All-vs-All or One-vs-All methods, coded 'all' and 'one', can be applied to deal with multiclass problem.

Under the 'mvrlm' option in a classification problem, lm is called with multivariate response, using cbind and dummy variables for class membership as the response. Since predictors are used to form polynomials, this should be a reasonable model, and is much faster than 'glm'.

Examples

# NOT RUN {
N <- 125
xyTrain <- data.frame(x1 = rnorm(N), 
                      x2 = rnorm(N),
                      group = sample(letters[1:5], N, replace=TRUE),
                      score = sample(100, N, replace = TRUE) # final column is y
                      )

pfOut <- polyFit(xyTrain, 2)

# 4 new test points
xTest <- data.frame(x1 = rnorm(4), 
                    x2 = rnorm(4),
                    group = sample(letters[1:5], 4, replace=TRUE))
  
predict(pfOut, xTest) # returns vector of 4 predictions

data(pef)
# predict wageinc
z <- polyFit(pef[,c(setdiff(1:6,5),5)],2)
predict(z,pef[2000,c(setdiff(1:6,5),5)])  # 56934.39
# predict occ
z <- polyFit(pef[,c(setdiff(1:6,3),3)],2,use='glm')
predict(z,pef[2000,c(setdiff(1:6,3),3)])  # '100', probs 0.43, 0.26,...

# }