A collection of functions for interpretation and presentation of regression analysis. These functions are used to produce the statistics lectures in < http://pj.freefaculty.org/guides>. Includes regression diagnostics, regression tables, and plots of interactions and "moderator" variables. The emphasis is on "mean-centered" and "residual-centered" predictors. The vignette 'rockchalk' offers a fairly comprehensive overview. The vignette 'Rstyle' has advice about coding in R. The package title 'rockchalk' refers to our school motto, 'Rock Chalk Jayhawk, Go K.U.'.

rockchalk package NEWS -- history of user-visible changes.

Copyright (C) 2013-8 Paul E. Johnson.

- version 1.8.114

** silence warnings cause by R update

- version 1.8.96 2015-12-15

** skewness and kurtosis added to summary stats for numeric variables.

- version 1.8.94 2015-11-01

** plotCurves. Replace hard coded value 40 for number of points at which to evaluate plotx. Tom Wenseleers suggested.

- version 1.8.91 2015-01-29

** outreg enhancements suggested by Irving Llamosas-Rosas, allow users pass in standard errors and p-values. Argments Blist, SElist, etc.

** plotSlopes and plotCurves allow level argument to pass on to predict methods. Suggested by Tom Wenseleers.

** New function: pctable for percentage tables.

** Learned how to document several functions in same file, reducing clutter in man directory.

** Re-organize package dependencies, shift some from required to import, when only examples need those packages

- version 1.8.1 2013-09-22

This is a minor bugfix update.

** genCorrelatedData2: calculations were correct, but printout of formula was in error (off by one in beta1[2:(d+1))] ** plotCurves: plotx always need to take series of points across axis until I learn way to ask formula if plotx has curvature in it.

- version 1.8 2013-07-30

This is the end of the Spring semester, so its time for the new rockchalk release.

** New, general, flexible framework for calculating marginal effects in regression models, linear or otherwise.

*** newdata function works. It can scan a regression, isolate the predictors, and then make a "mix and match" new data object for use with predict function. This is convenient for users but also very flexible.

*** The newdata framework is built on top of "divider" methods that can check whether a variable is numeric or categorical, and select example values according to user-specified criteria.

*** predictOMatic works dependably! Please try example(predictOMatic). The problem with single predictor models that bugged users of rockchalk 1.6.2 has been solved.

*** predictOMatic argument interval = c("none", "confidence", "prediction"). Guess what that is supposed to do? For glm, which does not provide a confidence interval, I've written code for an approximate Wald type CI, and hope to do better in future.

** Regression diagnostics.

*** getPartialCor: get partial correlations from a fitted model (student convenience).

*** getDeltaRsquare: Returns the change in estimated R-square observed when each predictor is removed. This is the squared semi-partial correlation coefficients (student convenience).

*** mcDiangose for multicollinearity diagnostics (student convenience)

** MeanCenter: add arguments to make selection of variables for centering more convenient when users don't like the automatic options centerOnlyInteractors.

** plotSlopes, plotCurves:

*** Added intervals argument, for confidence and prediction intervals.

*** Added opacity argument to determine darkness of interval regions (which use the transparency "alpha layer.").

*** A lot of fiddling under the hood to make colors consistent when levels of modx are altered to compare plots of a given model.

*** Can produce a simple regression prediction plot if modx argument is omitted. This is a widely requested feature.

Please run example(plotSlopes) and example(plotCurves)

*** Changes under the hood. The common plotting functions of plotSlopes and plotCurves are abstracted into a function plotFancy, so now this will be eaiser for me to maintain. The plotting ritual is the same, why keep 2 functions, you ask? plotCurves tolerates more complicated regression formula. plotSlopes leads to testSlopes, and hence to plot.testSlopes.

** addLines: communication between 2 dimensional regression plots and 3 dimensional plots from plotPlane. Run example(addLines).

** plot.testSlopes. Run testSlopes on an interactive model. For a model with 2 ocontinuous predictors that interact, this will generate an ABSOLUTELY EXCELLENT and highly informative plot displaying the effect of the interaction.

** outreg: LaTeX tables from regression models.

*** Reworked with new arguments to make tables for more types of regressions. There's quite a bit more room for users to customize the type of diagnostics they want to report.

The wide variety of output types from regression models is very bothersome. I refuse to write a separate outreg method for each different regression packages. If you want to use a package from an author who is willing to do that, consider the "texreg" package.

*** outreg2HTML. converts outreg to HTML markup and into a file. Intended for importation into word processor programs.

** New vignette Rstyle. Most of the source-code files have been reformatted to comply with my advice.

** genCorrelatedData2

*** genCorrelatedData2. For regression examples, suppose you want to have 6 columns of an MVN with a certain mean and covariance structure. And you want the regression formula to have interactions and squared-terms. No more hassle. This is a framework that works. Users set the mean, standard deviations, and correlation values in various ways. Run example(genCorrelatedData2).

*** To support that, there are more generally useful functions. lazyCor and lazyCov are flexible ways to create correlation and covariance matrices. As the names suggest, they are for lazy users who just want to specify some information and get the right thing. This requires a set of transformation functions, to receive vech and create matrices, and so forth. Check genCorrelatedData.R, for vech2Corr, makeVec, makeSymmetric, checkPosDef. The latter, which I am surprised not to find in the base of R itself, imitates code in the MASS package for ascertaining if a matrix is positive definite.

*** Small, almost microscopic, revision of MASS package mvrnorm function to assure replication of MVN draws when the sample size is adjusted. The first rows of the resulting MVN draw will be the same, no matter how the "n" argument is changed. The same change has been made in the mvtnorm package's MVN random generator. While this is a very small code change, it does solve some very mysterious simulation results that have been obtained with MASS mvrnorm in our lab.