Methods for manipulating regression models and for describing these in a style adapted for medical journals. Contains functions for generating an HTML table with crude and adjusted estimates, plotting hazard ratio, plotting model estimates and confidence intervals using forest plots, extending this to comparing multiple models in a single forest plots. In addition to the descriptive methods, there are add-ons for the robust covariance matrix provided by the 'sandwich' package, a function for adding non-linearities to a model, and a wrapper around the 'Epi' package's Lexis() functions for time-splitting a dataset when modeling non-proportional hazards in Cox regressions.
This package helps with building and conveying regression models. It has also a few functions for handling robust confidence intervals for the ols() regression in the rms-package. It is closely interconnected with the the Gmisc, htmlTable, and the forestplot packages.
Communicating statistical results is in my opinion just as important as performing the statistics. Often effect sizes may seem technical to clinicians but putting them into context often helps and makes you to get your point across.
The method that I most frequently use in this package is the
printCrudeAndAdjustedModel. It generates a table that has the full model coefficients together with confidence intervals alongside a crude version using only that single variable. This allows the user to quickly gain insight into how strong each variable is and how it interacts with the full model, e.g. a variable that shows a large change when adding the other variables suggests that there is significant confounding. See the vignette for more details,
I also like to use forest plots for conveying regression models. A common alternative to tables is to use a forest plot with estimates and confidence intervals displayed in a graphical manner. The actual numbers of the model may be better suited for text while the graphs quickly tell how different estimates relate.
Sometimes we also have situations where one needs to choose between two models, e.g. a Poisson regression and a Cox regression. This package provides a
forestplotCombineRegrObj function that allows you to simultaneously show two models and how they behave in different settings. This is also useful when performing sensitivity analyses and comparing different selection criteria, e.g. only selecting the patients with high-quality data and see how that compares.
plotHR function was my first attempt at doing something more advanced version based upon Reinhard Seifert's original adaptation of the
stats::termplot function. It has some neat functionality although I must admit that I now more commonly use ggplot2 for many of my plots as I like to have a consistent look throughout the plots. The function has though a neat way of displaying the density of the variable at the bottom.
Often much of our modeling ends up a little repetitive and this package contains a set of functions that I've found useful. The approach that I have for modeling regressions is heavily influenced by Frank Harrell's regression modeling strategies. The core idea consist of:
addNonLinearityfunction. The function tests using ANOVA for non-linearity and if such is found it maps a set of knots, e.g. 2-7 knots, of a spline function and then checks for the model with the lowest AIC/BIC value. If it is the main variable I do this by hand to avoid choosing a too complex model when the AIC/BIC values are very similar but for confounders I've found this a handy approach.
robcov_altmethod. In survival analyses the non-proportional hazards assumption can sometimes be violated where the
timeSplitterfunction helps you to set-up a dataset that allows you to build time-interaction models (see
NEWS for the Greg package