In umich-cphds/snif: Selection of Nonlinear Interactions by a Forward Stepwise Algorithm

snif

snif is a R package that implements "Selection of Nonlinear Interactions by a Forward Stepwise Algorithm". snif is currently in the middle of being tested and polished, and as such it is BETA software.

Installation

snif requires the packages purrr and rlang to be installed, and it is a good idea to update them to make sure the package works correctly.

install.packages(c("rlang", "purrr"))

# install.packages("devtools")
devtools::install_github("umich-cphds/snif")

Example

This section is based off the snif documentation.

The first argument to snif is a R formula that defines the initial model snif will use during the forward stage of the algorithm. the second argument, df, is a R data.frame, which is used for scoring, and the columns of df are the possible variables that snif will consider. In other words, if there variables in df that you don't want snif to consider, you should subset df first.

snif supports binary variables by setting type = "logistic". Then snif will use a glm to fit the data instead of a lm. The default for type is linear.

snif contains a toy dataset, snif.df which is used in this example. It has 500 rows and 11 columns.

library(snif)

names(snif.df)

In this data.frame, the response, y, is continuous so we will use a linear model by setting type = "linear". We would also like to use BIC scoring to determine which variable to add to the model, so we set method = "BIC". For the initial formula to fit, we do not know much about this data so we will start with the null model, y ~ NULL.

snif.out <- snif(formula = y ~ NULL, df = snif.df, type = "linear", method = "BIC")

summary(snif.out)

Alternatively, we can use a non null initial model

snif.out <- snif(formula = y ~ V2, df = snif.df, type = "linear",
                   method = "BIC")

summary(snif.out)

snif also supports interaction terms via : and basis expansions via bs.

snif.out <- snif(formula = y ~ bs(V2) + V2:V4, df = snif.df, type = "linear",
                   method = "BIC")

Note that bs implies that both V2 and the nonlinear expansion of V2, bs(V2)[,-1] are both included in the model. That is the following two formulas are equivalent: y ~ bs(V2) and y ~ V2 + bs(V2)[,-1] If you wish for only the non linear part of V2 to be included in snif, you need to drop the linear term from the expansion, that is, y ~ bs(V2)[, -1]. Finally, you may specify the degree of the expansion via the degree parameter in snif.

More detailed documentation is available on the snif help page, which provides rules for building initial snif formulas, and documents other optional parameters not mentioned here.

Reference

Narisetty, Naveen N. and Mukherjee, Bhramar and Chen, Yin-Hsiu and Gonzalez, Richard and Meeker, John D. Selection of nonlinear interactions by a forward stepwise algorithm: Application to identifying environmental chemical mixtures affecting health outcomes. Statistics in Medicine. 2019;38(9):1582-1600. \href{https://doi.org/10.1002/sim.8059}{10.1002/sim.8059}

Contact

Please email Alex Rix ([e-mail]) if you have any questions, comments or bugs to report. You may also open up an issue on github.

umich-cphds/snif documentation built on Nov. 5, 2019, 11:05 a.m.