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.
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")
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)
## Loading required package: splines
names(snif.df)
## [1] "y" "V2" "V3" "V4" "V5" "V6" "V7" "V8" "V9" "V10" "V11"
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)
## $formula
## y ~ bs(V2, degree = 3)[, -1] + V5 + V6 + V4 + V2
##
## $score
## [1] 2619.361
##
## $summary
##
## Call:
## object$model(formula = formula, data = object$df)
##
## Residuals:
## Min 1Q Median 3Q Max
## -10.6943 -2.0696 -0.1403 2.0606 9.4985
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) -8.4265 3.6224 -2.326 0.0204 *
## bs(V2, degree = 3)[, -1]2 13.3611 6.6660 2.004 0.0456 *
## bs(V2, degree = 3)[, -1]3 51.4487 7.5171 6.844 2.29e-11 ***
## V5 1.1324 0.1572 7.202 2.23e-12 ***
## V6 0.9949 0.1505 6.611 9.96e-11 ***
## V4 0.9261 0.1577 5.871 7.96e-09 ***
## V2 -11.0073 2.0782 -5.297 1.78e-07 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 3.183 on 493 degrees of freedom
## Multiple R-squared: 0.5475, Adjusted R-squared: 0.5419
## F-statistic: 99.4 on 6 and 493 DF, p-value: < 2.2e-16
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)
## $formula
## y ~ V2 + bs(V2, degree = 3)[, -1] + V5 + V6 + V4
##
## $score
## [1] 2619.361
##
## $summary
##
## Call:
## object$model(formula = formula, data = object$df)
##
## Residuals:
## Min 1Q Median 3Q Max
## -10.6943 -2.0696 -0.1403 2.0606 9.4985
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) -8.4265 3.6224 -2.326 0.0204 *
## V2 -11.0073 2.0782 -5.297 1.78e-07 ***
## bs(V2, degree = 3)[, -1]2 13.3611 6.6660 2.004 0.0456 *
## bs(V2, degree = 3)[, -1]3 51.4487 7.5171 6.844 2.29e-11 ***
## V5 1.1324 0.1572 7.202 2.23e-12 ***
## V6 0.9949 0.1505 6.611 9.96e-11 ***
## V4 0.9261 0.1577 5.871 7.96e-09 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 3.183 on 493 degrees of freedom
## Multiple R-squared: 0.5475, Adjusted R-squared: 0.5419
## F-statistic: 99.4 on 6 and 493 DF, p-value: < 2.2e-16
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.
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.
Please email Alex Rix (alexrix@umich.edu) if you have any questions, comments or bugs to report. You may also open up an issue on github.