Title: Model Wrappers for Multi-Level Models
Version: 1.0.0
Description: Bindings for hierarchical regression models for use with the 'parsnip' package. Models include longitudinal generalized linear models (Liang and Zeger, 1986) <doi:10.1093/biomet/73.1.13>, and mixed-effect models (Pinheiro and Bates) <doi:10.1007/978-1-4419-0318-1_1>.
License: MIT + file LICENSE
URL: https://github.com/tidymodels/multilevelmod, http://multilevelmod.tidymodels.org/
Depends: parsnip (≥ 1.0.0), R (≥ 2.10)
Imports: dplyr, lme4, purrr, rlang, tibble, withr
Suggests: covr, gee, ggplot2, knitr, nlme, rmarkdown, spelling, testthat, tidymodels
VignetteBuilder: knitr
Config/Needs/website: tidymodels/tidymodels, tidyverse/tidytemplate
Encoding: UTF-8
Language: en-US
LazyData: true
RoxygenNote: 7.2.0
NeedsCompilation: no
Packaged: 2022-06-17 11:38:17 UTC; hannah
Author: Max Kuhn ORCID iD [aut], Hannah Frick ORCID iD [aut, cre], RStudio [cph, fnd]
Maintainer: Hannah Frick <hannah@rstudio.com>
Repository: CRAN
Date/Publication: 2022-06-17 12:00:02 UTC

parsnip methods for hierarchical models

Description

multilevelmod allows users to use the parsnip package to fit certain hierarchical models (e.g., linear, logistic, and Poisson regression). The package relies on the formula method to specify the random effects.

Details

As an example, the package includes simulated longitudinal data where subjects were measured over time points. The outcome was the number of counts and the predictors are the time point as well as an additional numeric covariate.

We can fit the model using lme4::glmer(): 

library(tidymodels)
library(multilevelmod)
library(poissonreg) # current required for poisson_reg()

# The lme4 package is required for this model.

tidymodels_prefer()

# Split out two subjects to show how prediction works
data_train <- 
  longitudinal_counts %>% 
  filter(!(subject %in% c("1", "2")))

data_new <- 
  longitudinal_counts %>% 
  filter(subject %in% c("1", "2"))

# Fit the model
count_mod <- 
  poisson_reg() %>% 
  set_engine("glmer") %>% 
  fit(y ~ time + x + (1 | subject), data = data_train)

count_mod

## parsnip model object
## 
## Generalized linear mixed model fit by maximum likelihood (Laplace
##   Approximation) [glmerMod]
##  Family: poisson  ( log )
## Formula: y ~ time + x + (1 | subject)
##    Data: data
##       AIC       BIC    logLik  deviance  df.resid 
##  4474.553  4494.104 -2233.277  4466.553       976 
## Random effects:
##  Groups  Name        Std.Dev.
##  subject (Intercept) 0.9394  
## Number of obs: 980, groups:  subject, 98
## Fixed Effects:
## (Intercept)         time            x  
##     -0.5946       1.5145       0.2395

When making predictions, the basic predict() method does the trick: 

count_mod %>% predict(data_new)

## # A tibble: 20 × 1
##    .pred
##    <dbl>
##  1  1.19
##  2  1.42
##  3  1.65
##  4  1.83
##  5  2.04
##  6  2.66
##  7  2.96
##  8  3.43
##  9  3.94
## 10  4.64
## 11  2.21
## 12  2.60
## 13  2.97
## 14  3.38
## 15  4.16
## 16  4.90
## 17  5.45
## 18  6.20
## 19  7.55
## 20  8.64

Author(s)

Maintainer: Hannah Frick hannah@rstudio.com (ORCID)

Authors:

Other contributors:

See Also

Useful links:

GEE fitting function

Description

Custom fitting function to add GEE model with cluster variable to parsnip GEE function call.

Usage

gee_fit(formula, data, family = gaussian, ...)

Arguments

formula

Normal formula but uses the gee_formula() internal function for the id_var() specification for clustering.

data

Modeling data

family

a family object: a list of functions and expressions for defining link and variance functions. Families supported in gee are gaussian, binomial, poisson, Gamma, and quasi; see the glm and family documentation. Some links are not currently available: 1/mu^2 and sqrt have not been hard-coded in the cgee engine at present. The inverse gaussian variance function is not available. All combinations of remaining functions can be obtained either by family selection or by the use of quasi.

...

For additional parameters

Details

gee() always prints out warnings and output even when silent = TRUE. gee_fit() will never produce output, even if silent = FALSE.

Also, because of issues with the gee() function, a supplementary call to glm() is needed to get the rank and QR decomposition objects so that predict() can be used.

Value

A gee object

Simulated longitudinal Poisson counts

Description

Simulated longitudinal Poisson counts

Details

These are simulated data of 100 subjects each with 10 time points and an additional numeric covariate. The linear predictor has a random standard normal intercept per subject, a time coefficient of 1.50, and a covariate coefficient of 0.25.

Value

longitudinal_counts

a tibble

Examples

data(longitudinal_counts)
str(longitudinal_counts)

Measurement systems analysis data

Description

Measurement systems analysis data

Details

A biological assay (i.e. a lab test) was run on 56 separate samples twice. The goal is to measure what percentage of the total variation in the results is related to the measurement system and how much is attributable to the true systematic difference (sample-to-sample).

Value

msa_data

a tibble

Examples

data(msa_data)
str(msa_data)

Imipramine longitudinal data

Description

Imipramine longitudinal data

Details

These data are from a longitudinal clinical trial for depression. The outcome is the change in depression scores week-to-week. The endogenous column is an indicator for whether the subject fit the WHO Depression Scale classification of endogenous. The imipramine and desipramine columns are measurements of plasma levels for both substances.

Value

riesby

a tibble

Source

Reisby, N., Gram, L.F., Bech, P. et al. Imipramine: Clinical effects and pharmacokinetic variability. Psychopharmacology 54, 263-272 (1977).

Examples

data(riesby)
str(riesby)

mirror server hosted at Truenetwork, Russian Federation.