Title: Visualization the Effects of Collinearity in Distributed Lag Models and Other Linear Models
Version: 0.0.4
Description: Tool to assessing whether the results of a study could be influenced by collinearity. Simulations under a given hypothesized truth regarding effects of an exposure on the outcome are used and the resulting curves of lagged effects are visualized. A user's manual is provided, which includes detailed examples (e.g. a cohort study looking for windows of vulnerability to air pollution, a time series study examining the linear association of air pollution with hospital admissions, and a time series study examining the non-linear association between temperature and mortality). The methods are described in Basagana and Barrera-Gomez (2021) <doi:10.1093/ije/dyab179>.
Depends: R (≥ 4.1)
License: GPL-3
Encoding: UTF-8
LazyData: true
Imports: dlnm, graphics, grDevices, MASS, mgcv, nlme, stats, utils, VGAM
Suggests: knitr, rmarkdown, splines, xtable
RoxygenNote: 7.2.3
VignetteBuilder: knitr
NeedsCompilation: no
Packaged: 2023-09-18 14:18:30 UTC; jbarrera
Author: Jose Barrera-Gomez ORCID iD [aut, cre], Xavier Basagana ORCID iD [aut]
Maintainer: Jose Barrera-Gomez <jose.barrera@isglobal.org>
Repository: CRAN
Date/Publication: 2023-09-19 09:00:06 UTC

collin: Visualization the Effects of Collinearity in Distributed Lag Models and Other Linear Models

Description

Tool to assessing whether the results of a study could be influenced by collinearity. Simulations under a given hypothesized truth regarding effects of an exposure on the outcome are used and the resulting curves of lagged effects are visualized. A user's manual is provided, which includes detailed examples (e.g. a cohort study looking for windows of vulnerability to air pollution, a time series study examining the linear association of air pollution with hospital admissions, and a time series study examining the non-linear association between temperature and mortality). The methods are described in Basagana and Barrera-Gomez (2021) doi: 10.1093/ije/dyab179.

Author(s)

Maintainer: Jose Barrera-Gomez jose.barrera@isglobal.org (ORCID)

Authors:

Simulates effects from a distributed lag model pattern.

Description

Simulates results from a distributed lag model under an hypothetical effect pattern provided by the user, which can be linear or non-linear. The output is the passed to the plot method to visualize consequences of collinearity.

Usage

collindlnm(
  model,
  x,
  cb,
  at = 1,
  cen = 0,
  effect,
  type = c("coef", "risk"),
  shape = c("linear", "nonlinear"),
  nsim = 100,
  verbose = TRUE,
  seed = NULL
)

Arguments

model

a model that includes a crossbasis. Currently, models allowed are those of class "glm" or "lme", if shape = "linear"; or "glm", if shape = "nonlinear".

x

if shape = "linear", a matrix that includes the values of the predictor under study, in the first column, and the lagged values, up to the maximum lag considered, as the subsequent columns. If shape = "nonlinear", a numeric vector or 1-column matrix including the original (i.e. lag 0) values of the predictor.

cb

an object of class "crossbasis". The crossbasis included in model.

at

the increase(s) in the predictor under study to be considered to report the effects of the variable. If shape = "linear", at must be a single number. If shape = "nonlinear", at must be a numeric vector with at least two different values.

cen

a number. Reference value of the predictor under study, used to calculate effects. If shape = "linear", the value of cen is irrelevant (and it is internally set to 0).

effect

a vector or a matrix, depending on shape, including the hypothetical effect of the predictor under analysis. If shape = "linear", a vector including the linear effect at each lag (including lag 0). If shape = "nonlinear", a matrix including the effect at each lag (including lag 0) (columns) and for each value in at (rows).

type

a character. If type = "coef" (default), effect is supposed to be in the linear predictor scale (i.e. it is considered as regression coefficients in model). If type = "risk", effect is supposed to be in terms of relative risks (i.e. exp(coef), as ORs or RRs in logistic or Poisson families, respectively). If model is of class "lme", then it must be type = "coef" (default).

shape

the shape of the relationship between the linear predictor of the model and the outcome. Default is "linear".

nsim

number of simulations. Default is 100.

verbose

a logical value indicating output status messages. Default is TRUE.

seed

a number. Seed for reproducibility of results. Default is NULL (no seed).

Value

A list including the results of the simulations to be passed to the plot method.

See Also

crossbasis, glm, lme.

Examples

# For detailed examples:
browseVignettes("collin")

Function to get a vector lagged.

Description

Function to get a vector lagged.

Usage

lagpad(x, k)

Arguments

x

numeric vector to be lagged.

k

number of lags to be applied.

Value

A lagged vector.

Working memory test scores and air pollution.

Description

Simulated data for repeated measures of working memory test scores and residential PM_{2.5} levels in 2,221 children.

Usage

mempm25

Format

A data frame with 8,884 rows and 17 variables:

id

individual identifier

session

test session identifier

school

chool identifier

sex

sex of individual

agecen

age of individual, centered at the mean, in years

educ

maternal education

resses

residential neighborhood socioeconomic status indicator

pm25y0

estimated average concentration of PM_{2.5} (\mug/m^3) at residence address during pregnancy

pm25y1

estimated average concentration of PM_{2.5} (\mug/m^3) at residence address during year 1

pm25y2

estimated average concentration of PM_{2.5} (\mug/m^3) at residence address during year 2

pm25y3

estimated average concentration of PM_{2.5} (\mug/m^3) at residence address during year 3

pm25y4

estimated average concentration of PM_{2.5} (\mug/m^3) at residence address during year 4

pm25y5

estimated average concentration of PM_{2.5} (\mug/m^3) at residence address during year 5

pm25y6

estimated average concentration of PM_{2.5} (\mug/m^3) at residence address during year 6

pm25y7

estimated average concentration of PM_{2.5} (\mug/m^3) at residence address during year 7

wei

weight of the observation

wmemo

working memory test score

Details

This is a synthetic dataset generated with the synthpop package, based on a cohort study of children examining the association between PM_{2.5} concentrations during the prenatal period and the first seven postnatal years on cognitive tests taken at age 8 (see ‘References’).

Source

See ‘References’.

References

Rivas I et al. Association between Early Life Exposure to Air Pollution and Working Memory and Attention. Environmental Health Perspectives. 2019;127(5):57002. doi: 10.1289/EHP3169.

Examples

 # correlation of air pollution levels between two different periods:
 pm25yearly <- mempm25[, grep("pm25", names(mempm25))]
 print(cor(pm25yearly, use = "complete.obs"), digits = 2)

Visualizes effects of collinearity in distributed lag model under an hypothetical linear effect pattern.

Description

Visualize the results from a distributed lag model under an hypothetical linear effect pattern provided by the user, generated using the function collindlnm.

Usage

## S3 method for class 'collindlnmlinear'
plot(x, lags = NULL, ...)

Arguments

x

an object of class "collindlnmlinear", which is generated by the function collindlnm with shape = "linear".

lags

a number or a numeric vector indicating at what lags the results will be visualized. Default (NULL) shows all lags.

...

other parameters to be passed through to plot function.

Value

A plot showing a comparison between the results under the fitted model and the results under the hypothetical true effect.

See Also

collindlnm, plot.collindlnmnonlinear.

Examples

# For detailed examples:
browseVignettes("collin")

Visualize effects of collinearity in distributed lag model under an hypothetical non-linear effect pattern.

Description

Visualize the results from a distributed lag model under an hypothetical non-linear effect pattern provided by the user, generated using the function collindlnm. The number of plots shown is equal to the number of values passed by at in the function collindlnm. The way in which these plots are displayed is controlled by the user through the argument show.

Usage

## S3 method for class 'collindlnmnonlinear'
plot(
  x,
  lags = NULL,
  show = c("manual", "auto", "sequence"),
  addlegend = TRUE,
  varlegend = NULL,
  ...
)

Arguments

x

an object of class "collindlnmnonlinear", which is generated by the function collindlnm with shape = "nonlinear".

lags

a number or a numeric vector indicating at what lags the results will be visualized. Default (NULL) shows all lags.

show

character indicating how the multiple plots will be shown. If show = "manual", then it is expected that the user have previously set the graphical parameters to arrange them (i.e. setting mfrow) with par. If show = "auto", then the arrangement of the plots (i.e. the value of mfrow) is automatically set. If show = "sequence", then the plots are sequentially overlaid.

addlegend

logical indicating whether a legend indicating at what value (of at passed in collindlnm) the results correspond to.

varlegend

character indicating the label for the explored variable to be shown in the legend.

...

other parameters to be passed through to plot function.

Value

A plot showing a comparison between results under the fitted model and the results under the hypothetical true effect, for each of the different values of the variable of interest where effects were explored.

See Also

collindlnm, plot.collindlnmlinear.

Internal function to generate random numbers following over-dispersed Bernoulli distribution, based in the BetaBinomial distribution. It currently does not work and standard Bernoulli distribution is actually used instead.

Description

Internal function to generate random numbers following over-dispersed Bernoulli distribution, based in the BetaBinomial distribution. It currently does not work and standard Bernoulli distribution is actually used instead.

Usage

rbinomod(n, prob, d = 1)

Arguments

n

integer. Number of observations.

prob

numeric. Probability of success on each trial.

d

numeric. Dispersion parameter.

Value

a vector with the simulated sample.

Respiratory hospital admissions and air pollution.

Description

Simulated data for daily measures of hospital admissions for respiratory causes count and ambient air NO_2 concentrations during 10 years.

Usage

rhospno2

Format

A data frame with 3,652 rows and 7 variables:

date

date of the observation

t

numerical indicator of date

year

year indicator

dow

day of week

temp

ambient temperature in Celsius degrees

no2

NO_2 concentration in \mug/m^3

hresp

number of hospital admissions for respiratory causes

Details

This is a synthetic dataset generated with the synthpop package, based on true data on daily number of hospital admissions for respiratory causes and ambient air NO_2 concentrations in the city of Barcelona (Spain) for years 2006-2015.

Examples

 # time series:
 par(las = 1, mfrow = c(3, 1))
 with(rhospno2, plot(date, hresp, type = "l", lwd = 0.5))
 with(rhospno2, plot(date, no2, type = "l", lwd = 0.5))
 with(rhospno2, plot(date, temp, type = "l", lwd = 0.5))

Internal function to generate random numbers following over-dispersed Bernoulli or Poisson distribution. It wraps rpoisod and rbinomod functions.

Description

Internal function to generate random numbers following over-dispersed Bernoulli or Poisson distribution. It wraps rpoisod and rbinomod functions.

Usage

rod(n, mu, d, type = c("binomial", "quasibinomial", "poisson", "quasipoisson"))

Arguments

n

integer. Number of observations.

mu

numeric. Probability of success on each trial or mean, for (quasi)binomial and (quasi)Poisson, respectively.

d

numeric. Dispersion parameter.

type

character indicating the distribution.

Value

a vector with the simulated sample.

Internal function to generate random numbers following over-dispersed Poisson distribution based on simple cheat of using a standard negative binomial, but choosing the scale parameter to give the desired mean vs variance ratio at the given value of the mean. Taken from: https://stat.ethz.ch/pipermail/r-help/2002-June/022425.html

Description

Internal function to generate random numbers following over-dispersed Poisson distribution based on simple cheat of using a standard negative binomial, but choosing the scale parameter to give the desired mean vs variance ratio at the given value of the mean. Taken from: https://stat.ethz.ch/pipermail/r-help/2002-June/022425.html

Usage

rpoisod(n, lambda, d = 1)

Arguments

n

integer. Number of observations.

lambda

numeric. Mean.

d

numeric. Dispersion parameter.

Value

a vector with the simulated sample.

simulatedllm generic.

Description

Simulation procedure internally called by the collindlnm function for a given hypothetical effect.

Usage

simulatedlnm(x)

## S3 method for class 'linearlme'
simulatedlnm(x)

## S3 method for class 'linearglm'
simulatedlnm(x)

## S3 method for class 'nonlinearglm'
simulatedlnm(x)

Arguments

x

an object internally generated by the collindlnm function.

Methods (by class)