Package 'chest'

Change-in-Estimate Approach to Assess Confounding Effects

Description

In clinical trials and epidemiological studies, the association between an exposure and the outcome of interest in a study can be estimated by regression coefficients, odds ratios or hazard ratios depending on the nature of study designs and outcome measurements. We use a general term effect estimate here for any of those measurements in this document. Based on those measurements, we determine if a treatment is effective (or detrimental) or a factor is a risk factor. Imbalanced distributions of other factors could bias the effect estimates, called confounding. One way to assess the confounding effect of a factor is to examine the difference in effect estimates between models with and without a specific factor. 'chest' allows users quickly calculate the changes when potential confounding factors are sequentially added to the model in a stepwise fashion. At each step, one variable which creates the largest change (%) of the effect estimate among the remaining variables is added to the model. 'chest' returns a graph and a data frame (table) with effect estimates (95% CI) and change (%) values. The package currently has the following main functions: 'chest_lm' for linear regression, 'chest_glm' for logistic regression and Poisson regression, 'chest_clogit' for matched logistic regression, 'chest_nb' for negative binomial regression and 'chest_cox' for Cox proportional hazards models.

References

Zhiqiang Wang (2007) <https://doi.org/10.1177/1536867X0700700203>

Examples

? chest_glm
? chest_cox
? chest_clogit
? chest_lm
? chest_nb
? chest_plot
? chest_forest

---

Assessing confounding effects using conditional logistic regression models

Description

'chest_clogit' is used to fit many Conditional Logistic Regression models to assess confounding effects.

Usage

chest_clogit(
  crude,
  xlist,
  data,
  method = "exact",
  na_omit = TRUE,
  plus = "  + ",
  indicate = FALSE,
  ...
)

Arguments

crude	An object of formula for the initial model, generally crude model. However, any other variables can also be included here as the initial model.
xlist	A vector of characters with all variable names of potential confounders.
data	Data frame.
method	See 'clogit', default is the "exact" method.
na_omit	Remove all missing values, default: 'na_omit = TRUE'.
plus	Change the + sign before variable names.
indicate	indicate the calculation progress.
...	Further optional arguments.

Value

A table with effect estimates and their changes at all steps.

See Also

chest

'clogit' in 'survival'

Examples

vlist <- c("Age", "Sex", "Married", "Cancer", "CVD", "Education", "Income")
chest_clogit(
  crude = "Endpoint ~ Diabetes + strata(mid)",
  xlist = vlist, data = diab_df
)

---

Assessing confounding effects using Cox Proportional Hazards regression models

Description

'chest_cox' is used to assess confounding effects using Proportional Hazards Regression Model ('coxph' from 'survival' package). It presents the effect estimates (such as hazard ratios) for the association between exposure and outcome variables by adding other variables (potential confounders) to the model sequentially. The order of variables to be added is based on the magnitudes of the changes in effect estimates.

Usage

chest_cox(
  crude,
  xlist,
  data,
  na_omit = TRUE,
  plus = "  + ",
  indicate = FALSE,
  ...
)

Arguments

crude	An object of formula for initial model, generally crude model. However, any other variables can also be included here as the initial model.
xlist	A vector of characters with variable names of potential confounders.
data	Data frame.
na_omit	Remove all missing values, default: 'na_omit = TRUE'.
plus	Change the + sign before variable names.
indicate	indicate the progress.
...	Further optional arguments for forestplot.

Value

A table with effect estimates and their changes at all steps.

See Also

'survival'

Examples

vlist <- c("Age", "Sex", "Married", "Cancer", "CVD", "Education", "Income")

chest_cox(crude = "Surv(t0, t1, Endpoint) ~ Diabetes", xlist = vlist, data = diab_df)

---

Plot effect estimates and change-in-estimate values (forestplot type)

Description

'chest_forest' plots effect estimates and change-in-estimate values with forestplot package.

Usage

chest_forest(
  data,
  var_lab = "Variables",
  est_lab = "Estimate (95% CI)",
  change_lab = "Change, %",
  digits = "%.2f",
  digits_change = "%.1f",
  hrzl_lines = gpar(col = "#444444"),
  plus = "  + ",
  ...
)

Arguments

data	Object from chest_cox, chest_glm, chest_lm, chest_clogit, or chest_nb, including effect estimate values and change-in-estimate values.
var_lab	Character string for the column name of variables in the graph.
est_lab	Character string for the column name of effect estimates.
change_lab	Character string for the column name of "Changes".
digits	Set the display format for number in the graph other than the "Change" column. Default: "%.2f"
digits_change	Set the format for the "Change" column. Default: "%.1f"
hrzl_lines	A logic to include or remove horizontal line.
plus	Change the + sign before variable names.
...	Further optional arguments for forestplot.

Value

A table with effect estimates and their changes at all steps.

See Also

'forestplot'

Examples

vlist <- c("Age", "Sex", "Married", "Cancer", "CVD", "Education", "Income")
results <- chest_cox(crude = "Surv(t0, t1, Endpoint) ~ Diabetes", xlist = vlist, data = diab_df)
chest_forest(results)

---

Assessing confounding effects using Generalized Linear regression models

Description

Please note: 'chest_glm' is used to assess confounding effects using Generalized Linear Models, such as logistic regression and Poisson regression with 'glm'. It presents odds ratios or rate ratios for the association between exposure and outcome variables by adding other variables (potential confounders) to the model sequentially. The order of variables to be added is based on the magnitudes of the changes in effect estimates.

Usage

chest_glm(
  crude,
  xlist,
  data,
  family = "binomial",
  method = "glm.fit",
  na_omit = TRUE,
  indicate = FALSE,
  plus = "  + ",
  ...
)

Arguments

crude	An object of formula for initial model, generally crude model. However, any other variables can also be included here as the initial model.
xlist	A vector of characters with all variable names of potential confounders.
data	Data frame.
family	Description of the error distribution. Default is "binomial".
method	Method to detect for singularity.
na_omit	Remove all missing values, default: 'na_omit = TRUE'.
indicate	indicate progress.
plus	Change the + sign before variable names.
...	Further optional arguments for forestplot.

Value

A table with effect estimates and their changes at all steps.

See Also

'glm' 'stats'

Examples

chest_glm(
  crude = "Endpoint ~ Diabetes", xlist = c("Age", "Sex", "Married"),
  na_omit = TRUE, data = diab_df
)

---

Assessing confounding effects using Linear regression models

Description

'chest_lm' is used to assess confounding effects using Linear Regression Models. It presents linear regression coefficients as effect estimates and changes when other variables are added sequentially to the model.

Usage

chest_lm(
  crude,
  xlist,
  data,
  method = "qr",
  na_omit = TRUE,
  indicate = FALSE,
  plus = "  + ",
  ...
)

Arguments

crude	An object of formula for initial model, generally crude model. However, any other variables can also be included here as the initial model.
xlist	A vector of characters with all variable names of potential confounders.
data	Data frame.
method	The method to be used; see 'lm'.
na_omit	Remove all missing values.
indicate	indicate progress
plus	Change the + sign before variable names.
...	Further optional arguments.

Value

A table with effect estimates and their changes at all steps.

See Also

'lm' of 'stats'

Examples

vlist <- c("Age", "Sex", "Married", "Cancer", "CVD", "Education", "Income")
chest_lm(crude = "BMI ~ Diabetes", xlist = vlist, data = diab_df, na_omit = TRUE)

---

Assessing confounding effects using Negative Binomial regression models

Description

It fits negative binomial regression models to present rate ratios for the association between exposure and outcome variables by adding other variables (potential confounders) to the model sequentially. The order of variables to be added is based on the magnitudes of the changes in effect estimates.

Usage

chest_nb(
  crude,
  xlist,
  data,
  method = "glm.fit",
  na_omit = TRUE,
  indicate = FALSE,
  plus = "  + ",
  ...
)

Arguments

crude	An object of formula for initial model, generally crude model. However, any other variables can also be included here as the initial model.
xlist	A vector of characters with all variable names of potential confounders.
data	Data frame.
method	Method to detect for singularity.
na_omit	Remove all missing values, default: 'na_omit = TRUE'.
indicate	indicate progress.
plus	Change the + sign before variable names.
...	Further optional arguments.

Value

A table with effect estimates and their changes at all steps.

See Also

'glm.nb' of 'MASS'

Examples

library(MASS)
df <- quine
results <- chest_nb("Days ~ Lrn", xlist = c("Sex", "Age", "Eth"), data = df)
results$data

---

Plot effect estimate and change-in-estimate values (ggplot type)

Description

Plot effect estimate and change-in-estimate values (ggplot type)

Usage

chest_plot(
  data,
  no_values = FALSE,
  ylab = NULL,
  xlab = NULL,
  change_lab = "Change, %",
  digits = "%.2f",
  digits_change = "%.1f",
  plus = "  + ",
  nudge_y = 0.4,
  nudge_x = NULL,
  hjust = 0.5,
  height = 0.06,
  point_size = 3,
  point_shape = 15,
  vline_type = "dashed",
  vline_color = "grey50",
  ebar_color = "grey50",
  zero = 1,
  value_position = NULL,
  ...
)

Arguments

data	Object from chest_cox, chest_glm, chest_lm, chest_clogit, or chest_nb, including effect estimate values and change-in-estimate values.
no_values	Suppress effect estimate values in plot, default is FALSE.
ylab	Add y label.
xlab	Add x label.
change_lab	Character string for the column name of "Change" in the graph
digits	Set the display format for number in the graph other than the "Change" column. Default: "%.2f"
digits_change	Set the format for the "Change" column. Default: "%.1f"
plus	Change the + sign before variable names.
nudge_y	Adjust vertical distance between values and point marker.
nudge_x	Adjust horizontal distance between values and point marker.
hjust	Adjust horizontal alignment.
height	Change the height of error bars.
point_size	Change point marker size.
point_shape	Change point marker shape.
vline_type	Change vertical line type.
vline_color	Change vertical line color.
ebar_color	Change error bar color.
zero	x-axis coordinate for vertical non-effect line, see forestplot.
value_position	Change the position of value labels.
...	Further optional arguments for forestplot.

Value

a ggplot object.

See Also

'ggplot2'

Examples

vlist <- c("Age", "Sex", "Married", "Education", "Income")
results <- chest_glm(
  crude = "Endpoint ~ Diabetes",
  xlist = vlist, na_omit = TRUE, data = diab_df
)
chest_plot(results)

---

Example data: Health outcomes of 2372 adults with and without diabetes

Description

A data frame with 2372 rows and 14 variables with diabetes status diabetes and mortality status endpoint. For the purpose of demonstrate, assume that we are interested in the association between diabetes and endpoint. Other variables are considered as possible confounders. The purposes of this dataset is to illustrate those functions in chest package only. Therefore, we assume it is a cohort design for Cox Proportional Hazard regression, and a case-control design for logistic regression.

Usage

diab_df

Format

A data frame with 2372 rows and 14 variables:

Diabetes

diabetes status 1: with diabetes 0: without diabetes

Endpoint

mortality status 1: reached end point, and 0: survived

Age

Age, in years

Sex

sex, 1: male, 2: Female

BMI

Body mass index

Married

marital status 1: married, 0: not

Smoke

smoking status 1: smoker, 0: non-smoker

CVD

cardiovascular disease 1: yes 0: no

Cancer

cancer 1: yes, 0: no

Education

education 1: high, 0: low

Income

income 1: high, 0: low

t0

time (age) at the start of the follow-up

t1

time (age) at the end of the follow-up

mid

matched set id, for conditional logistic regression

---