| pointEstimate {causalPAF} | R Documentation |
Evaluates Point Estimates for Total PAF, Direct PAF, Indirect PAF and Path Specific PAF for a user inputted number of integral simulations. There is no bootstrap applied in this function.
Description
Evaluates Total PAF, Direct PAF, Indirect PAF and Path Specific PAF for a user inputted number of bootstraps and integral simulations
Usage
pointEstimate(
dataframe,
exposure = "phys",
mediator = c("subhtn", "apob_apoa", "whr"),
response = "case",
response_model_mediators = list(),
response_model_exposure = list(),
in_outArg,
Splines_outlist,
splinesDefinedIn_in_outDAG,
model_listArg,
weights = 1,
NumSimulation,
addCustom = FALSE,
custom = ""
)
Arguments
dataframe |
A wide format dataframe containing all the risk factors, confounders, exposures and outcomes within the causal DAG Bayesian network. |
exposure |
The name of the exposure column variable within dataframe in text format e.g. "phys". |
mediator |
The name of the mediator column variables within dataframe in text format. There can be more than one mediator of interest. It can be a vector of mediators names within the dataframe e.g. c("subhtn","apob_apoa","whr"). |
response |
The name of the response column variable within dataframe in text format e.g. "case". The cases should be coded as 1 and the controls as 0. |
response_model_mediators |
A regression model fitted for the response in a causal Bayesian network excluding “children” of the mediators in the causal Bayesian network. See example in tutorial.This model can be listed either as (1) an empty list ( response_model_mediators = list() ) or (2) the user can specify their own customised causal regression model(s) to use. When it is listed as an empty list the 'causalPAF' package will fit the response_model_mediators regression model automatically based on the causal DAG supplied by the user in in_outArg. Alternatively, the user can specify the exact model(s) that the user wishes to use, these model(s) must be in list format (list() where length(response_model_mediators) == length(mediator) ), the same length as the parameter, mediator, with the user customised model for each mediator listed in the same order as in the parameter, mediator, and if there is only one model, it must be listed each time within the list() so that length(response_model_mediators) == length(mediator). |
response_model_exposure |
A regression model fitted for the response in a causal Bayesian network excluding “children” of the exposure in the causal Bayesian network. This regression model will not adjust for mediators (exclude mediators) of the exposure in the regression model so that the total effect of the exposure on the response can be modelled. This model can be listed either as (1) an empty list ( response_model_exposure = list() ) or (2) the user can specify their own customised causal regression model to use. If specified as an empty list, list(), then the 'causalPAF' package will define and fit the model automatically based on the causal DAG defined by the in_outArg parameter. Alternatively, the user can specify the exact model that the user wishes to use, this model must be in list format (list() where length(response_model_exposure) == 1 ), of length 1, assuming only one exposure of interest (other exposures can be risk factors) and the model must be defined within a list() since the package assumes a list() format is supplied. See example in tutorial. E.G. If physical exercise ("exer") in the example given in the diagram is the exposure. Then the regression would include all parents of "exer" (i.e. sex, region, educ, age) as well as risk factors at the same level of the causal Bayesian network (i.e. stress, smoke, diet, alcoh). |
in_outArg |
This defines the causal directed acyclic graph (DAG). A list of length 2. It is defined as a two dimensional list consisting of, firstly, the first list, inlist, i.e. a list of the parents of each variable of interest corresponding to its column name in the data. Splines can be included here if they are to be modelled as splines. Secondly, the second list, outlist, contains a list of a single name of exposure or risk factor or outcome in form of characters i.e. a list of each variable of interest (risk factors, exposures and outcome) corresponding to its column name in the data. Splines should not be input here, only the column names of the variables of interest in the data. The order at which variables are defined must satisfy (i) It is important that variables are defined in the same order in both lists e.g. the first risk factor defined in outlist has its parents listed first in inlist, the second risk factor defined in outlist has its parents listed secondly in inlist and so on. The package assumes this ordering and will not work if this order is violated. (ii) Note it is important also that the order at which the variables are defined is such that all parents of that variable are defined before it. See example in tutorial. |
Splines_outlist |
A list defined of same size and order of variables as defined in in_outArg[[2]]. If splines are to be used for variables listed in in_outArg[[2]], then the splines should be defined in Splines_outlist in the same order as variables appear in in_outArg[[2]]. It is necessary to list variables in Splines_outlist the same as in in_outArg[[2]] without splines if no spline is to be applied. It should not be input as an empty list, list(), if no splines. A warning will show if input as an empty list requiring the user to populate Splines_outlist either the same as in_outArg[[2]] (if no splines) or in the same order as in_outArg[[2]] with splines (if splines). See example in tutorial. |
splinesDefinedIn_in_outDAG |
Logical TRUE or FALSE indicating whether the user has defined splines in the causal DAG, in_out, if TRUE. If FALSE and splines are defined in Splines_outlist_Var, then it is necessary for the package to populate the in_out DAG with splines listed in Splines_outlist_Var. |
model_listArg |
is a list of models fitted for each of the variables in in_outArg[[2]] (or in_outArg |
weights |
Column of weights for case control matching listed in the same order as the patients in the data e.g. weights = strokedata$weights. |
NumSimulation |
This is the number of simulations requested by the user to estimate integrals. The larger the number of simulations the more accurate the results but the longer the code takes to run. Therefore the user may wish to balance speed with accuracy depending on which is of more value in the specific context of interest. The integrals for continuous variables are estimated using simulation methods. |
addCustom |
Logical TRUE or FALSE indicating whether a customised interaction term is to be added to the each regression. The interaction term can include splines. |
custom |
text containing the customised interaction term to be added to each regression. The text should be enclosed in inverted commas. Splines can be included within the interaction terms. See tutorial for examples. |
Value
Estimates point estimates for 5 results that are:(1)Total Population Attributable Fraction (PAF),(2)Direct Effect Population Attributable Fraction (PAF) using alternative definition, (3)Indirect Effect Population Attributable Fraction (PAF) using alternative definition, (4)Path Specific Population Attributable Fraction (PAF), (5)Overall Direct Population Attributable Fraction (PAF)
Examples
# Loads some data (fictional Stroke data from the package 'causalPAF')
# In this example, we use a small data set called 'strokedata_smallSample' consisting of 5,000
# rows of fictional patient data. For more accurate results, a larger data set is available
# called 'strokedata'which contains 16,623 rows of fictional patient data. The methodology
# applied in the 'causalPAF' package is more accurate the larger the dataset. To use the larger
# 'strokedata' dataset, simply call
# stroke_reduced <- strokedata
stroke_reduced <- strokedata_smallSample
# Just shortening the name of a variable, "apob_apoa", to "apb" so the R code
# in document example is not truncated.
stroke_reduced$apb <- stroke_reduced$apob_apoa
# The data should contain a column of weights for case control matching.
# strokedata$weights
# Weigths are not needed for cohort/cross sectional designs.
# The data should have reference levels of all risk factors already set.
# This can be done as follows but has already been applied to the data so is not run here:
# levels(stroke_reduced$htnadmbp) <- c(0, 1)
# stroke_reduced$subhtn <- factor(stroke_reduced$subhtn,levels=c(1, 2))
# levels(stroke_reduced$nevfcur) <- c(1, 2)
# stroke_reduced$global_stress2 <- factor(stroke_reduced$global_stress2,levels=c(1,2))
# levels(stroke_reduced$whrs2tert) <- c(1, 2, 3)
# levels(stroke_reduced$phys) <- c(2, 1)
# levels(stroke_reduced$alcohfreqwk) <- c(1, 2, 3)
# stroke_reduced$dmhba1c2 <- factor(stroke_reduced$dmhba1c2,levels=c(1,2))
# stroke_reduced$cardiacrfcat <- factor(stroke_reduced$cardiacrfcat,levels=c(1,2))
# levels(stroke_reduced$ahei3tert) <- c(3,2,1)
# levels(stroke_reduced$apob_apoatert) <- c(1,2,3)
# The 'causalPAF' package assumes the data is either complete case data or that missing data
# analysis has already been performed.
# Next, define the causal structure or directed acyclic graph (DAG) of the causal Bayesian
# network defined by the data. We list the parents of each exposure or risk factor or outcome
# in a vector as follows:
# Note it is important that the order at which the variables are defined is such that all
# parents of that variable are defined before it. Please refer to the figure of the causal
# Bayesian network (with both direct and indirect effects) defined earlier as an example of this
# order.
in_phys <- c("subeduc","moteduc","fatduc")
in_ahei <- c("subeduc","moteduc","fatduc")
in_nevfcur <- c("subeduc","moteduc","fatduc")
in_alcohfreqwk <- c("subeduc","moteduc","fatduc")
in_global_stress2 <- c("subeduc","moteduc","fatduc")
in_subhtn <- c("subeduc","moteduc","fatduc","phys","ahei3tert","nevfcur","alcohfreqwk",
"global_stress2")
in_apob_apoa <- c("subeduc","moteduc","fatduc","phys","ahei3tert","nevfcur","alcohfreqwk",
"global_stress2")
in_whr <- c("subeduc","moteduc","fatduc","phys","ahei3tert","nevfcur","alcohfreqwk",
"global_stress2")
# Note splines can be fitted within the causal structure as shown below especially if splines
# are to be used in the fitted models.
# It is important that splines of parent variables are "typed" or "spelt" consistently
# (including spaces) throughout as 'causalPAF' can fit models automatically provided variables are
# spelt consistently. Also if a parent variable is a spline it should be defined in spline
# format in all occurences of the parent variable.
in_cardiacrfcat <- c("subeduc","moteduc","fatduc","phys","ahei3tert","nevfcur","alcohfreqwk",
"global_stress2",
"ns(apb,knots=quantile(apb,c(.25,.5,.75)),Boundary.knots=quantile(apb,c(.001,.95)))",
"ns(whr,df=5)","subhtn")
in_dmhba1c2 <- c("subeduc","moteduc","fatduc","phys","ahei3tert","nevfcur","alcohfreqwk",
"global_stress2",
"ns(apb,knots=quantile(apb,c(.25,.5,.75)),Boundary.knots=quantile(apb,c(.001,.95)))",
"ns(whr,df=5)","subhtn")
in_case <- c("subeduc","moteduc","fatduc","phys","ahei3tert","nevfcur","alcohfreqwk",
"global_stress2",
"ns(apb,knots=quantile(apb,c(.25,.5,.75)),Boundary.knots=quantile(apb,c(.001,.95)))",
"ns(whr,df=5)","subhtn","cardiacrfcat","dmhba1c2")
# Then we define a two dimensional list consisting of
# 1. inlist i.e. a list of the parents of each variable of interest corresponding to its column
# name in the data. Splines should be included here if they are to be modelled as splines.
# 2. outlist i.e. a list of each variable of interest corresponding to its column name in the
# data. Splines should not be input here, only the column names of the variables of interest in
# the data.
# Again the order is such that each variable is defined after all its parents.
in_out <- list(inlist=list(in_phys,in_ahei,in_nevfcur,in_alcohfreqwk,in_global_stress2,
in_subhtn,in_apob_apoa,in_whr,in_cardiacrfcat,in_dmhba1c2,in_case),
outlist=c("phys","ahei3tert","nevfcur","alcohfreqwk","global_stress2","subhtn",
"apb","whr","cardiacrfcat","dmhba1c2","case"))
# If splines are to be used for variables listed in in_out$outlist, then the splines should be
# defined in the same order as variables appear in in_out$outlist as follows. It is necessary to
# list variables in in_out$outlist without splines if no spline is to be applied.
# It is important that Splines_outlist is defined in the following format
# list(c("splinename1","splinename2","splinename3")) for the package to be applied correctly.
# And Splines_outlist should not be an empty list(). If there are no splines it should be
# defined the same as in_out[[2]] and in the same order as variables defined in_out[[2]].
Splines_outlist = list( c("phys","ahei3tert","nevfcur","alcohfreqwk","global_stress2","subhtn",
"ns(apb,knots=quantile(apb,c(.25,.5,.75)),Boundary.knots=quantile(apb,c(.001,.95)))",
"ns(whr,df=5)","cardiacrfcat","dmhba1c2","case") )
# To fit these models to case control data, one needs to perform weighted maximum likelihood
# estimation to imitate estimation using a random sample from the population. We chose weights
# of 0.0035 (for each case) and 0.9965 (for each control), reflective of a yearly incidence of
# first ischemic stroke of 0.35%, or 3.5 strokes per 1,000 individuals. These weights were
# chosen according to average incidences across country, age, group and gender within
# INTERSTROKE according to the global burden of disease.
w <- rep(1,nrow(stroke_reduced))
w[stroke_reduced$case==0] <- 0.9965
w[stroke_reduced$case==1] <- 0.0035
# It is important to assign stroke_reduced$weights to the updated weights defined in w.
# Otherwise if stroke_reduced$weights <- w is not set, the alternative weights supplied in the
# fictional data will be used. In this case, we want to use weigths as defined in w.
stroke_reduced$weights <- w
#The checkMarkovDAG() function in the 'causalPAF' package should be used before running
# causalPAFplot() to ensure:
#1. The causal Markov condition holds for the causal structure defined in the variable in_out.
#2. The variables in in_out are listed in the order so that no variable is defined before a
# parent or direct cause. Note: if this order does not hold, checkMarkovDAG() will automatically
# reorder the variables in, in_out, provided it is a Markov DAG.
#The causal analysis requires that the causal structure is a Markov DAG. The Causal Markov (CM)
# condition states that, conditional on the set of all its direct causes, a node is independent
# of all variables which are not direct causes or direct effects of that node. In the event that
# the structure of a Bayesian network accurately depicts causality, the two conditions are
# equivalent. However, a network may accurately embody the Markov condition without depicting
# causality, in which case it should not be assumed to embody the causal Markov condition.
# in_out is as defined above and input into this code.
if(checkMarkovDAG(in_out)$IsMarkovDAG & !checkMarkovDAG(in_out)$Reordered){
print("Your in_out DAG is a Markov DAG.")
} else if( checkMarkovDAG(in_out)$IsMarkovDAG & checkMarkovDAG(in_out)$Reordered ) {
in_out <- checkMarkovDAG(in_out)[[2]]
print("Your in_out DAG is a Markov DAG.The checkMarkovDAG function has reordered your
in_out list so that all parent variables come before descendants.")
} else{ print("Your ``in_out'' list is not a Bayesian Markov DAG so the methods in the
'causalPAF' package cannot be applied for non Markov DAGs.")}
# The pointEstimate() function evaluates Point Estimates for Total PAF, Direct PAF, Indirect PAF
# and Path Specific PAF for a user inputted number of integral simulations. There is no bootstrap
# applied in this function.
# Since bootstraps are not applied, the pointEstimate() function will run quicker than the
# alternative causalPAFplot() function which calculates bootstrap estimates which can take
# longer to run.
pointEstimate(dataframe = stroke_reduced,
exposure="phys",
mediator=c("subhtn","apb","whr"),
response="case",
response_model_mediators = list(),
response_model_exposure = list(),
in_outArg = in_out,
Splines_outlist = Splines_outlist,
splinesDefinedIn_in_outDAG = TRUE,
model_listArg = list(),
weights = w,
NumSimulation = 3,
addCustom = TRUE,
custom = "regionnn7*ns(eage,df=5)+esex*ns(eage,df=5)")
# The pointEstimate() function below has response_model_mediators, response_model_exposure and
# model_listArg pre fit. This allows the user to apply customised regressions instead of the
# default setting above, where the 'causalPAF' R package fitted these regressions automatically
# based on the causalDAG defined in in_outArg.
# Libraries must be loaded if fitting models outside of the 'causalPAF' R package.
library(MASS)
library(splines)
# Next we fit the, response_model_mediators and response_model_exposure, models outside of the
# 'causalPAF' package as an input into the package.
# It is important that response_vs_mediator is a list and it must be the same length as the
# parameter, mediator, i.e. length( response_vs_mediator ) == length( mediator). In this
# example, mediator=c("subhtn","apb","whr") so length( mediator) is 3, so we create a list
# with three models for "subhtn","apb" and "whr" respectively in that order. Note in this
# example, the model is the same for each mediator, but it must still be input 3 times within
# the list as follows:
response_vs_mediator <- list(
glm("case ~ regionnn7*ns(eage,df=5)+esex*ns(eage,df=5) +subeduc+moteduc+ fatduc+ phys+
ahei3tert+ nevfcur+ alcohfreqwk+ global_stress2+ subhtn +
ns(apb, knots = quantile(apb,c(.25,0.5,0.75)),
Boundary.knots = quantile(apb,c(.001,0.95)))+
ns(whr,df=5)",data = stroke_reduced,family='binomial',w = stroke_reduced$weights ),
# "subhtn" mediator model
glm("case ~ regionnn7*ns(eage,df=5)+esex*ns(eage,df=5) + subeduc+ moteduc+ fatduc+ phys+
ahei3tert+ nevfcur+ alcohfreqwk+ global_stress2+ subhtn +
ns(apb, knots = quantile(apb,c(.25,0.5,0.75)),
Boundary.knots = quantile(apb,c(.001,0.95)))+
ns(whr,df=5)",data = stroke_reduced,family='binomial',w = stroke_reduced$weights ),
# "apb" mediator model
glm("case ~ regionnn7*ns(eage,df=5)+esex*ns(eage,df=5) + subeduc+ moteduc+ fatduc+ phys+
ahei3tert+ nevfcur+ alcohfreqwk+ global_stress2+ subhtn +
ns(apb, knots = quantile(apb,c(.25,0.5,0.75)),
Boundary.knots = quantile(apb,c(.001,0.95)))+
ns(whr,df=5)",data = stroke_reduced,family='binomial',w = stroke_reduced$weights ) )
# "whr" mediator model
# Next we fit a customised response_model_exposure model rather than allowing the package fit it
# automatically as shown previously. This must be a list of length 1.
response_vs_phys <- list(glm("case ~ regionnn7*ns(eage,df=5)+esex*ns(eage,df=5)+subeduc+moteduc+
fatduc+ phys+ ahei3tert+ nevfcur+ alcohfreqwk+ global_stress2",data = stroke_reduced,
family='binomial',w= stroke_reduced$weights) )
# model_listArg is a list of models fitted for each of the variables in in_out$outlist based on
# its parents given in in_out$inlist. By default this is set to an empty list. Alternatively the
# user can supply their custom fitted, model_listpop, which should be consistent with the causal
# structure. model_listArg is defined earlier in this tutorial.
# Note it is important that model_listArg is defined as a list and in the same order and length
# as the variables defined in in_outArg[[2]].
model_listArgFit <- list(glm(formula = phys ~ subeduc + regionnn7 * ns(eage, df = 5) +
esex * ns(eage, df = 5) + moteduc + fatduc, family = "binomial", data = stroke_reduced,
weights = weights), # model 1 phys
polr(formula = ahei3tert ~ subeduc + regionnn7 * ns(eage, df = 5) + esex * ns(eage, df = 5) +
moteduc + fatduc, data = stroke_reduced, weights = weights), # model 2 ahei3tert
glm(formula = nevfcur ~ subeduc + regionnn7 * ns(eage, df = 5) + esex * ns(eage, df = 5) +
moteduc + fatduc, family = "binomial",data = stroke_reduced, weights = weights),
# model 3 nevfcur
polr(formula = alcohfreqwk ~ subeduc + regionnn7 * ns(eage, df = 5) +esex * ns(eage, df = 5) +
moteduc + fatduc, data = stroke_reduced,weights = weights), # model 4 alcohfreqwk
glm(formula = global_stress2 ~ subeduc + regionnn7 * ns(eage,df = 5) + esex * ns(eage, df = 5) +
moteduc + fatduc, family = "binomial",data = stroke_reduced, weights = weights),
# model 5 global_stress2
glm(formula = subhtn ~ subeduc + regionnn7 * ns(eage, df = 5) +esex * ns(eage, df = 5) +
moteduc + fatduc + phys + ahei3tert +nevfcur + alcohfreqwk + global_stress2,family = "binomial",
data = stroke_reduced, weights = weights), # model 6 subhtn
lm(formula = apb ~ subeduc + regionnn7 * ns(eage, df = 5) +esex * ns(eage, df = 5) +
moteduc + fatduc + phys + ahei3tert +nevfcur + alcohfreqwk + global_stress2,
data = stroke_reduced,weights = weights), # model 7 apob_apoa coded as apb
lm(formula = whr ~ subeduc + regionnn7 * ns(eage, df = 5) + esex *ns(eage, df = 5) + moteduc +
fatduc + phys + ahei3tert +nevfcur + alcohfreqwk + global_stress2, data = stroke_reduced,
weights = weights), # model 8 whr
glm(formula = cardiacrfcat ~ subeduc + regionnn7 * ns(eage, df = 5) +esex * ns(eage, df = 5) +
moteduc + fatduc + phys + ahei3tert +nevfcur + alcohfreqwk + global_stress2 +
ns(apb, knots = quantile(apb,c(0.25, 0.5, 0.75)),
Boundary.knots = quantile(apb,c(0.001, 0.95))) + ns(whr, df = 5) + subhtn,
family = "binomial",data = stroke_reduced, weights = weights), # model 9 cardiacrfcat
glm(formula = dmhba1c2 ~ subeduc + regionnn7 * ns(eage, df = 5) +esex * ns(eage, df = 5) +
moteduc + fatduc + phys + ahei3tert +nevfcur + alcohfreqwk + global_stress2 +
ns(apb, knots = quantile(apb,c(0.25, 0.5, 0.75)),
Boundary.knots = quantile(apb,c(0.001, 0.95))) + ns(whr, df = 5) + subhtn,
family = "binomial",data = stroke_reduced, weights = weights), # model 10 dmhba1c2
glm(formula = case ~ subeduc + regionnn7 * ns(eage, df = 5) +esex * ns(eage, df = 5) + moteduc +
fatduc + phys + ahei3tert +nevfcur + alcohfreqwk + global_stress2 +
ns(apb, knots = quantile(apb,c(0.25, 0.5, 0.75)),
Boundary.knots = quantile(apb,c(0.001, 0.95))) + ns(whr, df = 5) + subhtn +
cardiacrfcat +dmhba1c2, family = "binomial", data = stroke_reduced, weights = weights)
# model 11 case
)
pointEstimate(dataframe = stroke_reduced,
exposure="phys",
mediator=c("subhtn","apb","whr"),
response="case",
response_model_mediators = response_vs_mediator,
response_model_exposure = response_vs_phys,
in_outArg = in_out,
Splines_outlist = Splines_outlist,
splinesDefinedIn_in_outDAG = TRUE,
model_listArg = model_listArgFit,
weights = w,
NumSimulation = 3,
addCustom = TRUE,
custom = "regionnn7*ns(eage,df=5)+esex*ns(eage,df=5)")
# See examples from the function causalPAFplot() to see how bootstraps and confidence intervals
# can be calculated in addition to the pont estimates calculated from the pointEstimate()
# function. The causalPAFplot() function also shows examples of plotting these point estimate
# results with confidence intervals.