repGlm: Replication methods (JK1, JK2 and BRR) for linear regression...

In eatRep: Educational Assessment Tools for Replication Methods

Replication methods (JK1, JK2 and BRR) for linear regression models and trend estimation.

Description

Compute generalized linear models for complex cluster designs with multiple imputed variables based on the Jackknife (JK1, JK2) or balanced repeated replicates (BRR) procedure. Conceptually, the function combines replication methods and methods for multiple imputed data. Technically, this is a wrapper for the svyglm function of the survey package.

Usage

repGlm(datL, ID, wgt = NULL, type = c("none", "JK2", "JK1", "BRR", "Fay"), PSU = NULL,
        repInd = NULL, repWgt = NULL, nest=NULL, imp=NULL, groups = NULL,
        group.splits = length(groups), group.delimiter = "_",
        cross.differences = FALSE, trend = NULL, linkErr = NULL, formula,
        family=gaussian, forceSingularityTreatment = FALSE,
        glmTransformation = c("none", "sdY"), doCheck = TRUE, na.rm = FALSE,
        poolMethod = c("mice", "scalar"), useWec = FALSE,
        scale = 1, rscales = 1, mse=TRUE, rho=NULL, hetero=TRUE,
        se_type = c("HC3", "HC0", "HC1", "HC2", "CR0", "CR2"),
        clusters = NULL, crossDiffSE.engine= c("lavaan", "lm"),
        stochasticGroupSizes = FALSE, verbose = TRUE, progress = TRUE)

Arguments

datL	Data frame in the long format (i.e. each line represents one ID unit in one imputation of one nest) containing all variables for analysis.
ID	Variable name or column number of student identifier (ID) variable. ID variable must not contain any missing values.
wgt	Optional: Variable name or column number of weighting variable. If no weighting variable is specified, all cases will be equally weighted.
type	Defines the replication method for cluster replicates which is to be applied. Depending on type, additional arguments must be specified (e.g., PSU and/or repInd or repWgt).
PSU	Variable name or column number of variable indicating the primary sampling unit (PSU). When a jackknife procedure is applied, the PSU is the jackknife zone variable. If NULL, no cluster structure is assumed and standard errors are computed according to a random sample.
repInd	Variable name or column number of variable indicating replicate ID. In a jackknife procedure, this is the jackknife replicate variable. If NULL, no cluster structure is assumed and standard errors are computed according to a random sample.
repWgt	Normally, replicate weights are created by repGlm directly from PSU and repInd variables. Alternatively, if replicate weights are included in the data.frame, specify the variable names or column number in the repWgt argument.
nest	Optional: name or column number of the nesting variable. Only applies in nested multiple imputed data sets.
imp	Optional: name or column number of the imputation variable. Only applies in multiple imputed data sets.
groups	Optional: vector of names or column numbers of one or more grouping variables.
group.splits	Optional: If groups are defined, group.splits optionally specifies whether analysis should be done also in the whole group or overlying groups. See examples for more details.
group.delimiter	Character string which separates the group names in the output frame.
cross.differences	Either a list of vectors, specifying the pairs of levels for which cross-level differences should be computed. Alternatively, if TRUE, cross-level differences for all pairs of levels are computed. If FALSE, no cross-level differences are computed. (see examples 2a, 3, and 4 in the help file of the repMean function)
trend	Optional: name or column number of the trend variable. Note: Trend variable must have exact two levels. Levels for grouping variables must be equal in both 'sub populations' partitioned by the trend variable.
linkErr	Optional: name or column number of the linking error variable. If NULL, a linking error of 0 will be assumed in trend estimation.
formula	Model formula, see help page of glm for details.
family	A description of the error distribution and link function to be used in the model. See help page of glm for details.
forceSingularityTreatment	Logical: Forces the function to use the workaround to handle singularities in regression models.
glmTransformation	Optional: Allows for transformation of parameters from linear regression and logistic regression before pooling. Useful to compare parameters from different glm models, see Mood (2010). Note: This argument applies only if forceSingularityTreatment is set to 'TRUE'.
doCheck	Logical: Check the data for consistency before analysis? If TRUE groups with insufficient data are excluded from analysis to prevent subsequent functions from crashing.
na.rm	Logical: Should cases with missing values be dropped?
poolMethod	Which pooling method should be used? The “mice” method is recommended.
useWec	Logical: use weighted effect coding?
scale	scaling constant for variance, for details, see help page of svrepdesign from the survey package
rscales	scaling constant for variance, for details, see help page of svrepdesign from the survey package
mse	Logical: If TRUE, compute variances based on sum of squares around the point estimate, rather than the mean of the replicates. See help page of svrepdesign from the survey package for further details.
rho	Shrinkage factor for weights in Fay's method. See help page of svrepdesign from the survey package for further details.
hetero	Logical: Assume heteroscedastic variance for weighted effect coding? Only applies for random samples, i.e. if no replication analyses are executed.
se_type	The sort of standard error sought for cross level differences. Only applies if crossDiffSE == "wec" and hetero == TRUE and crossDiffSE.engine == "lm". See the help page of lm_robust from the estimatr package for further details.
clusters	Optional: Variable name or column number of cluster variable. Only necessary if weighted effecting coding should be performed using heteroscedastic variances. See the help page of lm_robust from the estimatr package for further details.
crossDiffSE.engine	Optional: Sort of estimator which should be used for standard error estimation in weighted effect coding regression. Only applies if useWec == TRUE. To date, only lavaan allows for stochastic group sizes.
stochasticGroupSizes	Logical: Assume stochastic group sizes for using weighted effect coding regression with categorical predictors? Note: To date, only lavaan allows for stochastic group sizes. Stochastic group sizes cannot be assumed if any replication method (jackknife, BRR) is applied.
verbose	Logical: Show analysis information on console?
progress	Logical: Show progress bar on console?

Details

Function first creates replicate weights based on PSU and repInd variables according to JK2 or BRR procedure. According to multiple imputed data sets, a workbook with several analyses is created. The function afterwards serves as a wrapper for svyglm implemented in the survey package. The results of the several analyses are then pooled according to Rubin's rule, which is adapted for nested imputations if the nest argument implies a nested structure.

Value

A list of data frames in the long format. The output can be summarized using the report function. The first element of the list is a list with either one (no trend analyses) or two (trend analyses) data frames with at least six columns each. For each subpopulation denoted by the groups statement, each dependent variable, each parameter and each coefficient the corresponding value is given.

group	Denotes the group an analysis belongs to. If no groups were specified and/or analysis for the whole sample were requested, the value of ‘group’ is ‘wholeGroup’.
depVar	Denotes the name of the dependent variable in the analysis.
modus	Denotes the mode of the analysis. For example, if a JK2 analysis without sampling weights was conducted, ‘modus’ takes the value ‘jk2.unweighted’. If a analysis without any replicates but with sampling weights was conducted, ‘modus’ takes the value ‘weighted’.
parameter	Denotes the parameter of the regression model for which the corresponding value is given further. Amongst others, the ‘parameter’ column takes the values ‘(Intercept)’ and ‘gendermale’ if ‘gender’ was the dependent variable, for instance. See example 1 for further details.
coefficient	Denotes the coefficient for which the corresponding value is given further. Takes the values ‘est’ (estimate) and ‘se’ (standard error of the estimate).
value	The value of the parameter estimate in the corresponding group.

If groups were specified, further columns which are denoted by the group names are added to the data frame.

References

te Grotenhuis, M., Pelzer, B., Eisinga, R., Nieuwenhuis, R., Schmidt-Catran, A., & Konig, R. (2017). When size matters: advantages of weighted effect coding in observational studies. International Journal of Public Health. 62, 163–167.

Examples

### load example data (long format)
data(lsa)
### use only the first nest
bt         <- lsa[which(lsa[,"nest"] == 1),]
### use only data from 2010
bt2010     <- bt[which(bt[,"year"] == 2010),]
## use only reading data
bt2010read <- bt2010[which(bt2010[,"domain"] == "reading"),]

### Example 1: Computes linear regression from reading score on gender separately
### for each country. Assume no nested structure.
mod1 <- repGlm(datL = bt2010read, ID = "idstud", wgt = "wgt", type = "jk2",
        PSU = "jkzone", repInd = "jkrep", imp = "imp", groups = "country",
        formula = score~sex, family ="gaussian")
res1 <- report(mod1, printGlm = TRUE)


### Example 2: Computes log linear regression from pass/fail on ses and gender
### separately for each country in a nested structure. Assuming equally weighted
### cases by omitting "wgt" argument
dat  <- lsa[intersect(which(lsa[,"year"] == 2010), which(lsa[,"domain"] == "reading")),]
mod2 <- repGlm(datL = dat, ID = "idstud", type = "JK2",  PSU = "jkzone",
        repInd = "jkrep", imp = "imp", nest="nest", groups = "country",
        formula = passReg~sex*ses, family = quasibinomial(link="logit"))
res2 <- report(mod2, printGlm = TRUE)

### Example 3: Like example 1, but without any replication methods
### trend estimation (without linking error) and nested imputation
dat  <- lsa[which(lsa[,"domain"] == "reading"),]
mod3 <- repGlm(datL = dat, ID = "idstud", wgt = "wgt", imp = "imp", nest = "nest",
        groups = "country",  formula = score~sex, trend = "year")
res3 <- report(mod3, printGlm = TRUE)

### Example 4: weighted effect coding to estimate whether a specific country's mean
### differs from the overall mean (whereas the overall population is a composite of 
### all countries). The procedure adapts the weighted effect coding procedures 
### described in te Grotenhuis (2017) for multiple imputation and replication methods. 
mod4 <- repGlm(datL = bt2010read, ID = "idstud", wgt = "wgt", type = "jk2",
        PSU = "jkzone", repInd = "jkrep", imp = "imp", formula = score~country,
        useWec=TRUE)
res4 <- report(mod4, printGlm = FALSE)

eatRep documentation built on March 31, 2023, 7:15 p.m.