LmImpute: INTERNAL FUNCTION: Regeression imputation.

In statisticsnorway/Kostra: Functions for Kostra

INTERNAL FUNCTION: Regeression imputation.

Description

Imputation by weighted regeression, using lm, allowing multiple explanatory variables and multiple response variables. Impute missing and wrong values (category 3) by the model based on representative data (category 1). Some data are considered correct but not representative (category 2).

Usage

LmImpute(
  data,
  model = "y~x",
  weights = NULL,
  limitModel = 2.5,
  limitIterate = 4.5,
  limitImpute = 50,
  maxiter = 10,
  returnIter = TRUE,
  returnYHat = FALSE,
  returnFirst = FALSE,
  returnLast = TRUE,
  returnFinal = FALSE,
  MultiFuction = function(x) {
     max(abs(x))
 },
  estimationGroup = TRUE,
  unfoldCoef = FALSE,
  category123 = NULL,
  forceCategory2 = rep(FALSE, N),
  BackTransform = NULL,
  warningEstimate = "estimate: Missing yImputed replaced by zero",
  removeEmpty = FALSE,
  NArStudHandling = warning,
  cvPercent = TRUE,
  returnSameType = FALSE
)

Arguments

data	Input data set (data.frame, data.table or list)
model	String with model formula
weights	NULL or string with weight expression
limitModel	Studentized residuals limit. Above limit -> category 2.
limitIterate	Studentized residuals limit for iterative calculation of studentized residuals.
limitImpute	Studentized residuals limit. Above limit -> category 3. No imputation when 0.
maxiter	Maximum number of iterations.
returnIter	When TRUE, iteration when observation was thrown outin output.
returnYHat	When TRUE, fitted values and corresponding estimates in output.
returnFirst	When TRUE, studentized residuals from first iteration in output.
returnLast	When TRUE, some results from last iteration in output.
returnFinal	When TRUE, extra results from final model in output.
MultiFuction	Transforming rStud for several responses into a single positive value.
estimationGroup	Total estimates will be be computed within each group. Default (and TRUE) is a single group (estimationGroup <- rep(1, N) ).
unfoldCoef	When TRUE several elements of coef will be spilt as several ouput elements. unfoldCoef=2 is a specialised variant used to ensure two coefficients in output (extra coefficient zero).
category123	When non-NULL, this is used directly with no iteration.
forceCategory2	Force category 2 (can be useful for elements imputed by another method)
BackTransform	When model contains transformation of y (e.g: "log(y)~x") a function (e.g: exp) can be supplied to transform back to original scale before calculation of leaveOutResid, yHat, yImputed, estimate, estimateYHat, estimateOrig and seRobust.
warningEstimate	Warning text when missing values. Use NULL to avoid warning.
removeEmpty	When TRUE empty elements will be removed from output.
NArStudHandling	Function (warning or stop) taking a message as input. Used when rStud in model (category 1) is missing.
cvPercent	When TRUE (default) cv output is in percent
returnSameType	When TRUE and when the type of input y variable(s) is integer, the output type of yImputed and estimate is also integer. Estimates/sums are then calculated from rounded imputed values.

Value

A list with separate elements. Each element can be a scalar, vector or a matrix. Possible elements are:

x	The input x variable
y	The input y variable
strata	The input strata variable
category123	The three imputation groups: representative (1), correct but not representative (2), wrong (3) and zero when x is missing.
yHat	Fitted values
yImputed	Imputed y-data
rStudFirst	Initial studentized residuals
rStud	The final (or last) studentized residuals
dffits	The final (or last) DFFITS statistic
hii	The final (or last) leverages (diagonal elements of hat matrix)
leaveOutResid	The final (or last) outside-model residual
iter	Iteration when observation was thrown out
N	Total number of observations (rows in data)
nImputed	Number of imputed observations
estimate	Totale estimate from imputed data
cv	Coefficient of variation = seEstimate/estimate. In percent when cvPercent=TRUE (default)
estimateYhat	Totale estimate based on model fits
estimateOrig	Estimate based on original data with missing set to zero
coef	The final (or last) model coefficient(s). Several variables when several parameters ("coef..Intercept.", "coef.x").
nModel	The final (or last) number of observations in model.
sigmaFirst	Initial square root of the estimated variance parameter
sigmaHat	The final (or last) square root of the estimated variance parameter
seEstimate	The final (or last) standard error estimate of the total estimate from imputed data
seRobust	Robust variant of seEstimate (experimental)

Examples

z = data.frame(  # Same example as in Thorud et.al (2010).
        x = c(1.1, 2.2, 3.3, 4.4, 5.5),
        y = c(2.3, 3.1, 3.2, 3.7, 4.5))
LmImpute(z)  # Simple regression
LmImpute(z, model = "y~x-1", weights = "1/x")  # Ratio model

rateData <- KostraData("rateData")               # Real Kostra data set
w <- rateData$data[, c(16, 5, 14, 15, 19)]
w <- w[is.finite(w[,"Ny.kostragruppe"]), ]       # Remove Longyearbyen
w[w[,"Ny.kostragruppe"]>13,"Ny.kostragruppe"]=13 # Combine small strata
names(w) = c("x", "y", "y14", "y15", "k")

# Ratio model within each strata assumming common variance
LmImpute(w, "y~x:factor(k)-1", weights = "1/x", estimationGroup = w$k)

# Similar to above, but two y variables
LmImpute(w, "cbind(y14,y15)~x:factor(k)-1", weights = "1/x", estimationGroup = w$k)

# Using transformation and "BackTransform"
LmImpute(w, "sqrt(y)~x",BackTransform = function(x) x^2,returnYHat = TRUE)

# Direct imputation of x
LmImpute(w, "I(y-x)~0",weights = "1/x",
  BackTransform = function(y){return(y+dynGet("data")$x)},
  limitModel = Inf, limitIterate = Inf, limitImpute = Inf,
  returnYHat = TRUE)

statisticsnorway/Kostra documentation built on Nov. 2, 2024, 6:40 p.m.