inst/doc/Handling_of_NA_and_domains.md

In tobiasschoch/robsurvey: Robust Survey Statistics Estimation

Handling of NA and domains

Tobias Schoch – 31-01-2025

1 Intro

2 Basic robust estimators

2.1 Bare-bone functions (R/robsurvey-interal.R)

.check_data_weights(x, w, na.rm = FALSE, check_NA = TRUE), file: robsurvey-interal.R

• if na.rm = TRUE, missing values are removed using complete.cases()
• observations with zero weights kept untreated

2.2 Survey methods (R/robsurvey-interal.R)

.check_formula(f, design, na.rm = FALSE, check_NA = TRUE), file: robsurvey-interal.R

• if na.rm = TRUE, missing values are removed using complete.cases()
• the design is subset to complete cases
• domains are identified with $w_i = 0$
• flag: domain = TRUE
• in_domain is a Boolean vector of length $n$, respectively, in_domain = NULL if domain = FALSE

The prototype of a call (for the example svymean_trimmed(); unimportant code has been left away)

    dat <- .check_formula(x, design, na.rm)

• if na.rm = TRUE in the call of svymean_, the argument na.rm = TRUE is passed on to .check_formula() and subsequently missing values are removed using complete.cases()

• also, the design object in dat$design is a subset using complete.cases()

• domains are identified with $w_i = 0$

• if any $w_i=0$, the flag domain = TRUE is set

• The observations of the domains are identified using in_domain, which is a Boolean vector of length $n$ (without domains: in_domain = NULL)

Population- vs. domain-level estimate using the dat object

    res <- if (dat$domain)
        weighted_mean_trimmed(dat$y[dat$in_domain], dat$w[dat$in_domain], LB,
                              UB, info = TRUE, na.rm = FALSE)
    else
        weighted_mean_trimmed(dat$y, dat$w, LB, UB, info = TRUE, na.rm = FALSE)

• weighted_mean_trimmed() is called with na.rm = FALSE because the NA treatment took place earlier

Variance estimation

    # influence function
    infl <- .infl_trimmed(res$model$y, res$model$w, LB, UB, res$estimate) *
                res$model$w / sum(res$model$w)
    if (dat$domain) {
        tmp <- numeric(dat$n)
        tmp[dat$in_domain] <- infl
        infl <- tmp
    }
    # variance
    design <- dat$design
    res$variance <- svyrecvar(infl, design$cluster, design$strata, design$fpc,
                              postStrata = design$postStrata)

• The influence function is computed using the res object, which results from weighted_mean_trimmed(), which is the result of .check_data_weights()
• design is taken from the dat object

==ISSUE== With NA present in the data, infl is vector of length $\vert r\vert$ (response set), whereas all vectors of dat$design are of length $\vert s\vert$ (sample).

    # return
    res$estimator$domain <- dat$domain

• the domain flag is added to the returned list

3 Regression

tobiasschoch/robsurvey documentation built on June 1, 2025, 10:10 p.m.