tests/testthat/test-extract-svyfgt-relm.R

In convey: Income Concentration Analysis with Complex Survey Samples

skip_on_cran()

# load libraries
library(survey)
library(laeken)
# library( vardpoor )

# run code for each level of g
for (this.g in c(0, 1, 2)) {
  # return test context
  context(paste0(
    "svyfgt-relm  g=",
    this.g ,
    " output survey.design and svyrep.design"
  ))

  ### test 1: test if funtion works on unweighted objects

  # load data
  data("api")
  # set up convey design
  suppressWarnings(dstrat1 <- svydesign(id =  ~ 1, data = apistrat))
  dstrat1 <- convey_prep(dstrat1)

  # perform tests
  test_that(paste0("svyfgt g=", this.g , " works on unweighted designs"), {
    expect_false(is.na (coef(
      svyfgt(
        ~ api00,
        design = dstrat1 ,
        g = this.g ,
        percent = .6 ,
        type_thresh = "relm"
      )
    )))
    expect_false(is.na (SE(
      svyfgt(
        ~ api00,
        design = dstrat1 ,
        g = this.g ,
        percent = .6 ,
        type_thresh = "relm"
      )
    )))
  })

  ### test 2: income data from eusilc --- data.frame-backed design object

  # collect and format data
  data(eusilc)
  names(eusilc) <- tolower(names(eusilc))

  # set up survey design objects
  des_eusilc <-
    svydesign(
      ids = ~ rb030 ,
      strata = ~ db040 ,
      weights = ~ rb050 ,
      data = eusilc
    )
  des_eusilc_rep <-
    as.svrepdesign(des_eusilc , type = "bootstrap" , replicates = 50)

  # prepare for convey
  des_eusilc <- convey_prep(des_eusilc)
  des_eusilc_rep <- convey_prep(des_eusilc_rep)

  # calculate estimates
  a1 <-
    svyfgt(
      ~ eqincome ,
      des_eusilc ,
      g = this.g ,
      percent = .6 ,
      type_thresh = "relm" ,
      deff = TRUE ,
      linearized = TRUE
    )
  a2 <-
    svyby(
      ~ eqincome ,
      ~ hsize,
      des_eusilc ,
      svyfgt ,
      g = this.g ,
      percent = .6 ,
      type_thresh = "relm" ,
      deff = TRUE ,
      covmat = FALSE # should be TRUE once we find a way to deal with full_design in svyby
    )
  a2.nocov <-
    svyby(
      ~ eqincome ,
      ~ hsize,
      des_eusilc ,
      svyfgt ,
      g = this.g ,
      percent = .6 ,
      type_thresh = "relm" ,
      deff = TRUE ,
      covmat = FALSE
    )
  b1 <-
    svyfgt(
      ~ eqincome ,
      des_eusilc_rep ,
      g = this.g ,
      percent = .6 ,
      type_thresh = "relm" ,
      deff = TRUE ,
      linearized = TRUE
    )
  b2 <-
    svyby(
      ~ eqincome ,
      ~ hsize,
      des_eusilc_rep ,
      svyfgt ,
      g = this.g ,
      percent = .6 ,
      type_thresh = "relm" ,
      deff = TRUE ,
      covmat = TRUE
    )
  b2.nocov <-
    svyby(
      ~ eqincome ,
      ~ hsize,
      des_eusilc_rep ,
      svyfgt ,
      g = this.g ,
      percent = .6 ,
      type_thresh = "relm" ,
      deff = TRUE ,
      covmat = FALSE
    )

  # calculate auxilliary tests statistics
  cv_diff1 <- abs(cv(a1) - cv(b1))
  se_diff2 <- max(abs(SE(a2) - SE(b2)) , na.rm = TRUE)

  # perform tests
  test_that("output svyfgt" , {
    expect_is(coef(a1) , "numeric")
    expect_is(coef(a2) , "numeric")
    expect_is(coef(b1) , "numeric")
    expect_is(coef(b2) , "numeric")
    expect_equal(coef(a1) , coef(b1))
    expect_equal(coef(a2) , coef(b2))
    # expect_lte( cv_diff1 , coef(a1) * 0.20 )         # the difference between CVs should be less than 5% of the coefficient, otherwise manually set it
    expect_lte(se_diff2 , max(coef(a2)) * 0.20)  # the difference between CVs should be less than 10% of the maximum coefficient, otherwise manually set it
    expect_is(SE(a1) , "matrix")
    expect_is(SE(a2) , "numeric")
    expect_is(SE(b1) , "numeric")
    expect_is(SE(b2) , "numeric")
    expect_lte(confint(a1)[1] ,  coef(a1))
    expect_gte(confint(a1)[2] , coef(a1))
    expect_lte(confint(b1)[, 1] , coef(b1))
    expect_gte(confint(b1)[2] , coef(b1))
    expect_equal(sum(confint(a2)[, 1] <= coef(a2)) , length(coef(a2)))
    expect_equal(sum(confint(a2)[, 2] >= coef(a2)) , length(coef(a2)))
    expect_equal(sum(confint(b2)[, 1] <= coef(b2)) , length(coef(b2)))
    expect_equal(sum(confint(b2)[, 2] >= coef(b2)) , length(coef(b2)))

    # check equality of linearized variables
    expect_equal(attr(a1 , "linearized") , attr(b1 , "linearized"))

    # check equality vcov diagonals
    expect_warning(expect_equal(diag(vcov(a2)) , diag(vcov(a2.nocov))))
    expect_warning(expect_equal(diag(vcov(b2)) , diag(vcov(b2.nocov))))

  })

  ### test 2: income data from eusilc --- database-backed design object

  # perform tests
  test_that("database svyfgt", {
    # skip test on cran
    skip_on_cran()

    # load libraries
    library(RSQLite)
    library(DBI)

    # set-up database
    dbfile <- tempfile()
    conn <- dbConnect(RSQLite::SQLite() , dbfile)
    dbWriteTable(conn , 'eusilc' , eusilc)

    # database-backed design
    dbd_eusilc <-
      svydesign(
        ids = ~ rb030 ,
        strata = ~ db040 ,
        weights = ~ rb050 ,
        data = "eusilc",
        dbname = dbfile,
        dbtype = "SQLite"
      )

    # prepare for convey
    dbd_eusilc <- convey_prep(dbd_eusilc)

    # calculate estimates
    c1 <-
      svyfgt(
        ~ eqincome ,
        dbd_eusilc ,
        g = this.g ,
        percent = .6 ,
        type_thresh = "relm" ,
        deff = TRUE ,
        linearized = TRUE
      )
    c2 <-
      svyby(
        ~ eqincome ,
        ~ hsize,
        dbd_eusilc ,
        svyfgt ,
        g = this.g ,
        percent = .6 ,
        type_thresh = "relm" ,
        deff = TRUE ,
        covmat = FALSE # should be TRUE once we find a way to deal with full_design in svyby
      )
    c2.nocov <-
      svyby(
        ~ eqincome ,
        ~ hsize,
        dbd_eusilc ,
        svyfgt ,
        g = this.g ,
        percent = .6 ,
        type_thresh = "relm" ,
        deff = TRUE
      )

    # remove table and close connection to database
    dbRemoveTable(conn , 'eusilc')
    dbDisconnect(conn)

    # peform tests
    expect_equal(coef(a1) , coef(c1))
    expect_equal(coef(a2) , coef(c2))
    expect_equal(SE(a1) , SE(c1))
    expect_equal(SE(a2) , SE(c2))
    expect_equal(deff(a1) , deff(c1))
    expect_equal(deff(a2) , deff(c2))
    expect_warning(expect_equal(vcov(a2) , vcov(c2)))
    expect_equal(suppressWarnings(vcov(a2.nocov)) , suppressWarnings(vcov(c2.nocov)))

    # check equality of linearized variables
    expect_equal(attr(c1 , "linearized") , attr(a1 , "linearized"))
    expect_equal(attr(c1 , "linearized") , attr(b1 , "linearized"))

    # check equality of indices
    expect_equal(attr(a1 , "index") , attr(c1 , "index"))

    # check equality of linearized variables
    expect_equal(attr(a1 , "linearized") , attr(b1 , "linearized"))

  })

  ### test 3: compare subsetted objects to svyby objects

  # calculate estimates
  sub_des <-
    svyfgt(
      ~ eqincome ,
      design = subset(des_eusilc , hsize == 1) ,
      g = this.g ,
      percent = .6 ,
      type_thresh = "relm" ,
      deff = TRUE ,
      linearized  = TRUE
    )
  sby_des <-
    svyby(
      ~ eqincome,
      by = ~ hsize,
      design = des_eusilc,
      FUN = svyfgt ,
      g = this.g ,
      percent = .6 ,
      type_thresh = "relm" ,
      deff = TRUE ,
      covmat = FALSE # should be TRUE once we find a way to deal with full_design in svyby
    )
  sub_rep <-
    svyfgt(
      ~ eqincome ,
      design = subset(des_eusilc_rep , hsize == 1) ,
      g = this.g ,
      percent = .6 ,
      type_thresh = "relm" ,
      deff = TRUE ,
      linearized = TRUE
    )
  sby_rep <-
    svyby(
      ~ eqincome,
      by = ~ hsize,
      design = des_eusilc_rep,
      FUN = svyfgt ,
      g = this.g ,
      percent = .6 ,
      type_thresh = "relm" ,
      deff = TRUE ,
      covmat = TRUE
    )

  # perform tests
  test_that("subsets equal svyby", {
    # domain vs svyby: coefficients must be equal
    expect_equal(as.numeric(coef(sub_des)) , as.numeric(coef(sby_des))[1])
    expect_equal(as.numeric(coef(sub_rep)) , as.numeric(coef(sby_rep))[1])

    # domain vs svyby: SEs must be equal
    expect_equal(as.numeric(SE(sub_des)) , as.numeric(SE(sby_des))[1])
    expect_equal(as.numeric(SE(sub_rep)) , as.numeric(SE(sby_rep))[1])

    # domain vs svyby: DEffs must be equal
    expect_equal(as.numeric(deff(sub_des)) , as.numeric(deff(sby_des))[1])
    expect_equal(as.numeric(deff(sub_rep)) , as.numeric(deff(sby_rep))[1])

    # domain vs svyby and svydesign vs svyrepdesign:
    # coefficients should match across svydesign
    expect_equal(as.numeric(coef(sub_des)) , as.numeric(coef(sby_rep))[1])

    # domain vs svyby and svydesign vs svyrepdesign:
    # coefficients of variation should be within five percent
    cv_diff <- abs(cv(sub_des) - cv(sby_rep)[1])
    expect_lte(cv_diff , .5)

    # check equality of linearized variables
    expect_equal(attr(sub_des , "linearized") , attr(sub_rep , "linearized"))

    # check equality of linearized variables
    expect_warning(expect_equal(vcov(sub_des)[1] , vcov(sby_des)[1, 1]))
    expect_equal(vcov(sub_rep)[1] , vcov(sby_rep)[1, 1])

  })

  ### test 4: compare subsetted objects to svyby objects

  # compare database-backed designs to non-database-backed designs
  test_that("dbi subsets equal non-dbi subsets", {
    # skip test on cran
    skip_on_cran()

    # load libraries
    library(RSQLite)
    library(DBI)

    # set up database
    dbfile <- tempfile()
    conn <- dbConnect(RSQLite::SQLite() , dbfile)
    dbWriteTable(conn , 'eusilc' , eusilc)

    # create database-backed design (with survey design information)
    dbd_eusilc <-
      svydesign(
        ids = ~ rb030 ,
        strata = ~ db040 ,
        weights = ~ rb050 ,
        data = "eusilc",
        dbname = dbfile,
        dbtype = "SQLite"
      )

    # create a hacky database-backed svrepdesign object
    # mirroring des_eusilc_rep
    dbd_eusilc_rep <-
      svrepdesign(
        weights = ~ rb050,
        repweights = attr(des_eusilc_rep , "full_design")$repweights ,
        scale = attr(des_eusilc_rep , "full_design")$scale ,
        rscales = attr(des_eusilc_rep , "full_design")$rscales ,
        type = "bootstrap" ,
        data = "eusilc" ,
        dbtype = "SQLite" ,
        dbname = dbfile ,
        combined.weights = FALSE
      )

    # prepare for convey
    dbd_eusilc <- convey_prep(dbd_eusilc)
    dbd_eusilc_rep <- convey_prep(dbd_eusilc_rep)

    # calculate estimates
    sub_dbd <-
      svyfgt(
        ~ eqincome ,
        design = subset(dbd_eusilc , hsize == 1) ,
        g = this.g ,
        percent = .6 ,
        type_thresh = "relm" ,
        deff = TRUE ,
        linearized = TRUE
      )
    sby_dbd <-
      svyby(
        ~ eqincome,
        by = ~ hsize,
        design = dbd_eusilc,
        FUN = svyfgt ,
        g = this.g ,
        percent = .6 ,
        type_thresh = "relm" ,
        deff = TRUE ,
        covmat = FALSE # should be TRUE once we find a way to deal with full_design in svyby
      )
    sub_dbr <-
      svyfgt(
        ~ eqincome ,
        design = subset(dbd_eusilc_rep , hsize == 1) ,
        g = this.g ,
        percent = .6 ,
        type_thresh = "relm" ,
        deff = TRUE ,
        linearized = TRUE
      )
    sby_dbr <-
      svyby(
        ~ eqincome,
        by = ~ hsize,
        design = dbd_eusilc_rep,
        FUN = svyfgt ,
        g = this.g ,
        percent = .6 ,
        type_thresh = "relm" ,
        deff = TRUE ,
        covmat = TRUE
      )

    # remove table and disconnect from database
    dbRemoveTable(conn , 'eusilc')
    dbDisconnect(conn)

    # perform tests
    expect_equal(coef(sub_des) , coef(sub_dbd))
    expect_equal(coef(sub_rep) , coef(sub_dbr))
    expect_equal(SE(sub_des) , SE(sub_dbd))
    expect_equal(SE(sub_rep) , SE(sub_dbr))
    expect_equal(deff(sub_des) , deff(sub_dbd))
    expect_equal(deff(sub_rep) , deff(sub_dbr))
    expect_warning(expect_equal(vcov(sby_des) , vcov(sby_dbd)))
    expect_equal(vcov(sby_rep) , vcov(sby_dbr))

    # compare database-backed subsetted objects to database-backed svyby objects
    # dbi subsets equal dbi svyby
    expect_equal(as.numeric(coef(sub_dbd)) , as.numeric(coef(sby_dbd))[1])
    expect_equal(as.numeric(coef(sub_dbr)) , as.numeric(coef(sby_dbr))[1])
    expect_equal(as.numeric(SE(sub_dbd)) , as.numeric(SE(sby_dbd))[1])
    expect_equal(as.numeric(SE(sub_dbr)) , as.numeric(SE(sby_dbr))[1])
    expect_equal(as.numeric(deff(sub_dbd)) , as.numeric(deff(sby_dbd))[1])
    expect_equal(as.numeric(deff(sub_dbr)) , as.numeric(deff(sby_dbr))[1])

    # check equality of linearized variables
    expect_equal(attr(sub_dbd , "linearized") , attr(sub_des , "linearized"))
    expect_equal(attr(sub_dbr , "linearized") , attr(sub_rep , "linearized"))

  })

}