Harmonized data as a survey design (experimental)

qes_design() turns the result of qes_harmonize() into a survey design object of the survey package (or a tbl_svy of srvyr), with the weight column that fits the targets, one stratum per study, and, in the long layout, the respondent as the sampling unit.

Usage

qes_design(
  x,
  weight = NULL,
  engine = c("survey", "srvyr"),
  pool = c("as_is", "equal")
)

Arguments

x: Harmonized data returned by qes_harmonize(), with its weight columns and attributes.
weight: The weight column: NULL (default, chosen from the waves of the targets, see Choosing a weight), "weight_pre" or "weight_post" (respondent layout) or "weight" (long layout).
engine: "survey" (default) returns a survey.design2 object of the survey package; "srvyr" a tbl_svy of the srvyr package. The package must be installed.
pool: "as_is" (default) keeps the weights of x; "equal" rescales them so that every study (every study and wave in the long layout; the polls of pooled polls count as one) has the same total.

Value

A survey.design2 (engine "survey") or tbl_svy (engine "srvyr") object whose data are the rows of x with a value of the weight, as a plain data frame, with the added column qes_stratum (the study, or <study>:<stratum> where the stratum column is set), and weight_auto when each study takes its own weight column (see Choosing a weight). The weight is the column named by weight, or weight_auto.

Choosing a weight

Each wave of a study has at most one recommended weight (see qes_spec("spec")$tables$weights). qes_harmonize() returns it as weight_pre (the pre-election wave) and weight_post (the post-election wave), or as weight in the long layout, normalized to mean 1 in each study and wave by default. A pre-election question such as vote intention is weighted with weight_pre, a post-election one such as reported vote with weight_post.

With weight = NULL, qes_design() looks at the wave each target of x was taken from, as attr(x, "qes_weight_guide") records it (targets that do not depend on the moment of the interview, such as the year of birth, do not count; a question that can be asked in any wave, such as sovereignty, counts through the wave that asked it): if all these waves call for the same column, it uses that column. If they call for both, but each study's call for one only (a pooled variable such as vote_choice takes the post-election recall of most studies and the pre-election intention of the CROP polls), each study gets its own, in a new column weight_auto that the design uses. If one study's targets call for both (the default targets = "core" does, for the studies with two waves), it stops and asks you to choose. With no such target, it uses the one weight column that has values, and asks you to choose when both have values. In the long layout the weight is the weight column.

Rows without a value of the weight (respondents outside the waves that have it, or waves whose weight needs review) are left out of the design, with a message that counts them by study.

Pooling studies

Normalized weights have mean 1 in each study and wave, so a pooled estimate gives each study a share proportional to its number of respondents. pool = "equal" rescales the weights so that each study (each study and wave in the long layout; the polls of pooled polls count as one) has the same total; it is the only place where qesR changes the relative size of studies. Whether either is meaningful depends on the question: studies differ in population, mode, wording and sampling, and the grades of qes_spec() say how comparable each study's question is.

Design

Studies are independent samples, so each study is a stratum. A study made of several independent samples is divided further, by its stratum column: the 1998 panel by polling firm (stratum 1 = CREATEC, 2 = CROP), the pooled CROP polls of 2007-2010 by poll (stratum is the poll's wave name, as in waves). In the respondent layout each row is its own sampling unit (ids = ~1). In the long layout a respondent interviewed in two waves appears in two rows of one study, so the respondent (qes_id) is the sampling unit and the variance accounts for the two answers of a panel respondent being related. The studies are opt-in online panels or telephone samples whose weights adjust to census margins; standard errors computed from these designs treat the weighted sample as a probability sample and are only as good as that assumption.

En français

qes_design() (expérimental) transforme le résultat de qes_harmonize() en plan de sondage du package survey (ou en tbl_svy de srvyr). Chaque étude est une strate ; en disposition longue, la personne (qes_id) est l'unité d'échantillonnage. Une étude formée de plusieurs échantillons indépendants est divisée selon sa colonne stratum : le panel de 1998 par firme de sondage (1 = CREATEC, 2 = CROP), les sondages CROP de 2007-2010 par sondage (stratum est le nom de la vague du sondage, comme dans waves). Avec weight = NULL, la pondération est choisie d'après la vague d'où vient chaque cible (attr(x, "qes_weight_guide")) : weight_pre pour des cibles de vagues préélectorales, weight_post pour des cibles de vagues postélectorales ; les cibles fixes, comme l'année de naissance, ne comptent pas ; si les études appellent des colonnes différentes mais chacune une seule (une variable regroupée comme vote_choice), chaque étude reçoit la sienne dans une nouvelle colonne weight_auto ; si une même étude mêle les deux, la fonction demande de choisir. Chaque vague a au plus une pondération recommandée. Les lignes sans valeur de pondération sont laissées hors du plan, avec un message. pool = "equal" donne le même total à chaque étude (et vague en disposition longue ; les sondages regroupés comptent pour une seule étude).

Examples

h <- qes_harmonize("qes_demo", targets = c("sov_indep", "vote_prov_recall"), quiet = TRUE)
# every target of the demonstration study was asked after the election
attr(h, "qes_weight_guide")
#>             target    study wave target_timing weight_column weight_var
#> 1 vote_prov_recall qes_demo post          post   weight_post       POND
#> 2        sov_indep qes_demo post           any   weight_post       POND
#>   weight_status
#> 1      reviewed
#> 2      reviewed
if (requireNamespace("survey", quietly = TRUE)) {
  d <- qes_design(h, weight = "weight_post")
  survey::svymean(~sov_indep, d, na.rm = TRUE)
}
#>                                          mean     SE
#> sov_indepYes                          0.41584 0.0718
#> sov_indepNo                           0.58416 0.0718
#> sov_indepWould not vote / would spoil 0.00000 0.0000