Flag suspicious survey responses via the Response Entropy Index (REI)

Computes each respondent's Response Entropy Index (Tawa, 2021) and flags unusually low or high values. Note that no rows are removed; entries are only flagged via the Suspicious column.

Usage

remove_outliers_REI(df, header = FALSE, variables = "", range = c(1, 5))

Arguments

df: Data frame containing the data.
header: Which columns enter the computation. TRUE: only the columns named in variables. FALSE (the default, kept for compatibility): every column of df – remove ID, timestamp and other non-item columns first. A warning names any column that does not look like a response: with header = FALSE a non-numeric column, and in either mode a column with values outside range.
variables: Which variables to consider when header = TRUE: either a single character string with names separated by commas ("var1,var2") or a character vector (c("var1", "var2")). Names are matched exactly (not as regular expressions), so survey-export names such as "G01Q01[SQ001]" work; names not found in df are ignored with a warning.
range: Numeric vector of length 2 specifying the range of the Likert scale (used to sanity-check the responses). Defaults to c(1, 5).

Value

A data frame with the calculated REI, the item columns used, Percentile, and a Suspicious flag ("No", "Maybe", "Yes"; NA for a respondent without answers).

Details

The REI is the Shannon entropy (base 10) of a respondent's distribution of answers over the response options, \(REI_i = -\sum_k p_{ki} \log_{10} p_{ki}\), where \(p_{ki}\) is the proportion of the respondent's answered items that received option \(k\). Low values indicate overly consistent answering (e.g. straight-lining), high values overly scattered answering; either can indicate careless responding. Because the index ignores item content, compute it on the items as answered, before reverse-coding (Tawa, 2021).

Missing responses are left out: proportions are taken over the items a respondent answered, so the same answer pattern yields the same REI however many items were skipped (dividing by the total number of items, as versions before 0.3.0 did, lowered the REI of anyone who skipped items). A respondent who answered nothing gets NA. Responses outside the declared Likert range trigger a warning (they often indicate mis-coded data, or a non-item column) but are still included in the REI computation.

Flags are relative to the sample: each REI is converted to a percentile of a normal distribution with the sample's mean and standard deviation. "Maybe" marks the outer 10% on either side (the preliminary guideline of Tawa, 2021), "Yes" the outer 5%. When all respondents have the same REI the percentiles are undefined: Percentile is NA, no row is flagged, and a warning says so.

References

Tawa, J. (2021). The Response Entropy Index: Comparative assessment of performance and cultural bias across indices of careless responding. Survey Research Methods, 15(3), 299–325. doi:10.18148/srm/2021.v15i3.7832

Examples

# \donttest{
df <- data.frame(
  id = 1:6,
  q1 = c(1, 5, 3, 3, 2, 4), q2 = c(1, 1, 4, 3, 2, 5),
  q3 = c(1, 4, 3, 3, 5, 4), q4 = c(1, 2, 4, 3, 1, 5)
)
# select the item columns; the ID column is no response
result <- remove_outliers_REI(df, TRUE, c("q1", "q2", "q3", "q4"), c(1, 5))
result
#>        REI q1 q2 q3 q4 Percentile Suspicious
#> 1 0.000000  1  1  1  1         13         No
#> 2 0.602060  5  1  4  2         91      Maybe
#> 3 0.301030  3  4  3  4         54         No
#> 4 0.000000  3  3  3  3         13         No
#> 5 0.451545  2  2  5  1         77         No
#> 6 0.301030  4  5  4  5         54         No
# }