In marleneweinauer/iatools:

Welcome to my very first R package! (still under construction).

If you want to use my package and there are issues, just contact me: [e-mail]

What is the package about?

Interview falsification has been occurring for decades on different levels but is still often insufficiently detected and treated in practice. This is a severe issue as interview fabrication can have major effects on final estimates.

This R package provides some tools to test for suspicious interviewer behaviour for CATI (telephone) interviews.

The underlying article for this package appears in the December 2019 issue of the Statisticial Journal of the IAOS under the title "Be a detective for a day: How to detect falsified interviews with Statistics" (https://content.iospress.com/articles/statistical-journal-of-the-iaos/sji190524). As revealing conspicuous interviewers sometimes feels like detective work, the article intentionally has a little detective side story: So does the package. The means of evidence are collected, investigated with various methods, the conclusive evidence is filtered out and finally I suggest how to close the case.

This package points out interviewers with suspicious behaivour - whether a suspicous interviewer is really a falsfier must be examined on a case-by-case basis. Also I want to point out that the aim of this package is NOT to point a finger at interviewers, whose work I respect and appreciate a lot. The aim of this package is the improvement of data quality only.

rm(list=ls())

#library(devtools)
#devtools::install_github("marleneweinauer/iatools")
library(iatools)

Data

Raw Data

With iatools two types of data can be analysed for suspicious behaivour:

• survey_data
• time_data

survey_data and time_data must each be a data.frame in wide-table-format with observations in rows and variables in collumns. The restriction to a specific mode is recommended: The tests are designed for CATI- telephone interviews.

This can be a bit annoying, but there are three things to obey in both data.frames:

• There must be a collumn called "ID", indicating the ID of the observation (e.g. person's ID, househould's ID or enterprise's ID).
• There must be a collumn called "INTERVIEWER", indicating the ID of the interviewer who has conducted the interview.
• If there is a multiple choice question with k answer options you must have k seperate collumns for each answer options. They must end with _A1 to _Ak. (e.g. for QUESTION_X with 4 options you need 4 collumns: QUESTION_X_A1, QUESTION_X_A2, QUESTION_X_A3, QUESTION_X_A4). Also - and this is very important - you must not have _Ai (i from 1: 100) in any other collumn name than for this specific reason.

In iatools there are two test data sets, fullfillying the three above mentioned conditions.

• survey_data_test
• time_data_test

They can be assesed via data():

data("survey_data_test")
data("time_data_test")

Within the first fivteen variables of survey_data_test you see the "ID" variable and the "INTERVIEWER" variable. Also you see that the variable IUDEV is a multiple choice question with four answer options, indicated -- according to condition 3 -- by the collumns IUDEV_A1 to IUDEV_A4.

head(survey_data_test[, 1:15])

Class iaclass

To use any other function of iatools your data needs to be converted to the class iaclass with the function create_iaclass. This function makes sure that your data fulfills the conditions stated above.

The parameter "variables" indicates which survey variables you want to use in the further analysis. In the shown example we use almost every variable.

The parameter "key_variables" indicates variables that are gate questions (-> that can -- if not read correctly -- help to shorten the path of the interview extremly) or other variables of key interest in the survey, e.g. "Are you employed?" in a survey about employment.

Make sure that you convert your time_data and your survey_data, respectively.

# time_data 

key_variables = c("IU", "IFU2", "GOV_A3", "IBUY")
variables <- colnames(time_data_test)[which(colnames(time_data_test ) == "IU") : 
                             which(colnames(time_data_test ) == "SEC_DBU")
                           ]

time_data <- create_iaclass(dat = time_data_test ,
                      variables = variables, 
                      key_variables = key_variables, 
                      type = "time")

survey_data <- create_iaclass(dat = survey_data_test ,
                      variables = variables, 
                      key_variables = key_variables, 
                      type = "survey")

Your data is now of class iaclass:

class(time_data)
class(survey_data)

Investigation

5 Methods to investigate for suspicious interviewer behaivour are implemented in iatools:

• "shares" Falsifying interviewers may show different patterns in the answers given in comparison to honest interviewers. Therefore, especially for questions that result in a list of follow-up questions ("gate questions"), deviant shares - leading to shorter questionnaire paths - are common in falsified interviews. The method "shares" tags shares that deviate from other interviewers' shares signficiantly.
• "median" Falsifying interviewers may have shorter time patterns. In this method, interviewers are tagged that show median question times below the lower bound ot the robost confidence interval.
• "q20" Again it's about time: With the median method only interviewers that falsify at least every second interview are tagaged. "q20" compares 20%-quantiles instead of medians to find interviews that may falsify between 20% anad 50% of their interviews.
• "q" Method "q" accounts for interviewers with overproportional many fast extreme times on specific questions.
• "cluster" Falsifying interviewers will make similar mistakes. Method "cluster" therefore shows dendograms of cluster analysis.

Fast investigation

With the function collect_evidence, the methods "q", "q20", "median" & "shares" can be applied. As ouptut you get the R data.table "conspi_DT.m".

conspi_DT.m <- collect_evidence(
  time_data = time_data, 
  survey_data = survey_data, 
  tools = c("q", "q20", "shares"), 
  min_of_fast_interviews = 15
)

conspi_DT.m is a data.table in long-format that indicates for each INTERVIEWER, each question and each method applied, whether conspicious patterns where found.

print(conspi_DT.m)

The object conpsi_DT.m can be plotted with display_evidence(). The produced table shows in how many questions each interviewer was conspicious in each method.

display_evidence(conspi_DT.m)

In this table you get e.g. the impression that Interviewer INT_56 is suspcious as there are high numbers in "q20_conspi" or "q_conspi", indicating speeding or that Interviewer INT_18 is suspcious as there are high numbers in "shares_conpsi" or "shares_filter_conspi" (which is "shares_conspi" restricted to the key variables).

If we want to have a compact profile of these interviewers we can produce it with create_profile() (This text is currently in german - sorry - , it will be translated soon) Here, is the created text for INT_18:

create_profile(conspi_DT.m = conspi_DT.m, 
                       interviewer = "INT_18") 

or here for INT_56

create_profile(conspi_DT.m = conspi_DT.m, 
                       interviewer = "INT_56") 

If more detailed information about an interviewer is necessary, ta detailed RMD for this specific interviewer is produced with render_specific_interviewer(). To be overseeable per default the RMD only contains conspicious information. If the Rmd shall contain all information on this interviewer the paramter conspi must be set FALSE.

If you want to run the example path below, you must replace "setapath" with a real path.

#render_specific_interviewer(survey_data = survey_data, 
#                            time_data = time_data, 
#                            path = "setapath", 
#                            interviewer = "INT_18"
#                            )

...TODO... ...This vignette is currently unter construction...

marleneweinauer/iatools documentation built on Jan. 13, 2020, 3:24 p.m.