This package is developed to evaluate the robustness of conjoint analysis results to the removal of small fractions of respondents and experimentally generated unique contests.
For any given Average Marginal Component Effect (AMCE) estimated with your conjoint experiment data, you can use our package to
-
detect the smallest proportion of respondents (or contests) that derives its sign or significance,
-
retrieve the set of those respondents (or contests) to further investigate their characteristics,
-
visualize the influences of respondents (or contests) to detect outliers.
Our package is built on zaminfluence introduced by Broderick, T.,
Giordano, R., & Meager, R. (2020) to calculate influence scores for
respondents and contests.
We provide a simple example to demonstrate how to use
ConjointSensitivity using the replication data for the non-partisan
YouGov experiment in Kirkland and Coppock (2018).
-
Broderick, T., Giordano, R., & Meager, R. (2020). An Automatic Finite-Sample Robustness Metric: When Can Dropping a Little Data Make a Big Difference?
-
Kirkland, P. A., & Coppock, A. (2018). Candidate choice without party labels: new insights from conjoint survey experiments. Political Behavior, 40, 571–591.
If you use ConjointSensitivity in your research, please cite our
paper:
Abramson, S., & Zafer, B. (2026) Do Voters Prefer Women and Young Candidates? Re-evaluating Evidence from Conjoint Experiments
You can install the development version of ConjointSensitivity: (Please let us know if you encounter any error.)
# install.packages("devtools")
devtools::install_github("zfrb/ConjointSensitivity")This example demonstrates how to use ConjointSensitivity using the replication data for the non-partisan YouGov experiment in Kirkland and Coppock (2018).
library(ConjointSensitivity)
library(dplyr, quietly = T, warn.conflicts = FALSE)
library(torch)
#> Warning: package 'torch' was built under R version 4.3.3
library(purrr)
#> Warning: package 'purrr' was built under R version 4.3.3
# Load the conjoint dataset
df <- read.csv("tests/testthat/testdata_KirklandCoppock_nonpartisan_yougov.csv")
df$cand_female = ifelse(df$Gender=="Female", 1, 0) # make sure the variable of interest is numeric
## Conjoint data should be in the long-form. # Fit the base linear model
df$Age = as.character(df$Age) # factor() is not allowed in formula when using AnalyzeRespondentSensitivity and AnalyzeContestSensitivity
fit.lm <- lm(
as.formula(win ~ cand_female + Age + Race + Job + Political),
data = df,
weights = weight,
x = TRUE,
y = TRUE
)
# Run the respondent-level analysis
respondent_results <- AnalyzeRespondentSensitivity(
fit.lm = fit.lm,
segroup = df$caseid, # survey respondent id
var_interest = "cand_female", # target AMCE
target = "sign", # or "significance"
dropped_resp_list = TRUE # if TRUE, function returns the set of respondent ids removed to reverse the target metric of the variable of interest.
)
# View main results
t(respondent_results$RespondentSensitivity)
#> [,1]
#> param_name "cand_female"
#> target "sign"
#> model_coef "0.03875656"
#> model_se "0.01932379"
#> model_p "0.04707456"
#> n_respondent "1146"
#> total_infl "-1.23122e-15"
#> median_infl "-4.175677e-06"
#> n_pos_infl "555"
#> n_neg_infl "591"
#> n_drop_auto "16"
#> n_drop "14"
#> rerun_coef "-0.001237848"
#> rerun_se "0.01630143"
#> rerun_pval "0.9395925"
#> reruns "4"
# View dropped respondents
respondent_results$DroppedRespondents
#> [1] 845 1133 1095 1157 1109 819 306 43 932 503 630 916 517 107
print(paste0(round(respondent_results$RespondentSensitivity$n_drop/respondent_results$RespondentSensitivity$n_respondent*100, 2), "% of respondents are removed to reverse the sign of the AMCE of cand_female."))
#> [1] "1.22% of respondents are removed to reverse the sign of the AMCE of cand_female."# Respondents removed to reverse the sign of the AMCE of "cand_female" are drawn in red.
# Median respondent is shown in green.
plot_cdf_respondent_influences(fit.lm = fit.lm,
segroup = df$caseid,
var_interest = "cand_female",
target = "sign",
ndrop = respondent_results$RespondentSensitivity$n_drop)
contest_results <- AnalyzeContestSensitivity(
formula = as.formula(win ~ cand_female + Age + Race + Job + Political),
data = as.data.frame(df),
respondent_id = "caseid",
contest_no = "contest_no",
var_interest = "cand_female",
target = "sign",
weights = "weight"
)
# View main results
t(contest_results$ContestSensitivity)
#> [,1]
#> param_name "cand_female"
#> target "sign"
#> model_coef "0.03875656"
#> model_se "0.01932379"
#> model_p "0.04707456"
#> total_infl "-1.248716e-15"
#> n_contest "2887"
#> median_infl "-1.698448e-06"
#> n_pos_infl "1416"
#> n_neg_infl "1471"
#> n_drop_auto "23"
#> n_drop "22"
#> rerun_coef "-0.001444463"
#> rerun_se "0.016424"
#> rerun_pval "0.9300392"
#> reruns "3"
# View dropped contests
contest_results$DroppedContests
#> [1] "201-870" "398-757" "350-664" "530-1246" "28-1138" "546-1214"
#> [7] "457-1236" "374-1035" "149-1036" "474-1135" "95-978" "190-1012"
#> [13] "201-773" "395-951" "634-706" "332-766" "600-1322" "220-1266"
#> [19] "103-937" "52-1001" "134-812" "57-954"
# View the mapping from profile attributes to contest ids.
# Note that contests are unordered pairs of profiles.
head(contest_results$ProfileKey)
#> profile_id profile
#> 1 584 0-65-Hispanic-Attorney-SchoolBoardPresident
#> 2 552 0-65-Black-Educator-CityCouncilMember
#> 3 291 0-45-Hispanic-Stay-at-HomeDad/Mom-StateLegislator
#> 4 433 0-55-Hispanic-Electrician-CityCouncilMember
#> 5 762 1-35-Hispanic-Educator-StateLegislator
#> 6 808 1-35-White-Electrician-RepresentativeinCongress