# Using CRTConjoint

library(CRTConjoint)

This vignette will go through examples cases listed in the package for all the main functions with details on how to use the inputs. Finally, the last section contains detailed steps on how to use CRTConjoint with the Amazon Web Services computing cluster.

## Understanding CRT_pval

To begin, we start with understanding the main function CRT_pval. All examples use the immigration choice conjoint experiment data from Hainmueller et. al. (2014). We first load this data from the package.

data("immigrationdata")

Each row consists a pair of immigrant candidates that were shown to respondents. For example, the first row shows that the left immigrant candidate was a Male from Iraq who had a high school degree while the right immigrant candidate was a Female from France with no formal education. The respondent who evaluated this task was a 20 year old college educated White Male, who voted for the left immigrant candidate.

In the first example, we aim to understand if the candidate’s education matters for immigration preferences. To test this, CRT_pval requires users to specify all factors and respondent factors of interest in the formula input along with the binary response. CRT_pval requires us to additionally specify which of the factors are the left and right profiles.

form = formula("Y ~ FeatEd + FeatGender + FeatCountry + FeatReason + FeatJob +
FeatExp + FeatPlans + FeatTrips + FeatLang + ppage + ppeducat + ppethm + ppgender")
left = colnames(immigrationdata)[1:9]
right = colnames(immigrationdata)[10:18]
left; right
#> [1] "FeatEd"      "FeatGender"  "FeatCountry" "FeatReason"  "FeatJob"
#> [6] "FeatExp"     "FeatPlans"   "FeatTrips"   "FeatLang"
#> [1] "FeatEd_2"      "FeatGender_2"  "FeatCountry_2" "FeatReason_2"
#> [5] "FeatJob_2"     "FeatExp_2"     "FeatPlans_2"   "FeatTrips_2"
#> [9] "FeatLang_2"

Users can see that the left and right profile factors are aligned, i.e., the first entry is the education for both the left and right profiles. It is important that they are expected to be aligned. We also note that the formula only contains the factors for the left profile. This is sufficient as the algorithm will take the left and right input to use both left and right profile attributes for testing the hypothesis. Lastly, we include all respondent characteristics (ppage, ppeducat, etc.) to boost power. We are ready to now run CRT_pval to test whether education matters.

education_test = CRT_pval(formula = form, data = immigrationdata, X = "FeatEd",
left = left, right = right, non_factor = "ppage", B = 100, analysis = 2)
education_test$p_val We again note X = "FeatEd" is sufficient to clarify which factor we are testing for. Because the function assumes all attributes in conjoint experiments are of class factor, if there are variables that are not factor class, for example respondent age (ppage), the function must know which are these non-factor variables. Furthermore, to save time we only run it for $$B = 100$$ resamples. Lastly, we set $$analysis = 2$$ to also allow the function to spit out two of the strongest interactions that contributed to the observed test statistic. We note that this is purely for exploratory purposes. The output should contain not only the $$p$$-value but also the observed test statistic, all the resampled test statistic, etc. The function will also show a progress bar to show the percentage of resamples the user is finished with. ### Using constrained randomization design Since the immigrant’s education was uniformly sampled, we did not need to specify the design because the default design was uniform. However, there are some factors that used more complex design. One such factor was job (FeatJob). For this experiment, the candidate’s occupation could only be financial analyst, computer programmer, research scientist, and doctor if their education degree was equivalent to at least some level of college. Because this constrained uniform design is popular in conjoint experiments, we allow the function to account for this design so long as the user specifies the constraints. We now show, using the same example, how to specify this constraint. constraint_randomization = list() # (Job has dependent randomization scheme) constraint_randomization[["FeatJob"]] = c("Financial analyst","Computer programmer", "Research scientist","Doctor") constraint_randomization[["FeatEd"]] = c("Equivalent to completing two years of college in the US", "Equivalent to completing a graduate degree in the US", "Equivalent to completing a college degree in the US") The constraint_randomization list is a list of length two. The first element contains the levels of job that can only be randomized with certain levels of education. The second element of the list contains the levels of education that allows the aforementioned jobs to have positive probability. The listed levels are assumed to match the levels in the supplied data. Additionally, the names of the list is also assumed to match the column names in the supplied data. We note that the user only has to supply the constraint for either the left or right factor and the function will assume the constraint randomization scheme is the same for both left and right factors, i.e., same for FeatJob_2 and FeatEd_2. job_test = CRT_pval(formula = form, data = immigrationdata, X = "FeatJob", left = left, right = right, design = "Constrained Uniform", constraint_randomization = constraint_randomization, non_factor = "ppage", B = 100) job_test$p_val

Once, we have the constraint list, we input it into the constraint_randomization input, after stating that the design = "Constrained Uniform". We supply other examples when a user has a nonuniform (but no constraints) design and how to force a variable to include as an interaction in the examples.

## Understanding extensions of CRT

We provide three other CRT functions that take similar inputs but tests regularity conditions often invoked in conjoint experiments. The first aims to test the profile order effect (CRT_profileordereffect), i.e., whether or not being in the left or right has any impact on the response. The syntax for such a test is straightforward.

profileorder_test = CRT_profileordereffect(formula = form, data = immigrationdata,
left = left, right = right, B = 100)
profileorder_test$p_val Testing the carryover effect and fatigue effect is slightly more complex. When testing the carryover effect, it requires resampling all the left and right factors. The default design = "Uniform" assumes that all factors were uniformly sampled. If that is not the case, (which the immigration example is not) we need to supply our own resamples. To do this, we build our resampling function, resample_func_immigration = function(x, seed = sample(c(0, 1000), size = 1), left_idx, right_idx) { set.seed(seed) df = x[, c(left_idx, right_idx)] variable = colnames(x)[c(left_idx, right_idx)] len = length(variable) resampled = list() n = nrow(df) for (i in 1:len) { var = df[, variable[i]] lev = levels(var) resampled[[i]] = factor(sample(lev, size = n, replace = TRUE)) } resampled_df = data.frame(resampled[[1]]) for (i in 2:len) { resampled_df = cbind(resampled_df, resampled[[i]]) } colnames(resampled_df) = colnames(df) #escape persecution was dependently randomized country_1 = resampled_df[, "FeatCountry"] country_2 = resampled_df[, "FeatCountry_2"] i_1 = which((country_1 == "Iraq" | country_1 == "Sudan" | country_1 == "Somalia")) i_2 = which((country_2 == "Iraq" | country_2 == "Sudan" | country_2 == "Somalia")) reason_1 = resampled_df[, "FeatReason"] reason_2 = resampled_df[, "FeatReason_2"] levs = levels(reason_1) r_levs = levs[c(2,3)] reason_1 = sample(r_levs, size = n, replace = TRUE) reason_1[i_1] = sample(levs, size = length(i_1), replace = TRUE) reason_2 = sample(r_levs, size = n, replace = TRUE) reason_2[i_2] = sample(levs, size = length(i_2), replace = TRUE) resampled_df[, "FeatReason"] = reason_1 resampled_df[, "FeatReason_2"] = reason_2 #profession high skill fix educ_1 = resampled_df[, "FeatEd"] educ_2 = resampled_df[, "FeatEd_2"] i_1 = which((educ_1 == "Equivalent to completing two years of college in the US" | educ_1 == "Equivalent to completing a college degree in the US" | educ_1 == "Equivalent to completing a graduate degree in the US")) i_2 = which((educ_2 == "Equivalent to completing two years of college in the US" | educ_2 == "Equivalent to completing a college degree in the US" | educ_2 == "Equivalent to completing a graduate degree in the US")) job_1 = resampled_df[, "FeatJob"] job_2 = resampled_df[, "FeatJob_2"] levs = levels(job_1) # take out computer programmer, doctor, financial analyst, and research scientist r_levs = levs[-c(2,4,5, 9)] job_1 = sample(r_levs, size = n, replace = TRUE) job_1[i_1] = sample(levs, size = length(i_1), replace = TRUE) job_2 = sample(r_levs, size = n, replace = TRUE) job_2[i_2] = sample(levs, size = length(i_2), replace = TRUE) resampled_df[, "FeatJob"] = job_1 resampled_df[, "FeatJob_2"] = job_2 resampled_df[colnames(resampled_df)] = lapply(resampled_df[colnames(resampled_df)], factor ) return(resampled_df) } This resampling function takes the data (x) as an input and given the indexes of the left and right profile attributes, it returns a completely new resampled dataframe of the same dimension as x. As stated above, because testing the carryover effect requires resampling all the attributes B times, we must supply all the manually resampled dataframes. To do this we store them in a list of length B, each containing a resampled data of all the left and right attributes. We store this in own_resamples. carryover_df = immigrationdata own_resamples = list() B = 100 for (i in 1:B) { newdf = resample_func_immigration(carryover_df, left_idx = 1:9, right_idx = 10:18, seed = i) own_resamples[[i]] = newdf } Lastly, the carryover test requires a column that indicates the task evaluation number for each row. In the immigration experiment, each respondent rated five tasks and the data is sorted by each task, i.e., the first five rows are the first five tasks for the first respondent. Consequently, we can define a new column, task, that iterates 1 to 5 and run the main function. NOTE: it is important that the task variable has no missing tasks, i.e., a respondent that only rated four tasks while all other respondents rated five tasks. If there is such a respondent, please drop it before using this function. J = 5 carryover_df$task = rep(1:J, nrow(carryover_df)/J)

carryover_test = CRT_carryovereffect(formula = form, data = carryover_df, left = left,
right = right, task = "task", supplyown_resamples = own_resamples, B = B)
carryover_test$p_val Lastly, to test the fatigue effect, we only need to additionally specify the respondent index. Like the task column, we similarly repeat 1 to 200 five times and store it in a new column called respondent. fatigue_df = immigrationdata fatigue_df$task = rep(1:J, nrow(fatigue_df)/J)
fatigue_df$respondent = rep(1:(nrow(fatigue_df)/J), each = J) fatigue_test = CRT_fatigueeffect(formula = form, data = fatigue_df, left = left, right = right, task = "task", respondent = "respondent", B = 100) fatigue_test$p_val

## Using CRTConjoint with Amazon Web Services

An optional argument for all CRT functions in this package is the num_cores input. Currently it is set at 2 cores as the default. Although 2 cores may be sufficient for someone who is doing exploratory work with $$B = 100$$, the final reported $$p$$-values is recommended to have a much higher value of $$B$$. A typical Mac laptop will only support up to num_cores = 4, which may be unsatisfactory for researchers. For researchers with their own computing cluster, this section is not applicable for them. However, for those with no easy access to computing cluster, we write this section to easily use our functions in powerful computing clusters provided by Amazon Web Services (AWS).

We will leverage the RStudio Server in AWS maintained and provided by: https://www.louisaslett.com/RStudio_AMI/. We will now list the steps needed to use AWS Rstudio.

We recommend that the user has already written all necessary code to run on the AWS. Since users will be charged hourly, it is recommended that they can directly run the code. If the user runs with 50 cores, the runtime for a typical conjoint experiment should not exceed more than 5-10 minutes for a single $$p$$-value with $$B = 2000$$, thus we do not believe any user to have to incur much fees. For any questions please do not hesitate to contact me at: daewoongham at g dot harvard dot edu