scdenney/open-science-skills/conjoint-cleaning

Conjoint Data Cleaning Expert

1. Qualtrics Export Settings and Metadata

Export format: The right export depends on the pipeline. cjoint::read.qualtrics() (Method A) hard-errors when choice columns contain text, so export "Use numeric values" for it. projoint::reshape_projoint() parses the chosen profile from a label at the end of the outcome string (default A/B), and manual reshaping benefits from readable labels, so "Use choice text" works for those. When unsure, export both; they differ only in response coding. If working with non-Latin scripts (Chinese, Korean, Arabic), export as XLSX rather than CSV to avoid UTF-8/ANSI encoding issues. On Windows with East Asian locales, read.csv() may still require Sys.setlocale() to match the file encoding before import (see ?cjoint::read.qualtrics East Asian Language Support).

Metadata rows: Current Qualtrics CSV exports include 3 header rows before respondent data: (1) variable identifiers, (2) question text/descriptions, (3) ImportId JSON. Legacy exports have 2 rows. The cjoint::read.qualtrics() parameter new.format = TRUE (set explicitly; default is FALSE) handles the 3-row format. For manual import via readxl::read_excel() or readr::read_csv(), skip the appropriate number of metadata rows after reading headers.

Randomization order columns: If "Export viewing order data" is enabled, Qualtrics adds _DO_ columns (e.g., Block1_DO) containing pipe-separated integers showing element display order. These are useful for task-order robustness checks but are not needed for the core reshape.

2. Qualtrics Conjoint Implementation Methods

Qualtrics conjoint experiments use one of three implementation methods, each producing different column naming conventions:

Method A — Conjoint Survey Design Tool (Strezhnev): Generates JavaScript that Qualtrics executes to randomize profiles. Column naming follows F-{task}-{profile}-{attribute} for attribute levels and F-{task}-{attribute} for attribute names. The cjoint R package's read.qualtrics() function is purpose-built for this format.

Method B — Custom JavaScript + Embedded Data: Researchers write JavaScript to randomize attributes and store values in Qualtrics embedded data fields. Column naming is researcher-defined. Two common conventions: (i) C{x}-F-{task}-{idx} for attribute names and C{x}-F-{task}-{profile}-{idx} for profile values; (ii) the Graham (2020) convention, choice{task}_{attr}{profile} with fixed attribute order (e.g., choice1_bread1) or c{task}_attrib{pos}_name / c{task}_attrib{pos}_sand{profile} when attribute order is also randomized. Requires manual reshaping (Section 4).

Method C — Loop & Merge: Each loop iteration represents one conjoint task. Embedded data fields are referenced via ${e://Field/variable_name} and displayed with ${lm://Field/N}. Column names reflect the embedded data field structure. Requires manual reshaping.

Identify the method before writing cleaning code from the actual column headers, the QSF survey definition, or the survey JavaScript. Naming conventions vary too much to assume one.

3. Existing R Packages for Conjoint Data Import

Check whether an existing package handles the format before writing custom reshaping code.

`cjoint::read.qualtrics()` — Purpose-built for Conjoint SDT exports (Method A). Reads Qualtrics CSV directly, handles metadata rows, outputs one row per profile with a selected column. Parameters: responses (choice column names), covariates (respondent-level variables), respondentID, new.format (TRUE for 3-row headers), ranks (for rank/rating/top-L designs). Supports binary forced choice, profile ranks, per-profile ratings, and top-L choices; see ?cjoint::read.qualtrics Details for the four response types. Requires PHP/JS output from the Conjoint Survey Design Tool.

`cjdata::reshape_conjoint()` — Lightweight alternative. Functions: read_Qualtrics() + reshape_conjoint(). Handles basic wide-to-long conversion. Requires the terminal character of each outcome string to be {"1","2"} or {"A","B"} (so "Candidate A" works; Japanese zenkaku digits supported). Respondent covariates merged separately.

`projoint::reshape_projoint()` — For measurement-error-corrected analysis per Clayton, Horiuchi, Kaufman, King, and Komisarchik (2023). Built-in support for repeated tasks (IRR estimation), missing-agreement imputation (.fill = TRUE), and bias-corrected AMCEs. Outcome column names must contain task-ID digits, and the repeated-task outcome must be the last element of .outcomes. Expects wide columns named K-{task}-{attribute} and K-{task}-{profile}-{attribute} by default (.alphabet = "K"); selected profile is parsed from the final character of each outcome string via .choice_labels (default c("A","B")). Specify .flipped = TRUE when the repeated task presents profiles in reversed left/right order (see exampleData1 vs. exampleData2 in the manual); this changes how agreement is computed. Trap: projoint::read_Qualtrics() hard-codes a 2-row metadata skip (legacy format). For current 3-row Qualtrics exports, pre-strip the third metadata row or read manually via readr::read_csv(skip = 3) before calling reshape_projoint().

`cregg::cj_tidy()` — Reshapes wide data across the three-level respondent/task/profile hierarchy via two named lists: profile_variables (features and profile-specific outcomes that vary within a task) and task_variables (variables that vary by task but not across profiles within it). A choice variable that names the chosen profile ("left"/"right") goes in task_variables and must be recoded after reshaping, whereas per-profile "chosen" indicators go in profile_variables — getting this wrong silently corrupts the outcome. Constraint handling is not a cj_tidy feature; two-way design constraints are specified downstream via * in the amce()/cj() formula.

Package decision rule (default then escape hatch):

• Method A (SDT) export, no measurement-error correction needed → cjoint::read.qualtrics()
• Measurement-error / IRR correction via a repeated task → projoint::reshape_projoint()
• Simple Qualtrics CSV with binary outcome and default column naming → cjdata::reshape_conjoint()
• Complex wide data with non-standard column names that still map cleanly to profile/task variables → cregg::cj_tidy()
• Method B/C exports with custom embedded-data naming, language translation, attribute-level merges, or pilot-data exclusion that existing packages cannot handle → manual reshape (Section 4)

4. Manual Wide-to-Long Reshaping

When existing packages cannot handle the data format, reshape manually. The goal: one row per respondent x task x profile, one column per attribute.

Step 1: Build a long table of (ResponseId, task, profile, attribute_name, attribute_value)

Iterate over tasks, profiles, and attribute positions. For each combination, read the attribute name from the name column and the corresponding value from the value column. This naturally handles randomized attribute order.

# n_tasks / P / K = tasks, profiles per task, attributes per profile.
# Never name the task count `T`: it aliases TRUE in R, so seq_len(T) would
# silently run exactly one task instead of erroring.
rows <- vector("list", n_tasks * P * K)
i <- 0L
for (task in seq_len(n_tasks)) {
  name_cols <- paste0(prefix, "-F-", task, "-", seq_len(K))
  for (profile in seq_len(P)) {
    val_cols <- paste0(prefix, "-F-", task, "-", profile, "-", seq_len(K))
    for (idx in seq_len(K)) {
      i <- i + 1L
      rows[[i]] <- data.frame(
        ResponseId      = data$ResponseId,
        task            = task,
        profile         = profile,
        attribute_name  = data[[name_cols[idx]]],
        attribute_value = data[[val_cols[idx]]],
        stringsAsFactors = FALSE
      )
    }
  }
}
long <- data.table::rbindlist(rows)

For Graham-style embedded fields (e.g., choice1_bread1, c1_attrib1_name/c1_attrib1_sand1), a tidyr::pivot_longer(names_pattern = ...) one-shot is often cleaner than the triple loop — match the numeric indices into task/profile/attribute-position columns, then pivot back wide on attribute_name. Use data.table::rbindlist() for performance on large datasets.

Step 2: Filter missing data. Remove rows where attribute_name or attribute_value is NA — these indicate respondents who skipped the conjoint section.

Step 3: Apply attribute name and level merges before pivoting. Fix typos, encoding variants, or synonymous levels.

Step 4: Pivot to wide-by-attribute using data.table::dcast() or tidyr::pivot_wider():

dcast(long, ResponseId + task + profile ~ attribute_name, value.var = "attribute_value")

If dcast warns about duplicate row/column combinations, two positions in the same task share an attribute name for some respondents — investigate the name merge.

5. Choice Variable Mapping

Choice variables are separate Qualtrics questions (MC type), one per task, placed after each conjoint display.

Identify choice columns: These are not part of the embedded data fields. Map each choice question to its task number (Q17 -> task 1, Q19 -> task 2, etc.). Inspect the QSF or survey flow to confirm the mapping.

Text vs. numeric encoding: Text exports produce labels like "Person A"/"Person B" or "Profile 1"/"Profile 2"; numeric exports produce 1/2. Check which one the file contains.

Create the binary outcome: Prefer matching the terminal character of the choice string so the same code handles "Person A"/"Person B", "Profile 1"/"Profile 2", "Candidate A"/"Candidate B", and "Sandwich 1"/"Sandwich 2" uniformly. If numeric exports are used, coerce directly.

last_char <- stringr::str_sub(raw_choice, -1L)
chosen_profile <- dplyr::case_when(
  last_char %in% c("A", "1") ~ 1L,
  last_char %in% c("B", "2") ~ 2L,
  TRUE                       ~ NA_integer_
)
# Merge by (ResponseId, task), then:
chosen <- as.integer(profile == chosen_profile)

Handle missing choices: Drop rows where chosen is NA. Some tasks may have higher dropout rates than others (especially the last task).

6. Ratings and Secondary DVs

Ratings may be per-profile (one per profile per task, usable as a continuous conjoint DV) or per-task (one rating of the chosen profile — descriptive only, not a standard conjoint DV).

Endpoint label recoding: Qualtrics text exports encode scale endpoints as text labels (e.g., "Strongly disagree" = 1, "Strongly agree" = 7). Recode these before converting to numeric. Intermediate scale points are already numeric.

rating_num <- dplyr::case_when(
  raw_rating == "Strongly disagree" ~ 1,
  raw_rating == "Strongly agree"    ~ 7,
  TRUE                              ~ suppressWarnings(as.numeric(raw_rating))
)

7. Attribute-Level Translation and Factor Ordering

Use `recode_factor()` with deliberate level ordering. The order of arguments sets the factor level order, which determines:

• The AMCE reference category (first level = baseline, shown at zero)
• The display order in figures

For Stata/SPSS/Qualtrics exports that already carry labels (e.g., loaded via haven::read_stata()), haven::as_factor() converts all labelled columns to factors in one pass (Heiss 2023). Combine with forcats::fct_relevel() to set the baseline without re-typing every level. Use recode_factor() when you also need translation or level merges.

Reference category principles:

• Ordinal attributes (time, income): lowest level, so AMCEs show the benefit of higher levels
• In-group/out-group attributes (ethnicity): majority/in-group, so AMCEs show the out-group penalty
• Status/legal attributes: lowest status, so AMCEs show the benefit of higher status
• Nominal with a neutral option: neutral/no-opinion