Perception and norms

Scope and status

clean_perception! is an optional respondent-level transformation for RCT perceptions of community behavior and social approval, together with related personal attitudes and reported behavior. It is loaded and exported by HondurasTools, but neither clean_respondent nor make_data.jl calls it. This guide documents the current implementation; it does not establish that every recode is appropriate for every wave or analysis.

Keep this step explicit. The function assumes a broad historical delivery schema, mutates its input, and makes decisions about denominators and missing responses that need to be checked against the questionnaire. Synthetic end-to-end regression tests cover refusal handling, numeric inputs, missing gender, and categorical coding, but do not validate every substantive recode or real delivery. Integration into the main pipeline should wait for wave-specific input contracts and regression tests for these decisions.

These are perceptions of community practices and approval. They are distinct from Cognitive Social Structures reports about who is connected to whom.

Distinguish the measures

A perceived norm is not a village-level estimate of actual behavior. Likewise, personal knowledge and approval are not interchangeable. Suffixes describe the implementation's naming convention, not a validated measurement model.

Question polarity matters. For example, prenatal-care and fathers' involvement approval items ask about not performing the behavior; the baby-bathing approval item asks about not bathing immediately. The cleaner retains the question's approval direction rather than harmonizing every item to a common desirable-behavior scale. Check the exact wording before pooling, reversing, or comparing scores.

Codebook and wave coverage

The bundled codebook records 49 J-prefixed entries in each of RCT waves 1, 2, and 3, and none in wave 4. This counts metadata entries, including non-measurement entries; it is not a count of validated scales. Examples such as j0100, j0300, j0400, j2000a, j2900, and j3000 have survey labels w1 baseline, w2 baseline, and all, respectively. A wave entry does not imply every respondent was asked the question.

Inspect metadata and options before interpreting a value:

using HondurasTools
cb = load_codebook()
variable_info(cb, "j0300"; study="rct")
variable_options(cb, "j0300"; study="rct", wave=3)
variable_info(cb, "j0400"; study="rct", wave=3)

Companion I-, E-, F-, and C-section questions can have different eligibility rules, formats, and wave coverage. Empty metadata for repeated-question fields is not evidence that a cleaner's assumption is valid. See Codebooks and provenance for the source release and external guides.

Inputs and invocation

The intended input has wave suffixes removed, the raw question IDs listed below still present, and the respondent fields child and gender prepared. The implementation expects child to contain "Yes"/other labels and uses gender values "man" and "woman" when resolving birth-decision responses. Missing or unrecognized gender leaves "Partner"/"Respondent" resolutions missing.

The function unconditionally renames the crosswalk columns. It also reads f5100a–f5100i, i2500a–i2500k, f6000m, c0900a, and c0700d. Missing columns cause errors; inserting all-missing columns is not a safe substitute because several recodes interpret missing as false. Numeric fields accept numbers or numeric strings; fields requiring integers reject fractional values. Unexpected response labels also need review.

clean_respondent performs general respondent preparation, but its output is not a guaranteed ready-to-use input for this optional cleaner. Inspect retained question IDs, types, response labels, and wave eligibility first. In particular, a failed perception-cleaning call may leave a partially renamed table.

After those checks, the explicit invocation is:

# prepared_resp is a delivery-specific DataFrame satisfying the requirements above.
perception = deepcopy(prepared_resp)
name_map = Dict{Symbol,Symbol}()
clean_perception!(perception; namedict=name_map)

# Optional: apply supported categorical orderings after inspecting output values.
code_variables!(perception)

Keep using perception after the call. Do not assign the function's return value to it: the current implementation ends with a vector of variable names, not the modified DataFrame. Calling it twice on the same table also fails because the raw columns have already been renamed. name_map maps renamed output columns back to raw IDs; it is not a complete lineage map for derived variables.

Recoding and missingness

The implementation generally replaces "Don't know", "Don't Know", "Dont_Know", "Refused", and "Removed" with missing, but subsequent operations can turn those missing values into false. The following are existing behaviors to audit, not recommended universal rules:

Retain original responses and compare distributions by wave and eligibility group. In particular, distinguish a structurally unasked question, an unselected checkbox, a refusal, and an observed negative response before adopting these results.

code_variables! applies Never < Rarely < Sometimes < Always to frequency-perception columns and Bad < Neither < Good to approval-perception columns, including skin-to-skin, laxative frequency, chupon approval, and fathers' care of sick children approval. Boolean and nominal outputs keep their distinct representations. It does not reverse negative-polarity items or build composite norm scores. The cb keyword does not make these hardcoded perception orderings wave-specific.

The exported percvars list is a convenience list, not an exhaustive output schema: it omits the diarrhea-treatment block and some derived indicators. Inspect names(perception) when selecting outputs.

Implemented source-to-output crosswalk

This table records the active renames in rename_perception!. Derived columns and additional component inputs described above are not represented as one-to-one renames.

Before making this a default pipeline step

Establish the supported question set for each wave and delivery, preserve skip/refusal distinctions, verify the direction and denominator of each measure, and test the complete path from raw input through categorical coding. Until then, use this as a separately reviewed analysis transformation.

See the API reference for the function entry and Processing and analysis for the general cleaning workflow.