Method at a glance
- Modality: multimodal (semantic self-report about words, not about the rater’s own body)
- Measures: The average judged position of a large set of words on a set of semantic/affective dimensions — e.g. how far each word is ‘related to internal body sensations’ — yielding per-word norms that can be regressed against each other to ask which dimensions constitute a category.
- Procedure: A large word set is split across questionnaire versions of ~240 items each. Each version collects one variable, introduced by a one-sentence instruction and a 7-point (or 5/9-point) Likert scale, and is completed by at least ~20 raters. Calibrator words spanning the scale range are presented first to anchor the scale; control words repeated across all versions of a variable allow cross-version reliability checks; raters may mark a word as unknown. Ratings correlating < .10 with the version mean are discarded and replaced. Per-word means become the norms; ICCs and cross-version control-word correlations quantify reliability.
- Also called: normative rating study, semantic feature ratings, word norms, affective norms, prototypicality ratings, Emopro
Strengths
- Scale. Ferré et al. norm 1,286 words on 9 variables from 651 raters — an item count no experimental paradigm in this wiki approaches, giving regression models real power (R2 = .582 on 1,283 observations).
- Decomposes a category judgment into its constituents. Asking raters to consider each emotion component separately reveals which ones the holistic judgment was made of — information no amount of holistic rating recovers.
- Cheap, reusable, and cumulative: norms are published as open datasets (Emopro; https://osf.io/a2hcu) and re-analysed by later studies. Ferré et al. is itself a second pass over Pérez-Sánchez et al.’s (2021) word set.
- Reliability is checkable per variable, via ICCs and repeated control words across questionnaire versions — a transparency most self-report in this wiki lacks.
Limitations
- Measures words, not people. A norm is the average semantic judgment about a lexical item. It licenses claims about what a concept contains and no claims about what anyone’s body does — the interpretive line that matters most, and the easiest to cross in a discussion section.
- Interoception norms are ratings of a word’s relatedness to internal body sensations, not interoceptive accuracy, sensibility, or attention. They belong to a different measurement family than the Heartbeat detection task or the Multidimensional Assessment of Interoceptive Awareness (MAIA) despite sharing the term.
- Reliability varies sharply by variable and the weak ones are not always the ignorable ones: interoception was the least agreed-upon variable in Ferré et al. (control-word r(22) = .654 vs .824-.955 for the rest) and also its headline predictor.
- One instruction sentence carries the entire construct. ‘Rate whether the word is related to internal body sensations’ is the whole interoception manipulation; raters’ readings of it are unaudited.
- Language-bound. Norms are per-language and per-population, and the standard sample is undergraduates (Ferré et al.: 85.3% women, mean age 22.2).
- Multicollinearity is endemic — affective variables correlate heavily (valence/evaluation r = .712), forcing stepwise or penalised models whose entry order can tell a different story than their coefficients.
Connections
- Key studies: Ferré et al. (2024)
- Related methods: emBODY (topographical bodily sensation mapping), Multidimensional Assessment of Interoceptive Awareness (MAIA)
The psycholinguist’s workhorse, and new to this wiki with Ferré et al. (2024). Take a large word set, split it across questionnaires, have ~20 people rate every word on one dimension, average, publish. The resulting norms — valence, arousal, concreteness, age of acquisition, and now interoception — become shared infrastructure that other labs regress against for decades (ANEW; Warriner et al. 2013; Stadthagen-Gonzalez et al. 2017; Emopro).
What kind of measure this is
The placement problem is the same one emBODY (topographical bodily sensation mapping) has, one level further removed, and it is worth being blunt about.
| method | asks | ground truth |
|---|---|---|
| Heartbeat detection task | can you count your heartbeats? | physiology |
| Multidimensional Assessment of Interoceptive Awareness (MAIA) | how do you relate to your body? | none (sensibility) |
| emBODY (topographical bodily sensation mapping) | where do you feel this emotion? | none (felt topography) |
| rating norms | is this word about internal body sensations? | none (semantics) |
emBODY at least asks about the rater’s own felt body. Rating norms ask about a word — a fact about the lexicon, averaged over people. On the Interoceptive taxonomy (Farb et al. Box 1) the method does not fit anywhere, because it is not measuring the rater’s interoception at all. It is measuring what a linguistic community takes interoception to be relevant to.
That sounds like a fatal disqualification for an interoception wiki and is not, for one reason: the semantic layer is where the word-cue objection lives. If you want to know whether word-cued bodily maps could be driven by conceptual knowledge, you need to know what the concepts contain — and that is a question about words, answerable only by a method like this one. See Emotion prototypicality.
The design details that carry weight
Three features distinguish a serious norming study from a survey, all present in Ferré et al.:
- Calibrators. Seven words spanning the full scale range are shown first, so raters do not discover the extremes halfway through and silently rescale. Chosen per variable (for interoception: hambriento/hungry at the high end, imaginación/imagination at the low).
- Repeated control words. Twenty-two words appear in every version of a variable, so the versions can be correlated against each other. This is the study’s real validity check and the reason its interoception weakness is visible at all — without control words, an ICC of .80 would have looked fine.
- Correlation-based trimming. Raters correlating < .10 with their version’s mean are dropped and replaced (1% of data here); near-zero correlations indicate idiosyncratic responding, negative ones indicate a reversed scale.
The reliability lesson
Worth generalising beyond this paper. Ferré et al. report two reliability metrics that disagree: interoception’s ICC (.80) is unremarkable and its cross-version control-word correlation (.654) is the worst in the study by a distance. ICC and cross-version agreement are answering different questions — the ICC asks whether raters within a version rank items consistently, the control-word correlation asks whether different groups of raters converge on the same values for the same items.
For a norming study, whose entire product is per-item means intended for reuse by other people, the second is the one that matters, and it is the one that is reported in prose rather than in the reliability table. When a method’s headline variable is its least reproducible one, that belongs in the limitations section.