
Affective methods: asking "do you like it?"
Consumer (affective) methods measure the subjective response: acceptance, preference, emotions. Here the "instrument" is not a calibrated panel but a real consumer, so hundreds of respondents and honest statistics of means are needed. The methodological framework is ISO 11136:2014 (+Amd 1:2020; Russian counterpart GOST ISO 11136-2017).
The main tool is Peryam's nine-point hedonic scale (1957): from "like extremely" to "dislike extremely". Analysis: means ± confidence intervals, ANOVA (product as a fixed factor, respondent as a random one), Tukey post-hoc, then preference segmentation by hierarchical clustering (Ward's method): the "average consumer" often does not exist — there are 2–3 segments with opposite tastes.
How much did you like the sample overall?
Preference
- Paired preference: "which of the two samples do you prefer?" The statistic is a two-sided binomial test with p₀ = 1/2 (Module 04 can compute it too — in the Sensostat demo it is a separate method).
- Ranked preference (ISO 8587:2006): respondents rank 3–6 samples; analysis by Friedman's test with post-hoc comparisons (see Module 09).
- The purchase intent scale (5 points: "definitely would buy" … "definitely would not buy") — a bridge between sensory work and the marketing forecast.
JAR scales and penalty analysis
The JAR scale (just-about-right) is a bipolar five-point scale for a specific attribute: 1 — "too little", 3 — "just about right", 5 — "too much". It answers the technologist's question: "what should be tweaked in the recipe?".
Penalty analysis links deviations from JAR to the loss of overall acceptance: mean drop = the mean liking of the "JAR" group minus the mean liking of the deviation group. Reading rules:
- analyse an attribute if ≥ 20 % of respondents are outside JAR (with a smaller share the conclusions are unstable);
- mean drop significance — a t-test, with small groups — a bootstrap (e.g. 10 000 resamples);
- improvement priority = a combination of the share of dissatisfied respondents and the size of the penalty.
How the penalty is computed — a micro-example (teaching data, snack bar, "sweetness" attribute, 120 respondents):
| JAR group | n | Share | Mean liking | Mean drop |
|---|---|---|---|---|
| "Not sweet enough" (1–2) | 34 | 28 % | 5.9 | −1.5 |
| "Just about right" (3) | 64 | 53 % | 7.4 | — |
| "Too sweet" (4–5) | 22 | 18 % | 6.7 | −0.7 |
Reading: 28 % find the bar not sweet enough, and their mean acceptance is 1.5 points below the "JAR" group — at the 20 % threshold this is a priority improvement (raise sweetness). The "too sweet" group is below the threshold (18 %) with a moderate penalty — conclusions for it are unstable: record it, but do not act.
CATA and RATA: consumer profiling
CATA (check-all-that-apply) — the respondent ticks all applicable descriptors from a list of 15–20 terms. It requires no training and works with 60–100 consumers. RATA (rate-all-that-apply) adds an intensity rating for the ticked attributes.
CATA analysis:
- a frequency table "attribute × product";
- Cochran's Q test for each attribute — do the products differ in mention frequency;
- pairwise comparisons — McNemar's test with a multiple-comparison adjustment;
- a "products–attributes" map — correspondence analysis (CA) with bootstrap stability ellipses.
A reproducible example: 80 respondents, 4 snacks, the "too sweet" attribute. Mention frequencies: A = 8, B = 34, C = 12, D = 29. Cochran's Q = 31.31 (df = 3), p < 0.0001 — the products differ significantly in perceived sweetness. Pairwise McNemar: B vs C — p ≈ 0.0002; the B/D vs A/C difference is confirmed. Interpretation: "B and D are perceived by consumers as excessively sweet".
CATA correspondence analysis map (illustrative)
Consumer test design rules
- Sample size: screening — 30–50 respondents, a full test — 100+, for segmentation — 150+. CATA/JAR stabilise from ~50; for publications — 80–100+.
- Recruiting from the target group: real consumers of the category, not "whoever walked by"; quotas by sex/age if the product is segmented.
- Serving: sequential monadic (one product at a time with a full questionnaire); blind codes; order by a Latin square.
- Do not ask consumers analytical questions ("rate the intensity of the isoamyl acetate note") — they will answer, but the answer will be noise. The consumer answers for "liking"; the panel answers for "what and how much".
Further reading
- Jaeger S. R. et al. Check-All-That-Apply (CATA) Questions for Sensory Product Characterization by Consumers. Food Quality and Preference, 2015 — how many terms to put on a CATA list and why it hardly affects the outcome. article
- Meyners M., Castura J. C., Carr B. T. Existing and New Approaches for the Analysis of CATA Data. Food Quality and Preference, 2013 — all the statistics of CATA: from Cochran's Q to penalty-lift. article
- Ares G. et al. Penalty Analysis Based on CATA Questions to Identify Drivers of Liking. Food Quality and Preference, 2014 — penalty analysis without JAR scales, directly from CATA with an "ideal product". article
- Meyners M., Jaeger S. R., Ares G. On the Analysis of Rate-All-That-Apply (RATA) Data. Food Quality and Preference, 2016 — how to properly analyse the RATA variant with intensities. article
Test yourself
Three questions on consumer methods
1. In a JAR test, the "sweetness" attribute is outside the JAR zone for 12 % of respondents, and the penalty is significant. What should you do?
The ≥ 20 % rule: on a small group, the mean drop is computed from a handful of respondents and is unstable. Test significance does not override the sample adequacy rule.
2. How does CATA differ from QDA descriptive analysis?
Different data — different statistics: frequencies versus scale means. CATA does not measure intensity; it records the "noticeability" of an attribute.
3. Why cluster consumers after a hedonic test?
The classic example is coffee: a mean liking of 6.2 hides a "strong coffee lovers" segment (8.5) and a "mild coffee lovers" segment (4.1). The average product pleases no one.
Consumer testing without manual spreadsheets
The Sensostat demo already runs paired preference and the 9-point hedonic scale with automatic statistics and conclusions in the client's language.