01
Survey Design
Following work is done precursory to shortlisting of a questionnaire:
SERVQUAL (Parasuraman, Zeithaml & Berry, 1988) — the classic service-quality instrument. Five dimensions (the "RATER" model): Reliability, Assurance, Tangibles, Empathy, Responsiveness. Each dimension gets multiple items, often asked twice — once for expectations, once for perceptions — with the gap between them as the quality score. CONSUMATICS' 2–3 items per attribute mirrors this multi-item-per-construct approach, but skips the expectation/perception gap and just measures perceived experience directly.
ACSI — American Customer Satisfaction Index (Fornell et al., 1996) — a structural model: perceived quality + perceived value → overall satisfaction → loyalty/complaints, estimated via partial least squares (PLS) regression. This is the closest analogue to CONSUMATICS' Stage 4: regressing overall satisfaction on attribute scores to derive data-driven weights, rather than assuming attributes matter equally. ACSI's use of PLS specifically to handle collinear predictors is exactly why CONSUMATICS falls back to ridge/PLS when VIF > 5.
Kano Model (Kano, 1984) — categorizes attributes as must-be (absence causes dissatisfaction but presence doesn't delight, e.g. reliability), performance (linear — more is better, e.g. value/price), and attractive (unexpected delighters, e.g. design flourishes). CONSUMATICS doesn't currently categorize attributes this way — it treats all five as linear performance drivers. This is a real methodological gap you could mention in the paper: Kano-style categorization would explain why "Reliability" often has a low regression weight (it's must-be, low variance once basic quality is met) even though it's critical to consumers.
Net Promoter Score (Reichheld, 2003) — single-item "would you recommend" loyalty measure. Not part of the current design but often paired alongside attribute batteries as the "overall satisfaction" dependent variable that key-driver regressions predict — which is precisely the role your Stage 4 "overall satisfaction" rating plays.
02
Verified Collection
Response verification (Stage 2) is the quality-filtering layer that runs on every submitted response before it's allowed into scoring. It checks four things:
Attention check — a planted instruction item (e.g. "select 'Agree' here") catches respondents who aren't reading questions. Fail it, response is voided.
Straight-lining — same answer on every item (zero variance) signals disengagement, not opinion. Flagged and dropped.
Speeding — completion time under 1/3 of the survey's median duration is treated as too fast to have been read genuinely. Removed.
Duplicate identity — backstopped by the Google-auth requirement at capture, re-checked at the database level.
Only responses passing all four proceed to Stage 3 (attribute scoring) — nothing failing verification ever touches the score.
03
Bayesian Weighted Scoring
Bayesian Weighted Scoring (Stage 6 — Bayesian shrinkage) solves a small-sample problem: a brand with only 12 responses shouldn't outrank a brand with 900 just because a handful of enthusiasts gave it 5 stars. It pulls each brand's raw score toward the category average, proportional to how little data backs it up.
The formula:
score = (n / (n + C)) × brand_mean + (C / (n + C)) × category_mean
n = the brand's response count
C = the shrinkage constant (default 40) — effectively "how many responses count as enough evidence to trust the brand's own mean"
brand_mean = the brand's raw weighted score from Stage 5
category_mean = the average score across all brands in that sub-category
As n grows relative to C, the brand's own mean dominates; when n is small, the category mean pulls the score back toward the pack. Concretely: at n=10 the brand's own mean only carries 20% weight (80% is category mean); at n=40 it's an even split; by n=120 the brand's mean carries 75%; by n=400, 91%.
This is a standard technique from Bayesian statistics called shrinkage estimation (empirical Bayes) — it treats category_mean as a prior and brand_mean as the observed data, blending them based on how much evidence (n) supports the observation over the prior. It's the same logic behind "wilson score" or "add-k smoothing" ranking systems used for things like review-site star ratings, adapted here with a tunable constant rather than a fixed pseudo-count.