CPG Demo · Beer Concept Test (n=803)

Model Builder

Find which measured factors are associated with the outcome you choose

Pick one of six outcomes — purchase intent, package appeal, brand fit, high-quality perception, premium perception, or most-often-purchased brand — then choose predictor variables across eight categories and see which ones move the headline you've selected.

Five of the six are a top-2-box recode of a 5-point scale: 1 if the respondent picked the top two boxes (e.g. Extremely / Very likely for purchase intent, or Strongly / Somewhat agree for premium perception); 0 otherwise. The sixth, most-often-purchased brand, is an unordered multinomial outcome across seven brand categories — the six most common brands plus an Other group.

The simulators and stress test are scoped to the Purchase Intent outcome. To build explanatory models for the other five outcomes, use this Model Builder. The Survey Explorer works across all six outcomes.

You can also open the Bayesian Network Analysis (BETA) directly — no prior model needed. Pick an outcome and it fits a constrained network on its own, mapping conditional dependencies among the selected measured factors. It shows which factors are connected to the outcome in the fitted network and whether those connections are represented as direct or indirect paths. Those paths do not establish causal or temporal direction.

How to Use This Tool

Pick the outcome you want to model, then check the predictor variables you want to test and click Run Analysis to fit. Variable cards start empty — for Purchase Intent, click Load default model to drop in the published model’s six variables as a starting point, or let Auto-Build search for you.

1. Pick an outcome

Eight outcomes are available. Purchase Intent (Q21, the case study's headline) appears three ways: Top-2-Box (logistic, the default), Top-1-Box (logistic on the strictest definition), and ordered 5-point (an ordered logit on the full scale, still reporting the headline as the top-2 probability). Four more are modeled on their top-2-box: Package Appeal (Q12), Brand Fit (Q16), High-Quality Perception (Q18r6), and Premium Perception (Q18r8). Most-Often-Purchased Brand (Q10) is fit as a multinomial logit across seven brand categories (headline = the predicted Bud Light share). Switch outcomes anytime with the chooser at the top of the variable section.

2. Pick predictors

Use the search box and category buttons to find variables: Test Design, Package Reaction, Competitor Intent, Brand Knowledge, Beer Preferences, Behavior, Background, and Demographics. Cards start empty — for Purchase Intent (Top-2-Box), Load default model loads the six published Package Reaction predictors; the other outcomes have no published model, so start from scratch or use Auto-Build. Run Analysis is capped at 20 predictors; Auto-Build searches the full set and isn't capped.

3. Fit and read

Click Run Analysis to fit the model with whatever variables you've checked. Or use Auto-Build Standard to let the algorithm search all variables automatically, or Auto-Build Actionable Predictors to weight selection toward factors you can act on — things you can change through package design or marketing, plus the demographic segments you can target. That classification is a planning aid, not evidence that changing a factor would change the outcome.

Optional: subgroup

By default, the model is fit on all 803 respondents. Optionally restrict it to one subgroup at a time using the Whose data? panel — e.g. just younger respondents, or just heavy-category buyers — to see how the model behaves within that group. Subgroup levels with fewer than 100 respondents are hidden.

Where to start. Variable cards start empty, so you are never anchored to one model. Three ways to begin: click Load default model to load a published starting model (the six-variable Purchase Intent model; the other outcomes have no published model), use Auto-Build to let the tool search the full predictor set, or just check the variables you want. Switching outcomes clears the grid so each model starts clean.

Which factors can you act on? BETA The Bayesian Network Analysis is a standalone tool — you don't need to build or Auto-Build a model here first. Pick an outcome (and optionally a subgroup) and it resolves the factor set for you, fits a constrained Bayesian network, and ranks each factor by its model-implied do()-style projection on the outcome — separating factors with a direct path to the outcome in the fitted network from those connected only indirectly or not at all. A what-if simulator lets you set a factor and read the projected change. These are conditional projections under the fitted structure, not causal effects; the arrows do not establish direction.

After you run — what to expect

Review results

Results show how well the model predicts the outcome and how much each predictor contributes. The Other Variables in This Survey section lists every variable not yet in your model — each card shows whether adding it would likely improve fit.

Use the simulator

Click Launch in the Simulator panel to open an interactive tool. In Pin mode, set any combination of survey responses — e.g. high package appeal and low brand fit — and watch the predicted headline probability update. For ordered outcomes you also see the full outcome distribution shift — where each rating's share starts versus where your scenario moves it — and a Distribute mode lets you reshape response shares with sliders that sum to 100%.

Iterate and compare

Each run is saved in the Saved Analyses tray. Click Load to restore a run, or Pin two runs to view them side by side. Each card shows AUC, Brier, and either Tjur R² (logistic) or McFadden R² (ordered logit) — higher R²/AUC and lower Brier mean a better-performing model.

If you ran with a synthetic variable, results show three sections:

Full Model

Complete model including all selected predictors and the synthetic variable.

Base Model

Model performance without the synthetic variable.

Synthetic variable impact analysis

How well the synthetic met its specifications and how it changed model fit (versus the base model).

What are you trying to predict?

Choose the outcome your model will try to explain.

Required

Loading outcomes…

Survey Questions to Include in Your Model

Download CPG Questionnaire (PDF)

Search variables

Filter by topic

View Group by category

Quick actions

Hide variable cards

Loading variables…

Outcome: …

0 of … predictor variables selected

Synthetic-Variable Sensitivity Module — optional vulnerability check using a hypothetical omitted factor

Enable the check — targets the selected outcome

What this checks. Adds a hypothetical omitted factor with a specified relationship to the outcome and a cap on its overlap with the model's predictors, then shows how much the model's fit diagnostics and predicted margins change. The exercise does not show that the factor exists or identify what it represents; it tests how vulnerable the published explanation would be to a missing factor with those properties.

Target Correlation with Outcome

r =

-0.8 0.8

Max Correlation with Any Predictor

r =

0 0.8

# Response Categories

2 (Binary) 5 (Max)

The model you're about to fit

Predictors (X)

No predictors selected yet. Check variable cards above or click Auto-Build below.

Outcome (Y)

…

A logistic regression model uses the predictors on the left to estimate the probability of the outcome on the right. This summary updates as you change your selection — and shows exactly what gets submitted when you click Run Analysis. Auto-Build ignores this selection and chooses predictors for you.

Run Analysis

Build a model from exactly the predictors you've selected above. Full analyst control.

Returns: model fit, predictor strengths, calibration, and an interactive simulator.

Auto-Build

Ignores your selection above. Searches the full set of available predictors and chooses a subset under the selected rule. Standard prioritizes held-out predictive fit; Actionable Predictors applies an additional movability classification. That classification is a planning aid, not evidence of causal effect.

Prioritizes predictors classified as more movable — a planning filter, not causal evidence.

Returns: a chosen subset of predictors plus full model fit, strengths, and an interactive simulator.

Bayesian Network Analysis

Runs a Bayesian network on the predictors you've selected above — which are directly vs. indirectly connected to the outcome, and how those connections are represented in the fitted network.

Returns: direct vs. indirect connections, link confidence, and a what-if simulator.

Auto-Building Optimal Model… 0s

Initializing…
Screening candidate predictors
Running forward stepwise selection
Fitting final model
Computing diagnostics & margins

Full Model Performance

Complete model including all predictor variables and the synthetic variable

Model fit metrics Tap to view Predicted Probability, Pseudo R², Tjur R², AUC, Brier, AIC, LL, LR test

Predicted Probability

—

Fit index — how it is built Scoring Details

The index averages the fit statistics available for this model: AUC, Brier, Tjur R², and pseudo R². The likelihood-ratio statistic and its nominal p-value are shown only as descriptive weighted-pseudolikelihood comparisons and are excluded from the index; under the analysis weight, the p-value is not interpreted as a significance test. The AIC improvement over the null model is also excluded because it measures distance from the null rather than predictive fit. On a binary outcome the four indexed measures describe closely related aspects of the same fit, so read the index as one summary rather than as four independent dimensions.

How the score is built

Each of the four measures is scored against the cut-offs below — the same ones its tile uses, plus an excellent mark above Strong. The score turns the four into a single number out of 100: Moderate on all four gives 60, Strong on all four gives 80, excellent on all four gives 100.

Discrimination (AUC)

Weak below 0.70 · Moderate 0.70 · Strong 0.80 · Excellent 0.90

Prediction error (Brier), lower is better

Weak above 0.18 · Moderate 0.18 · Strong 0.12 · Excellent 0.06

Separation (Tjur R²)

Weak below 0.10 · Moderate 0.10 · Strong 0.20 · Excellent 0.35

Pseudo R²

Weak below 0.20 · Moderate 0.20 · Strong 0.30 · Excellent 0.40

Values between two cut-offs score in between, in proportion; a measure past its excellent mark counts as excellent. Zero means chance or no better than a model with no predictors (AUC 0.50, Brier 0.25, Tjur and pseudo R² 0).

Aggregation Method

Clamp to [0,1], average equally over the metrics present, map to a 0–100 index. The cuts are the ones the metric tiles on this card already use for their Weak, Moderate and Strong chips, plus an excellent mark above Strong: each metric scores 0 at its floor, 0.6 at the Moderate cut, 0.8 at the Strong cut and 1 at the excellent mark. A model sitting on every Strong cut therefore scores 80, one on every Moderate cut scores 60, and the printed bands describe the tiles themselves, so the index cannot disagree with the chips beside it. The cuts and the bands are display conventions calibrated to this demonstration, not standards. The index summarizes fit only: it carries no information about whether the model contains the right predictors.

Display bands (fit only)

≥80

Strong

60–79

Moderate

<60

Weak

—

Model Fit

Pseudo R²

—

Higher is better

Benchmark: — (strong ≥ .30)

Separation (Tjur R²)

—

Higher is better

Benchmark: — (strong ≥ .20)

Discrimination (AUC)

—

0.5 = random, 1 = perfect

Benchmark: — (strong ≥ .80)

Prediction Error (Brier)

—

Lower is better

Benchmark: — (good ≤ .12)

Each tile rates one statistic against its own display cut-off used in this prototype. The overall verdict above combines them with held-out fit, so it can read lower.

Information Criteria

AIC Comparison

Null Model

—

Change

Final Model

—

Lower is better

Log Likelihood Comparison

Null Model

—

Change

Final Model

—

Higher is better

LR χ²

—

p-value

—

Synthetic-variable impact Whether adding the synthetic variable improved fit, and how it shifted predicted margins

Synthetic variable impact analysis

A sensitivity analysis: how the model's diagnostics change when adding a synthetic variable — a hypothetical unmeasured factor built to the strength you specified, with its overlap with your predictors held under your cap — versus the base model.

Δ McFadden R²

—

Δ Tjur R²

—

Δ AIC

—

Δ Log Likelihood

—

Δ AUC

—

Δ Brier

—

Summary:

Adding the synthetic variable produces small changes across the displayed fit diagnostics, with the largest change in model fit as measured by AIC.

Model Builder

How to Use This Tool

1. Pick an outcome

2. Pick predictors

3. Fit and read

Optional: subgroup

Review results

Use the simulator

Iterate and compare

If you ran with a synthetic variable, results show three sections:

What are you trying to predict?

Whose data?

Survey Questions to Include in Your Model

Run a null model?

Full Model Performance

Predicted Probability

How the score is built

Aggregation Method

Display bands (fit only)

Model Fit

Information Criteria

Sample Completeness

Detailed Results

Predicted Margins (Table)

Predicted Margins (Chart)

Base Model — Detailed Results

Predicted Margins (Base Model)

Synthetic variable impact analysis

Synthetic-variable change in predicted margins: Full (less) Base