Electric Insights
JFK Case Study All-or-Nothing Fine-Tuning Non-Response Explore Build
Advanced DIY Tools

Model Builder

Build and auto-select models, configure synthetic predictor variables, and create custom simulations.

How to Use This Tool

Follow these steps to build and analyze your models

1. Select Predictors

Check relevant variables. Hover info icons for definitions. The six predictor variables in the published Approval model are pre-selected as a default. Only relevant for Run Analysis — Auto-Build tests all predictors automatically and overrides your selection. Maximum 20 variables for Run Analysis.

2. Add Synthetic Variable (Optional)

Optionally create a synthetic variable to explore alongside your selected inputs. This helps test how a hypothetical factor might affect the outcome—and reveals which real-world variables show similar patterns. The algorithm adjusts the synthetic variable iteratively to match the parameters you specify. Works with both Run Analysis and Auto-Build.

3. Select Outcome

Choose what to model: Presidential Approval, Vote Intention, or Tax Cut Support. Presidential Approval is selected as a default.

4. Run Analysis or Auto-Build

Run Analysis estimates a model using your selected predictors. Auto-Build Optimal Model ignores your predictor selections and instead tests every available predictor, adding each one that meaningfully improves Tjur R² (forward stepwise). Use Auto-Build as a ceiling benchmark or starting point, then refine manually with Run Analysis. Note: Auto-Build does not use a synthetic variable — enable it afterwards and re-run with Run Analysis if desired.

Auto-Build vs. the published model: Auto-Build optimizes on Tjur R² and is best understood as a predictive benchmark. The published model selects predictors on theoretical grounds first — fit is a behavior check, not the selection criterion. A theory-driven model may score lower on fit and still be the more defensible choice for public release. The step log is designed to prompt that comparison.

5. Review Results

Examine probabilities, margins, and diagnostics. The Variables to Improve Your Model section scores and ranks candidates not yet in your model — click Add to selection on any card to check that variable, then re-run. After Auto-Build, the step log shows which predictor was added at each stage and the Tjur R² gain.

6. Build Simulator

Set hypothetical values to see their effects, including values for any synthetic variables you configured.

7. Iterate

Each run is saved automatically to the Saved Analyses tray above the results. Click Load on any card to restore its predictors, outcome, and synthetic setting and re-run. Compare runs side-by-side using the Tjur R², AUC, and Brier metrics shown on each card.

Understanding Your Results

After running your analysis, results are organized into up to three sections:

1

Full Model

Complete model including all selected predictors and the synthetic variable (if configured)

2

Base Model

Model performance without the synthetic variable (appears only when synthetic is configured)

3

Synthetic Variable Performance

How well the synthetic variable met your specifications and its impact on model performance (appears only when synthetic is configured)

Available Predictor Variables

Download Harris/Newsweek Questionnaire (PDF)

Select Outcome Variable

What outcome would you like to model?

Required
r =
-0.8 0.8
r =
0 0.8
2 (Binary) 5 (Max)
Modeling: Presidential Approval
0 of 20 variables selected
Auto-Building Optimal Model… 0s
  • Initialising…
  • Screening candidate predictors
  • Running forward stepwise selection
  • Fitting final model
  • Computing diagnostics & margins

Section 1: Full Model Performance

Complete model including all predictor variables and the synthetic variable

Predicted Probability

The average predicted probability across all observations

0.42

± 0.03 (95% CI)

Model Strength Overview Scoring Details

Overall score blends discrimination, calibration, separation, and explanatory power. When available, we also include evidence against the null and parsimony .

Rescaling to 0–1

  • AUC: (AUC − 0.50)/0.50
  • Brier: 1 − (Brier/0.25)
  • Tjur: Tjur/0.35
  • Pseudo R²: R²/0.40
  • p-value: −log10(p)/6
  • ΔAIC: ΔAIC/10

Aggregation Method

Clamp to [0,1], average equally over present metrics, map to a 0–100 score.

Interpretation Guide

≥80
Strong
60–79
Moderate
<60
Weak
—

Model Fit

Goodness-of-fit metrics for the model

Pseudo R²

Measures how well the model explains the variance in the outcome compared to a null model. Higher is better.

0.32

Higher is better

Rating: —

Tjur R²

A measure of discrimination. The mean predicted probability for outcome-positive cases (e.g., approvers) minus the mean for outcome-negative cases (e.g., non-approvers). Higher is better.

0.28

Higher is better

Rating: —

AUC (ROC)

The probability that the model assigns a higher predicted probability to a random positive case (e.g., an approver) than to a random negative case (e.g., a non-approver). Ranges from 0.5 (chance) to 1.0 (perfect); higher is better.

0.82

0.5 = random, 1 = perfect

Rating: —

Brier Score

Average squared difference between each predicted probability and the actual outcome (1 = approver, 0 = non-approver). Ranges from 0 (perfect) up to 1; lower is better.

0.15

Lower is better

Rating: —

Information Criteria

Metrics balancing fit and complexity

AIC Comparison

An information score for comparing models: how well the model fits the data after penalizing additional estimated parameters (e.g., each added predictor, each added level for a categorical variable, interaction term). Lower is better.

Null Model

–642.5

Change

Final Model

–498.3

Lower is better

Log Likelihood Comparison

Measures how well the model predicts observed outcomes. For each case: if y = 1, add ln(p̂); if y = 0, add ln(1 − p̂). Sum across all cases. Higher (less negative) means better fit.

Null Model

–642.5

Change

Final Model

–498.3

Higher is better

LR χ²

Likelihood ratio chi-square statistic comparing the fitted model to the null model

288.4

p-value

Statistical significance of the likelihood ratio test

<0.001

Sample Completeness

Sample retention after preprocessing
Retained 850 85%
Dropped 150 15%

Total: 1,000

The model shows statistically significant improvement over the null model (p < 0.001) with good predictive accuracy.