NBA 2014-15 · Shot Logs · 33,362 Three-Point Shots
How Model-Dependent Is the Estimate?
Triangulation check
The simulator reports what the published model implies under a stated scenario. This page asks whether the same scenario produces a similar result through two structurally different routes: propensity-score subclassification, and the fitted Bayesian network where it can represent the scenario. The results appear side by side.
Agreement narrows one concern: that the result is an artifact of the published logit specification. Material disagreement points the other way: the scenario should be described as model-form dependent rather than presented as though the alternative route supported it.
More detail
Frederick Mosteller suggested that a difficult quantity be estimated through several structurally different models (triangulation), on the reasoning that similar outcomes from differing structures are themselves evidence.
The propensity route standardizes the observed outcome among records that actually have the selected values to the full sample, whereas the published model assigns the selected values to every record and averages the predicted probabilities. The network route can read the scenario under the same rule as the published model, holding the other factors at each record’s observed values.
Where a scenario factor reaches the outcome only through another factor, the network reports a projection that lets the other factors respond, labeled as a different estimand and comparable on direction rather than size. Where a factor has no path to the outcome, it reports no network estimate.
The check uses the network fitted on the same specification: the same predictors and subgroup scope. The check does not address omitted measurement, nonresponse, causal direction, or other shared vulnerabilities, and similar point estimates are not treated as formal equivalence.
Loading simulator…
Couldn't load the simulator
Modeling on this page is restricted to binary (yes/no) outcomes.
Auto-Build Actionable found no movable predictors for this outcome, so this simulator uses the best Standard model instead — its predictors may not be classified as directly movable.
Set the scenario
The number on each card is the variable's incremental contribution — the gain in Tjur R² it adds once the other variables are included, the same quantity reported elsewhere on this site. Bars and ordering use it too, largest first. Hover a number to see how much of the variable's one-at-a-time separation that represents.
Predicted outcome
How did you change the shot mix?
Each pair of bars compares the actual distribution (gray) to the distribution under your scenario (blue where you changed it). This is what you changed — not the predicted impact, just the input.
How certain is this result?
Every prediction has wiggle room — these histograms show how much. The gray bars are the plausible answers for the baseline; the blue bars are the plausible answers for your scenario. Where the colors overlap, the two answers are close enough that the model can't cleanly tell them apart.
Set-everyone-to-X table
Click to show predicted outcomes for every level of every variable
Set-everyone-to-X table
Click to show predicted outcomes for every level of every variable
For each variable's level, the predicted outcome if every respondent had that response, all else unchanged. This is the analytic view of the simulator.
Support and calibration
This panel answers two questions. First, is there enough survey data behind your scenario? It shows how many people actually gave the answers you picked, and warns you if any of those answers was given by too few people to trust. Second, how well does the model match the survey?
Explore each factor
For each variable, see how the predicted outcome would change if you switched just that one selection — holding all your other selections fixed. This shows which individual factor settings produce the largest model-implied changes in the current scenario.
Run a scenario above and we’ll cross-check that exact change here — the simulator’s number beside a structurally different propensity estimate.