Lewsearch Public Forecasts

Audience Construction

National weighting targets

Both forecasts use a U.S. adult synthetic panel aligned to national targets. We publish the weighting targets so readers can see the post-stratification frame. We do not publish subgroup response forecasts in this release because those require separate validation against comparable Pew crosstabs.

Methodology note

Lewsearch forecasts are AI-generated estimates, not interviews with human respondents and not probability samples. Public forecasts are anchored to the latest published Pew toplines and use conservative movement estimates for the next comparable wave.

We publish point forecasts and uncertainty bands so the claim is testable. When Pew publishes the next comparable wave, this page will be updated with the error audit.

The public forecast uses baseline plus capped shrinkage of the Lewsearch movement signal. This prevents uncalibrated response-scale concentration from becoming the published point estimate.

The Risk label reflects expected forecast difficulty, not only interval width. It incorporates issue volatility, partisan divergence, ambiguity in comparable wording, and model confidence metadata.

Lewsearch does not manually select, smooth, or override individual item forecasts after the study is frozen. If a forecast moves down while most move up, it remains published as-is so the later audit evaluates the actual model output.

The studies were fielded as independent item-level runs, not as Pew's exact matrix/battery order. That design reduces UI and queue constraints, but it means these are item-level forecasts rather than exact mode/order replications of Pew's questionnaire.