Quality alerts raised while preparing a recipe

prep() computes non-fatal quality alerts at every stage of the cascade, not only at the end. Each alert names a specific risk, its trigger and a remedy. Alerts are always stored on the prepped object (read them with weighting_alerts() / has_alerts()) and shown in the HTML report; with prep(warn = TRUE) they are also raised as R warnings. Every alert is tagged with the step that produced it, e.g. "[step_calibrate] ...".

Details

This page catalogues the alerts by theme. Thresholds are the prep() defaults unless noted; min_cell_n and max_factor are arguments of prep().

Weight distribution

Negative calibration weights: an unbounded linear/GREG calibration produced weight(s) below zero. They stay active (the totals are honest), but a negative survey weight is rarely wanted and makes the estimator erratic. Use a bounded distance (calfun = "logit" with bounds) or step_trim_calibrated().
Sub-one weights: unit(s) with a final weight below 1, so they represent less than themselves. Usually a sign of an over-aggressive adjustment.
Extreme g-factors: a calibration adjustment factor outside the Deville-Sarndal range [0.1, 10]. Consider bounds, a ridge penalty, or coarser auxiliaries.

Adjustment cells

Small cells: an adjustment cell (weighting class or post-stratum) with fewer than min_cell_n (default 30, Kalton and Flores-Cervantes 2003) cases. Collapse cells or switch to raking.
Empty cells: a cell with no unit to adjust to (no respondents, or all of unknown eligibility); the affected units are set to weight 0.
Large adjustment factors: a per-cell factor above max_factor (default 2.5), which inflates the variance.

Nonresponse and response propensities

Tiny propensities: a minimum fitted response propensity below 0.01 blows up the 1/p weights; check the model or trim.
Collapsed propensity classes: the fitted propensities were nearly constant, so the requested num_classes could not be formed and the class correction did nothing.
Miscalibrated propensities: a calibration slope far from 1; honest probabilities matter for 1/p. Use num_classes (rank-robust) or revise the model.
Non-positive nonresponse g-weights: an implied response probability of zero or below; use a bounded distance or other auxiliaries.

Calibration

Ill-conditioned system: near-collinear auxiliaries (large condition number) make the weights unstable; drop a redundant auxiliary or set a ridge penalty.
Did not converge: raking, linear or bounded calibration stopped without meeting every target; increase maxit or check the margins are mutually consistent.
Reconciled control totals: the tidy totals did not all sum to the same population size and were rescaled to a common N (informative, not an error).

Machine-learning adjustments

Learner without cross-fitting: a flexible learner (tree / forest / boost) run without crossfit can understate the variance through same-sample prediction. Add crossfit to estimate each unit out of sample.

Two-phase (double) sampling

Few phase-2 sampling units: fewer than 30 units were subsampled at the second phase. The recipe-aware bootstrap resamples the phase-2 variance component (V2) at this level, so few units leave V2 with few degrees of freedom and an unstable phase-2 standard error. Inspect the split with two_phase_variance().
Tiny phase-2 probability: a minimum phase-2 selection probability below 0.02 expands the subsampled weights sharply (up to 1/pi2), inflating V2. Check the phase-2 design or trim the expanded weights.

Finalising

Rounded to zero: rounding pushed a small or negative weight to exactly 0, so the unit left the active set and collect_weights(); round to more decimals or resolve those weights first.