The estimator | The birth rotation group, and why there are two imputations | Where each symbol lives in the call | The recursion | What it buys | The variance trap | The fat carry | Equal representation across rotation groups | References | Where to look next

The problem is not the weights | A map of what the steps are for | The same strategy as one object | The shortest real recipe | The object is the documentation | Where to go next

The two files | A volunteer sample that is wrong in four different ways | The reference is a sample, so give it its replicates | Step one: the pseudo-weight | What the sum of the weights is telling you | Step two: one calibration that carries every variable | Where it stops working: the rank ceiling | What sharing costs | What the reference contributes | Dissemination | How large is this sample, really? | Hold something out, because the weights cannot see selection bias | The whole thing | In short | References

The sequence, and which verb does each step | The longitudinal file is an intersection | Three ways to be missing, and only one of them is nonresponse | Attrition compounds | Calibrate to the first wave | Gross flows | Cross-sectional estimates from a longitudinal file are reference only | Where to look next

The structure comes first | Net change between two periods | Chaining: how production actually runs | The estimands are declared, not programmed | The chain | What has to be saved | Combinations over the chain | The composite estimator | Gross flows: who moved | Where to look next

What has to travel | The four cases | Reading $strata | The covariance is the pairing | previous is a list, and the lag is a design property | What coordination does not touch | Where to look next

The idea | In weightflow, with a random forest | Validation: the constraints are met exactly | Does it help the estimate? | What you need: unit-level auxiliaries | The hybrid: model term and consistency constraints | Equal weights within a household | Relationship to the literature | See also

When there is a second phase of sampling | A worked example | The two-phase variance | The cascade is re-run on every replicate | Calibrating the subsample to the first-phase sample | Scope | References

Post-stratification | Raking | Calibration distances | Linear (chi-square) distance -- the GREG estimator | Raking (exponential) distance -- always-positive weights | Logit distance -- bounded g-weights | Distribution of the adjustment factors | Integrative calibration (one weight per household) | Domain (partitioned) calibration | Estimates agree too | Agreement summary | Beyond survey: two checks against ReGenesees | Trimmed calibration matches trimcal() | Net change matches the analytic Berger-Priam estimator | What weightflow adds

The recipe is the artifact | A programmatic quality gate | Reproducibility | The report is the quality document | Variance and dissemination | Where it sits | See also

A reference survey is a weighted data frame | Estimated totals carry variance, so propagate it | Only the bootstrap propagates this component | When to use which

What calibration does | Why calibration reduces variance: the regression link | Post-stratification | Raking | Linear calibration (GREG) | Bounded calibration | Equal weights within a cluster | Which to use | See also | References

A worked recipe | What the report contains | Arguments

Every step has a stable id | The quality-alert gate | Unit by unit: what did a step do? | Domain by domain: is each domain reliable? | HTML report vs programmatic QC

Weighting classes | Response propensities | Logistic regression | Trees, forests and boosting | Flexibility, overfitting, and cross-fitting | Weighting the propensity model | Person or household level | Nonresponse by calibration | Which to use | See also

Why the adjustments matter for variance | Method 1: a PSU bootstrap that re-applies the recipe | Estimates with bootstrap standard errors | Method 2: hand the weights to the survey package | Replicate weights for a tidyverse workflow | Method 3: a delete-a-PSU jackknife that re-applies the recipe | Lonely PSUs and parallel replicates | When the finite-population correction matters | Confidence intervals: normal, t and percentile | Estimated control totals | Which one to use

Post-stratification | One variable | Several variables crossed | Raking | Linear / GREG calibration | Why the totals are usually awkward | The tidy way | Mixing categorical and continuous auxiliaries | Domain (partitioned) calibration | Model calibration | Validation and messages | Which one should I use? | See also

Automatic caps: step_trim_weights() | Ratio caps: step_trim() | Calibration-preserving caps: step_trim_calibrated() | Bounds that differ by subgroup | Which one should I use? | References

Flexible propensities | The overfitting problem | Cross-fitting | Ridge (penalized) calibration | Potter (MSE-optimal) trimming | Redistributing the trimmed mass | Calibration-preserving trimming | Putting it together | References

The disposition tree | What each branch means, and when it happens | Why there is unused and out-of-scope sample: sizing the design | What your input data needs | Order of operations | A full pipeline on the multistage sample | The same recipe with the ready-made indicator columns | References

The starting point: design weights | The cascade | Unknown eligibility | Ineligible units | Within-household selection | Nonresponse | Calibration | Trimming, rounding, rescaling | Reading the cascade: the design effect | Why the order matters