Animal Welfare and Policy Risk Index (AWPRI)

Overview

The AWPRI places 24 countries, from 2010 to 2022, on three quantities that can be counted from public sources. The first quantity is the volume of animal use a country carries, measured as animals slaughtered per person, meat supplied per person, the animal share of dietary protein, and the aquatic share of flesh supply. The second is the binding animal welfare law a country has enacted, together with the institutional conditions under which that law is enforced and contested. The third is the division of a country's research effort between automating animal agriculture and studying animal welfare. The AWPRI is a composite of 12 indicators in three layers, and the composite is arithmetic throughout, so that no part of the published score is a model prediction.

The index documented on this page is the second version of the AWPRI. The first version was submitted to the Journal of Applied Animal Welfare Science and rejected with an invitation to resubmit. The revision changed the measurement and not the wording. Three variables were withdrawn for having no usable source, two were rebuilt on documented legislation instead of a general rights index, one country and eight years were dropped because the underlying statistics do not cover them, and the second layer was renamed because the earlier name did not describe what the layer measured. What changed and why is set out below. Where a claim in the published paper no longer holds, this site follows the data and not the paper.

What the Index Measures, and What It Does Not

Five concepts are routinely run together in this literature, and the first version of this index ran the five together as well. The five are separated here.

A high AWPRI score therefore means the following, and only the following. The country uses many animals, or has little binding welfare law and weak conditions for enforcing that law, or directs its research towards automating animal agriculture and away from animal welfare, in each case relative to the other 23 countries in the same year. A high score does not mean that more animals are suffering in that country, and no result on this site establishes that.

How the Score Is Built

Weights

The 12 weights sum to one, and the pipeline asserts that at import; if they ever stop summing to one the code refuses to run. A Layer 3 variable carries 0.1111 of the composite and a Layer 2 variable 0.0667, so a variable can sit further from the panel mean than another and still move a country's score less. Every breakdown on this site is weighted accordingly.

Codebook

The codebook lists every variable in the index together with the source the variable comes from, the exact query used to draw the variable, the unit, the point at which the variable is inverted, the weight the variable carries, and the route by which the variable is claimed to bear on animal welfare. That last field is graded in three levels. Direct means the variable counts animals or counts the law that covers animals. Indirect means the variable measures a condition under which welfare law is enforced or contested. Assumed means the link to welfare outcomes is asserted by this project and has not been demonstrated. All three variables in Layer 3 are graded as assumed, and Layer 3 should be read on that footing.

L1. Animal-use exposure

How many animals a country uses, and how intensively, which sets how much welfare is at stake before any question of protection arises.

L2. Governance capacity

Whether a country has binding animal welfare law and the institutional conditions under which such law is enforced and contested.

L3. AI amplification

How much of a country's research effort goes into automating animal agriculture, read against how much goes into animal welfare itself.

This table is generated from data/final/codebook.csv, which the pipeline writes from a single Python module. The site cannot state a label, weight, layer or source that differs from the one the analysis used, because it has no copy of its own to drift from. In the first version it did, and by the time the paper was reviewed the dashboard was describing 15 variables in three differently-named layers while the manuscript described something else.

The Panel and Its Coverage

The panel holds 24 countries over the years 2010 to 2022, giving 312 country-years with no gaps. The sources set that frame and the frame was not chosen for convenience, for the three reasons below.

• Why 2010 and not earlier. FAO revised the Food Balance Sheet methodology in 2020 and re-estimated the series only back to 2010. Splicing the old and new series would put a methodological break inside the panel. The published version ran from 2004 and did exactly that.
• Why 24 countries. Japan is absent from the FAOSTAT Food Balance Sheets and cannot be given values for three Layer 1 variables without inventing them, so it was dropped; the published version carried it. The remaining 24 are the countries for which every one of the 12 variables is available in every year. Availability in every year is the whole selection rule, and the rule is applied mechanically. The earlier version described its panel as countries with "documented welfare legislation advancement" and "substantial agricultural AI adoption" without defining either, which left the selection unreproducible.
• What that frame costs. A span of 13 years is too short for a fitted ARIMA to be identified, so the projections on this site are ordinary least squares trends. The same span also rules out the EU welfare directives with 2012 and 2013 transposition deadlines as difference-in-differences treatments, because the panel opens two years before the earlier of those two deadlines. The one event study that does run uses a later instrument.

Regions are the World Bank classification, used unmodified. The published version used a six-region scheme of its own with no stated principle, which put Australia and New Zealand in an "Oceania" group of two and reported regional means over it.

What the Score Is Made Of

The published version reported a random forest trained to predict AWPRI from its own 12 inputs, and read its feature importances as evidence that civil liberties "dominate" animal welfare risk. That model was fitting an identity. AWPRI is by construction an exact linear function of the normalised variables with known weights, so any model predicting the composite from those variables will reach near-perfect fit, and the feature importances describe how the algorithm split a deterministic function and describe nothing about animals. The random forest has been removed.

What can honestly be said is how the variance of the composite decomposes across its inputs. For each variable the share is w_i · cov(x_i, AWPRI) / var(AWPRI), and the shares sum to one exactly. A variable earns a large share by being both heavily weighted and widely spread across countries. In 2022 the largest share belongs to Rule of law. Three of the shares are negative, which is not an error. A share is negative when a variable runs against the composite across countries, so that countries scoring high on that variable tend to score low overall. All three are in the animal-use exposure layer, which is why the shares within that layer largely cancel. The countries that use the most animals per person are not for the most part the countries with the weakest governance.

These are accounting identities, not causal estimates. A variable with a large share is one the composite's spread is built from; nothing follows about what would happen to animals if the underlying quantity changed. Principal components are reported in data/ml/pca_loadings.csv for anyone who wants them, and are likewise a description of the index's own correlation structure.

Robustness

The 2022 ranking was rebuilt 46 ways, namely under z-scoring instead of min-max normalisation, under bounds pooled across all years instead of within each year, under four alternative layer weightings, dropping each variable in turn, dropping each layer in turn, and dropping each country in turn. Every rebuilt ranking is compared with the ranking as built by Spearman correlation of the country order.

What this shows. The normalisation choice barely matters. Z-scoring within year gives ρ = 0.9939 and moves no country more than two places. The weighting choice matters more. The weakest of all 22 structural checks is "without L2 entirely" at ρ = 0.6809, moving one country 10 places, and dropping a whole layer changes the ranking substantially, as it should, since the three layers measure different things. The most consequential single variable is the animal welfare research base. Dropping that variable takes ρ down to 0.8652 and moves the United States 10 places.

Sample dependence. Because the min-max bounds are set by the panel, every score depends on which countries are in the panel. The 24 leave-one-out checks measure the size of that dependence, and the answer is reassuring. Of the 24 countries, 18 can be removed without changing the order of the rest at all, and the worst case is CN (China), where ρ = 0.9941 and the largest move is two places. The full table is in data/final/robustness_summary.csv.

External Validation

The only cross-national benchmark that covers all 24 of these countries is World Animal Protection's Animal Protection Index, whose 2020 grades were converted to a numeric scale and correlated with the index. Spearman ρ is reported raw and partialling out log GDP per capita, since both measures are related to national income.

How to read the table. The composite is not significantly associated with the WAP grade (ρ = -0.264, p = 0.213). The earlier version of the index reported a stronger association (ρ = -0.451, p = 0.027), but that came from variables since withdrawn, and it did not survive adjustment for income (p = 0.467). This site does not describe the index as validated.

Taken as a result and not a failure, the divergence is informative and it is the reason the index is now framed as one of exposure and capacity, not of protection. WAP grades what a state has committed to on paper. AWPRI's first layer counts animals, and it is the layer that does correlate with the WAP grade (ρ = -0.484, p = 0.016, and it holds net of income at p = 0.027). The direction of that association is that countries WAP grades well tend to use fewer animals per person. The governance layer, which contains the legislation counts, is unrelated to the WAP grade (ρ = -0.105, p = 0.624). Counting instruments in FAOLEX and grading policy commitments are not measuring the same thing, and the strongest correlate of the WAP farm-animal grade in the whole dataset is rule of law (ρ = -0.780), which is not about animals at all.

Still outstanding. The Animal Protection Index is itself a measure of policy commitment, so agreement between the two indices would not have demonstrated that either index tracks animal outcomes. The validation the AWPRI needs is against inspection coverage, non-compliance rates and enforcement actions. The realistic source for those three quantities is the official controls reporting of the European Union, which covers roughly nine of these 24 countries. That validation has not been carried out, and no claim on this site rests on the AWPRI having been validated against welfare outcomes.

The One Event Study, and Its Limits

The published version described its findings as "quasi-experimental evidence". The findings were nothing of the kind. That design compared groups of countries that were never assigned to any treatment, over a window that opened after the policies the design named had already taken effect. The comparison has been rebuilt around the one instrument the panel can bracket in time, Regulation (EU) 2017/625 on official controls, which applied from December 2019. The United Kingdom is the treated country, and the specification carries eight pre-periods and four post-periods around a 2019 reference year.

The outcome is the composite score excluding the legislation variables. Excluding those variables is the point of the specification, because leaving the legislation variables in would test whether a regulation changed a count of regulations, which is mechanical. The result is a null. None of the eight pre-period coefficients and none of the four post-period coefficients differs from zero at the five per cent level. Two placebo specifications are reported alongside the main specification in data/final/event_study.csv. The first placebo drops the treated country and the second dates the event to 2020.

A null over four post-periods on 24 countries is weak evidence of anything. The null is reported because the alternative is to report nothing and leave the earlier causal language standing. The United Kingdom case study in the published paper was presented as showing the index "performs as designed"; a single country moving in an expected direction over a window that contains no identified intervention shows nothing of the kind, and that claim has been withdrawn.

What Changed From the First Version

The published index had 15 variables over 25 countries and 19 years. Three variables were withdrawn outright for having no usable source, and those three variables are listed below.

Four further changes were made. The V-Dem animal rights variable and the year-on-year change in that variable were withdrawn because v2carig_ord measures the rights of civil and political groups, not animals, and were replaced by counts of animal welfare instruments in FAOLEX as a stock and a five-year flow. The plant protein variable was replaced by the animal share of protein supply and moved into the exposure layer, since it measures how much of a diet comes from animals, and not any policy direction. Layer 3 was rebuilt on OpenAlex work counts with annual GDP and population denominators, in place of the previous mixture of hand-typed dictionaries and backward extrapolation. Layer 2 was renamed from "Policy Trajectory" to "Governance capacity", because the layer holds a cross-section of institutional conditions in a given year, and calling a level a trajectory invited the reading that a rising Layer 2 score meant policy was improving. Where this site shows a trajectory, the word means the ordinary least squares trend of the composite score over time, and the page says so at the point where the trend is shown.

The headline claims of the published paper do not all survive. The claim that governance dominates risk rested on the random forest described above. The 51.6 per cent first principal component, the k = 4 archetypes and the ARIMA forecasts were all computed on the 15-variable panel and are not carried forward. Current figures are on this site and in the repository, and the arXiv paper describes the first version.

Data Sources

Google Trends and PATSTAT/LENS appeared in the published version's source table and supply nothing to this index; the variables they were said to support have been withdrawn. V-Dem is used under CC BY 4.0; FAOSTAT and FAOLEX are open access; OpenAlex is CC0. The assembly code, the codebook and every panel file are in the repository, and the panel can be rebuilt end to end from the raw sources.

Citation

That paper describes the first version of the index. The construction documented on this page is the revision, and where the two disagree this page is current.