Data and methods

A cross-national claim is only as good as the comparability of its measure. This page summarizes how the book builds its construct, what it links that construct to, and — at the end — what the design is not able to show.

The construct

Authoritarian Governance Preference is estimated from three items in the World Values Survey and European Values Study governance battery. Each presents a governance arrangement and asks the respondent to rate it on a four-point scale from very good to very bad.

Two items are reverse-coded so that all three point the same way: a higher score always means stronger preference for unconstrained, unaccountable authority. The battery contains other candidates — rule by experts, for instance — and they are deliberately left out. Endorsing technocratic decision-making is adjacent to AGP rather than constitutive of it: a citizen may want expert governance alongside democratic accountability rather than instead of it, and including a weaker indicator would have diluted the latent variable.

The measurement model is a single-factor confirmatory factor analysis with three ordinal indicators, estimated with the WLSMV estimator and fitted separately for every country-wave cell. AGP is defined as a single dimension of institutional evaluation — not general political dissatisfaction, not ideological conservatism, not populism, and not nationalist sentiment, none of which a high score implies on its own.

Making it comparable

Full scalar invariance across seventy-odd countries does not hold, and testing for it the usual way would mean discarding most of the sample to find a subset where it does. The book uses alignment optimization instead. Alignment starts from a configurally invariant model, lets loadings and intercepts vary freely by group, and then adjusts factor means and variances to minimize total measurement non-invariance — while reporting which parameters are non-invariant and by how much.

That gives a diagnostic rather than an assumption. Three criteria govern whether the aligned scores are trustworthy: non-invariant parameters below roughly 25 percent of the total, item-mean regressions on the aligned factor means above R² = 0.90, and correlations with simple country means above 0.95. Chapter 9 reports the diagnostics. The book is explicit throughout that cross-national comparisons on this measure are approximate rather than exact — that is the constraint the method is designed to accommodate, not one it removes.

What is being predicted

The outcome comes from V-Dem’s Episodes of Regime Transformation dataset: the onset of an autocratization episode, defined as a sustained deterioration on the Liberal Democracy Index meeting stated magnitude and duration thresholds. It is a deliberately conservative outcome. It excludes the subtler erosions — declining subnational democratic practice, press freedom narrowing without regime change, the gradual judicialization of political competition — which means a demand-side model trained only on onset events may understate how much AGP matters for pressure that never crosses the threshold.

The hazard model

Chapter 6 estimates a discrete-time hazard model on the at-risk country-year panel. Each at-risk country-year contributes one observation; the outcome is 1 in the year of onset and 0 otherwise.

The discrete-time form was chosen over a Cox model because it takes time-varying covariates naturally, needs no proportionality assumption, and yields predicted probabilities directly. Its known weakness is that it does not account for differences in time at risk across an unbalanced panel; Chapter 9 re-estimates with a complementary log-log link and with annual dummies in place of the linear trend.

The forecast

Chapter 7 projects forward to 2033 using an augmented specification that adds the Chapter 5 mechanism variables — insecurity, trust, emancipative values, and their interaction — plus polity age, to the Chapter 6 baseline. The parsimonious baseline was built to preserve power on a small linked sample; the forecast wants the full mechanism structure.

Projecting AGP forward is the hard part, and the book treats it as a structured approximation rather than a forecasting engine. The WVS/EVS gives at most seven wave observations per country over four decades, and most countries have three to five. Fitting time-series models to sequences that short produces weakly identified parameters and sampling variation that swamps any trend. So the projection estimates a simple linear trend per country and shrinks it toward the regional mean trend with empirical Bayes weighting, giving more weight to the country slope as the number of waves grows; countries with fewer than three observations take the regional trend in full. Intervals widen substantially beyond the last observed wave.

Insecurity, trust, and emancipative values have no annual coverage in the survey files, so annual proxies are built from administrative sources and calibrated to the WVS/EVS metric. Annual hazards convert to ten-year cumulative risk under absorbing-transition logic, and uncertainty comes from a 2,500-draw parametric bootstrap over the coefficient covariance and the projection intervals together. Three scenarios bound the range: a baseline, elevated insecurity, and institutional erosion, with a recovery scenario added in Chapter 8 to evaluate the ebb hypothesis.

The forecasts are scenario-conditional. They do not assert what will happen. They estimate the probability of onset under an explicit set of assumptions, and the assumptions are printed alongside the numbers.

Weighting, missing data, and samples

Design weights are applied in all descriptive and individual-level analyses — W_WEIGHT for Wave 7, pwght for within-country work on the joint file, gwght where the estimand is a population-weighted global figure. Country-level analyses use population-weighted country means, since re-applying the weight at the country level would double-count it. Item nonresponse on the AGP items runs 1 to 3 percent and is treated as missing at random; a country-wave cell with fewer than 50 valid AGP observations is dropped. Chapter 9 re-runs the primary complete-case results under multiple imputation.

What this design cannot establish

Chapter 10 sets out five limitations. They are reproduced here because a summary that dropped them would be a different book.

The instrument is constrained by the item set. The three WVS items were not designed to measure AGP. They capture a narrow slice: bypassing parliament, military rule, and rejection of democracy as a system. They do not measure tolerance for executive overreach inside a working democracy, preference for concentrated media power, or willingness to accept electoral manipulation when your own side benefits.

Causal identification is not achieved. The claim that AGP affects autocratization rather than correlating with its preconditions rests on lagged specifications, subsample analyses, and an instrumental-variables approach whose exclusion restriction is imperfect. Insecurity and trust are theoretically motivated predictors of AGP, but they are also causal mechanisms in their own right. What the analysis provides is consistent independent predictive information, and it should be read as that.

Survey coverage is uneven. Sub-Saharan Africa, parts of Southeast Asia, and several post-Soviet contexts have sparse or widely spaced wave coverage, so their AGP estimates rest on fewer cells and carry more uncertainty. The forecast is most reliable where coverage is dense — Western Europe, Latin America, parts of East Asia — and more speculative elsewhere, even though the sparse regions may include the highest-risk environments.

The micro-level updating process is not modeled. The analysis establishes correlations between insecurity, trust, and AGP, and infers ebb dynamics at the aggregate level. It does not trace how an individual’s prior orientations update after exposure to improved economic security or institutional performance. That needs longitudinal panel data in which the same people are re-interviewed across shocks.

The outcome measure is coarse. ERT onset is a discrete threshold event. Continuous V-Dem component indices — clean elections, freedom of expression, judicial constraints on the executive — would show how demand-side pressure translates into institutional change across the full severity spectrum rather than only at the point where it becomes an episode.

These are the terms on which the book asks to be read. Every estimate behind them can be reproduced.

• Replication materials and data sources