Why
Judges in most countries are not directly elected, yet courts still depend on public trust to function. Rulings carry weight only when the public accepts them, and contested elections are only resolved peacefully when the losing side accepts the court's verdict. The public is therefore the repository of judicial legitimacy, and tracking how this varies across countries and over time tells us where the foundation is robust and where it may be eroding.
The data
Millions of responses are pooled from every national and cross-national programme that asks about trust or confidence in courts. These include Gallup, the World Values Survey, the regional Barometers, the European Social Survey, and ISSP. The programmes use very different response scales, ranging from binary through four-point and ten- or eleven-point ladders to continuous composites. Before fitting, every response is collapsed to a yes/no question: did the respondent give an above-the-middle answer? That lets one model accept every programme without having to decide which scales are mutually comparable.
Responses also carry their survey's own design weights, so a fielding that over- or under-represents part of the population is corrected before it informs the estimate. The weights enter as an effective (Kish) sample size for each survey cell, so a heavily reweighted survey counts for less than its raw number of respondents.
The model
For each country we estimate a latent trust level that evolves year by year as a smooth random walk, following the dynamic latent-trait framework of Claassen (2019). The trusting answers to each survey item are modelled as a beta-binomial count, so a single noisy fielding carries less weight than a large, consistent one. Every survey item has its own difficulty and its own sensitivity to the underlying trust level, and each item-country pairing carries a small bias, which lets programmes on very different scales be pooled without assuming they are interchangeable:
Here yikt is the number of trusting answers out of sikt respondents to item k in country i and year t; θit is the country's latent trust level in that year; λk and γk are the item's difficulty and slope; δik is the item-country bias; and φ sets how far surveys scatter around the latent proportion. The numbers shown throughout the app put this latent scale on a probability by evaluating η at a common World Values Survey confidence anchor item (with its item-country bias set to zero), giving P(trust) = logit−1(λanchor + γanchor·θ). Fitting is by Hamiltonian Monte Carlo (NUTS) in NumPyro.
Claassen, Christopher. 2019. “Estimating Smooth Country–Year Panels of Public Opinion.” Political Analysis 27(1): 1–20. doi:10.1017/pan.2018.32.
Coverage
Loading…
The estimate rests mainly on the trust questions. The most common is the
generic question about trust or confidence in the courts in general
(j_trust_gen), alongside trust in the supreme court
(j_trust_sc) and the constitutional court
(j_trust_cc). It also accounts for other domestic-court
items, including judgements of how well the courts perform and whether
they are fair, independent, and even-handed. International and European
courts are left out, so the estimate stays about a country's own
judiciary.
Every answered item enters as its own observation, so a respondent who answers several court questions contributes several. Counting all of a person's answers, rather than keeping only the highest-priority one, uses more of the available evidence and sharpens how each item is calibrated, at the cost of treating one person's answers as independent.