The report applies a logistic regression model estimating the probability that an individual i surveyed in country c expresses trust in the national government, local government, the national legislature, or the civil service. Specifically, the following model is applied:
where Pr(=1) is the probability that individual i trusts the institution.
Trust is treated as a binary variable equal to 1 if the respondent scored 6 or above on a 0–10 scale (where 0 indicates no trust and 10 indicates full trust), and 0 if the respondent scored between 0 and 4. Therefore, trust is modelled as a binary indicator distinguishing low from high trust. The binary coding simplifies interpretation and increases statistical power compared to using an ordinal scale. The midpoint category (5) is excluded from the baseline specification, as it might ambivalence rather than a clear trust position.
The main explanatory variables Driversic capture citizens' perceptions of government performance across the core dimensions of the OECD Framework on Drivers of Trust in Public Institutions. These reflect respondents' perceptions of governments' responsiveness, reliability, openness, integrity, and fairness, and also encompass perceptions of government action on intergenerational and global challenges. Overall, 27 perception variables are included in the model for OECD countries, all of which are listed in the results table below. For the OECD accession countries, 22 variables are included because questions about 5 were not part of the OECD Trust Survey in Latin America and the Caribbean and were thus not collected for Brazil and Peru. To facilitate interpretation, these drivers are measured on a 0–10 scale and standardised to have a mean of zero and a standard deviation of one at the country level.
The vector Xic captures demographic, socio-economic, and political characteristics of each respondent that may also influence institutional trust. These controls include interpersonal trust (standardised with mean zero and standard deviation one), demographic characteristics (age and age squared; men and women); educational attainment (low, medium, and high), financial concerns at the household level, a sense of belonging to a discriminated-against group, and support for the governing party
Country fixed effects ωc control for time-invariant country characteristics. Probability weights are used to ensure national representativeness, and heteroskedasticity-robust standard errors are applied throughout.
Because the estimated coefficients β from a logistic regression describe effects on the log-odds of trust, they cannot be directly interpreted as changes in probabilities. To facilitate interpretation, results are reported as average marginal effects (AMEs). Equation 2 computes the AME of driver x by averaging, across all N respondents, the product of each individual's predicted probability of trust, one minus that probability, and the estimated coefficient x. For example, the AME for the perception that the government makes evidence-based decisions (Table A.1 A.4, column 1) can be interpreted as follows: a one standard-deviation increase in this perception is associated with a six-percentage point increase in the probability of trust (0.061 × 100 = 6.1), significant at the one per cent level.
Robustness is assessed through several alternative specifications. First, additional control variables are included, namely satisfaction with the health and education systems. Second, responses of 5 on the 0–10 trust scale are recoded as indicating low trust. Third, missing values in the driver variables are replaced with the weighted country-level mean for the respective driver, and a corresponding missingness indicator is included for each imputed variable. Robustness is also assessed using an OLS specification, although these results are not reported.
Overall, the results remain largely consistent across specifications. Similar patterns are also found in pooled regressions (not reported) using data from the 2023 and 2025 Trust Survey waves, although some differences emerge due to the more limited set of regressors available for the pooled analysis.
There are several caveats to this model that are worth noting. First, the inclusion of a large number of drivers and control variables leads to a substantial reduction in the sample size in the baseline regression, as individuals with at least one missing value on any of these variables are excluded from the regression. This issue is further compounded by the exclusion of individuals reporting neutral trust. As a result, the share of missing observations is high with 20 to 40% of observations. This concern is addressed by imputing missing values and by conducting a robustness check in which respondents in category 5 of the trust variable are reclassified as low-trusting individuals.
Second, the estimated coefficient signs should be interpreted conditionally on the full set of control variables included in the model. Some regressors exhibit negative coefficients, but these effects are generally small in magnitude and typically not significant at the 0.05 level, with confidence intervals that often include zero. As a result, there is insufficient evidence to determine whether the true effects are positive or negative. Moreover, these coefficients likely reflect partial correlations, capturing the effect of a given variable while holding many related factors constant rather than indicating a genuinely negative relationship in isolation.
Third, the regression results require careful interpretation. Statistical significance does not imply that the associated variables causally increase trust. All explanatory variables are correlated, and the relationship between trust and perceptions of public governance is likely to be reciprocal.
