Political freedom, education, and value liberalization and deliberalization: A cross-national analysis of the world values survey, 1981-2014

ABSTRACT

Since World War II, value liberalization has been a worldwide trend, but more recently, there has been a resurgence of conservativism. Modernization and cultural theories have difficulty explaining the shift to liberal or illiberal values, but the political environment, an underestimated contextual factor, could shed light on the mechanisms driving it. This study used hierarchical linear modeling on all six waves of the World Values Survey data (1981–2014) to demonstrate that political freedom helps to explain the rise and fall of liberal attitudes when controlling for societal affluence, inequality, and cultural backgrounds. It finds that political freedom conditions the effects of education: although education is usually considered a liberalizing force, its effect is much weaker in non-free than free societies. The findings remind us of the importance of a free political environment for a robust democracy and point to the complex nature of educational effects, namely the ability of education to socialize people in liberal or illiberal directions depending on the political context and the regime’s agenda.

KEYWORDS:

• Political Freedom
• Education
• Liberal values
• Democratization
• Modernization theory

Highlights

• Political freedom is associated with liberal attitudes;

• Political freedom moderates education’s ostensibly liberalizing effect on attitudes;

• Reduced political freedom and regimes’ agentic control of education systems could lead to value deliberalization.

Notes

¹ Multiple imputation was executed with the R statistical package Amelia II (Honaker, King, & Blackwell, 2011). The Amelia II package employs EMB (expectation maximization with bootstrapping) methods in estimation. The literature disagrees on the optimal number of imputations. Hershberger and Fisher (2003) believe five multiple imputations are not enough, and more (potentially hundreds) is ideal. Yet Von Hippel (2005) argues that using five to ten is more than sufficient and does not cause considerable loss in precision. The study adopted m = 5 as the final number for multiple imputations. More (m = 10, 20) were tested, but the differences in their estimates were trivial.

² After multiple imputations, there was no missing information for the individual-level predictors, such as age, gender, marital status, occupation, and educational attainment. However, there was some missing information for the dependent variables and national-level statistics, and these data should not be imputed. Fortunately, the missing rate was not high for the dependent variables. More details on the numbers of observations and descriptive statistics are in Table 2.

³ For instance, the UK was surveyed in 1998 and 2005. In 1998, the UK had a GDP per capita of 33344.01 dollars (measured in constant 2005 US dollars), a Gini coefficient of 34.36%, and a Freedom House Index of 1.5. In 2005, the three numbers were 39934.78 dollars, 34.88%, and 1.0, respectively.

⁴ The coding by Freedom House needs some attention in the interpretation of results, as it uses smaller values to represent more freedom and larger values to represent less freedom. However, the study did not change the direction of coding to respect the original coding and to ensure consistency with other studies using the same dataset.

⁵ For this variable, a multi-level multinomial logistic regression model would be more appropriate. However, to simply the discussion and table presentation, the study treated it like the other three dependent variables. When fitted with both types of models, the main findings, especially the interaction effects addressed by this study, were consistent. Details of codes and results can be requested from the author.

⁶ To ensure comparability across countries, the study calculated Cronbach’s alpha for the scale of each country-year (Alemán & Woods, 2016). Of the 225 country-year observations, 178 (79.1 percent) had alpha values at or above the conventional 0.7 cut-off point (Nunnally, 1978). Another 45 country-year observations had alpha values between 0.6 and 0.69. Only two cases (0.9 percent of country-years) – Venezuela in 2001 and Tanzania in 2000 – had alphas below 0.6. The study retained the country-year observations with a low Cronbach’s alpha for the following reasons: first, they represented a small proportion of the observations; second, a comparison of the regression modelling results before and after the removal of cases revealed only trivial differences; third, psychometricians consider 0.6 a permissible cut-off for scales with fewer than ten items and items with fewer than seven response options if the scales are valid and theoretically justified, as was the case here (Loewenthal, 2001).

⁷ Details and codes available upon request.