Skip to content

Measuring the impact of self-censorship on political party support in Afrobarometer data using machine learning

Measures of political party support are viewed suspiciously in hybrid and authoritarian regimes where the two core indicators of political support – elections and public opinion polling – cannot be readily trusted. More accurately, the degree of measurement error is unknown, though scholars presume support for incumbent leaders and parties is exaggerated. This paper scrutinizes the degree to which political opinions on Afrobarometer surveys can be trusted by re-estimating voting preferences for 34 African countries. First, a range of demographic and interview metadata are regressed upon vote choice to see if they systematically influence vote-preference responses. Next, key markers of potential self-censorship are identified and used to create pools of “clean” and “polluted” data. Clean data are used to train a machine learning model to predict voting intentions, and then all data with self-censorship markers are recalculated. The results show that ruling party support is broadly similar in most countries but exaggerated by more than 10 percentage points in Sudan and Uganda and by between 5 and 10 points in Burundi, Tanzania, Togo, and Zimbabwe. Declines in ruling party support due to self-censorship are concurrent with increases in both non-responses and opposition party support.