🔍 What Was Tested
This evaluation assesses how well multilevel regression and poststratification (MRP) can reconstruct country-level estimates from pooled Eurobarometer surveys when full country-representative samples are treated as the truth.
🧩 How the Test Worked
- Repeatedly drew subsets from the full Eurobarometer sample across EU member states.
- Applied MRP using census covariates to produce adjusted country means from each subset.
- Compared those MRP estimates to the "true" country means computed from the complete Eurobarometer samples.
- Conducted the procedure for ten survey items drawn from various Eurobarometer waves.
📈 Key Findings
- MRP typically produces estimates that are highly correlated with the true values (mean correlation = 0.90).
- Despite high correlation, MRP is less capable of accurately reconstructing the relative rankings of countries and of capturing the full plausible range of individual country means.
- The majority of the adjustment arises from the modeling of country (state) means rather than from the poststratification weighting step.
- Samples pooled with population-weighting perform no worse than samples constructed with equal country shares of respondents.
💡 Why This Matters
- MRP is a reliable tool for recovering average country-level attitudes from pooled surveys, but caution is warranted when using it to infer precise country rankings or extremes.
- Emphasis should be placed on how state means are modeled; poststratification and different pooling schemes appear less decisive in these tests.
