Welfare is the largest expenditure category in all advanced democracies. Consequently, much literature has studied partisan effects on total and policy-specific welfare expenditure. Yet, these results cannot be trusted: the methodological standard is to apply time-series cross-section regressions to annual observation data. But governments hardly change annually. Thus, the number of observations is artificially inflated, leading to incorrect estimates. While this problem has recently been acknowledged, it has not been convincingly resolved. This article proposes mixed-effects models (also known as 'multilevel models' or 'hierarchical models') as a solution, which allows decomposing variance into different levels and permits complex cross-classification data structures. It is argued that mixed-effects models combine the strengths of existing methodological approaches while alleviating their weaknesses. Empirically, partisan effects on total and on disaggregated expenditure in 23 OECD countries in the period 1960-2012 are studied using several measures of party preferences and revealing several substantially relevant findings.