Lost in Aggregation: Improving Event Analysis with Report-Level Data

Abstract

Most measures of social conflict processes are derived from primary and secondary source reports. In many cases, reports are used to create event-level data sets by aggregating information from multiple, and often conflicting, reports to single event observations. We argue that this pre-aggregation is less innocuous than it seems, costing applied researchers opportunities for improved inference. First, researchers cannot evaluate the consequences of different methods of report aggregation. Second, aggregation discards report-level information (i.e., variation across reports) that is useful in addressing measurement error inherent in event data. Therefore, we advocate that data should be supplied and analyzed at the report level. We demonstrate the consequences of using aggregated event data as a predictor or outcome variable, and how analysis can be improved using report-level information directly. These gains are demonstrated with simulated-data experiments and in the analysis of real-world data, using the newly available Mass Mobilization in Autocracies Database (MMAD).