Trade is unpredictable — mirror statistics

The so-called mirror statistics evidence such measurement uncertainties. Trade flows are in principle recorded twice, once as exports by the sending economy and once as imports by the receiving one. The IMF's Direction of Trade Statistics (DOTS) database,1 the most widely used resource for bilateral trade statistics in IR research, provides both figures. If they were very similar, “mirror statistics” would not raise significant questions. Yet, discrepancies in mirror statistics are large and persistent, even between countries with highly developed statistical systems. In our global sample, on average discrepancies are almost as large as trade volume estimates themselves.2 Dyads with modest trade are important drivers of these huge differences. But also for dyads that trade a lot, the differences remain substantial: among the dyads in the top decile in trade volume (more than USD 92.5 million in trade per year according to importing-country records), the median discrepancy, relative to the average of two mirror flows, is still 23.1 percent (with a mean of 40.2 percent). This “mirror problem,” as we shall call it, reveals the substantial uncertainty in trade statistics, challenging analyses of the character, origins, or implications of trade (cf. Morgenstern 1963; International Monetary Fund 1987; Schultz 2015; Linsi and Mügge 2019). Trade statisticians and economists have long recognized mirror discrepancies (Ely 1961; Morgenstern 1963; Bhagwati 1964, 1967; Yeats 1978, 1990; Gaulier and Zignago 2010). Some have proposed statistical remedies, such as estimating mirror averages weighted by inferred reporter reliability in the BACI3 or OECD Balanced International Merchandise Trade Statistics (BIMTS)4 datasets. While we recognize and build on these efforts, their coverage remains too limited for many IR analyses—most of which ignore trade data defects altogether.5 Using the most widely available trade statistics, often based on import values alone, most IR scholars implicitly trust those measures to be the “right” ones, discarding potentially valuable information from parallel export records. Yet, as we argue below, the assumption that estimates based on import records are more reliable frequently does not hold. Leveraging information from both sides of the mirror—rather than mechanically relying on only one side—can improve trade volume estimates and inferences drawn from them.