Asynchronous Fieldwork in Cross-Country Surveys: An Application to Physical Activity

Multi-country surveys often aim at cross-country comparisons. A common quality standard is conducting these surveys within a common fieldwork period, across all participating countries. However, the rate the target sample is achieved within that fieldwork period in each country varies substantially. Thus, the distribution of the interview month often varies substantially in the final sample. This may lead to biased estimates of cross-country differences, especially if the variable of interest exhibit a non-constant trend over time. This paper aims at demonstrating when such a problem cause biased estimates of country differences in physical activity. We demonstrate the implications of such an asynchronous fieldwork in cross-country surveys, using the European Social Survey Round 7. Our analytical sample focuses on 6 countries with data collected between September 2014 and January 2015. We present results for modelling physical activity using regression analysis. We compare unadjusted and adjusted regression coefficients accounting for fieldwork month. Moreover, we present a set of different postestimation predictions obtained from such pooled cross-country analyses. We found that physical activity varies across interview month, with the highest activity reported in September, decreasing thereafter, reaching the lowest level in January. Thus, countries with more observations during autumn were upward-biased, compared to countries with more observations during winter. Our results demonstrate how comparisons between countries are affected when interview month is omitted. This is prevalent using both unweighted and weighted regression techniques. Studies using pooled samples of cross-country surveys are commonplace. While a common fieldwork period accounts for severe biases in country comparison, often the bias remains when the outcome of interest has substantial seasonal variation. Our study suggests how accounting for interview month in analyses is an easy way to mitigate this problem.