ABSTRACT
This paper uses textual analysis to examine how European corporations assess sanctions in their annual reports. Using observations from a panel of almost 11,500 corporate annual reports from 2014 to 2017, we document significant cross-country variation in how firms perceive Russia-related sanctions, even after controlling for many firm-level characteristics, sectoral differences, and time trends. The cross-country differences also remain for sentiments about sanctions and contexts in which sanctions are mentioned. We also examine the role of macroeconomic linkages in explaining these differences. We show that Russia’s inward and outward FDI stocks and high levels of imports and exports with Russia explain about half of the cross-country variation, leaving a nontrivial share of variation unexplained.
Disclosure statement
No potential conflict of interest was reported by the author(s).
Notes
1. From the legal point of view, the EU sanctions only prohibit entities based in the EU from engaging in specified economic activities with Russian counterparts. The US measures, in contrast, are extra-territorial in nature.
2. For analysis of sanction effects on the Russian oil sector, see Mitrova, Grushevenko, and Malov (Citation2018); for the gas sector, see Sun (Citation2020); and for Russia’s defense sector, see Juola et al. (Citation2019).
3. Correspondingly, a search window too long may cause false positives. However, our results are not sensitive to the choice of a 413-word search window for Mention. As a robustness check, we replicated our main results using a narrower 100-word window, which corresponds to the length of a median text paragraph in our sample. The results remain quantitatively unchanged, although the narrower search window resulted in increased false negatives.
4. Specifically, we use the canonical logit link function for the dependent variables Mention, First, and Pages. The log link function is used in the case of the TFIDF variable.
5. As a robustness check, we tried alternative covariance structures, including full and log-Cholesky, but found that the alternative specifications do not materially change our results. These results are available upon request.
6. As a robustness check, we estimated the MACRO specification using data on FDI stocks in 2012. The estimation results (available upon request) were essentially unchanged from using the 2017 data.
7. These results are not reported for the sake of brevity, but are available upon request.
8. Loughran and McDonald’s (2011) lexicon has been widely used to measure tone, for example, in newspaper articles/columns and corporate press releases. For an overview on textual analysis and use of alternative lexicons, see the survey by Loughran and McDonald (Citation2016).
9. Specifically, we estimate the SECTOR model specification of EquationEquation (1)(1) (1) but control for heterogeneity across firms with firm random effects instead of financial ratios.
10. Our analysis groups those countries with a few observations that share a common geographic or economic area. Lichtenstein, for example, is grouped with Switzerland and reported as CH. Gibraltar is grouped with Great Britain. Excluding the countries with an insufficient number of observations does not change our results.
11. Full estimation results of the GLME model (2) are not reported, but available upon request.
12. Results of cross-validation tests, perplexity, and log likelihood measures are available upon request.
13. Full estimation results of the GLME model in EquationEq. (3)(3) (3) are not reported, but available upon request.
14. Full estimation results of the GLME model in EquationEq. (4)(4) (4) are not reported, but available upon request.
15. As our dataset does not include information on the ultimate owners, we leave this question for future research.
16. These results are available upon request.