Does bilateral trust across countries really affect international trade and factor mobility?

Trade is often characterised by incomplete contracts as it is too costly to take into account or even know all contingencies when establishing them. Furthermore, it can be difficult to negotiate, monitor, and enforce contracts, especially in international trade where the commercial partners are established in different jurisdictions (Rodrik 2000). As a result, profitable trade opportunities might not be realised, unless the parties trust each other (Akerlof 1970; Arrow 1972; Putnam 1993; Greif 1993, 2000; Coleman 1994; Kollock 1994; Fukuyama 1995; Knack and Keefer 1997; Paldam 2000; Dyer and Chu 2003).

GSZ studied the effect of trust on international capital flows, in addition to trade flows, but did not consider international labour migration. Future research may attempt replicating the findings of GSZ with respect to international capital mobility.

On trust and institutions, see also La Porta et al. (1997), Alesina and La Ferrara (2000), Bjornskov (2006), Tabellini (2008, 2010), Bloom et al. (2009), and Aghion et al. (2010).

For a short overview of the origin of the gravity model and the corresponding literature, see Anderson (1979, 2011), Anderson and van Wincoop (2003), Baldwin and Taglioni (2006), Head and Mayer (2013), and Felbermayr et al. (2015).

Aggregated commodity export flows are taken from the UN Comtrade Database; http://​comtrade.​un.​org/​db/​default.​aspx. Unfortunately, data on trade in services cannot be included as it has only been collected since the year 2000. There are no zero-trade observations for the country sample included in the analysis.

See http://​zacat.​gesis.​org/​webview/​. In 1996, citizens of 17 European countries were asked to indicate the trust they had in citizens of 25 EU and non-EU countries.

These dyadic dummy variables come from the CEPII Gravity Data set generated by Head, Mayer and Ries (2010, 2013); see www.​cepii.​fr.

The measure is based on the data from www.​factiva.​com, which collects and archives information made available by over 30,000 newspapers, journals, magazines, Web pages, etc., on a broad range of contents from over 200 countries. It is constructed as follows: ‘In Factiva, we searched the newspaper with the highest circulation for each country. For each pair of countries i and j, we recorded the number of articles in the newspaper of countries that mentioned country j or its citizens in the headline. We divided this number by the number of total news stories on foreign countries’ (GSZ; p. 1106).

We employ the prices of shipping a 1,000 kg unspecified freight type load with no special handling in June 2011, as provided at http://​importexportwiza​rd.​com. This measure is based on Giuliano et al. (2006).

Data are drawn from www.​ethnologue.​com, also see Lewis et al. (2014).

See www.​kellogg.​northwestern.​edu/​faculty/​sapienza/​htm/​somaticdistance.​zip.

For some countries, it is very difficult to decide which trait prevails, especially when focusing on the different categories of cephalic index. For example, in Fig. 2, northern Germany falls into category 3, ‘79.0–82.9’, while the other half of Germany falls into category 4, ‘83.0–86.9’. Guiso et al. (2008a) do not indicate how they decide which one of these categories prevails in such situations. We partially succeed in replicating their somatic distances when we decide visually (based on Fig. 2) which trait covers a larger area and assume that it is the dominant characteristic. However, this procedure is somewhat arbitrary, especially when ignoring the distribution of the population. As the German population is approximately equally distributed, we would ignore the characteristics of half of the population if we arbitrarily decided that either category 3 or 4 prevails. To account for this, our two measures of somatic distance allow a country to be home to two categories of traits, depending on the distribution of the population. In the case of Germany, we find that the categories of cephalic index 3 (‘79.0–82.9’) and 4 (‘83.0–86.9’) roughly share the German territory and population. Therefore, we decide to assign it the score of 3.5. This measure is certainly not flawless, but it allows us to further explore the robustness of the results published in GSZ.

See Stanners and Bourdeau (1994) or www.​eea.​europa.​eu/​publications/​92-827-5122-8/​page008.​html for the chart on population density in 1989 and Roto (2011) or www.​nordregio.​se/​en/​Maps--graphs/​ for the population density in 2010.

See the online appendix for a correlation matrix.

Regarding the latter, in a sensitivity analysis (Sect. 3.4), we also employ other language-related measures provided by Melitz and Toubal (2014).

Compare to Table IV in GSZ (pp. 1116 f.).

The endogeneity of bilateral trust is verified with a control function approach (see Wooldridge 2010, p. 127) and an endogeneity test that is robust to arbitrary heteroscedasticity, following Baum et al. (2007). Both tests reject the null hypothesis that bilateral trust is exogenous at conventional levels.

The Kleibergen–Paap \(rk\) LM statistic is the efficient first-stage statistic used to verify the relevance of the instruments when non-i.i.d. disturbances are assumed. Rejection of the null hypothesis suggests that the model is identified, i.e. that the instruments are relevant.

In the presence of i.i.d. disturbances, weak identification problems are detected with the Cragg–Donald F-statistic, which is compared to the critical values published by Stock and Yogo (2005). However, in case of non-i.i.d. disturbances, the Kleibergen–Paap \(rk\) Wald statistic is the efficient statistic (Kleibergen and Paap 2006; Kleibergen and Schaffer 2007; Baum et al. 2007). So far, no critical values have been computed for this statistic, and in practice, it is usually compared to the threshold number of 10 recommended by Staiger and Stock (1997); see also Stock et al. (2002). As a robustness test, we compute the limited information maximum likelihood (LIML) estimates of all our 2SLS regressions and find that the bilateral trust coefficients change only slightly in their size and that the levels of statistical significance are identical to the IV estimates. The results of this sensitivity analysis are available in the online appendix.

See Anderson and Rubin (1949), Dufour (2003), and Chernozhukov and Hansen (2008) for a formal explanation of this alternative test and Angrist and Krueger (1991, 2001) for an application of this method. See also Baum et al. (2007) for an implementation of the Anderson–Rubin test in Stata. This test verifies whether the instruments are significant in the reduced-form equation of the dependent variable.

Only the coefficients on DtS trust, somatic distance, and religious similarity are reported. Complete tables including the estimates of the coefficients on the control variables are available in the online appendix.

For several country pairs, we did not manage to do so, which may explain the diverging results.

Melitz and Toubal (2014) emphasise that a measure of common linguistic roots based on the language trees provided by the Ethnologue is problematic, as it does not allow the comparison of languages that belong to different trees.

The regression results and the correlations are reported in the online appendix.

This is a special case of Bertoli and Moraga (2013) that was proposed by Ortega and Peri (2013).

Ortega and Peri (2013) do not allow for destination country fixed effects to vary over time.

To complete the data set, Ortega and Peri (2009, 2011) interpolate observations when the missing value is situated between two years for which the observations are available and compute the net immigration flows. They correct for the outflow of foreign citizens using the International Migration Database and the data set on emigration stocks for the years 1990 and 2000 collected by Docquier et al. (2007). However, these net immigration flows are less precise than the gross flows and only have a limited coverage.