Revisiting the trade effects of the euro: data sources and various samples

Abstract

New evidence in the literature on trade effects of the euro often reports different estimates. In this paper, I investigate the impact of trade data, instead of methodology, on the estimation of the key coefficient. In particular, I apply both the log-linearized least squares (OLS) estimator and the Poisson pseudo-maximum likelihood (PPML) estimator to the structural gravity model and compare these estimates by using trade data from two of the most widely used sources (IMF DOTS and UN Comtrade) and by varying samples. One surprising result is that the OLS estimator yields coefficients of the euro with opposite signs for the two data sources, when a sample covering all countries is applied. It is as expected that the PPML estimator is much less sensitive to sample size than the OLS estimator, taking a data source as given. However, the variation in estimates caused by data sources and sampling is consistent for both estimators. It indicates that both estimators are not free from the measurement error issue. More findings include: (1) the discrepancy in OLS estimates derived for the two datasets persists across samples, but the magnitude varies; (2) the magnitude of the discrepancy in PPML estimates from the two datasets is more stable to sampling; (3) both OLS and PPML estimators are sensitive to sample compositions for a given sample size.

Appendices

Appendix A: Data preparation and summary statistics

The raw data for the period of 1992–2015 are downloaded from the official websites of both datasets, the Direction of Trade Statistics²² and the UN Comtrade. Unless indicated otherwise, I use the import data reported by importers to mirror the export data of exporters, which is reported at the CIF (cost, insurance, freight) price. After excluding trade unit defined at more aggregated level like “World,” “Euro Area,” “Africa” or “Special Categories,” etc., the DOTS raw data cover 216 exporters, 215 importers while the Comtrade covers 251 exporters, 204 importers for the mirrored export data.

To prepare for the final dataset, I first rename countries with different names in the two datasets. Then, I drop observations for countries/territories pairs covered only by one dataset. Table 7 lists those countries/territories and the mean of export when the corresponding countries are the exporter. Figure 6 shows the share of export of the dropped observations to the total export for Comtrade.

Table 7 Countries/territories only covered by one dataset

Fig. 6

Source: UN Comtrade, CIF value

The export share of dropped observations to total export

During the period of 1992–2015, a few countries have experienced unification or independence. In the third step, I process the statistics for those countries. Table 8 presents the summary statistics for the observations of them. Following Glick and Rose (2016) and Mika and Zymek (2018), I treat Belgium and Luxembourg as one unit, as there were no separate statistics for them before 1997. Similarly, I aggregate Serbia, Rep. of and Montenegro into one unit, as they were only not in one union until 1996. For simplicity, I drop observations if one trade partner is Czechslovakia. From 1993 on, the statistics are reported separately for Czech Republic and Slovakia.

Table 8 Statistically troublesome countries/territories

Figure 7 presents total export values contained by the two datasets, both in FOB (Freight on Board) and CIF (Cost, Insurance and Freight)²³ prices. As CIF price is usually higher than the corresponding FOB price, we should expect a higher value of exports measured by CIF price than that measured by FOB, no matter in which dataset. However, we only see this gap in Comtrade. Moreover, exports measured by CIF and FOB in DOTS are almost identical, both of which have been slightly more than CIF exports in Comtrade since 2002. It might be due to the projection procedure for statistics in DOTS while not in Comtrade. And there is very likely to be an overestimate of export value reported by exporters (FOB) in DOTS (Tables 9, 10, 11; Figs. 7, 8, 9, 10, 11, 12).

Fig. 7

Source: DOTS and Comtrade, CIF value in US$

Total export value in the two datasets. Note The values are based on observations which exist either in DOTS or Comtrade after the data cleaning process

Table 9 Frequency of country pairs with given number of high gap observations

Table 10 Country pairs with most frequent high relative gaps

Table 11 Country pairs with most frequent high absolute gaps

Fig. 8

Countries most frequently reported trade data with high relative gaps. Note High relative gap means that this gap measure is bigger than 1, different from the threshold in Table 10. The frequency is calculated at country level for the year of 2015. It is the number of observations that a given country, as importers, reported with high relative gaps. To save space, countries with frequency less than 4 are not included in the figure. These countries include Cameroon, United Kingdom, Azerbaijan, Macao, Paraguay, Serbia and Montenegro, Algeria, Australia, Brazil, China, Egypt, Georgia, Guatemala, India, Israel, Japan, Peru, Uruguay, Zambia. In total, there are 956 observations with high relative gap in their trade statistics for DOTS and Comtrade

Fig. 9

Countries most frequently reported trade data with high absolute gaps. Note High absolute gap means that this gap measure is bigger than 1 billion US dollars, different from the threshold in Table 10. The frequency is calculated at country level for the year of 2015. It is the number of observations that a given country, as importers, reported with high absolute gaps. Denmark, Italy, Japan, Lao PDR, Oman, Russian Federation, Sudan, Sweden, Zimbabwe also reported trade statistics with absolute gap more than 1 billion US dollars in 2015, but only for one trade flow

Fig. 10

Top 30 countries with the highest ratio of frequencies reported trade data with high relative gaps. Note “Ratio of frequencies” is the ratio of frequency with high relative gaps as defined in Fig. 8 to total number of observations that a country reported as importers in 2015

Fig. 11

Top 30 countries with the highest ratio of frequencies reported trade data with high absolute gaps. Note “Ratio of frequencies” is the ratio of frequency with high absolute gaps as defined in Fig. 9 to total number of observations that a country reported as importers in 2015

Fig. 12

Number of observations with high gaps in trade data across time. Note The left y-axis displays the number of observations with high relative or absolute gaps in trade data. The right y-axis displays the number of those observations as a share of total number of observations for a given year. High relative gap means that this gap measure is bigger than 10. High absolute gap means that this gap measure is bigger than 464 millions US$

Appendix B: A description of EU countries

See Tables 12, 13 and Fig. 13.

Table 12 Population (thousands person, 2016)

Table 13 GDP 2016, gross domestic product at market prices, current price (billions euro)

Fig. 13

Source: DOTS, CIF

The ratio of export to emu initial members. Note Export share is the ratio of export value to 11 EMU initial members to total export of each country. “Followers” refers to EU countries which joined EMU after 1999. This group includes Greece, Slovakia, Slovenia, Estonia, Cyprus and Malta (ranked by the nominal GDP in 2016, the same in the following). The 11 EMU “initial member” countries are Germany, France, Italy, Spain, Netherlands, Belgium, Austria, Ireland, Finland, Portugal, Luxembourg. Belgium and Luxembourg are treated as one unit, as there is no data for the separate countries before 1999. “Non-EMU” countries include the rest of EU countries. As Lithuania and Latvia joined EMU only in 2014 and 2015, I group them into Non-EMU. The export share for EMU initial members is slightly overestimated since the denominator does not include the export to itself

Appendix C: Supplement materials

See Tables 14, 15, 16, 17 and Figs. 14, 15, 16, 17.

Table 14 Benchmark comparison with FOB value reported by exporters

Table 15 Summary statistics of export for the sample of EU countries

Table 16 Benchmark results for comparison: without gravity variables

Table 17 Comparison of coefficients for the sample of 28 EU countries

Fig. 14

Source: DOTS, Comtrade and Eurostat, CIF

Ratio of export among EMU initial members to export among EU countries. Note The 11 EMU “initial member” countries are Germany, France, Italy, Spain, Netherlands, Belgium, Austria, Ireland, Finland, Portugal, Luxembourg. Belgium and Luxembourg are treated as one unit, as there is no separate data for the two countries before 1999

Fig. 15

Sample restricted by the percentile of absolute gaps in statistics. Note The x-axis variable is the percentile of the absolute gap distribution in trade data for the two data sources, as specified across years. In each sample, 28 EU countries are always covered

Fig. 16

Distribution of the key coefficient over sample sizes: OLS, DOTS. Note “Size 1” means that except for 28 EU countries, extra 20 randomly selected countries from the rest of world are included as exporters or importers. Size 2–6 adds 20 more randomly selected countries gradually than the number of countries in the correspondingly previous sample. To get the distribution for each sample size, I implement the selection process for 1000 times for each size

Fig. 17

Distribution of discrepancy in OLS estimates over sample sizes. Note The discrepancy is constructed from OLS estimates derived for DOTS minus those for Comtrade for each random draw of a sample. “Size 1” means that except for 28 EU countries, extra 20 randomly selected countries from the rest of world are included as exporters or importers. Size 2–6 adds 20 more randomly selected countries gradually than the number of countries in the correspondingly previous sample. To get the distribution for each sample size, I implement the selection process for 1000 times for each size