Information and Communication Technology (Evidence from Turkey) Part 2

5. EMPIRICAL RESULTS We first present the descriptive statistics and the correlation matrix for the variables of interest in Tables 3-6. The correlation matrices indicate that the IDI index and three ICT subindices have positive relationships with export and import variables. As a visual inspection, we also provide the plots of export and import versus each ICT indices in Figs.1-2. There appears a positive relationship between each ICT indices and export/ import (see Figs.1-2). Table 7 reports the results of the export model. For comparison purpose, we also present the results of pooled OLS model along with those of time fixed effects model. We conduct the Hausman (1978) test to make a selection between random effects and fixed effect models and the results favored the fixed effects model.3 Further, the F test results indicate the presence of both time effects and country (individual) fixed effects for both models.4 We estimate four regressions, one for each ICT indices. The first regression (col.1) uses IDI index, the second (col.2) regression employs ICT access index, the third (col. 3) regression utilizes ICT use index, and the fourth (col. 4) regression includes ICT skills index as the ICT variable5. 3 However, it is not necessary to conduct a Hausman test since the countries in our sample were not randomly selected, i.e. they are the most traded partners of Turkey. Also, country and time effects are mostly accepted as fixed instead of random in gravity models (see Shepherd, 2012). 4 F tests results are available upon request from the corresponding author. We did not include country fixed effects in our main models because including country effects leads to unexpected coefficient results for binary dummy variables as these dummy variables may pick up the some impacts of the country fixed effects. We estimate one-way fixed effects models including only time fixed effects. However, in the sensivity analyses, we also included country fixed effects and the impacts of ICT indices on trade remain unchanged. Besides, due to dummy variables in the models, we applied “Least square dummy variables (LSDV)” approach to estimate the fixed effects model instead of using a fixed effects-Within estimator. 5 We also estimated robust standard errors that are corrected against to autocorrelation and heteroscedasticity problems by using approaches suggested by Driscoll-Kray (1998) and Beck and Katz (1995). The results of these approaches are also available upon request from the corresponding author. All variables were mostly found significant at 1% level in these procedures, as well. Variables 1 Pooled .367a ( 0343) .344 a (.0421) -.634 a (.062) .845 a (.129) .888 a (.119) -1.478 a (.396) -.516 a (.107) .296 a (.065) -.520 a (.060) -1.111 a (.382) 35 GDP Population Distance ICT Border Language Colony RTA Landlocked Island Number of country Obs. 513 Time effects ----Adjusted R0.711 square HausmanTest Table 7: Export Model Estimation Results 2 3 Fixed Pooled Fixed Pooled Fixed time time time effects effects effects .278 a .356 a .326 a .406 a .353 a (.060) (.035) (.054) (.031) (.045) a a a a .423 .346 .373 .320 .381 a (.070) (.043) (.064) (.042) (.062) -.595 a -.615 a -.609 a -.693 a -.722 a (.064) (.063) (.065) (.061) (.065) .775 a .930 a 1.135 a .287 a .468 a (.396) (.139) (.343) (.043) (.131) .904 a .889 a .890 a .906 a .927 a (.120) (.119) (.122) (.119) (.123) a a a a -1.539 -1.484 -1.490 -1.477 -1.501 a (.395) (.395) (.399) (.397) (.398) -.482 a -.524 a -.521 a -.540 a -.536 a (.110) (.108) (.111) (.107) (.110) .276 a .287 a .281 a .287 a .255 a (.066) (.064) (.065) (.068) (.065) -.528 a -.532 a -.534 a -.534 a -.540 a (.062) (.061) (.063) (.060) (.063) a a a a -1.079 -1.122 -1.117 -1.129 -1.115 a (.372) (.379) (.377) (.378) (.382) 35 35 35 35 4 Pooled Fixed time effects .414 a .415 a (.036) (.032) a .304 .294 a (.041) (.042) -.602 a -.593 a (.064) (.060) 1.099 a .924 a (.248) (.286) .859 a .860 a (.125) (.124) a -1.420 -1.438 a (.399) (.405) -.499 a -.508 a (.114) (.113) .364 a .347 a (.067) (.064) -.465 a -.473 a (.068) (.063) a -1.064 -1.077 a (.385) (.388) 35 35 513 Yes 0.715 513 ----0.700 513 ----0.711 29.14 a [0.000] Notes Cluster-robust standard errors are reported in 513 Yes 0.714 21.26 a [0.000] a, parentheses. 513 ---0.712 513 Yes 0.715 22.92 a [0.000] b, c 513 Yes 0.711 46.54 a [0.000] indicate 1%, 5% and 10% significance levels 2 respectively. χ Statistics are reported for Hausman test, and its probabilities are provided in brackets. All regressions include a constant term; however, we do not report it. In Hausman test, fixed effects model is consistent under the null and alternative hypotheses, while random effects model is inconsistent under the alternative hypothesis and efficient under the null hypothesis. Table 7 shows that the results from the models of pooled OLS and fixed effects are consistent with each other. ICT indices, as the primary independent variables of this study, have significant and positive impacts on Turkish exports in all regressions. Therefore, this result corroborates the hypothesis of this study that ICT decreases trade-related costs, particularly more important in international trade. A 1% increase in IDI index, ICT access index, ICT use index, and ICT skill index, respectively, appears to raises Turkish export around by 0.7%, 1.1%, 0.4%, and 0.9% respectively. Among the three sub- ICT indices, ICT access seems to have the strongest impact on Turkey’s export, while ICT use index appears to have the smallest effect. This result shows that ICT infrastructure and individuals’ access to basic ICTs are so crucial, while ICT intensity and usage are less necessary to spurring important for galvanizing Turkey’s export. GDP mass variable, as expected, have significant and positive impacts on the bilateral export of Turkey in all four regressions. Therefore, as stated in gravity model, economic size of country pairs is a boosting factor for export. Population mass, as a proxy for country size, affects Turkey’s exports to these countries positively and significantly. This result seems to indicate that population growth leads to higher domestic production and in turn higher export by increasing labor supply in Turkey. The distance between Turkey and its trading partners, as expected, affects Turkey’s export negatively. The larger the distance between Turkey and its partners, the lower is the export from Turkey to those countries. With respect to the results of the binary dummy variables, Turkey appears to export more to its border countries. However, sharing a common language seems to have a negative and significant impact on Turkey’s exports.6 This result is not an odd one as Turkey does not have a common language with its most importers such as EU member countries. As regard the colonial past, the results indicate that Turkey exports less to its imperial partners such as Egypt, Bulgaria, Syria, Libya, Cyprus, and Tunisia. In gravity model, the colonial heritage is expected to have a positive and significant impact on bilateral trade. However, there are also some studies that found the destructive impact of colonial past on trade. For instance, Head et al. (2010) obtained 6 In our sample, Cyprus and Bulgaria are the only two importers of Turkey that share a common language with Turkey. The impact of language may be biased due to this case. that, on average, trade between a colony and its colonizer declines by about 65% during the first 40 years of independence as the result of the deterioration of business networks or similar institutions. Also, independence may boost trade costs and lowers trade between metropole and its dependencies by terminating some specific formal relationships such as preferential trade agreements and informal connections such as business networks (Lavallee et al., 2012). Among other dummy variables, RTA appears to have, as expected, a positive and significant impact on Turkish exports to its trading partners. In particular, the Customs Union agreement, which went into force between Turkey and EU members in January 1996, has reinforced the trade linkages. EU member countries are now the largest trading partners of Turkey. Landlocked and island dummies have negative impacts on bilateral export of Turkey as expected. This result implies that Turkey exports less to its landlocked and island trading partners due to higher transportation cost. In Table 8, we report the results for the import model. For comparison purpose, we also report the results of the pooled OLS models, and as in the export models, we run four regressions, one for each ICT index7. The four ICT indices affect the bilateral import of Turkey positively. For instance, in the fixed effects model, a 1% rise in IDI index, ICT access index, ICT use index, and ICT skills index lead to more import around 3.3%, 2.6%, 0.8%, and 2.7%, respectively. Among ICT-sub indices, ICT skills index has the largest impact while the ICT use index has the least effect on Turkish bilateral import from its trading partners. ICT skills index reflects the outcomes of more efficient and effective ICT use such as higher secondary and tertiary gross enrollment rates. ICT skills index appears to be one of the pushing factors for 7 We also corrected standard errors against to autocorrelation and heteroscedasticity problems by using approaches suggested by Driscoll-Kray (1998) and Beck and Katz (1995). In these estimations, most variables were found significant at the 1%level along with only a few exceptional cases. These results are available upon request from the corresponding author. Turkey’s imports. However, the level of use of ICTs in the society (ICT use index) seems less important for Turkish imports compared to other ICT indices. The GDP mass as a proxy for economic size has a significant and positive impact on Turkey’s bilateral import, as expected. The population mass variable as a measure of country size raises Turkey’s import. The growing population implies more consumers and more demand for imported goods. Distance, as in the export model, has a negative and significant impact on Turkey’s import. Thus, Turkey imports fewer goods from its distant exporters due to high transportation costs. With respect to the binary dummy variables, border variable is significant and positive, reflecting that Turkey imports more from its neighboring countries because sharing a border decreases transportation costs. Contrary to the export models, sharing a common language seems to impact Turkey’s import positively as theoretically expected.8 The pastcolonial heritage appears to have an adverse and significant impact on Turkey’s import as in the export models, indicating that Turkey imports less from its exporters with which it shares a colonial past. The explanations for this result might be the same as those for the export models. As for RTA, a negative and significant result is obtained and it is contrary to our expectations. If there is an RTA, there should be more trade between the trading partners who happen to be members of the RTA. However, this result may have been driven by the fact that RTA has a high negative correlation with the distance dummy. When we exclude the distance variable from the import model, the impact of RTA turns to be affirmative and significant as expected. Thus, distance seems to have picked up some of the effects of RTA on trade.9 Finally, regarding the 8 In our sample, Bulgaria is the only exporter of Turkey that has a common language with Turkey. Therefore, this case may cause biased results. We also excluded language dummy and re-run the regressions and the results didn’t change. 9 There is a -0.6635 negative correlation between distance and RTA. In the import model, the most trading partners of Turkey are EU member countries, In other words. They are located in a specific region, Europe, and there is a particular physical distance between Turkey and them. Also, these countries have a customs union agreement with Turkey. Therefore, it is reasonable to be a correlation between RTA and distance. When we exclude RTA from the effects of landlocked and island binary variables, Turkey appears to import less from its landlocked and island exporters due to higher transportation costs. Overall, the baseline models for both export and import indicate that all ICT indices affect bilateral trade between Turkey and its trading partners positively. Furthermore, ICT has a stronger impact on Turkey’s imports than its exports because the coefficients of ICT indices are higher in the import models. Besides, ICT use index has the least impact on both bilateral exports and imports of Turkey. However, ICT access index has the largest impact on exports while ICT skills index has the strongest role on import. From these results, we may conclude that access to ICT is highly crucial to improve Turkey’s exports to its main trading partners. import model, the coefficients of all variables remain unchanged and stable, too. We also check the correlation between RTA and distance for the export model; however, there is only a small correlation, with a value of -0.1631 Table 8: Import Model Estimation Results Variables 1 Pooled GDP Population Distance ICT Border Language Colony RTA Landlocked Island Country Time effects Obs. Adjusted Rsquare Hausman Test .556a (.035) .270 a (.037) -.956 a (.071) 1.056 a (.112) .725 a (.089) 1.270 a (.114) -1.568 a .106 -.321 a (.103) -.218 a (.080) -.519 a (.118) 34 ---510 0.767 2 Fixed time effects .201 a (.067) .646 a (.072) -.897 a (.066) 3.330 a (.362) .937 a (.090) 1.025 a (.121) -1.610 a (.103) -.339 a (.101) -.131 (.080) -.549 a (.087) 34 Yes 510 0.806 21.16 a (0.000) Pooled .537 a (.034) .276 a (.037) -.930 a (.072) 1.195 a (.127) .693 a (.087) 1.266 a (.121) -1.566 a (.103) -.311 a (.105) -.248 a (.080) -.509 (.112) 34 --510 0.770 3 Fixed time effects .320 a (.053) .493 a (.055) -.866 a (.069) 2.640 a (.298) .775 a (.089) 1.132 a (.136) -1.583 a (.103) -.307 a (.105) -.231 a (.079) -.522 a (.089) 34 Yes 510 0.793 Pooled .621 a (.034) .221 a (.036) -1.046 a (.075) .345 a (.042) .743 a (.089) 1.276 a (.116) -1.589 a (.109) -.373 a (.107) -.257 a (.080) -.543 a (.108) 34 ---510 0.762 26.07 a (0.000) Notes: Cluster-robust standard errors are reported in parentheses. significance levels, respectively. χ 2 Fixed time effects .476 a (.045) .393 a (.051) -1.126 a (.078) .819 a (.107) .903 a (.089) 1.119 a ( .119) -1.648 a (.108) -.454 a (.110) -.261 a (.080) -.584 a (.085) 34 Yes 510 0.781 4 Pooled Fixed time effects .594 a .481 a (.035) (.042) .254 a .364 a (.035) (.051) -.877 a -.793 a (.039) (.063) a 1.630 2.724 a (.070) (.327) .666 a .693 a (.187) (.100) a 1.395 1.367 a (.104) (.114) -1.510 a -1.503 a (.113) (.119) a .-.250 -.209 b (.122) (.098) -.0718 .0352 (.105) (.085) -.517 a -.482 a (.083) (.106) 34 34 --Yes 510 510 0.756 0.782 16.84 a (0.002) 20.57 a (0.000) a, b, c indicate 1%, 5% and 10% Statistics are reported for Hausman test, and the probabilities are provided in brackets. All regressions include a constant term; however, we did not report it. In Hausman test, fixed effects model is consistent under the null and alternative hypotheses, while random effects model is inconsistent under the alternative hypothesis and efficient under the null hypothesis. 5.1 SENSIVITY ANALYSES 5.1.1 System GMM Estimation To check the robustness of the results, we first apply a different estimation method, the System Generalized Method of Moments (GMM), developed by Arellano and Bover (1995). GMM is particularly useful for correcting the endogeneity problem. There could be an endogeneity problem because ICT and export/import might be determined simultaneously. Furthermore, the causality may run in the reverse direction. That is, export/import may also stimulate ICT use. Besides, there may be additional reverse causality running from other control variables such as GDP and population to export/import. Finally, changes in ICT indices (and other explanatory variables) and export/import may be caused by a third factor that is not observable (see Liu and Nath, 2016). In the related literature, different solutions have been offered to address the endogeneity problem. For instance, Freund and Weinhold (202) use twicelagged internet variable as an independent variable in the regression. Clarke and Wallsten (2006) utilize the regulation of data services as an instrument for the Internet use and apply two-stage least squares. Vemuri and Siddiqi (2009) apply an instrumental variable approach of Hausman and Taylor (1981). Besides, as in this paper, GMM approach is also suggested as a solution to the endogeneity problem (see Choi, 2010; Liu and Nath, 2013; Lin, 2015). There are two different GMM approaches: Difference GMM (Arellano and Bond, 1991) and System GMM (Arellano and Bover, 1995). Arellano and Bond transform all regressors by differencing and use the Generalized Method of Moments (Hansen 1982). However, ArellanoBover estimator assumes that the first differences of instrumental variables are uncorrelated with the fixed effects, and thus it augments the Arellano-Bond approach. As stated by Roodman (2006), the System GMM approach may improve efficiency by using more instruments. It includes a system of two equations, i.e. an original equation and a transformed one. Along with the System GMM estimator, we also use one-year lag of ICT indices in the fixed effects model as another solution to the endogeneity problem. Their results are presented in Table 9.10 Table 9: Results from the System GMM and the Fixed Effects Model with One-Year Lagged ICT Indices System GMM Estimation Panel A First lag of dependent variable ICT indices AR(1) AR(2) Hansen’s Jstatistic Country Obs. Panel B Firs lag of ICT indices Country Obs. Adjusted Rsquare Export Model 1 2 3 .522a .434a .434a (0.000) (0.000) (0.000) .191b (0.039) -2.53b (0.011) -0.76 (0.447) 30.39 (0.732) 35 478 .410c (0.056) -2.02b 0.043 -0.92 0.358 30.04 (0.615) 35 478 1 1.464 a [ .395] 35 479 0.718 2 1.098 a [.348] 35 479 0.717 .089c (0.084) -1.89c (0.059) -0.77 (0.439) 29.96 (0.186) 35 478 4 .399 a (0.001) 1 .628 a (0.000) Import Model 2 3 .543 a .498 a (0.000) (0.000) 1.605 c .297a .528a (0.000) (0.008) (0.056) -2.03b -2.57a -2.39b (0.042) (0.010) (0.017) -0.91 -1.41 -1.32 (0.159) (0.363) ( 0.189) 14.31 31.16 32.58 (0.939) (0.698) (0.632) 35 34 34 478 476 476 One-year Lagged ICT Indices 3 4 1 2 .480 a .798 a 3.118a 2.455a [.353] [.298] [.132] [.282] 35 35 34 34 479 479 476 476 0.718 0.714 0.803 0.791 4 .604 a (0.000) .370a (0.000) -2.33b (0.020) -1.51 (0.131) 33.29 (0.598) 34 476 .378a (0.000) -2.72a (0.007) -0.65 (0.514) 31.73 (0.134) 34 476 3 .785a [.110] 34 476 0.782 4 2.542a [.321] 34 476 0.782 Notes: Probabilities are reported in parentheses. a, b, and c represent significance levels at 1%, 5%, and 10% levels respectively. The null hypothesis of Hansen’s J test is that the instruments are valid. All other control variables are also included in regressions. Among the independent variables, ICT, GDP, and population are treated as endogenous variables; two- and three years lag of endogenous variables are used as instruments. Binary dummy variables are accepted as exogenous. Robust standard errors are presented in brackets. Time effects are included in regressions. We run four regressions for – one for each ICT index for exports and four more regressions for imports. ICT indices have significant and positive impacts on export and import at different significance levels, indicating that the impact of ICT on export/import are robust. However, the coefficients of ICT indices are quite distinct in their magnitudes from those of 10 We could not report the coefficient results of other variables in the models to conserve space. However, the results are available from the corresponding author upon request. fixed effects model. However, it should be kept in mind that GMM and the fixed effects are two separate approaches, and thus comparing the sizes of their coefficients is not appropriate. The autocorrelation test results indicate the presence of the first-order autocorrelation, as expected, while there is no second-order autocorrelation. Hansen’s (1982) J-test results support the validity of instruments, i.e. the satisfaction of orthogonality condition. Apart from the system-GMM estimation, we also include one-year lagged ICT indices as independent variables instead of current ICT indices to address simultaneity between ICT and trade. As reported in Panel B of Table 9, one-year lagged ICT indices also have positive and significant impacts on both export and import. Overall, these findings provide evidence favoring the robustness of our results. 5.1.2 Additional Regressors As a part of the sensitivity analysis, we run additional regressions by including additional control variables in the baseline time fixed effects models and by dropping or including some specific groups of countries.11 The results are reported in Table 10. First, the difference between GDP per capita of Turkey and its trading partners was included as a control variable. By doing so, we also examine empirical evidence in support of the trade theories of Heckscher (1919)Ohlin (1933) and Linder (1961) for Turkey. The results provided clear evidence for the H-O theory in the case of import and for the Linder’s theory in the case of export. Turkey appears to import more goods from countries that have achieved different levels of economic development whereas it exports more to countries that have similar levels of economic development. Besides, the sign and significance of the coefficients of ICT indices do not change after including the differences in GDP per capita variable into the export/import models. Additionally, including real exchange rate, tariff rates, railway, and political openness variables does not lead to 11 We didn’t report the results of the coefficients of control variables, goodness of fit, and observation numbers of the regressions to conserve space. However, they are available upon request from author. Besides, tariff rates and railway data are available only for a small number of countries in both models. significant changes in the coefficients of ICT indices.12 These findings indicate that ICT indices are not picking up the effects of control variables on export and import. Table 10: Results from Additional Regressions Variables Exchange rate Tariff rates Railway Difference GDP per capita Political openness For developed countries For developing countries Drop EU countries Drop landlocked countries For (20092014) Country fixed effects Import Model 1 2 3.147a 2.447 a (.373) (.297) a 3.222 2.581 a (.405) (.338) 2.693a 2.149 a (.355) (.298) a 2.944 2.955 a (.538) (.641) 3 .764 a (.107) .779 a (.121) .602 a (.108) .834 a (.170) 4 2.455 a (.313) 2.605 a (.371) 2.034 a (.359) 3.334 a (.622) 1 1.162a (.352) 1.055a (.364) 1.091a (.270) 1.811a (.390) Export Model 2 3 .927a .428a (.319) (.120) b .627 .177 (.293) (.145) .503a .160c (.192) (.083) a 1.327 .509a (.332) (.128) 4 .504c (.263) 1.172a (.417) .994a (.278) 1.164a (.334) 3.724a (.330) 3.185a (.469) 2.836 a (.292) 3.246a (.529) .871 a (.111) .673a (.118) 3.377 a (.330) 1.794a (.490) 1.542a (.435) .462a (.203) 1.228a (.372) .606a (.293) .483a (.145) .155a (.069) .638c (.327) .765a (.273) 4.654a (.686) 4.181a (.539) 1.201a (.190) 2.845a (.488) 1.243a (.171) 1.282a (.184) .372a (.056) 1.932a (.239) 3.866 a (.475) 3.337 a (.353) 3.249 a (.364) 2.688 a (.286) .928 a (.137) .878 a (.107) 4.186 a .452 3.264 a (.338) 1.061a (.389) 1.972a (.438) .661c (.383) 1.779a (.388) .039 (.168) .297b (.150) 1.818a (.272) 1.919a (.316) 5.574a (.553) 1.547 a (.349) 5.823 a (.654) .977 a (.213) 2.024 a (.285) .489 a (.084) 2.676 a (.508) 2.133 a (.415) 2.980a (1.067) .748a (.250) 5.371a (1.345) .408b (.172) .746 (.525) .079 (.078) -.524 (.511) .743a (.275) Notes: Cluster-robust standard errors are reported in parentheses. a, b and c indicate 1%, 5%, and 10% significance levels, respectively. Time effects are included in all regressions; control variables are used in their natural logarithmic forms. 1, 2, 3, and 4 indicate regressions that use IDI index, ICT access index, ICT use index and ICT skills index as independent variables, respectively. We also run regressions for some specific groups of countries. First, we estimate the models for only developed countries (that are trading partners of Turkey). There are 19 developed 12 Real exchange rate has a positive and significant impact on import while it has a negative impact on export in all regressions, as expected. Tariff rate has negative and significant impacts on both import and export exceptions with only few cases in which it is insignificant. Also, railway is mostly positive and significant in the both model along with few exceptional cases in which it is insignificant. Political openness has also positive and significant effects on import/export exceptions with few cases. countries in the import model and 17 developed countries in the export model. In case of developed countries, ICT still has positive impact on both export and import. When we run regressions only for the sample of developing countries, ICT indices have still positive and significant impacts on both export and import. However, the impact of ICT seems to be stronger for the developing countries than for developed countries. This result might be explained by the diminishing marginal return to ICT. Developed countries have had a high level of ICT infrastructure since the early 1990s. However, developing countries have started to invest in their ICT sectors in the post-1990s. Therefore, the same percentage increase in ICT indices might have created more impact on trade for developing countries. On the other hand, as suggested by Liu and Nath (2016), it is likely that the most ICT-induced trade growth in developed countries occurred in the 1990s, before the period of this study. Therefore, there may be some other factors that are relatively more important for exports to and import from developed countries during the post-1990s period. In contrast, developing countries may have just started to benefit from ICT with the proliferation of technology in these countries in the post-1990s. As yet another sensitivity check, we excluded the EU member countries, the top trading partners of Turkey. There are 17 and 18 EU member economies in the import and export models respectively. Excluding EU countries from the sample does not lead to any significant change and the positive impact of ICT remains except for the case of ICT use index in the export model. As another sensitivity check, five landlocked countries from both import and export samples are dropped. However, ICT still keeps its positive and strong impact on trade. We test the robustness of our findings for a shorter period as well, 2009-2014. This period corresponds to the aftermath of the last global financial crisis of 2008. Turkey’s international trade was heavily affected in this period due to recession in EU member countries. Since then, Turkey has started to look for new trading partners, particularly from the Middle East and Africa, to prevent the destructive impacts of the crisis on its international trade and economic growth. When we re-run the regressions only for the period of 2009-2014, all ICT indices still appear to affect Turkey’s imports positively and significantly. However, for the export model, ICT use and ICT skills indices lose their positive effects while IDI and ICT access indices keep their positive and significant influences. Finally, including country fixed effects into all regressions do not cause any significant change on the impacts of ICT indices, except with the ICT use, losing its significance in the export model. Overall, all these sensitivity analyses provide clear evidence of the robustness of our results except for only a few cases. Based on these results, it might be asserted that ICT is a leading factor that boosts bilateral trade between Turkey and its trading partners. By reducing traderelated costs such as transportation cost, information and communication costs, ICT fosters the international trade of Turkey. 6. CONCLUSIONS AND POLICY IMPLICATIONS In this study, we estimate the impact of ICT on bilateral trade flows between Turkey and its trading partners based on an augmented panel gravity model. We estimate two different models for exports and imports. The export model includes the top 35 importers of Turkish goods while the import model includes 34 top exporters to Turkey. The sample period extends from 2000 to 2014. As a primary estimation method, we employ panel fixed effects model along with GMM and conduct several sensitivity analyses to check the robustness of the results. As a proxy for ICT, we compute four different ICT indices: ICT development index (IDI), ICT access index, ICT use index, and ICT skills index. The first of these indices subsumes the rest three. The findings from the panel fixed effects model indicate that ICT has positive and significant effects on both Turkish bilateral export and import volumes. Furthermore, ICT appears to have larger impact on Turkish import than Turkish export. Among ICT indices, ICT access seems to be the most crucial for Turkish bilateral export whereas ICT skills index is the most influential one for Turkish bilateral import. Furthermore, ICT use index has the least impact on both export and import volumes of Turkey. As a robustness check, we also employ a dynamic panel data estimator (GMM) to address the endogeneity problem. The results from GMM also support our findings from the panel fixed effects model. We run additional regressions for a smaller time period, some specific country groups and by including control variables as sensitivity analyses. Our results appear to be robust. These results indicate that by lowering trade costs such as transportation costs, market entry costs, and communication and information costs, ICT development in Turkey as well as in its trading partner countries stimulates trade flows and thus has a trade-enhancing effect. In other words, ICT provides new communication channels leading to new or improved trading stimuli (see Vemuri and Siddiqi, 2009). As Freund and Weinhold (2000) argue, due to improvements in ICT industry, importing countries can buy goods at lower prices and have a lot to gain in consumer surplus, while exporting countries can reduce the costs of gaining access to new markets. Although the onus of developing ICT in trading partner countries rests on the respective countries, the results presented in this paper have some implications for policymakers of Turkey. For instance, Turkey should increase its trade linkages with countries that have high levels of ICT endowments. In particular, the trade relations with countries that have high levels of ICT infrastructure and access to basic ICTs will foster Turkey’s bilateral export flows. Furthermore, the Turkish government, NGOs and public and private agencies should invest more into ICT sector. Note that one sided development in ICT would not help bilateral exchanges. As its trading partners invest in ICT development, Turkey should also improve and implement ICT. The priority of policymakers should be to build the Internet infrastructure and to adopt ICT to reduce the unnecessary trade impediments (Biswas and Kennedy, 2016). The policies that facilitate and encourage the adoption and the deployment of ICT will go a long way in boosting trade in Turkey. As we discussed before, Turkey has recently shifted its orientation in international trade to reduce the risk associated with trade relationships primarily with western countries. As it looks for new trading partners, ICT development in the potential partner countries could be an important criterion. In particular, as Turkey expands its trade with emerging markets and other developing economies, this should be of crucial importance as there are substantial variations in ICT endowments across the developing world.