- Email: [email protected]
The role of insurance in international shipping costs
Economics Letters 155 (2017) 116–120
Contents lists available at ScienceDirect
Economics Letters journal homepage: www.elsevier.com/locate/ecolet
The role of insurance in international shipping costs Adrian Wolanski Department of Economics, Indiana University, United States
highlights • • • •
Define and estimate a model of insurance costs in international shipping. Results show distance is not a parameter of insurance costs. Transportation costs are per-unit rather than ad-valorem. GDP per capita is a weak instrumental variable in transportation cost models.
article
info
Article history: Received 16 October 2016 Received in revised form 20 March 2017 Accepted 21 March 2017 Available online 22 March 2017
a b s t r a c t We estimate insurance costs for international shipments, and determine that distance does not affect insurance costs. This corroborates empirical observations of the Alchian–Allen effect. We also show that GDP per capita is endogenously related to unit value through insurance costs. © 2017 Elsevier B.V. All rights reserved.
Keywords: Transportation costs Iceberg costs Alchian–Allen effect
1. Introduction A recent focus of international trade research has been on modeling transportation costs. While there has been substantial work on transportation costs as a whole, the various subcomponents have not received as much attention. Specifically, there has been very little investigation into determinants of insurance costs and their role in total transportation costs. Since Hummels (2001) demonstrates that explicitly measured costs make up the majority of transportation costs, it is important to develop a better understanding of these explicit factors–including insurance costs. The likely reason so little work has been done is due to a lack of data. Firms typically report insurance costs and freight costs together, and there is no established methodology for assigning shares of total transportation costs to the separate components of transportation costs. According to Hummels et al. (2009), the only real information available about shipping insurance is that insurance cost is highly related (elasticity coefficient of 0.88) to the price of the object being shipped. We use Chilean import data from 2007, which reports freight and insurance separately, to gain information about the determinants of insurance costs. In particular, we find that insurance costs are distance insensitive E-mail address: [email protected]. http://dx.doi.org/10.1016/j.econlet.2017.03.025 0165-1765/© 2017 Elsevier B.V. All rights reserved.
(nearly distance invariant). In addition, our results show that insurance declines as a share of total transportation cost as distance increases, which indicates that freight costs are per unit rather than iceberg. This corroborates Hummels and Skiba (2004), who find empirical evidence supporting the theorem proposed by Alchian and Allen (1964). We also find endogeneity problems between price and GDP per capita (a commonly used instrument for price), which validates growing concerns about the empirical utility of instrumental variable estimation. 2. Materials We investigate Chilean imports for 2007 using the variables insurance, unit value, quantity, weight, importer, product, distance, and GDP per capita. One unique feature of this data set is that Chile has a somewhat special place in economic geography–rich trade partners like the United States, Europe, and Japan, are far away while neighboring trade partners tend to be much poorer. In Section 5, we consider the effects that this might have on transportation and insurance costs. The main feature of this data set is the insurance variable, which distinguishes insurance costs from freight costs. However, there is a problem with the way that the data set reports insurance costs. The data set reports costs by item, even though the data was recorded by shipment. Therefore, in shipments containing more than one item, the items were simply
A. Wolanski / Economics Letters 155 (2017) 116–120
117
Table 1 Full sample data summary. Variable name
All methods
Maritime shipments only
Air shipments only
Imports (USD, Millions) Freight Cost (USD, Millions) Total Insurance Cost (USD, Millions) Number of Importers Number of Shipments Number of Entries HTS Code Observations
39,000 2,730 116 31,366 811,020 2,500,725 6,508
28,400 1,990 61.8 13,802 287,162 1,020,650 5,903
4,990 292 33.5 18,100 363,658 979,969 5,372
assigned an insurance cost proportional to the value share of the item in the shipment. Since there is so little information available about the determinants of insurance cost, we treat insurance as a black box instead of assuming that insurance costs for shipments are linear combinations of the items in the shipments. We therefore aggregate the data back to the shipment level for our analysis. There are some problems with this method; namely, we cannot include product-specific fixed effects on shipments that contain multiple products. We compensate by first considering the shipments that contain only one observation in the overall data set, and then compare these results to the entire data set. We first provide some descriptive statistics of the total data set in Table 1. The most important thing to note is that freight costs are about 7 percent of total import value, and that insurance costs are about 0.2 percent of total import value. It is also important to note that these numbers are explicit freight and insurance costs only, which is why they are much smaller than the estimate by Anderson and van Wincoop (2004) of total transportation costs at 170 percent of import values. Since the subset of shipments containing only one item is heavily used throughout our analysis, we have included some descriptive statistics of that subsample in Table 2. As table two shows, this subset comprises just over half of the total number of shipments and about 64 percent of the total import value for Chile in 2007. In addition, about 80 percent of importers and about 90 percent of product classifications present in the main data set are also present in this subset. Since there is also a similar distribution of value and shipments across transportation methods, this subset provides reasonable coverage of the main data set despite not being representative due to a lack of composite shipments. 3. Model specification The theory behind this model is similar to equation 10 in Hummels and Skiba (2004), which provides a model for estimating freight cost as a function of distance. We replace unit freight cost with ad-valorem insurance cost to produce Eq. (1): log(insurancev ) = α1 + α2 + log(distance) ∗ β1
+ log(unit v alue) ∗ β2 + log(unit w eight) ∗ β3 + ϵ.
(1)
Where insurancev is the insurance cost as a proportion of unit value, α1 is fixed effects for importers, α2 is fixed effects for product and importer combinations, and ϵ is an error term. At most one fixed effect at a time will be used when estimating an equation, so the regression tables in the next section will indicate which fixed effects are being used in which regressions. We then define Eq. (2) with the same right hand side variables, but the dependent variable is insurance cost as a percentage of explicit freight costs, rather than as a percentage of value. log(insurancec ) = α1 + α2 + log(distance) ∗ β1
+ log(unit v alue) ∗ β2 + log(unit weight) ∗ β3 + ϵ.
(2)
Where insurancec is insurance as a percentage of explicit freight costs, and all other variables are the same as above. We estimate both equations first on the set of shipments containing only one item in the main data set, and then repeat the specifications on the full data set. We estimate all models using both OLS and IV regression with GDP per capita of the exporting country as an instrument for price. We also report estimates for all transportation methods together, as well as separately for maritime and air transportation. Finally, we include importer–product and importer fixed effects for the partial sample, and importer fixed effects for the whole set. 4. Results In Table 3, we provide the regression results of Eq. (1). The OLS estimates show that increases in distance have a slight positive effect on insurance cost, but that the inclusion of product-specific fixed effects reduces the magnitude of distance effects and makes the results statistically insignificant for all transportation methods together and also for just maritime transport. The results from the IV regression generally agree with the OLS results when including importer–product fixed effects, and are generally statistically insignificant. Therefore, the results from Table 3 indicate that insurance cost is not a function of distance. In Table 4, we estimate Eq. (2) and find that insurance cost declines as a portion of total freight cost as distance increases. While this shows that freight cost is more sensitive to distance than insurance cost, the previous result that insurance cost is generally unaffected by distance indicates that the relationship captured by this model could simply be caused by the increase in freight cost over distance. We then repeat the specifications on Eqs. (1) and (2) using the full data set. The results from the full samples are similar to the partial sample results, with Table 5 showing that insurance costs are not a function of distance and Table 6 showing that insurance declines as a share of total explicit freight costs as distance increases. While this provides direct information on the behavior of insurance costs, these results also link to other debates in international trade. Specifically, the results corroborate previous research that freight costs are primarily per unit, rather than iceberg. One of the justifications for iceberg freight cost models is that more expensive goods require more insurance to ship, so freight cost should be strongly proportional to price. We showed earlier, however, that insurance makes up less than 10 percent of freight costs. Additionally, the large negative parameter estimates from Tables 4 and 6 demonstrate that insurance is significantly less sensitive to changes in distance than total transportation costs. We therefore conclude that insurance costs are not the main factor governing transportation costs, which corroborates the results in Hummels and Skiba (2004) that freight cost is per unit rather than iceberg. 5. Robustness checks To support our results, we provide a series of robustness checks. First, all regressions in this paper are done with heteroscedasticity– robust standard errors that are also clustered by fixed effects.
118
A. Wolanski / Economics Letters 155 (2017) 116–120 Table 2 Subsample of shipments containing only one item data summary. Variable name
All methods
Maritime shipments only
Air shipments only
Imports (USD, Millions) Freight Cost (USD, Millions) Total Insurance Cost (USD, Millions) Number of Importers Number of Shipments HTS Code Observations
25,300 1,890 60 24,489 433,710 5,895
19,400 1,420 32.7 9,786 141,906 4,826
2,350 141 1.66 14,587 200,303 4,616
Table 3 Eq. (1) estimates for the subsample of shipments containing only one item. OLS
IV
All shipment methods log(distance) log(unit value) log(unit weight) R2
.0414 (.02)** .0333 (.0077)*** −.1061 (.012)*** .0272
−.0027 (.0159) −.0785 (.0054)***
.0530 (.0237)**
.0289 (.0313) −.0765 (.0199)*** .0569 (.0190)*** .0114
.0057 (.0244) .1732(.0566)*** −.1956(.0585)***
.027 (.032)
.0433 (.0178)** −.1143 (.0057)*** .0446 (.0069)*** .0224
.036 (.0397) .0946 (.0474)*** −.1463 (.0426)***
.0307 (.0179)* .0123 (.0372) −.0539 (.0297)*
.0323 (.0061)*** .0049
−.0324 (.0175)* .2168 (.0386)*** −.2659 (.03820***
−.0162 (.0142) −.0942 (.0486)** −.1095 (.0397)***
Maritime shipments only log(distance) log(unit value) log(unit weight) R2
−.0317 (.0148)** .0144 (.0182) .0088
Air shipments only log(distance) log(unit value) log(unit weight) R2 α1 used α2 used
.0869 (.0424)**
−.1033 (.0084)*** .0196 (.0121) .0233 Yes
Yes Yes
Yes
Table 4 Eq. (2) estimates for the subsample of shipments containing only one item. OLS
IV
All shipment methods log(distance) log(unit value) log(unit weight) R2
−.1997 (.015)*** .3655 (.008)*** −.312 (.0125)*** .1615
−.2394 (.0174)*** .4286 (.0063)*** .4286 (.0063)*** .1597
−.2395 (.0174)*** .2836 (.0295)*** −.2407 (.0291)***
−.2348 (.0158)***
−.2426 (.0302)*** .4044 (.0157)*** −.4117(.0172)*** .1379
−.2947 (.0272)*** .5168 (.066)*** −.5496 (.0688) ***
−.2453 (.0301)***
−.4086 (.0170)*** .5039 (.0066)*** −4864 (.0081)*** .3372
−.4034 (.0303)*** .4874 (.0427)*** −.4641 (.0366)***
−.3957 (.0170)***
.3686 (.0561)***
−.3602 (.0461)***
Maritime shipments only log(distance) log(unit value) log(unit weight) R2
−.2789 (.0242)*** .4481 (.0147)*** −.4792 (.0199)***
.779 (.1508)***
−.7499 (.139)***
Air shipments only log(distance) log(unit value) log(unit weight) R2 α1 used α2 used
−.4139 (.0275)*** .5281 (.0098)*** −.4982 (.0117)*** .3370 Yes
.3735 (.0418)
−.3851 (.0326)
Yes Yes
Because our data only covers one country in South America, geography is potentially a problem for our models since large trading partners such as the United States, Europe, and China are relatively far away. It is possible that these areas are more secure, in which case our estimates would accidentally attribute the change in insurance costs to the changes in distance. We therefore repeat the specifications and include GDP per capita as an independent variable to determine if a relationship between country income (as a proxy for security) and distance is driving our results. Previous literature, however, suggests that this should not be the case since GDP per capita is commonly used as an instrument for price when estimating freight cost models and its validity as an instrument necessitates no correlation with insurance. First, we define Eq. (3), which is Eq. (1) with GDP per capita added as an explanatory
Yes
variable. log(insurancev ) = α1 + α2 + log(distance) ∗ β1
+ log(unit v alue) ∗ β2 + log(unit weight) ∗ β3 + log(GDPcapita) ∗ β4 + ϵ.
(3)
We then estimate Eq. (3) with OLS and provide the results in Table 7. The OLS parameter estimates and standard errors for advalorem insurance do not change significantly when we include GDP per capita as an explanatory variable. We see, however, that GDP per capita has a small positive effect on insurance costs. While the correlation is rather weak, it raises concerns about the validity of GDP per capita as in instrument when estimating freight costs.
A. Wolanski / Economics Letters 155 (2017) 116–120 Table 7 Eq. (3) estimates on the subsample of shipments containing only one item.
Table 5 Eq. (1) estimates with full sample. OLS
IV
All shipment methods log(distance) log(unit value) log(unit weight) R2
−.0445 (.1263) 2.4048 (10.1433) .8312 (3.8108)
log(distance) log(unit value) log(unit weight) log(GDPcapita) R2
.0293 (.0203) .1576 (.1405) .0424 (.0687)
Maritime shipments only
.1636
Maritime shipments only .0249 (.0187) −.1165 (.0073)*** −.0903 (.0052)*** .0576
Air shipments only log(distance) log(unit value) log(unit weight) R2 α1 used
.0338 (.0312) −.2541 (.0145)*** −.1496 (.0089)*** .1081 Yes
.0374 (.0326)
Yes
OLS
IV
−.1211 (.0141)***
−.2954 (.6152)
.1455 (.0098)***
12.3721 (49.9542) 4.5753 (18.7716)
−.0728 (.0742) −.0832 (.0275)***
Table 6 Eq. (2) estimates with full sample.
All shipment methods log(distance) log(unit value) log(unit weight) R2
−.0124 (.0051)** .0617
Maritime shipments only log(distance) log(unit value) log(unit weight) R2
−.2715 (.0226)*** −.0401 (.0129)*** −.1029 (.0083)***
−.2528 (.0434)*** 1.1268 (.3655)*** .4624 (.1870)**
.0970
log(distance) log(unit value) log(unit weight) log(GDPcapita) R2
.0374 (.02)* .0261 (.0076)*** −.0992 (.0116) *** .0631 (.0112)*** .0316
−.0063 (.0155) −.0811 (.0054)*** .0339 (.0061)*** .0273 (.0076)*** .0037
.0561 (.0239)** −.0366 (.0145)** .0182 (.018) .0344 (.0087)*** .0053
.0267 (.0314) −.0812 (.0204)*** .0611 (.0195) *** .0277 (.0134)** .0093
.084 (.042)** −.1066 (.0083)*** .022 (.0119)* .045 (.0095)*** .0255 Yes
.0415 (.0178)** −.1158 (.0057)*** .0452 (.0069)*** .0281 (.006)*** .0236
Air shipments only log(distance) log(unit value) log(unit weight) log(GDPcapita) R2 α1 used α2 used
Yes
Table 8 Eq. (4) estimates on the subsample of shipments containing only one item. OLS All shipment methods log(distance) log(unit value) log(unit weight) log(GDPcapita) R2
−.1979 (.0146)*** .3687 (.0081)*** −.3151 (.0124)*** −.0282 (.0102)*** .1634
−.2382 (.0168)*** .4295 (.0064)*** −.4101 (.008)*** −.0095 (.0089) .1602
−.2778 (.0242)*** .4465 (.0146) *** −.4779 (.0198)*** .0115 (.0104) .1358
−.2457 (.0295)*** .3977 (.0157) *** −.4057 (.0168)*** .0396 (.0157)** .1365
−.4133 (.0258)** .5287 (.001)*** −.4987 (.0118)*** −.0093 (.0098) .3371 Yes
−.4068 (.0169)*** .5054 (.0067)*** −.4870 (.0081)*** −.0215 (.0067)*** .3371
Maritime shipments only
Air shipments only log(distance) log(unit value) log(unit weight) R2 α1 used
OLS All shipment methods
−.0074 (.0155) −.1993 (.0109)*** −.1459 (.0056)***
log(distance) log(unit value) log(unit weight) R2
119
−.3494 (.0367)***
−.3426 (.0390)***
.2604 (.0141)***
−.0261 (.0080)***
.6101 (.1079)*** .1019 (.0396)***
.1785 Yes
Yes
We now define Eq. (4), which is simply Eq. (2) with GDP per capita added as an explanatory variable. log(insurancec ) = α1 + α2 + log(distance) ∗ β1
+ log(unit v alue) ∗ β2 + log(unit w eight) ∗ β3 + log(GDPcapita) ∗ β4 + ϵ. (4) We then estimate Eq. (4) with OLS and provide the results in Table 8. Like Table 7, Table 8 shows that adding GDP per capita as an explanatory variable does not change the OLS parameter estimates for insurance costs as a share of total freight cost. Therefore, our interpretation of Tables 3 and 4 is robust to including GDP per capita as an explanatory variable. 6. Conclusion We find that freight cost is more sensitive to changes in distance than insurance cost is, and that insurance cost is likely distance invariant. Additionally, the data suggest that as distance increases, freight cost is more likely to be per-unit rather than ad-valorem because of this difference in relative sensitivity. This is consistent with findings by Hummels and Skiba (2004) on the behavior of freight costs. Finally, GDP per capita is endogenously related to unit value through insurance costs, which presents problems for using GDP per capita as an instrument for unit value. Since a key
log(distance) log(unit value) log(unit weight) log(GDPcapita) R2 Air shipments only log(distance) log(unit value) log(unit weight) log(GDPcapita) R2 α1 used α2 used
Yes
result of this paper is that distance is not a parameter of transportation insurance, future research should focus on determining other variables that might be relevant parameters. Future research should also determine how insurance cost changes when multiple items are shipped together, which would allow more controls to be accurately included in insurance cost models. Acknowledgments We would like to thank Volodymr Lugovskyy for providing many helpful comments and suggestions, as well as the data set (acquired from datamyne.com). We would also like to thank an anonymous referee and the students in Indiana University’s spring 2016 E490 class for their helpful comments and suggestions. Appendix A. Supplementary data Supplementary material related to this article can be found online at http://dx.doi.org/10.1016/j.econlet.2017.03.025.
120
A. Wolanski / Economics Letters 155 (2017) 116–120
References Alchian, Armen A., Allen, William R., 1964. University Economics. Wadsworth, Belmont, Calif., pp. 74–75. Anderson, James E., van Wincoop, Eric, 2004. Trade costs. Journal of Economic Literature 42, 691–751. Hummels, David, 2001. Toward A Geography of Trade Costs, mimeo, University of Chicago.
Hummels, David, Skiba, Alexandre, 2004. Shipping the good apples out? An empirical confirmation of the alchian-allen conjecture. J. Polit. Econ. 112 (6), 1384– 1402. Hummels, David, Lugovskyy, Volodymr, Skiba, Alexandre, 2009. The trade reducing effects of market power in international shipping. J. Dev. Econ. 89 (1), 84–97.
Contents lists available at ScienceDirect
Economics Letters journal homepage: www.elsevier.com/locate/ecolet
The role of insurance in international shipping costs Adrian Wolanski Department of Economics, Indiana University, United States
highlights • • • •
Define and estimate a model of insurance costs in international shipping. Results show distance is not a parameter of insurance costs. Transportation costs are per-unit rather than ad-valorem. GDP per capita is a weak instrumental variable in transportation cost models.
article
info
Article history: Received 16 October 2016 Received in revised form 20 March 2017 Accepted 21 March 2017 Available online 22 March 2017
a b s t r a c t We estimate insurance costs for international shipments, and determine that distance does not affect insurance costs. This corroborates empirical observations of the Alchian–Allen effect. We also show that GDP per capita is endogenously related to unit value through insurance costs. © 2017 Elsevier B.V. All rights reserved.
Keywords: Transportation costs Iceberg costs Alchian–Allen effect
1. Introduction A recent focus of international trade research has been on modeling transportation costs. While there has been substantial work on transportation costs as a whole, the various subcomponents have not received as much attention. Specifically, there has been very little investigation into determinants of insurance costs and their role in total transportation costs. Since Hummels (2001) demonstrates that explicitly measured costs make up the majority of transportation costs, it is important to develop a better understanding of these explicit factors–including insurance costs. The likely reason so little work has been done is due to a lack of data. Firms typically report insurance costs and freight costs together, and there is no established methodology for assigning shares of total transportation costs to the separate components of transportation costs. According to Hummels et al. (2009), the only real information available about shipping insurance is that insurance cost is highly related (elasticity coefficient of 0.88) to the price of the object being shipped. We use Chilean import data from 2007, which reports freight and insurance separately, to gain information about the determinants of insurance costs. In particular, we find that insurance costs are distance insensitive E-mail address: [email protected]. http://dx.doi.org/10.1016/j.econlet.2017.03.025 0165-1765/© 2017 Elsevier B.V. All rights reserved.
(nearly distance invariant). In addition, our results show that insurance declines as a share of total transportation cost as distance increases, which indicates that freight costs are per unit rather than iceberg. This corroborates Hummels and Skiba (2004), who find empirical evidence supporting the theorem proposed by Alchian and Allen (1964). We also find endogeneity problems between price and GDP per capita (a commonly used instrument for price), which validates growing concerns about the empirical utility of instrumental variable estimation. 2. Materials We investigate Chilean imports for 2007 using the variables insurance, unit value, quantity, weight, importer, product, distance, and GDP per capita. One unique feature of this data set is that Chile has a somewhat special place in economic geography–rich trade partners like the United States, Europe, and Japan, are far away while neighboring trade partners tend to be much poorer. In Section 5, we consider the effects that this might have on transportation and insurance costs. The main feature of this data set is the insurance variable, which distinguishes insurance costs from freight costs. However, there is a problem with the way that the data set reports insurance costs. The data set reports costs by item, even though the data was recorded by shipment. Therefore, in shipments containing more than one item, the items were simply
A. Wolanski / Economics Letters 155 (2017) 116–120
117
Table 1 Full sample data summary. Variable name
All methods
Maritime shipments only
Air shipments only
Imports (USD, Millions) Freight Cost (USD, Millions) Total Insurance Cost (USD, Millions) Number of Importers Number of Shipments Number of Entries HTS Code Observations
39,000 2,730 116 31,366 811,020 2,500,725 6,508
28,400 1,990 61.8 13,802 287,162 1,020,650 5,903
4,990 292 33.5 18,100 363,658 979,969 5,372
assigned an insurance cost proportional to the value share of the item in the shipment. Since there is so little information available about the determinants of insurance cost, we treat insurance as a black box instead of assuming that insurance costs for shipments are linear combinations of the items in the shipments. We therefore aggregate the data back to the shipment level for our analysis. There are some problems with this method; namely, we cannot include product-specific fixed effects on shipments that contain multiple products. We compensate by first considering the shipments that contain only one observation in the overall data set, and then compare these results to the entire data set. We first provide some descriptive statistics of the total data set in Table 1. The most important thing to note is that freight costs are about 7 percent of total import value, and that insurance costs are about 0.2 percent of total import value. It is also important to note that these numbers are explicit freight and insurance costs only, which is why they are much smaller than the estimate by Anderson and van Wincoop (2004) of total transportation costs at 170 percent of import values. Since the subset of shipments containing only one item is heavily used throughout our analysis, we have included some descriptive statistics of that subsample in Table 2. As table two shows, this subset comprises just over half of the total number of shipments and about 64 percent of the total import value for Chile in 2007. In addition, about 80 percent of importers and about 90 percent of product classifications present in the main data set are also present in this subset. Since there is also a similar distribution of value and shipments across transportation methods, this subset provides reasonable coverage of the main data set despite not being representative due to a lack of composite shipments. 3. Model specification The theory behind this model is similar to equation 10 in Hummels and Skiba (2004), which provides a model for estimating freight cost as a function of distance. We replace unit freight cost with ad-valorem insurance cost to produce Eq. (1): log(insurancev ) = α1 + α2 + log(distance) ∗ β1
+ log(unit v alue) ∗ β2 + log(unit w eight) ∗ β3 + ϵ.
(1)
Where insurancev is the insurance cost as a proportion of unit value, α1 is fixed effects for importers, α2 is fixed effects for product and importer combinations, and ϵ is an error term. At most one fixed effect at a time will be used when estimating an equation, so the regression tables in the next section will indicate which fixed effects are being used in which regressions. We then define Eq. (2) with the same right hand side variables, but the dependent variable is insurance cost as a percentage of explicit freight costs, rather than as a percentage of value. log(insurancec ) = α1 + α2 + log(distance) ∗ β1
+ log(unit v alue) ∗ β2 + log(unit weight) ∗ β3 + ϵ.
(2)
Where insurancec is insurance as a percentage of explicit freight costs, and all other variables are the same as above. We estimate both equations first on the set of shipments containing only one item in the main data set, and then repeat the specifications on the full data set. We estimate all models using both OLS and IV regression with GDP per capita of the exporting country as an instrument for price. We also report estimates for all transportation methods together, as well as separately for maritime and air transportation. Finally, we include importer–product and importer fixed effects for the partial sample, and importer fixed effects for the whole set. 4. Results In Table 3, we provide the regression results of Eq. (1). The OLS estimates show that increases in distance have a slight positive effect on insurance cost, but that the inclusion of product-specific fixed effects reduces the magnitude of distance effects and makes the results statistically insignificant for all transportation methods together and also for just maritime transport. The results from the IV regression generally agree with the OLS results when including importer–product fixed effects, and are generally statistically insignificant. Therefore, the results from Table 3 indicate that insurance cost is not a function of distance. In Table 4, we estimate Eq. (2) and find that insurance cost declines as a portion of total freight cost as distance increases. While this shows that freight cost is more sensitive to distance than insurance cost, the previous result that insurance cost is generally unaffected by distance indicates that the relationship captured by this model could simply be caused by the increase in freight cost over distance. We then repeat the specifications on Eqs. (1) and (2) using the full data set. The results from the full samples are similar to the partial sample results, with Table 5 showing that insurance costs are not a function of distance and Table 6 showing that insurance declines as a share of total explicit freight costs as distance increases. While this provides direct information on the behavior of insurance costs, these results also link to other debates in international trade. Specifically, the results corroborate previous research that freight costs are primarily per unit, rather than iceberg. One of the justifications for iceberg freight cost models is that more expensive goods require more insurance to ship, so freight cost should be strongly proportional to price. We showed earlier, however, that insurance makes up less than 10 percent of freight costs. Additionally, the large negative parameter estimates from Tables 4 and 6 demonstrate that insurance is significantly less sensitive to changes in distance than total transportation costs. We therefore conclude that insurance costs are not the main factor governing transportation costs, which corroborates the results in Hummels and Skiba (2004) that freight cost is per unit rather than iceberg. 5. Robustness checks To support our results, we provide a series of robustness checks. First, all regressions in this paper are done with heteroscedasticity– robust standard errors that are also clustered by fixed effects.
118
A. Wolanski / Economics Letters 155 (2017) 116–120 Table 2 Subsample of shipments containing only one item data summary. Variable name
All methods
Maritime shipments only
Air shipments only
Imports (USD, Millions) Freight Cost (USD, Millions) Total Insurance Cost (USD, Millions) Number of Importers Number of Shipments HTS Code Observations
25,300 1,890 60 24,489 433,710 5,895
19,400 1,420 32.7 9,786 141,906 4,826
2,350 141 1.66 14,587 200,303 4,616
Table 3 Eq. (1) estimates for the subsample of shipments containing only one item. OLS
IV
All shipment methods log(distance) log(unit value) log(unit weight) R2
.0414 (.02)** .0333 (.0077)*** −.1061 (.012)*** .0272
−.0027 (.0159) −.0785 (.0054)***
.0530 (.0237)**
.0289 (.0313) −.0765 (.0199)*** .0569 (.0190)*** .0114
.0057 (.0244) .1732(.0566)*** −.1956(.0585)***
.027 (.032)
.0433 (.0178)** −.1143 (.0057)*** .0446 (.0069)*** .0224
.036 (.0397) .0946 (.0474)*** −.1463 (.0426)***
.0307 (.0179)* .0123 (.0372) −.0539 (.0297)*
.0323 (.0061)*** .0049
−.0324 (.0175)* .2168 (.0386)*** −.2659 (.03820***
−.0162 (.0142) −.0942 (.0486)** −.1095 (.0397)***
Maritime shipments only log(distance) log(unit value) log(unit weight) R2
−.0317 (.0148)** .0144 (.0182) .0088
Air shipments only log(distance) log(unit value) log(unit weight) R2 α1 used α2 used
.0869 (.0424)**
−.1033 (.0084)*** .0196 (.0121) .0233 Yes
Yes Yes
Yes
Table 4 Eq. (2) estimates for the subsample of shipments containing only one item. OLS
IV
All shipment methods log(distance) log(unit value) log(unit weight) R2
−.1997 (.015)*** .3655 (.008)*** −.312 (.0125)*** .1615
−.2394 (.0174)*** .4286 (.0063)*** .4286 (.0063)*** .1597
−.2395 (.0174)*** .2836 (.0295)*** −.2407 (.0291)***
−.2348 (.0158)***
−.2426 (.0302)*** .4044 (.0157)*** −.4117(.0172)*** .1379
−.2947 (.0272)*** .5168 (.066)*** −.5496 (.0688) ***
−.2453 (.0301)***
−.4086 (.0170)*** .5039 (.0066)*** −4864 (.0081)*** .3372
−.4034 (.0303)*** .4874 (.0427)*** −.4641 (.0366)***
−.3957 (.0170)***
.3686 (.0561)***
−.3602 (.0461)***
Maritime shipments only log(distance) log(unit value) log(unit weight) R2
−.2789 (.0242)*** .4481 (.0147)*** −.4792 (.0199)***
.779 (.1508)***
−.7499 (.139)***
Air shipments only log(distance) log(unit value) log(unit weight) R2 α1 used α2 used
−.4139 (.0275)*** .5281 (.0098)*** −.4982 (.0117)*** .3370 Yes
.3735 (.0418)
−.3851 (.0326)
Yes Yes
Because our data only covers one country in South America, geography is potentially a problem for our models since large trading partners such as the United States, Europe, and China are relatively far away. It is possible that these areas are more secure, in which case our estimates would accidentally attribute the change in insurance costs to the changes in distance. We therefore repeat the specifications and include GDP per capita as an independent variable to determine if a relationship between country income (as a proxy for security) and distance is driving our results. Previous literature, however, suggests that this should not be the case since GDP per capita is commonly used as an instrument for price when estimating freight cost models and its validity as an instrument necessitates no correlation with insurance. First, we define Eq. (3), which is Eq. (1) with GDP per capita added as an explanatory
Yes
variable. log(insurancev ) = α1 + α2 + log(distance) ∗ β1
+ log(unit v alue) ∗ β2 + log(unit weight) ∗ β3 + log(GDPcapita) ∗ β4 + ϵ.
(3)
We then estimate Eq. (3) with OLS and provide the results in Table 7. The OLS parameter estimates and standard errors for advalorem insurance do not change significantly when we include GDP per capita as an explanatory variable. We see, however, that GDP per capita has a small positive effect on insurance costs. While the correlation is rather weak, it raises concerns about the validity of GDP per capita as in instrument when estimating freight costs.
A. Wolanski / Economics Letters 155 (2017) 116–120 Table 7 Eq. (3) estimates on the subsample of shipments containing only one item.
Table 5 Eq. (1) estimates with full sample. OLS
IV
All shipment methods log(distance) log(unit value) log(unit weight) R2
−.0445 (.1263) 2.4048 (10.1433) .8312 (3.8108)
log(distance) log(unit value) log(unit weight) log(GDPcapita) R2
.0293 (.0203) .1576 (.1405) .0424 (.0687)
Maritime shipments only
.1636
Maritime shipments only .0249 (.0187) −.1165 (.0073)*** −.0903 (.0052)*** .0576
Air shipments only log(distance) log(unit value) log(unit weight) R2 α1 used
.0338 (.0312) −.2541 (.0145)*** −.1496 (.0089)*** .1081 Yes
.0374 (.0326)
Yes
OLS
IV
−.1211 (.0141)***
−.2954 (.6152)
.1455 (.0098)***
12.3721 (49.9542) 4.5753 (18.7716)
−.0728 (.0742) −.0832 (.0275)***
Table 6 Eq. (2) estimates with full sample.
All shipment methods log(distance) log(unit value) log(unit weight) R2
−.0124 (.0051)** .0617
Maritime shipments only log(distance) log(unit value) log(unit weight) R2
−.2715 (.0226)*** −.0401 (.0129)*** −.1029 (.0083)***
−.2528 (.0434)*** 1.1268 (.3655)*** .4624 (.1870)**
.0970
log(distance) log(unit value) log(unit weight) log(GDPcapita) R2
.0374 (.02)* .0261 (.0076)*** −.0992 (.0116) *** .0631 (.0112)*** .0316
−.0063 (.0155) −.0811 (.0054)*** .0339 (.0061)*** .0273 (.0076)*** .0037
.0561 (.0239)** −.0366 (.0145)** .0182 (.018) .0344 (.0087)*** .0053
.0267 (.0314) −.0812 (.0204)*** .0611 (.0195) *** .0277 (.0134)** .0093
.084 (.042)** −.1066 (.0083)*** .022 (.0119)* .045 (.0095)*** .0255 Yes
.0415 (.0178)** −.1158 (.0057)*** .0452 (.0069)*** .0281 (.006)*** .0236
Air shipments only log(distance) log(unit value) log(unit weight) log(GDPcapita) R2 α1 used α2 used
Yes
Table 8 Eq. (4) estimates on the subsample of shipments containing only one item. OLS All shipment methods log(distance) log(unit value) log(unit weight) log(GDPcapita) R2
−.1979 (.0146)*** .3687 (.0081)*** −.3151 (.0124)*** −.0282 (.0102)*** .1634
−.2382 (.0168)*** .4295 (.0064)*** −.4101 (.008)*** −.0095 (.0089) .1602
−.2778 (.0242)*** .4465 (.0146) *** −.4779 (.0198)*** .0115 (.0104) .1358
−.2457 (.0295)*** .3977 (.0157) *** −.4057 (.0168)*** .0396 (.0157)** .1365
−.4133 (.0258)** .5287 (.001)*** −.4987 (.0118)*** −.0093 (.0098) .3371 Yes
−.4068 (.0169)*** .5054 (.0067)*** −.4870 (.0081)*** −.0215 (.0067)*** .3371
Maritime shipments only
Air shipments only log(distance) log(unit value) log(unit weight) R2 α1 used
OLS All shipment methods
−.0074 (.0155) −.1993 (.0109)*** −.1459 (.0056)***
log(distance) log(unit value) log(unit weight) R2
119
−.3494 (.0367)***
−.3426 (.0390)***
.2604 (.0141)***
−.0261 (.0080)***
.6101 (.1079)*** .1019 (.0396)***
.1785 Yes
Yes
We now define Eq. (4), which is simply Eq. (2) with GDP per capita added as an explanatory variable. log(insurancec ) = α1 + α2 + log(distance) ∗ β1
+ log(unit v alue) ∗ β2 + log(unit w eight) ∗ β3 + log(GDPcapita) ∗ β4 + ϵ. (4) We then estimate Eq. (4) with OLS and provide the results in Table 8. Like Table 7, Table 8 shows that adding GDP per capita as an explanatory variable does not change the OLS parameter estimates for insurance costs as a share of total freight cost. Therefore, our interpretation of Tables 3 and 4 is robust to including GDP per capita as an explanatory variable. 6. Conclusion We find that freight cost is more sensitive to changes in distance than insurance cost is, and that insurance cost is likely distance invariant. Additionally, the data suggest that as distance increases, freight cost is more likely to be per-unit rather than ad-valorem because of this difference in relative sensitivity. This is consistent with findings by Hummels and Skiba (2004) on the behavior of freight costs. Finally, GDP per capita is endogenously related to unit value through insurance costs, which presents problems for using GDP per capita as an instrument for unit value. Since a key
log(distance) log(unit value) log(unit weight) log(GDPcapita) R2 Air shipments only log(distance) log(unit value) log(unit weight) log(GDPcapita) R2 α1 used α2 used
Yes
result of this paper is that distance is not a parameter of transportation insurance, future research should focus on determining other variables that might be relevant parameters. Future research should also determine how insurance cost changes when multiple items are shipped together, which would allow more controls to be accurately included in insurance cost models. Acknowledgments We would like to thank Volodymr Lugovskyy for providing many helpful comments and suggestions, as well as the data set (acquired from datamyne.com). We would also like to thank an anonymous referee and the students in Indiana University’s spring 2016 E490 class for their helpful comments and suggestions. Appendix A. Supplementary data Supplementary material related to this article can be found online at http://dx.doi.org/10.1016/j.econlet.2017.03.025.
120
A. Wolanski / Economics Letters 155 (2017) 116–120
References Alchian, Armen A., Allen, William R., 1964. University Economics. Wadsworth, Belmont, Calif., pp. 74–75. Anderson, James E., van Wincoop, Eric, 2004. Trade costs. Journal of Economic Literature 42, 691–751. Hummels, David, 2001. Toward A Geography of Trade Costs, mimeo, University of Chicago.
Hummels, David, Skiba, Alexandre, 2004. Shipping the good apples out? An empirical confirmation of the alchian-allen conjecture. J. Polit. Econ. 112 (6), 1384– 1402. Hummels, David, Lugovskyy, Volodymr, Skiba, Alexandre, 2009. The trade reducing effects of market power in international shipping. J. Dev. Econ. 89 (1), 84–97.