Gravity with scale effects

Journal of International Economics 100 (2016) 174–193

Contents lists available at ScienceDirect

Journal of International Economics journal homepage: www.elsevier.com/locate/jie

Gravity with scale effects夽 James E. Anderson a, b, * , Mykyta Vesselovsky c , Yoto V. Yotov d, e, f a

Department of Economics, Boston College, Chestnut Hill, MA 02467, USA NBER, USA Global Affairs Canada, Ottawa, ON K1A OG2, Canada d School of Economics, Drexel University, Philadelphia, PA 19104, USA e CESifo Research Network, Munich, Germany f Economic Research Institute, Bulgarian Academy of Sciences, Sofia, Bulgaria b c

A R T I C L E

I N F O

Article history: Received 10 March 2014 Received in revised form 9 January 2016 Accepted 7 March 2016 Available online 19 March 2016 JEL classification: F14 F15 F16

A B S T R A C T This paper extends the structural gravity model to incorporate scale effects and exchange rate passthrough. Bilateral scale effects in cross-border trade are inferred from the difference in distance elasticities between cross border and inter-provincial bilateral trade in a majority of 28 goods and services sectors for Canada’s provinces. Bilateral-specific relationship investment is a possible explanation. Incomplete passthrough of large exchange rate changes from 1997 to 2007, amplified by scale effects, produces direct effects on bilateral trade for 12 of 19 goods sectors but none of 9 services sectors. © 2016 Elsevier B.V. All rights reserved.

Keywords: Gravity Exchange rates Goods and services

1. Introduction We find evidence of scale effects in cross-border bilateral trade in this paper. Differences between cross-border and domestic gravity elasticities are interpreted to reflect scale effects in an extended structural gravity model applied to Canadian provincial bilateral trade flow data for 28 goods and services sectors over the decade 1997–2007.Wesuggest an explanation based on bilateral relationship-specific investment, but in the absence of direct evidence on such investment the scale effect remains a black box. Estimated scale effects are economically substantial. The absolute

夽 We thank Mario Larch for useful comments and discussions. The paper has also benefitted from comments at seminars at Oxford and Pompeu Fabra Universities, and at the 2014 Meetings of the European Economic Association. We are indebted to Emily Yu for invaluable help with the data. The research was supported by the Department of Foreign Affairs, Trade and Development, Canada. All errors are our own. An earlier version circulated as “Gravity, Scale and Exchange Rates”, NBER WP No. 18807. * Corresponding author at: Department of Economics Boston College 140 Commonwealth Avenue Chestnut Hill, MA 02467, USA.

http://dx.doi.org/10.1016/j.jinteco.2016.03.003 0022-1996/© 2016 Elsevier B.V. All rights reserved.

value of the distance elasticity of cross border bilateral trade exceeds that of domestic bilateral trade statistically and quantitatively significantly in a majority of 28 sectors for Canada’s provinces (23 of 28 cases for inward trade and 13 of 28 cases for outward trade). In aggregate goods trade a 100% rise in imports lowers Canadian trade costs by 12.3% and lowers US trade costs by 6.1%, assuming an elasticity of substitution equal to 6.13 (based on Head and Mayer, 2014). For aggregate services imports of Canada, the corresponding reduction is 9.3% while for the US the estimated elasticity is not significantly different from 0. The lower US destination scale effects satisfy the intuition that the order of magnitude larger US market tends to exhaust scale effects in cross-border trade. Our black box empirical model of scale effects can be interpreted in terms of the recent literature on network links in export dynamics. Chaney (2014) is a prime example. Bilateral links are specific capacities formed and maintained by specific buyer–selleragent interactions. These occur for example between employees of firms’ purchasing and marketing departments or via intermediaries’ services purchased by buyers and sellers. The total number of such individual bilateral links is an aggregate bilateral capacity that naturally

J. Anderson, M. Vesselovsky, Y. Yotov / Journal of International Economics 100 (2016) 174–193

suggests bilateral scale effects on trade costs.1 The volume changes suggested by this mechanism can be on either the intensive or extensive margin. Lacking firm level data, we are unable to discriminate between these. Alternatively, bilateral scale effects measured by the model could reflect fixed costs associated with designing products tailored to different destinations, e.g. Manova and Zhang (2012), fixed costs that are increasing with distance, e.g. Krautheim (2012), or firm search costs for buyers, e.g. Eslava et al. (2015). Yi (2010) influentially argues that fragmentation can explain nonlinear effects of trade costs on trade. Dis-agglomeration in the sense of Jones et al. (2005) emphasizes fragmentation of vertically integrated production driven by external increasing returns to scale in transportation and communication. While each of the above channels could potentially contribute to explaining our findings,the scale effects of this paper remain ‘dark’ like all gravity costs in Head and Mayer’s (2013) cosmological metaphor. Stepping back from the model, our empirical results pose a distance/border puzzle in addition to the time series non-declining distance puzzle emphasized by Disdier and Head (2008).2 Head and Mayer (2014) suggest that distance elasticities may vary bilaterally with market size, but we find little support for this pattern. The starting point of the model is to allow trade costs to vary tractably with bilateral volume and do so differently in cross-border than in domestic trade. Invariance to volume is assumed in the standard gravity literature, a limiting case consistent with long run equilibrium investment of various sorts including that by atomistic agents in bilateral relationships. Cross-border trade, size-adjusted, is much smaller than domestic trade and tends to have a much shorter history. This difference suggests our specification that identifies a cross-border scale effect normalized by any effective domestic scale effect, i.e. the scale effects in our theory (and empirics) are defined and should be interpreted as relative to domestic scale effects in intra-national trade.3 Other potential reasons for different gravity elasticities within and across borders are plausibly neutralized for US–Canadatrade, where mode choice plays little role. Most goods trade moves by road or rail, both within and across borders. Mode choice in tourism is dominated by distance, equally within and across borders. In contrast, mode choice plays an important role in other interregional and international trade, as emphasized by Hummels (2007) and Hillberry and Hummels (2008) among others. Our data from 1997 to 2007 contains dramatic Canada– US exchange rate variation: an 11% depreciation followed by a 45% appreciation.4 To control for this source of volume changes we develop a treatment of the effect of exchange rate changes with incomplete passthrough in combination with scale effects acting on the structural gravity model. Previous gravity applications could not simultaneously measure exchange rate effects and control for multilateral resistance with importer-time and exporter-time fixed effects. The availability of inter-provincial and cross-border trade

1 Bilateral trade scale effects are distinct from scale effects in supply to all destinations (classic external scale economies as in Antweiler and Trefler, 2002) or demand from all origins (non-homothetic preferences across sectors as in Fieler, 2011). Any such aggregate scale effects are captured in the empirical application by country–sector–time fixed effects. 2 Yotov (2012) notes that the non-declining distance puzzle goes away when crossborder and domestic distance elasticities are allowed to differ, as in this paper. 3 In the sensitivity analysis we capitalize on a unique feature of our data, which enables us to distinguish between intra-provincial and inter-provincial trade. This offers an opportunity for us to actually test for possible scale effects within Canada. We do not find such effects for goods trade, but they are present for services. 4 In 1997 the exchange rate stood at 0.72, then it fell to 0.64 in 2003, and in 2007 it was at 0.93.

175

applied to our extended model resolves this indeterminacy.5 We concentrate here on bilateral trade between the US and Canada, suppressing relationships with the Rest of the World (ROW).6 Sensitivity experiments include bilateral trade between Canada and Mexico. Our estimates of scale elasticities and other trade cost parameters turn out to be insignificantly affected by our treatment of exchange rate changes, though exchange rates separately contribute significantly to explaining variation in trade flows. The parametric scale elasticity assumed in this paper is a simplification. Support for the simplification is provided by our finding that the inferred elasticities are constant over time in a decade in which cross-border volume varies substantially. Further support comes from results of estimating simple alternative specifications. No universal constancy is suggested, because we find directional asymmetry in the bilateral cross-border scale effects and sectoral variation of scale effects. An obvious caveat is that the scale and passthrough elasticities are black box parameters. The variation of estimates across sectors suggests a payoff to opening the boxes. For scale elasticities, the caveat applies especially to a few sectors where the results suggest a mis-specified trade cost equation. As for exchange rate passthrough elasticities, there is ample evidence that exchange rate passthrough is incomplete over horizons of several years (Goldberg and Knetter, 1997) but it is unlikely to be constant.7 Following much of the literature, we do not model incomplete exchange rate passthrough in this paper,8 nor the exchange rate itself. The modeling innovations of this paper may be useful in other settings. The data must contain both international and intra-national bilateral trade flows in order to identify scale elasticities on crossborder trade. A time series dimension to the preceding intra-national and international dimensions is required to examine passthrough implications of relative price changes at the border andtheir interaction with scale effects. The same modeling treatment applies to any change in cross-border frictions such as tariff reforms (potentially incompletely passed through) or free trade agreements. These are absent from Canada–US trade in 1997–2007 but would be relevant for gravity applications to other data sets. Section 2 sets out the theoretical foundation. Section 3 develops the econometric specification and describes the data. Section 4

5 Some previous empirical gravity models have inserted real exchange rates into gravity equations without a theoretical foundation. The standard practice in these studies is to include a real exchange rate variable in a traditional version of the empirical gravity model, with no country-time fixed effects to control for multilateral resistance and with country mass variables represented by GDP and population. See for example Griffoli (2006), Kim et al. (2003) and Martínez-Zarzoso and Nowak-Lehmann (2003). A prominent but tangentially related literature considers the effect of exchange rate regimes such as currency unions on bilateral trade patterns. See Baldwin (2006) for a review of the literature on the effects of exchange rate regimes. 6 We suppress ROW for three reasons. First, the trade cost function we develop below is unlikely to plausibly approximate such a heterogeneous aggregate region. Second, aggregation may bias our inferences regarding scale effects and controlling for exchange rate effects on trade with such a large region. Third, introducing ROW data does not have any effect on our model of bilateral trade or its estimated results due to the separable fixed effects estimation structure that we use. 7 Goldberg and Knetter conclude that “While the response varies by industry, a price response equal to one-half the exchange rate change would be near the middle of the distribution of estimated responses for shipments to US” (p. 3). We abstract from explaining high frequency trade movements (within a year) because these may reflect random shocks and dynamic adjustment that have yet to be integrated with the gravity model. Differences in currency invoicing practices and length of contract terms affect high frequency price responses to exchange rate changes. It is possible that such differences across sectors may induce differing passthrough rates that persist in the medium run. In that case differing invoicing and contracting practices may help explain part of the differences in results we report across sectors. 8 A search for evidence of pricing-to-market using our industry level data produced no informative results. One reason for this is the insensitivity of markups to exchange rates under CES preferences, as explained in Technical Appendix B, which is available by request.

176

J. Anderson, M. Vesselovsky, Y. Yotov / Journal of International Economics 100 (2016) 174–193

presents the main results, quantitative implications, and robustness checks. Section 6 concludes. Appendix A describes the data. Appendix B (available on request) gives technical notes. 2. Theoretical foundation A review of structural gravity is followed by a development of iceberg trade costs to include external (to firms) scale effects that operate at international borders. The treatment of structural gravity is based on the ArmingtonCES demand system due to Anderson (1979).9 The structural gravity model (Anderson and van Wincoop, 2003) specifies that in each sector k the share of the world’s trade in k that flows from origin i to destination j is given by ⎛ Xijk = Y k ski bkj ⎝

tijk Pki Pjk

⎞1−sk ⎠

,

(1)

where, Xijk is the bilateral shipment, Yk is the world shipment from all origins to all destinations, ski = Yik /Y k is the share of world shipments coming from origin i, bkj = Ejk /Y k is the share of world shipments going from all origins to destination k, and ski bkj is the predicted pattern of trade in a frictionless world economy. All shipments are assumed to be valued at destination prices. The term in brackets gives the effect of frictions that drive trade away from the predicted frictionless pattern ski bkj . Bilateral iceberg trade costs tijk ≥ 1 melt away an initial shipment x so that only x/t arrives at its destination. Outward multilateral resistance Pki and inward multilateral resistance Pjk , are implied by the market clearance and budget constraint systems that lead to Eq. (1):

(Pki )1−sk =

 1−sk  tijk Pjk

j

(Pjk )1−sk =

 i



tijk

bkj

(2)

ski .

(3)

1−sk

Pki

Gravity equations are typically estimated using origin- and destination-time fixed effects to control for the shares ski , bkj and the

unobserved multilateral resistances Pki and Pjk . The total shipment Yk is typically not believably observed, so it is controlled for with a fixed effect (constant term). The controls for unobservable tijk are discussed below in Section 2.1. Gravity is estimated as Xijk = ck xki mkj (tijk )1−sk + 4ijk

2.1. Bilateral trade costs with scale effects Trade cost factors in the gravity literature are modeled as loglinear functions of a set of proxies. The main novelty of this paper is to bring scale effects in bilateral trade costs into the structural gravity model. The scale effects are external to the agents making trade volume decisions to avoid dealing with market power.10 Differential volume effects of scale are posited to potentially operate at international relative to intra-national entry points. A multiplicative component of bilateral trade costs is thus modeled as 0

Vij ij

(5)

where Vij is the physical volume shipped from i to j and the sectoral superscript k is suppressed to reduce notational clutter. 0ij = 0j Bij , where Bij is a dummy variable equal to 1 when i and j are separated by an international border and equal to 0 for intra-national trade. Destination specific scale elasticity 0j is negative in the increasing returns case, positive in the diminishing returns case and equal to zero in the standard constant returns case. Pair specific elasticities are a possibility but cannot be identified because pair fixed effects absorb all bilateral cost sources. Constancy of elasticity 0j over time turns out to be a less restrictive specification than it appears, since we find no evidence of its variation over time despite variation of bilateral volume. A variety of potential stories might be reflected in specification (5). We lean toward investment in capacities that are bilateralrelationship-specific. On the one hand, the bilateral-specific fixed factor suggests diminishing returns, hence positive 0j estimates.11 On the other hand, the econometric model identifies 0j off the variation over exporters of bilateral trade to j, so a positive correlation of observable bilateral trade cost proxies such as distance with the unobservable cost-reducing capacities results in negative 0j estimates. Moreover, if middlemen provide network links, Marshallian external economies of scale are plausible. (See the fascinating description in Urbina and Bradsher (2013) of the bilateral economies of scale provided by Li & Fung’s specialized network of middlemen.) Without the fully developed models and rich data that could sort out various causes of 0j = 0, an appropriate first step is to leave Eq. (5) as a black box. Specification (5) implies that exporters from i selling to j face a common component of bilateral trade costs that varies with aggregate bilateral volume Vij . While the standard iceberg trade cost is invariant to volume, 0j = 0, the key iceberg property remains in that 0

initial shipment x results in delivered goods x/t. But now Vij ij is a component of t. That is, resources used in shipment are in the same proportion as resources used in production.

(4) 2.2. Exchange rate effects on trade

where ck is the constant term controlling for Yk and mean measurement error in the Xijk ’s, xki and mkj are exporter and importer fixed ski

(Pki )1−sk

bkj

(Pjk )1−sk ,

and / respectively, and effects controlling for / 4ijk is an error term (which is multiplicative in some treatments but additive in ours, as we justify below). Because the full set of importer and exporter fixed effects is perfectly collinear with the constant vector, it is necessary (and harmless) to omit a base country so that mk0 xk0 for some country 0 is factored into the constant term c, along with the scaling term Yk and the mean measurement error in the trade flow data.

9 Gravity models of trade flows have a variety of consistent theoretical foundations that lead to equivalent representations at the sectoral level. See Anderson (2011) and Costinot and Rodriguez-Clare (2014) for details.

The time variation of Canadian trade with the US between 1997 and 2007 helps identify trade costs and scale elasticities but raises the question whether the dramatic movement of the exchange rate had a confounding effect on bilateral trade. Appreciation of the exchange rate is commonly treated theoretically as equivalent to a subsidy on imports and a tax on exports. The incomplete passthrough

10 More specifically, the rents to the fixed factors responsible for scale effects are external to the trade decisions. We regard market power as important in much of bilateral trade. However, the approximately constant markup property of CES structures minimizes the role of market power. 11 If all bilateral capacities are efficient, the envelope theorem implies 0j = 0. But trade patterns suggest that inefficiently low investment in cross-border trade is pervasive.

J. Anderson, M. Vesselovsky, Y. Yotov / Journal of International Economics 100 (2016) 174–193

177

literature (e.g. Goldberg and Knetter, 1997) implies that the equivalence is to that portion of exchange rate appreciation passed through to the import or export price.12 The portion of the appreciation of the currency that is passed through is modeled with parametric passthrough elasticity qj for some generic good from region i sold to region j.13 Including the scale effect specified in Eq. (5), the per unit trade cost is given by

delivered.15 Apply this deflator and the deterministic portion of the right hand side of Eq. (4) for Xij and combine with Eq. (6) to solve for volume as a reduced form function of the exogenous variables:

qj ri 0 tij = tij Vij ij , rj

Substituting Eq. (7) into Eq. (6) and simplifying,16 the reducedform bilateral trade cost function is

(6)

where ri is the appreciation factor of the currency for region i with respect to some base period and similarly for rj , hence ri /rj is the bilateral appreciation factor raising the cost of i s shipments to j. tij is the standard bilateral iceberg trade cost factor, a (log-linear) function of the standard set of trade cost gravity variables such as bilateral distance (see below). The direct effect (holding volume constant) of exchange rate appreciation on expenditures in j on goods from i in the CES system is given by the exchange rate component of bilateral trade costs (ri /rj )(1−s)qj . (ri /rj )qj is the passthrough to price paid by j, and is recognized as a component of bilateral trade cost tij when i and j are separated by a border. Multilateral resistances are also affected by (ri /rj )(1−s)qj acting in Eqs. (2)–(3). In the uniform passthrough case qj = q∀j, the multilateral resistances completely absorb the exchange rate changes and exchange rates have no real effects.14 In the non-uniform passthrough case, exchange rates have real effects in the gravity model through shifting (tijk /Pki Pjk )1−sk . Exchange rates act both directly through (ri /rj )qj and through their effect on volume via the scale effect. 2.3. Reduced form trade flows Volume depends on trade cost and trade cost depends on volume, but the log-linear specification leads to a tractable reduced form gravity equation. Volume is given by Vij = Xij /tij where the deflator tij removes both the effect of exchange rate appreciation (that raises ‘factory gate’ prices pi in terms of the numeraire currency used to convert trade flows to common value units) and the ‘volume’ used up in trade costs, thus specifying in Eq. (6) that trade cost is a function of volume

12 Exchange rate variation is plausibly exogenous (to sectoral bilateral trade) variation, but passthrough rates are much less so. The CES demand structure common to structural gravity and monopolistic competition suggests parametric markup and hence passthrough as approximately valid. Sensitivity experiments confirm the robustness of our main findings to endogeneity concerns. The experiments use lagged exchange rates and alternatively employ the average treatment effect methodology from Baier and Bergstrand (2007), who use it to successfully address trade policy endogeneity. The monopolistic competition literature suggests endogenous exchange rate passthrough due to Pricing to Market (PTM) by firms that rationally price discriminate. We do not address PTM in this paper based on industry level data. We failed to find meaningful evidence of PTM when using industry data along with strong symmetry assumptions about the unobserved distribution of firms. 13 This constant elasticity form is standard in empirical trade analysis; see Feenstra (2004), Chapter 7 for example. We can allow for different elasticity parameters in different regions and at different times, but always as exogenous parameters. 14 Divide Eq. (1) through by Yi Ej /Y to obtain size-adjusted trade on the left hand side. Substitute for tij on the right hand side using Eq. (6). The resulting version of the right hand side of system (1)–(3) is invariant to replacing initial values of 1/P0i Pj0 with exchange rate appreciated values

1 q

q

(P0i /ri )Pj0 rj

.

Then exchange rate changes have no real effects, and the purely nominal effects of exchange rate movements on Xij Y/Yi Ej on the left hand side of Eq. (1) cancel out in numerator and denominator.

 1/(1+s0ij ) Vij = xi mj tij−s (ri /rj )−1−qj s .

(7)

 1/(1+s0ij ) tij = tij (xi mj )0ij (ri /rj )qj −0ij , ∀i, j.

(8)

In the empirical analysis, the real effects of exchange rate movements are identified at the border where ri = rj in contrast to internal and interprovincial trade. The remainder of Eq. (8) differs from the usual gravity specification in that (i) tij is modified by being raised to a destination-specific power 1/(1 + s0ij ) and (ii) there is a cost-changing scale term (xi mj )0ij /(1+s0ij ) . The stability condition for the plausible quantity-adjustment mechanism in the bilateral trade market is 1 + s0ij > 0, i.e. IRS cannot be too strong. (Appendix B, available on request, presents the argument.) The trade flow model is completed by substituting the reduced form trade cost function (8) into the deterministic part of the structural gravity Eq. (4), for a generic good (1−s)/(1+s0ij )

Xij,t =c(1+0ij )/(1+s0ij ) (xi,t mj,t )(1+0ij )/(1+s0ij ) tij ×(ri,t /rj,t )(qj −0ij )(1−s)/(1+s0ij ) .

(9)

The determinants of tij comprise the usual list of geographic variables, each entering Eq. (9) with a coefficient to be inferred that combines with the exponent of tij , (1 − s)/(1 + s0ij ). Since 0ij = 0 for i, j in the same country, the variation of this exponent between intra-national and cross-border trade indicates scale effects.17 The rightmost term in Eq. (9) is the direct exchange rate effect. For qj − 0ij ≥ 0, appreciation of the bilateral exchange rate increases trade costs tij and thus decreases the value of trade at delivered (user) prices Xij or at origin (‘factory gate’) prices Xij /tij . Scale economies 0j < 0 amplify the effect of passthrough in decreasing trade because the decline in trade volume raises trade costs. (Notice that qj −0ij > 1 is possible, so standard models assuming 0ij = 0 may puzzlingly find passthrough elasticity greater than 1.) In contrast, decreasing returns 0j > 0 damp the effect of passthrough. For qj − 0ij < 0, a potential case we do not observe, appreciation decreases trade costs and thus increases the value of trade at delivered prices Xij or at origin prices Xij /tij . The intuition is that less of the iceberg melts: volume going through shipment falls by Eq. (7) and with qj < 0ij the net effect is a reduction in loss due to shipping costs, resulting in a gain in both factory gate priced exports and user priced exports.

15 Specifying Eq. (6) with volume defined at the origin, Xij /(ri /rj ), differs inessentially for purposes below. Also, switching from assuming value is measured in the data based on origin currency prices to measurement based on destination currency prices makes no essential difference; the deflation is by (ri /rj )qj . 16 At the first ste

 0ij /(1+0ij s) tij = tij (ri /rj )qj cxi mj tij−s (ri /rj )−1−qj s .

Collecting the exponents of ri /rj and tij and simplifying yields Eq. (8) below. The specification identifies cross-border scale effects that are normalized by any effective domestic scale effects. Thus, the scale effects in our theory (and empirics) are defined and should be interpreted as relative to domestic scale effects in intra-national trade. In principle, it is possible that intra-national trade flows may also be subject to scale effects. We test for such possibility in the sensitivity analysis. 17

178

J. Anderson, M. Vesselovsky, Y. Yotov / Journal of International Economics 100 (2016) 174–193

3. Econometric specification An important simplification preserves degrees of freedom. Let R = {CA, US} denote the destination country. Impose common scale and passthrough parameters across all Canadian destination provinces: 0ij = 0iR = 0R Bij .∀j ∈ R.18 For intra-national trade (within Canada or the US or ROW) Xij,t = cxit mjt tij1−s ; the term 0ij disappears from Eq. (9) because the international trade indicator variable Bij = 0 for those observations. In subsequent steps below, it is convenient to omit Bij = 1 in modeling typical trade flows as international, while leaving Bij = 0 for observations for internal and interprovincial trade to be implicit. For international trade flows it is convenient to rewrite Eq. (9) as follows, ∀j ∈ R:

1 + 0R 1−s Xij,t = exp kR + (ln xi,t + ln mj,t ) + ln tij 1 + s0R 1 + s0R  (qR − 0R )(1 − s) + ln(ri,t /rj,t ) 1 + s0R

(10)

where kR = (1 + 0R )/(1 + s0R ) ln c and R ∈ {US, CA}. The right hand side of Eq. (10) exponentiates a sum of three terms: the fixed effects term (including the constant), the trade costs term exclusive of exchange rate changes, and the exchange rate term. We develop the econometric treatment of these terms in order of their appearance. 3.1. Fixed effects specification 1+0R The fixed effects term in Eq. (10) is 1+s0 (ln xi,t + ln mj,t ) where R R refers to importer j s country when i and j are separated by a border (e.g. Canada when j is Ontario). The nonlinearity of the fixed effects term (in logs) is approximated by expanding a Taylor’s series about 0R = 0 and the average values of ln xi,t and ln mj,t . For example, for a typical Canadian importing province replace the theoretically exact value (ln xi,t + ln mj,t )(1 + 0can )/(1 + s0can ) ∀j ∈ CAN with ¯ (ln xi,t + ln mj,t ) + (ln x¯ + ln m)[(1 − s)0can /(1 + s0can )], where  ln x¯ = i,t xi,t /NT where N is the number of regions and T is the ¯ For a typical Canadian number of time periods and similarly for m. exporting province the analogous procedure yields an expression in which the roles of i and j exchange and 0can is replaced by 0usa .19 Due to collinearity the data is too weak to identify directional border fixed effects reflecting the potential difference between 0can and 0usa , hence in our application the two border fixed effects are combined. The specification of the fixed effects term in Eq. (10) under the considerations above simplifies to:

1 + 0R (ln xi,t +ln mj,t ) = gi,t +hj,t +bbrdr USA CAN, 1 + s0R

R = USA, CAN. (11)

Here, gi,t and hj,t are time-varying exporter and importer fixed effects, respectively. USA_CAN is a dummy variable equal to one for bilateral

18 For the US as a destination there is no state variation because our services data is available only for aggregate US origin and destination, and we impose the same aggregation on the goods data for comparability. 19 The first order Taylor’s series approximation for the exporter time fixed effect is

1 + 0R ln xi,t ≈ ln x¯ + [ln xi,t − ln x¯ ] + ln x¯ [(1 + 0R )(1 + s0R ) − 1] 1 + s0R = ln xi + ln x¯ (1 − s)0R /(1 + s0R ). The expansion is simplified to time invariance by imposing a single mean ln x¯ . In principle this can be relaxed to allow time-varying means, ln x¯ t . Allowing for time-varying means is inconsequential for our estimations because those will be absorbed by the time-varying fixed effects in our estimating equations.

trade between the US and Canada. bbrdr = busa,can + bcan,usa , where busa,can and bcan,usa are the directional border effects for US exports to Canada and for US imports from Canada, respectively. Finally, we note that the origin and destination fixed effects in Eq. (11) will absorb completely the destination-specific constant term kR in specification (10). 3.2. Trade cost specification As a first step, tij is specified as a function of bilateral distance, contiguity and borders. All border effects will be absorbed into the fixed effects specification described in Section 3.1. Three important implications of scale effects deviate from standard gravity treatments of these variables’ contribution to trade costs. First, owing to destination-specific trade-volume effects (i.e. 0can = 0usa ), the effects of international distance could differ across importers. Therefore, we split the international distance variable into its directional components. DIST CAN USA = ln DIST × can usa is defined as the product between the logarithm of bilateral distance between Canada and the United States (ln DIST) and a dummy variable (can_usa) that takes a value of one for Canadian exports to the United States and is equal to zero otherwise. Thus, DIST_CAN_USA is equal to zero for intra-national trade and for US exports to Canada. DIST USA CAN = ln DIST × usa can is defined as the product between the logarithm of bilateral distance between Canada and the United States (ln DIST) and a dummy variable (usa_can) that takes a value of one for US exports to Canada and is equal to zero otherwise. Thus, DIST_USA_CAN is equal to zero for intra-national trade and for Canadian exports to the US. Second, the effects of internal distance within each region, INTERNAL_DIST, can be different from the effects of international distance.20 Third, the model implies potential directional asymmetries in the effects of contiguity between Canadian provinces and US states, hence we split CONTIG in its directional components CONTIG_PR_ST, for provincial exports, and CONTIG_ST_PR, for provincial imports and allow for asymmetric scale effects that can strengthen or weaken the effects of contiguity. The effect of trade costs on volume (apart from pure border effects and exchange rates) becomes for R = {USA, CAN}: 1−s c1 (1 − s) ln tij =c1 (1 − s)INTERNAL DIST + DIST CAN USA 1 + s0R Bij (1 + s0can ) c1 (1 − s) c2 (1 − s) DIST USA CAN + CONTIG PR ST (1 + s0usa ) (1 + s0usa ) c3 (1 − s) (12) + CONTIG ST PR. (1 + s0can ) +

3.3. Exchange rate effects specification The exchange rate effect on log imports of j from i is ln(ri,t /rj,t ), ∀j ∈ R, i ∈ / R. In the essentially two coun-

(qR −0R )(1−s) 1+s0R

try applications,the abstract ri,t /rj,t is conveniently replaced with a bilateral exchange rate appreciation/depreciation. Define rcan,t as Canadian dollars per US dollar at time t relative to the same ratio in the base year, 1997. Thus a rise is a depreciation in the Canadian −0usa )(1−s) dollar. If (qusa1+s0 > 0, a depreciation of the Canadian dollar usa should lead to an increase in Canadian exports to US. The exchange rate effect on log imports for province i depends on the negative of −0can )(1−s) ln rcan,t . Thus if (qcan1+s0 > 0, a depreciation of the Canadian can dollar should be associated with a fall in Canadian imports from US. To model exchange rate effects on trade in both directions, let CAN_USA and USA_CAN be dummy variables for Canadian exports

20 In the sensitivity analysis we also experiment by allowing for differential effects of distance on intra-provincial vs. inter-provincial trade within Canada.

J. Anderson, M. Vesselovsky, Y. Yotov / Journal of International Economics 100 (2016) 174–193

(to US) and for Canadian imports (from US), respectively. The exchange rate effect on the generic flow is given by (qusa − 0usa )(1 − s) CAN USA × ln rcan,t 1 + s0usa (qcan − 0can )(1 − s) − USA CAN × ln rcan,t . 1 + s0can

ber

exp ER

CA EXP + ber

imp ER

CA IMP,

ea5 CONTIG

(13)

where ER CA EXP = CAN USA × ln rcan,t and ER CA IMP = USA CAN × ln rcan,t . The interactive fixed effect structure of Eq. (13) combines with directional border fixed effects in the full model Eq. (10) to imply that the estimates of the exchange rate effects in each direction are deviations from the corresponding directional border fixed effect esti−0usa )(1−s) mates for CA–UStrade. ber exp = (qusa1+s0 − bcan,usa measures usa the ER effects on Canadian exports relative to the corresponding directional border effect, bcan,usa , which is defined above following −0can )(1−s) − busa,can measures the Eq. (11). Similarly, ber imp = − (qcan1+s0 can relative ER effects on Canadian imports. The two exchange rate terms in Eq. (13) and the time-varying directional province fixed effects in Eq. (11) are perfectly collinear. To allow estimation, we drop the ER term for Canadian imports from the US. The interpretation of the fixed effects is that the dropped term is subtracted from the ‘true’ fixed effect, and the estimate of the ER effect on Canadian exports to the US includes the ER term for Canadian imports. Eq. (13) as it is estimated becomes: ber ER CA EXP,

(14)

where, making use of the definitions of the directional ER estimates,

(qusa − 0usa )(1 − s) − bcan,usa 1 + s0usa

(qcan − 0can )(1 − s) − busa,can . + − 1 + s0can

ber = ber

exp

+ ber

imp

=

econometric specification (for a generic sector) that incorporates exchange rates, in addition to the standard set of gravity covariates:

Xij,t =ea0 +a1 INTERNAL

In terms of estimated coefficients, the preceding expression is rewritten as:

(15) Eq. (15) is an important relationship between the relative border and the relative ER estimates, used in the empirical section to recover some of the structural parameters in our model.21 Note that identification of the exchange rate effect in the structural gravity setting depends on its differential action between interprovincial trade (where it is shut off) and cross-border trade. The same differential action characterizes our treatment of scale effects. Both effects are identified using interregional and international trade data simultaneously.

179

DIST+a2 DIST CAN USA+a3 DIST USA CAN+a4 CONTIG PR ST

ST PR+bbrdr USA CAN+ber ER CA EXP+hj,t +gi,t

+ 4ij,t .

∗

(16)

Specification (16) is estimated with the Poisson Pseudo Maximum Likelihood (PPML) technique recommended by Santos Silva and Tenreyro (2006).22 c1 (1−s) The coefficients a 1 = c1 (1 − s), a2 = (1+s0 , and a3 = ) c1 (1−s) (1+s0usa )

can

capture the effects of distance on trade. If the estimate of the effect of internal distance, a 1 , is smaller in absolute value (all distance elasticities are negative) than the effects of international distance, a 2 and a 3 , this is evidence for economies of scale in bilateral trade. A statistically significant difference in the magnitudes of the estimates of a 2 and a 3 rejects the hypothesis of scale neutrality. With independent information on the elasticity of substitution s, the scale parameters are identified from the estimates of a 1 , a 2 , a 3 and the structural restrictions. c2 (1−s) c3 (1−s) a4 = (1+s0 and a5 = (1+s0 capture the effects of contiusa ) can ) guity between a province and a state. Estimates of these coefficients are expected to be positive and we can test whether scale effects contribute to directional asymmetries, if any. The structural model also yields a test of whether passthrough is uniform. Rearranging Eq. (15) and using bbrdr = busa,can + bcan,usa

ber + bbrdr =

(qusa − 0usa )(1 − s) (qcan − 0can )(1 − s) − . 1 + s0usa 1 + s0can

When the estimated bˆ er + bˆ brdr does not differ significantly from zero and the scale neutrality hypothesis aˆ 3 − aˆ 2 = 0 also cannot be rejected, the implication of the preceding equation is that passthrough uniformity cannot be rejected because the right hand side is equal to zero if qusa = qcan and 0usa = 0can .23 Non-uniform passthrough in the setting of this paper affects the distribution of goods within a sector. Departures from uniformity (neutrality) in any sector also imply relative price changes between sectors that shift sales and expenditure shares. These additional real effects on the economy are controlled for in the sectoral gravity equations with origin and destination time fixed effects.

3.5. Data This project features a comprehensive data set that covers most of Canada’s economy at the sectoral level for a total of 28 industries including agriculture, fuels, 17 manufacturing sectors, and 9

3.4. Full econometric specification Substitute Eqs. (11), (12) and (14) into the deterministic gravity Eq. (10) and add an error term 4ij,t to obtain the following

21 To build intuition and trust in the structural use of collinearity in Eq. (15), in the empirical analysis (see Table 2) we show that estimating the two coefficients ber exp and ber imp through the expedient of dropping another origin country or province fixed effect yields a sum equal to the estimate of ber as described above. The same alternative regression demonstrates that bbrdr = bus,ca + bca,us .

22 We rely on the PPML estimator because it simultaneously addresses the issue of heteroskedasticity in trade data and takes advantage of the zero trade flows in the data. We refer the reader to Head and Mayer (2014) for an insightful discussion on the appropriate choice among alternative gravity estimators including OLS, Poisson PML, and Gamma PML. Following the recommendations of Head and Mayer (2014), in analysis that is available by request, we demonstrate that the three estimators deliver similar results with our sample. 23 Given uniformity, Eq. (15) can be rearranged to yield a relationship between the uniform q and s. Thus, given s the uniform q value is estimated or given the value of uniform q the value of s is estimated. Our coefficient estimates under this procedure yield unreported results that are reasonable for either q or s in most sectors.

180

J. Anderson, M. Vesselovsky, Y. Yotov / Journal of International Economics 100 (2016) 174–193

service categories for the period 1997–2007.24 The choice of the 1997–2007period is due to coverage limitations of our services data set. Our sample consists of a total of 13 regions including 12 Canadian provinces and territories and the United States.25 In order to estimate gravity, we use industry-level data on bilateral trade flows and output for each trading partner (including all Canadian provinces and territories and US), all measured in current (’00,000) Canadian dollars, as well as other variables which we describe below. The data notably includes consistent production, expenditure and trade data on inter-provincial, intraprovincial and international trade flows from Statistics Canada. As described in Genereux and Langen (2002), every effort has been made by the developers of the dataset to ensure completeness and internal consistency of trade, production, and expenditure data. Despite the efforts of Statistics Canada, there are still some possible data caveats that may affect our results. For example, international trade flows are traditionally better measured than internal trade flows. One possible explanation for that is that import tariff revenues have long been the most important source of revenues for many countries. Another possible explanation is that, by construction, there are two potential sources of measurement error for intra-national trade as they are obtained as the difference between total gross output and total exports. A third potential source of measurement error is that the original data come from various sources. For example, Genereux and Langen (2002) and Agnosteva et al. (2014) point to some deficiencies of the Canadian Annual Survey of Manufacturers (ASM), which is collected at the establishment level and is the primary source of data for interprovincial trade flows of manufactured goods. “For instance, the ASM does not collect destination information for small manufacturing establishments, which account for close to 8% of all Canadian manufacturing shipments and vary by province. Instead, in the ASM, these shipments are recorded as purchased by consumers within the province of the establishment. While, it is reasonable to assume that most of the shipments of the small firms are indeed local, it is unrealistic to assume that all shipments are local. This implies that, by construction, our sample may overstate the importance of intra-provincial trade.” (Agnosteva et al., 2014).26 We test for this possibility in the sensitivity analysis, where we allow for scale effects in interprovincial vs. intra-provincial trade. We find no evidence of such effects in the goods sectors in our sample, but we cannot exclude such possibility in the case of services. Next, we briefly discuss the main variables in our sample and the data sources used to construct them. Trade flows data. Statistics Canada’s Table 386-0002 is the original data source for intra-provincial and interprovincial trade flows

24

The sector selection was based on (but is not completely identical to) the S-level of aggregation as classified in the Statistics Canada’s Hierarchical Structure of the I-O Commodity Classification (Revised: November 3, 2010). Our sample covers 28 sector categories. A table with our sectoral coverage and a detailed description of each of the sector categories in our sample are presented in Appendix A. The few commodities missing from the complete S-level I-O Commodity Classification spectrum are Forestry Products, Fish, Metal Ores, and Tobacco and Beverages. Reliable bilateral trade data ware not available for those products. Finally, we sometimes aggregate all goods (GOODS) and all services (SRVCS). 25 We aggregate the Northwest Territories and Nunavut in one unit, even though they are separate since April 1st, 1999. In the sensitivity analysis, we extend our sample to also include sectoral trade in manufacturing goods between Canada and Mexico. 26 For a detailed description of the challenges, advantages, and caveats with the construction of the Canadian intra-provincial, interprovincial, and international trade data, we refer the reader to Genereux and Langen (2002) and to Agnosteva et al. (2014), who use these data to study intra-national trade costs in Canada.

for both goods and services.27 Data on shipments between Canadian provinces and the United States are from the Trade Data Online web interface of Industry Canada, which provides access to Canadian and US trade data by product classified according to NAICS; the NAICS sectors were then matched or aggregated to the S-level. Internal trade for US and for the Mexican manufacturing sectors, which we use in the sensitivity analysis, isobtained as the difference between output and total exports. Constructing and using intranational trade data for the US and for the Mexican economies as a whole, instead of using trade within and between states areanother potential source of measurement error. While we do not have access to inter-state and intra-state Mexican trade data, in principle, it is possible to construct inter-state trade data for the United States from the US Commodity Flow Survey. However, from previous experience we know that the US Commodity Flow Survey data are contaminated with measurement error due to the inclusion of entrepôt trade and other known issues. Therefore, we decided not to use the latter. However, in analysis that is available by request, we check the robustness of our results by dropping the intra-national observations for U.S. to find that the distance gravity estimates and the corresponding estimates of the scale parameters with and without intra-national US trade are virtually identical. Output data. Provincial output, defined here as the value of production plus shipments out of the inventories of producers, wholesalers and retailers is from Statistics Canada’s Table 386-0002. All zero values and blank cells in the output data are treated as missing information and linearly interpolated.28 Output data for the United States come from several sources. Manufacturing data are from the UNIDO Industrial Statistics database, which reports industry-level output data at the 3- and 4-digit level of ISIC code. Output for Agriculture and Mineral Fuels, 1997–2003,isfrom Anderson and Yotov (2010). The original sources of these data are the United Nations Food and Agriculture Organization (FAOSTAT) web page, which provides data on agricultural output, and the Energy Information Administration, which provides official energy statistics on the value of fuel production (including oil, natural gas, and coal). Finally, services output data are from Anderson et al. (2014). The US Bureau of Economic Analysis is the original source for US service production data. Other variables. We use the bilateral distances data from Anderson and Yotov (2010), who follow Mayer and Zignago (2006) to obtain population-weighted bilateral distances. This procedure is consistent with respect to calculating both internal and bilateral distances. See Anderson and Yotov (2010) for more details. Exchange rates data are from the Federal Reserve Bank of Saint Louis’ web site at http:// research.stlouisfed.org/fred2/categories/15. Finally, we construct a series of border and regional dummy variables, which are described in the text.

27 The actual services data used here (including trade, output and expenditures) is from Anderson et al. (2014). Please see their data section and data appendix for further details. The services data were constructed and provided by Statistics Canada and these data do not include any missing values. Sectoral goods data include missing values. 28 Two reasons drove our decision to recover missing output data. First, output data is needed in order to construct the values for intra-provincial trade. The latter is important for identification in the paper. Therefore, we wanted to have a complete set of observations for internal trade. Second, more importantly, we had to interpolate missing output values in order to be able to aggregate the individual goods and services sectors into aggregate categories. It may be desirable to treat missing output (internal trade) values as missing at the sectoral level, but these values are automatically treated as zeros in summing sectoral data to obtain the aggregate goods and services categories. The better alternative is to interpolate the missing values rather than to treat them as zeros under the certainty of positive values. Thus, for consistency the constructed missing values are also used at the sector level. In analysis that is available by request we demonstrate that sectoral results obtained without filling in the missing output values are not statistically different from our main findings.

J. Anderson, M. Vesselovsky, Y. Yotov / Journal of International Economics 100 (2016) 174–193

181

Table 1a Sectoral PPML Panel Gravity Estimates, 1997–2007.

A. Gravity estimates INTERNAL_DIST DIST_USA_CAN DIST_CAN_USA CONTIG_PR_ST CONTIG_ST_PR ER_CA BRDR_USA_CAN N B. Scale effects specification tests a1 − a2 a1 − a3 a2 − a3

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

GOODS

AGRIC

FUELS

FOOD

LETHR

TXTLE

APPRL

WOOD

FRNTR

PAPER

−0.728 (0.038)∗∗ −2.994 (0.307)∗∗ −1.166 (0.277)∗∗ 0.281 (0.301) 1.467 (0.292)∗∗ −0.002 (0.215) 16.235 (2.327)∗∗ 1014

−1.194 (0.040)∗∗ −1.344 (0.518)∗∗ −0.996 (0.361)∗∗ 0.703 (0.412)+ 4.121 (0.450)∗∗ 3.702 (0.844)∗∗ −6.886 (4.201) 835

−0.839 (0.239)∗∗ −0.010 (1.342) 2.832 (2.644) 1.965 (0.925)∗ −0.121 (1.043) 1.086 (1.113) −38.570 (20.321)+ 435

−0.621 (0.030)∗∗ −2.529 (0.409)∗∗ −0.116 (0.228) −0.387 (0.289) 2.482 (0.418)∗∗ −0.647 (0.272)∗ 3.615 (3.295) 964

−0.679 (0.036)∗∗ −3.479 (0.306)∗∗ −2.077 (0.257)∗∗ −1.183 (0.245)∗∗ 2.826 (0.323)∗∗ 0.205 (0.142) 26.925 (2.640)∗∗ 888

−0.469 (0.034)∗∗ −2.847 (0.280)∗∗ −1.876 (0.422)∗∗ 0.017 (0.282) 3.246 (0.295)∗∗ −0.738 (0.378)+ 21.099 (3.661)∗∗ 812

−0.558 (0.034)∗∗ −2.477 (0.381)∗∗ −1.215 (0.437)∗∗ 2.512 (0.437)∗∗ 3.976 (0.369)∗∗ −1.449 (0.489)∗∗ 8.296 (4.576)+ 705

−0.816 (0.037)∗∗ −0.836 (0.708) −0.331 (0.270) 0.923 (0.276)∗∗ 4.818 (0.745)∗∗ 1.664 (0.263)∗∗ −11.124 (5.985)+ 931

−0.667 (0.041)∗∗ −3.886 (0.514)∗∗ −1.146 (0.243)∗∗ 1.085 (0.310)∗∗ 2.984 (0.460)∗∗ 0.160 (0.325) 20.336 (3.975)∗∗ 789

−0.663 (0.029)∗∗ −2.773 (0.372)∗∗ −0.193 (0.233) 0.476 (0.203)∗ 2.839 (0.537)∗∗ −0.083 (0.135) 6.584 (3.020)∗ 865

0.150 (0.516) −0.198 (0.361) −0.347 (0.710)

−0.830 (1.465) −3.672 (2.497) −2.842 (3.213)

1.908 (0.404)∗∗ −0.505 (0.234)∗ −2.414 (0.509)∗∗

2.800 (0.303)∗∗ 1.398 (0.246)∗∗ −1.402 (0.439)∗

2.378 (0.287)∗∗ 1.406 (0.412)∗∗ −0.972 (0.530)+

1.919 (0.378)∗∗ 0.657 (0.419) −1.263 (0.545)∗

0.020 (0.701) −0.485 (0.246)∗ −0.505 (0.730)

3.219 (0.502)∗∗ 0.479 (0.242)∗ −2.739 (0.603)∗∗

2.111 (0.371)∗∗ −0.469 (0.230)∗ −2.580 (0.484)∗∗

2.266 (0.300)∗∗ 0.438 (0.281) −1.828 (0.500)∗∗

Notes: This table reports PPML gravity estimates of the effects of exchange rate fluctuations on Canadian goods trade. Column (1) includes estimates for aggregate goods and the numbers in the next nine columns are for individual sectors. All estimates are obtained with time-varying, directional (importer and exporter) fixed effects and the years included in our sample are 1997, 1999, 2001, 2003, 2005 and 2007. The dependent variable is nominal exports. Standard errors are clustered by pair and are in parentheses. See text for more details. + p < 0.10. ∗ p < .05. ∗∗ p < .01.

4. Estimation results and analysis

4.1. Gravity estimates

We estimate Eq. (16) with the Poisson pseudo-maximumlikelihood (PPML) estimator. Santos Silva and Tenreyro (2006) propose PPML to simultaneously address the prominent presence of zeroes and of heteroskedasticity in bilateral trade flows data. The Poisson error term is assumed to be independently drawn across time, but we use 2-year intervals rather than a simple panel because Cheng and Wall (2005) note that “[f]ixed-effects estimations are sometimes criticized when applied to data pooled over consecutive years on the grounds that dependent and independent variables cannot fully adjust in a single year’s time.”(p.8).29 Tables 1a–1c report results from estimating Eq. (16) for each sector. The first column of Table 1a presents estimates for all goods and the remaining columns of Tables 1a and 1b present estimates for the 19 individual goods sectors. Similarly, the first column of Table 1c presents aggregate estimation results for all service sectors, and the remaining columns in the table report estimates for the 9 individual service sectors. Table 2 reports results of dropping alternative dummy variables from the overall goods trade regression of Table 1a. These results illustrate the collinearity structure of the data that is used to interpret results. Coefficient estimates from Tables 1a–1c are used with theoretically based identifying restrictions to recover scale parameters and to infer substitution and passthrough elasticities. Results are reported in Tables 3a, 3b and 3c. Finally, sensitivity experiments are offered in Tables 4, 5, and 6a–6b.

Overall, the PPML estimates from Tables 1a–1c give the usual good fit of gravity for both disaggregated goods and services. The coefficient estimates of each of the gravity covariates are discussed in the order in which they appear in econometric specification (16). Internal Distance. Distance is a significant impediment to internal trade, just as it is for international trade. All the estimates of the effects of internal distance on trade are statistically significant at any level and for each sector in our sample. Variation of the effects of internal distance over sectors is mostly intuitive. For example, the largest estimates among the goods sectors are for Agriculture, Printing and Minerals, while the largest estimates among the services categories are for Health services, Education services, Finance services and Other services (beauty and personal care, funeral, child care, household, automobile repairs and recreation among others). Most activities here are strongly locally biased. Comparison between the aggregate distance elasticities for goods (see column 1 of Table 1a) and for services (see column 1 of Table 1c) reveals that the latter are, on average, larger in magnitude: Services are on average more localized. Finally, we note that our estimates of the effects of internal distance are not sensitive to the exclusion of US, the single largest region in the sample. International Distance. Most of the estimates on the international distance variables are significant at any conventional level of statistical significance. Notable exceptions, where the estimates of the effects of international distance are not statistically significant, are some resource sectors such as Fuels, Petroleum and Coal Products and Wood Products. More novel and important, we find significant asymmetries in the effects of international distance between Canada’s exports and

29 Olivero and Yotov (2012) confirm the relevance of this issue by experimenting with various intervals in a dynamic gravity setup.

182

J. Anderson, M. Vesselovsky, Y. Yotov / Journal of International Economics 100 (2016) 174–193

Table 1b Sectoral PPML Panel Gravity Estimates, 1997–2007.

A. Gravity estimates INTERNAL_DIST DIST_USA_CAN DIST_CAN_USA CONTIG_PR_ST CONTIG_ST_PR ER_CA BRDR_USA_CAN N B. Scale effects specification tests a1 − a2 a1 − a3 a 2 -a 3

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

PRNTG

METL1

METL2

MCHNS

VHCLS

ELCTR

MNRLS

PETRL

CHMCL

MISCL

−1.022 (0.032)∗∗ −4.804 (0.524)∗∗ −0.243 (0.466) 2.488 (0.578)∗∗ 2.709 (0.458)∗∗ 0.239 (0.181) 11.768 (4.951)∗ 911

−0.741 (0.065)∗∗ −2.509 (0.268)∗∗ −0.645 (0.481) 2.627 (0.614)∗∗ 2.424 (0.418)∗∗ −2.000 (0.249)∗∗ 5.262 (3.654) 821

−0.794 (0.030)∗∗ −2.960 (0.286)∗∗ −1.757 (0.205)∗∗ 0.320 (0.281) 1.852 (0.403)∗∗ 0.489 (0.273)+ 17.842 (2.122)∗∗ 943

−0.757 (0.048)∗∗ −2.616 (0.252)∗∗ −2.365 (0.286)∗∗ 0.619 (0.256)∗ 2.160 (0.266)∗∗ 0.423 (0.464) 22.157 (1.986)∗∗ 903

−0.747 (0.030)∗∗ −4.412 (0.212)∗∗ −2.491 (0.201)∗∗ 1.343 (0.210)∗∗ 0.348 (0.262) 0.136 (0.254) 37.220 (1.576)∗∗ 891

−0.434 (0.033)∗∗ −2.874 (0.203)∗∗ −0.984 (0.181)∗∗ 1.071 (0.330)∗∗ 1.555 (0.270)∗∗ −0.581 (0.204)∗∗ 17.948 (2.050)∗∗ 866

−1.037 (0.038)∗∗ −3.289 (0.411)∗∗ −1.469 (0.354)∗∗ 1.843 (0.402)∗∗ 3.104 (0.463)∗∗ 2.514 (0.285)∗∗ 12.121 (3.402)∗∗ 837

−1.016 (0.056)∗∗ −0.297 (0.696) −0.919 (0.653) 0.292 (0.919) −0.022 (0.592) 0.211 (0.622) −9.893 (4.826)∗ 805

−0.585 (0.039)∗∗ −3.293 (0.295)∗∗ −1.146 (0.295)∗∗ 1.350 (0.376)∗∗ 2.508 (0.583)∗∗ −0.938 (0.242)∗∗ 17.555 (2.784)∗∗ 911

−0.687 (0.040)∗∗ −4.215 (0.301)∗∗ −1.072 (0.298)∗∗ 0.766 (0.512) 2.414 (0.290)∗∗ −0.462 (0.236)+ 22.426 (2.751)∗∗ 887

3.783 (0.516)∗∗ −0.778 (0.470)+ −4.561 (0.758)∗∗

1.768 (0.274)∗∗ −0.096 (0.455) −1.864 (0.597)∗

2.166 (0.278)∗∗ 0.963 (0.207)∗∗ −1.204 (0.413)∗

1.860 (0.257)∗∗ 1.608 (0.274)∗∗ −0.252 (0.468)

3.665 (0.202)∗∗ 1.744 (0.202)∗∗ −1.921 (0.351)∗∗

2.440 (0.195)∗∗ 0.550 (0.178)∗ −1.889 (0.270)∗∗

2.251 (0.401)∗∗ 0.432 (0.359) −1.819 (0.627)∗

−0.719 (0.672) −0.096 (0.676) 0.622 (1.201)

2.708 (0.288)∗∗ 0.561 (0.299)+ −2.147 (0.472)∗∗

3.528 (0.286)∗∗ 0.386 (0.308) −3.142 (0.482)∗∗

Notes: This table reports PPML gravity estimates of the effects of exchange rate fluctuations on sectoral goods trade in Canada. All estimates are obtained with time-varying, directional (importer and exporter) fixed effects and the years included in our sample are 1997, 1999, 2001, 2003, 2005 and 2007. The dependent variable is nominal exports. Standard errors are clustered by pair and are in parentheses. See text for more details. + p < 0.10. ∗ p < .05. ∗∗ p < .01.

Canada’s imports. Our estimates suggest that distance is a larger impediment to trade for Canadian imports of both goods and services. See Panels B of Tables 1a–1c, where we obtain statistically significant differences between the effects of distance on Canadian imports and exports for fourteen of the nineteen goods sectors in our sample and for five of the nine services sectors. These findings lend support to our theoretical predictions for destination-specific distance effects. Specifically, based on the structural definitions of the distance coefficients (a 1 = c1 (1 − s), c1 (1−s) c1 (1−s) a2 = (1+s0 , a3 = (1+s0 ), the estimates on DIST_CAN_USA and can ) usa ) DIST_USA_CAN suggest an IRS relationship between trade volume and trade costs (0R < 0, R ∈ {CA, USA}), which is more pronounced for Canada’s imports, 0can < 0usa . See Tables 3a–3c and corresponding discussion below for details. This implication should be treated cautiously, because it is sensitive to imprecisely estimated distance elasticities from Canada’s provinces to the single US market in contrast to the more precisely estimated distance elasticities for Canada’s imports. The greater imprecision for the US destination is due to less variation in the bilateral distance data. Service trade data limitations required aggregating trade to the US destination. As a result, only 3 of 9 service sectors and 12 of 19 goods sectors have significant distance elasticities for export to the US. The only two sectors for which we obtain negative and significant estimates on DIST_USA_CAN are Transportation and Accommodation.30 Aggregation bias is not a glaring issue

30 The estimate on DIST_USA_CAN for Health services is positive and statistically significant, while the corresponding estimate in the opposite direction (on DIST_CAN_USA for Health) is the largest of all negative distance estimates. Both findings indicate that the trade cost function for Health is mis-specified, especially the positive distance elasticity that violates the stability condition. For purposes of this study it is useful to maintain a single specification of the trade cost function but a serious treatment of gravity for health services trade should alter the specification.

because all bilateral distance variables are consistently aggregated by construction as population-weighted aggregates of city-pair distances. Panels B of Tables 1a–1c also report statistically significant differences between the effects of internal and international distance, a 1 − a 2 and a 1 − a 3 . We obtain statistically and economically smaller effects of internal distance as compared to the effects of distance on Canadian imports, a 1 − a 2 , for fifteen of the nineteen goods sectors and for five of the nine services categories. The findings from comparing the effects of internal distance and the effects of distance on Canadian exports to US are mixed. We find that the effects of internal distance, a 1 − a 3 , are smaller for eight of the nineteen goods sectors and for none of the nine services sectors. We also estimate smaller effects of distance on CA exports for four goods and four services sectors. Usually, these results are driven by an insignificant estimate of the effects of distance on CA exports to US. The much smaller (in absolute value) estimated effects of internal distance in most cases suggest that the scale effects introduced in Section 2 are indeed operational. The results suggest an IRS relationship (0R < 0, R ∈ {CA, USA}) between trade volume and trade costs, confirmed in Tables 3a–3c discussed in Section 4.2 below. Notable exceptions are Fuels, Petroleum and Coal Products and Wood Products, where the estimates of the effects of international distance are not only smaller in magnitude but also not statistically significant. The specific modes of transportation in these sectors in cases where pipelines are used may be a natural explanation, but the finding also suggests that the cost function is too crude to accurately represent the reality. Contiguity. Most of the contiguity estimates from Tables 1a–1c are positive and statistically significant in each direction of Canadian trade. We also find evidence for directional asymmetries in the effects of contiguity. In the case of goods trade, contiguity raises Canada’s imports but not its exports on average, evidenced by the

J. Anderson, M. Vesselovsky, Y. Yotov / Journal of International Economics 100 (2016) 174–193

183

Table 1c Sectoral PPML Panel Gravity Estimates, 1997–2007.

A. Gravity estimates INTERNAL_DIST DIST_USA_CAN DIST_CAN_USA CONTIG_PR_ST CONTIG_ST_PR ER_CA BRDR_USA_CAN N B. Scale effects specification tests a1 − a2 a1 − a3 a2 − a3

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

SRVCS

TRNSP

CMNCN

WHLSL

FNNCE

BUSNS

EDCTN

HELTH

ACMDN

OTHER

−1.162 (0.055)∗∗ −2.712 (0.520)∗∗ −0.478 (0.527) 0.954 (0.405)∗ 0.247 (0.357) −0.055 (0.171) −0.541 (2.477) 1014

−0.944 (0.043)∗∗ −1.563 (0.404)∗∗ −1.341 (0.401)∗∗ 0.598 (0.356)+ 0.777 (0.354)∗ 0.308 (0.142)∗ 0.892 (2.332) 1014

−1.044 (0.059)∗∗ −1.734 (0.523)∗∗ −0.364 (0.483) 0.352 (0.374) 0.145 (0.370) −0.745 (0.342)∗ −7.365 (2.060)∗∗ 1014

−0.911 (0.039)∗∗ −2.300 (0.375)∗∗ −0.428 (0.334) −0.006 (0.348) −0.077 (0.399) 0.153 (0.299) −3.691 (2.439) 1014

−1.372 (0.095)∗∗ −4.382 (0.780)∗∗ 0.123 (0.692) 1.442 (0.532)∗∗ −0.464 (0.511) −0.819 (0.219)∗∗ 3.870 (3.070) 1014

−1.143 (0.054)∗∗ −3.287 (0.514)∗∗ −0.350 (0.471) 1.123 (0.363)∗∗ −0.000 (0.401) 0.751 (0.430)+ 4.213 (2.844) 1014

−1.764 (0.087)∗∗ −2.840 (1.013)∗∗ −0.191 (1.015) 0.700 (0.692) −0.086 (0.681) −0.228 (0.353) −8.924 (3.912)∗ 1014

−2.745 −1.336 −1.382 (0.099)∗∗ (0.041)∗∗ (0.066)∗∗ −6.028 −0.895 −2.520 (0.766)∗∗ (0.437)∗ (0.650)∗∗ 2.059 −1.780 −0.137 (0.862)∗ (0.438)∗∗ (0.660) 1.180 0.248 1.282 (0.733) (0.352) (0.476)∗∗ −0.888 1.079 0.549 (0.639) (0.383)∗∗ (0.432) 1.488 −0.298 −0.154 (0.215)∗∗ (0.251) (0.565) −15.907 −4.497 −6.513 ∗∗ (4.559) (3.395) (3.311)∗ 1014 1014 1014

1.550 (0.506)∗ −0.684 (0.541) −2.234 (0.998)∗

0.620 (0.411) 0.398 (0.391) −0.222 (0.746)

0.689 (0.503) −0.680 (0.502) −1.369 (0.970)

1.389 (0.366)∗∗ −0.483 (0.342) −1.872 (0.636)∗

3.010 (0.711)∗∗ −1.495 (0.761)∗ −4.505 (1.418)∗

2.144 (0.508)∗∗ −0.793 (0.478)+ −2.937 (0.917)∗

1.076 (1.013) −1.573 (1.020) −2.649 (1.968)

3.284 (0.731)∗∗ −4.804 (0.899)∗∗ −8.087 (1.526)∗∗

−0.441 (0.434) 0.444 (0.434) 0.885 (0.754)

1.139 (0.620)+ −1.245 (0.690)+ −2.383 (1.241)+

Notes: This table reports PPML gravity estimates of the effects of exchange rate fluctuations on Canadian services trade. Column (1) includes estimates for aggregate services and the numbers in the next nine columns are for individual sectors. All estimates are obtained with time-varying, directional (importer and exporter) fixed effects and the years included in our sample are 1997, 1999, 2001, 2003, 2005 and 2007. The dependent variable is nominal exports. See text for more details. + p < 0.10. ∗ p < .05. ∗∗ p < .01.

aggregate goods estimates from column (1) of Table 1a. There is a large, positive and statistically significant estimate on CONTIG_ST_PR, capturing contiguity effects on Canada’s imports, but a statistically insignificant estimate on CONTIG_PR_ST for contiguity effects on Canada’s exports. Exactly the opposite is true for services trade, where we obtain a positive and statistically significant estimate on CONTIG_PR_ST for Canadian services exports but an insignificant estimate on CONTIG_ST_PR. See column (1) of Table 1c. In the next section, we analyze the contribution of trade-volume effects for the magnitudes of the contiguity estimates and for the directional asymmetries between them. International Borders. The novelty of generalizing the trade cost function to have distance and contiguity responses that potentially vary by destination has parallel potential consequences for the border effect coefficient estimates. Based on specification (11), the border coefficient is a time-invariant sector specific trade balance effect. This explains the large magnitude and varying signs of the border coefficient estimates in Tables 1a–1c. The estimates imply that Canadian imports from US, all else equal, are larger than Canadian exports to US in all but five goods sectors: Agriculture, Fuels, Wood, and Petroleum and Coal Products, with Food being statistically insignificantly affected. The picture is quite different for services, where most of the sectoral border estimates of bbrdr are not statistically significant in column (1) of Table 1c. The four significant services border estimates are all negative, which suggests larger Canadian exports in Communication, Education, Health and Other services, all else equal. It is useful for interpreting results to demonstrate that the estimates of bbrdr from Tables 1a–1c capture the border effects on Canadian imports from US plus the border effects on Canadian exports to US. In particular, Table 2 demonstrates empirically that bbrdr = busa,can + bcan,usa . In column (1) of Table 2, we reproduce

our main gravity results for aggregate goods trade from the first column of Table 1a.31 In column (2), we include a dummy variable for Canada’s exports to US, gcan,usa , in addition to the border dummy for Canada’s imports from the main specification, gusa,can . In order to estimate both directional border coefficients busa,can and bcan,usa , we drop one of the exporter, time-varying fixed effects. Then, at the bottom panel of Table 2, we show that the sum of the directional border estimates from column (2) is exactly equal to the relative border estimate from column (1), i.e. bbrdr = busa,can + bcan,usa . Finally, in column (3), we reproduce the experiment after omitting a different exporter-time fixed effect. As expected, the directional border estimates from columns (2) and (3) are different, however, their sum is the same and, once again, equal to the relative border estimate from column (1). Exchange Rates. There is wide variability in the relative ER effects across sectors. For some industries, such as Agriculture, Wood, Minerals and Health services, we obtain large, positive and significant estimates on ER_CA_EXP, which suggest that the ER effects on Canadian exports dominate the corresponding effects on Canadian imports. For other categories, such as Apparel, Raw Metals, Chemical Products and Finance Services, the ER effects on imports are stronger. We conclude by demonstrating that the estimates of ber from Tables 1a–1c are estimates of the exchange rate effects on Canadian exports relative to Canadian imports that also net out border effects. The relative ER estimates from column (1) of Table 2 are taken from column 1 of Table 1a. Compare these in Table 2 to the sum of the directional ER estimates from columns (4) and (5). The latter is obtained simultaneously in each column at the expense of

31 For brevity, we only report the estimates of the border effects and the exchange rate effects.

184

J. Anderson, M. Vesselovsky, Y. Yotov / Journal of International Economics 100 (2016) 174–193

one dropped exporter-time fixed effect. The comparison reveals that ber = ber imp + ber exp , regardless of the choice of omitted fixed effect. 4.2. Inferred structure A key question of this paper is whether scale economies in bilateral trade costs are present. The structural model links these to differences in distance elasticities and contiguity effects between intraand inter-national trade. The operation of scale economies is also linked to real effects of exchange rate changes when passthrough elasticities are non-uniform. Theoretical restrictions of the model partially identify the structural parameters from the coefficient estimates from Tables 1a–1c. Identification is completed with assumed parameter values for either the elasticity of substitution or the elasticity of exchange rate passthrough. The theoretical restrictions are: aˆ 1 = c1 (1 − s)

(17)

aˆ 2 =

c1 (1 − s) (1 + s0can )

(18)

aˆ 3 =

c1 (1 − s) (1 + s0usa )

(19)

(qusa − 0usa )(1 − s) (qcan − 0can )(1 − s) bˆ er + bˆ brdr = − , 1 + s0usa 1 + s0can

(20)

where Eq. (20) is based on Eq. (15) and utilizes the relationship between the directional border estimates, bˆ brdr = bˆ usa,can + bˆ can,usa . First we obtain estimates of the trade cost scale parameters for Canada and the US from Eqs.(17)–(19) assuming a substitution elasticity s. The scale parameters are solved from Eqs.(17)–(19) as: 0can =

a1 − a2 sa2

and

0usa =

a1 − a3 . sa3

(21)

Next, these estimates are substituted into Eq. (20) to isolate qusa as a function of qcan and known variables: qusa =

aˆ 2 0usa − 0can bˆ er + bˆ brdr aˆ1 qcan + + . 1 + s0can 1−s aˆ 3 aˆ 3

(22)

If aˆ 2 = aˆ 3 then the middle term on the right is equal to zero. Then if the third term on the right is equal to zero, we have uniform passthrough of the US and Canada. Thus, given a 1 < 0 (which intuitively may be taken as certain), a simple test for passthrough uniformity is aˆ 2 − aˆ 3 = 0 and bˆ er + bˆ brdr = 0. Passthrough uniformity is consistent with exchange rate neutrality in the direct, partial equilibrium sense. The general equilibrium effects of exchange rate changes would still operate through the country fixed effects. Scale Elasticities. Substituting the estimates aˆ i ; i = 1, 2, 3 from Tables 1a–1c into Eq. (21) and setting s = 6.13 yields the tradevolume parameter estimates of 0can and 0usa reported in panel A of Tables 3a–3c, where we also report the difference 0can − 0usa along with standard errors.32

32 s = 6.13 comes from Head and Mayer (2014), who survey the related literature and report average values and standard deviations of 744 estimates of e = 1−s, which are obtained from a sample of 32 papers. Head and Mayer (2014) split the elasticity estimates according to several criteria. For consistency with our estimation methods, which rely on the theoretical gravity model, we chose e = −5.13 (translating into s = 6.13) as the mean estimate of s from Head and Mayer (2014) when the selection criteria is ‘structural gravity’ estimation. We also experimented with e = −5.03, which is Head and Mayer’s preferred estimate. The corresponding scale estimates were almost identical to those from the new version. We refer the reader to Head and Mayer (2014) and to Costinot and Rodriguez-Clare (2014) for informative discussions related to s.

The majority of the scale parameters are small in magnitude but statistically significantly less than zero in 36 of the 56 destinationcountry/sector cases. In Health services for Canada’s exports to the US the statistically significant estimate of 0usa is suspicious because it results from the positive elasticity of distance to the US aˆ 3 . We interpret this result (which violates the stability condition)33 as implying a mis-specified trade cost function that does not appropriately control for trade in health services. Similarly, the insignificant scale parameters for Fuels are due to its presumptively mis-specified trade cost equation. Apart from these suspicious sectors, scale effects on average for all goods reported in column(1) of Table 3a show that a 10% rise in trade volume will lower trade costs to Canada by 1.23% and to the US by 0.61%. In most goods sectors 0can < 0usa , scale economies are more pronounced for Canadian imports (i.e., US exports to Canada). This general finding is in accord with our intuitive explanation for the difference between internal and cross-border scale economies: bigger markets tend to be closer to exhausting the scale economies in the network formation process. Canadian provincial market destinations are smaller than the US market destinations, so are further from exhausting scale economies. This rationale also is consistent with the finding of larger scale effects on cross-border trade relative to domestic trade. An alternative hypothesis about trade elasticities varying with size relates to selection of the most productive firms due to meeting fixed export cost. Head and Mayer (2014), using a framework related to Chaney (2008), predict that larger markets will have larger trade elasticities (in absolute value) due to having higher proportions of low productivity firms. We looked for evidence of correlation between 0s and market size in cross-border trade using the variation across sectors, but found no strong evidence of non-zero correlation. The relatively small variation in 0 estimates is partly responsible.34 Generally the US exhibits more cases where we cannot reject constant returns to scale and not all differences between the directional scale effects are statistically significant, e.g. Machines. This may partly be due to our data structure that aggregates all US states. In the case of Finance, the somewhat anomalous point estimate of the scale elasticity for the US destination along with its large standard error suggests an omitted regulatory barrier. The general characterization that scale economies are more pronounced for Canadian imports apparently fails in one goods sector and in one service sector. The one goods sector is Fuels and Petroleum. We believe this is due to our mis-specification of the trade cost for these products. An alternative specification that reflects their physical distribution infrastructure and the peculiarities of US export policy could do better but our specification aimed at a uniform treatment of sectors. The one services sector where the scale elasticity ranking characterization fails to hold is Health. As noted earlier, one possible explanation for this result could be a mis-specified trade cost function for trade in Health services. Anderson et al. (2014) offer an alternative explanation based on the differences between the health systems in US and in Canada. They note that Canadian import of health services is a combination of Canada’s own supply congestion and limited access to specialists. Canadians, in response, look for alternatives and find them in the close-proximity and high-quality

33 Stability of the bilateral trade equilibrium depends on the supply schedule cutting the demand schedule from above. The quantity demanded for the CES case has inverse elasticity equal to −1/s. The quantity supplied has elasticity equal to 0. The stability condition is 1/s > −0. With 0 ≈ 0.1, the stability condition is met for s < 10. This condition is met except for Health services. 34 For goods trade the correlation is 0.005. For services trade the correlation is -0.81, offering some evidence that larger services markets have bigger scale effects. Overall, combining the two, the correlation is -0.20, weak support for a systematic relationship between scale elasticity and market size.

J. Anderson, M. Vesselovsky, Y. Yotov / Journal of International Economics 100 (2016) 174–193 Table 2 Border and Exchange Rates Collinearity Analysis. (1)

(2)

(3)

(4)

(5)

Main

BRDR1

BRDR2

ER1

ER2

16.235 (2.327)∗∗

16.235 (2.327)∗∗

28.784 (9.933)∗∗ −28.785 (9.993)∗∗ 1014 −0.002 (0.215)*

0.625 (8.852) −0.627 (8.914) 1014 −0.002 (0.215)+

16.235 (2.327)∗∗

bbrdr busa,can bcan,usa

−0.002 (0.215)

ber ber

exp

ber

imp

N ber

imp

1014 + ber

23.570 (2.583)∗∗ −7.334 (2.546)∗∗ −0.002 (0.215)

17.188 (2.450)∗∗ −0.953 (2.330) −0.002 (0.215)

1014

1014

16.235 (2.327)∗∗

16.235 (2.327)∗∗

exp

bcan,usa + busa,can

Notes: This table reveals collinearity relationships between the border variables and the exchange rate variables, respectively. Estimates of the rest of the gravity variables are omitted for brevity. Standard errors are clustered by pair and are in parentheses. See text for more details. + p < 0.10. ∗ p < .05. ∗∗ p < .01.

US market. In contrast, while Canada’s medical care is also of a high standard, it is government-controlled and rationed, which limits access for anyone who is not a member of the government-run, provincial health care plans. Furthermore, no Canadian health care provider would accept US domestic health insurance and Medicare coverage does not extend outside the United States.

185

In order to visualize and more easily compare scale effects across sectors, in panel A of Tables 3a–3c we report estimates, 0symm , that are obtained after imposing symmetry on the scale effects in each direction of Canada’s trade. Overall, our symmetric industry-specific estimates are as expected. Without exception, the symmetric scale parameters fall within the bounds of their directional (exports and imports) counterparts. In addition, three of the symmetric indexes (including Fuels, Communication and Education) are not statistically significant, even though one of their directional counterparts is. This result points to the importance of allowing for directional asymmetries. In terms of magnitude, the symmetric sectoral estimates are relatively homogeneous across goods and across services, however, the estimates for services are on average smaller. The outliers on the Goods side include Agriculture, Wood, and Petroleum and Coal products, with insignificant estimates, as well as Fuels, with an estimate that is more than three times larger as compared to the average across all goods. The result for Fuels is driven by the large estimate of 0usa for this sector, while the estimates for Agriculture, Wood, and Petroleum and Coal products are consistent with the insignificant estimates of the asymmetric scale effects. Scale Elasticity and Contiguity. Scale elasticities contribute to the asymmetries in contiguity parameter estimates. To ‘remove’ the scale effects, we multiply the estimates of CONTIG_PR_ST and CONTIG_ST_PR by (1+s0can ) and by (1+s0can ), respectively. Results are presented in Panels B of Tables 3a–3c, where we also report the original estimates on CONTIG_PR_ST and CONTIG_ST_PR. Three properties stand out. First, the majority of contiguity effects are smaller after the ‘removal’ of the trade-volume effects; tradevolume effects contribute to mostly larger contiguity estimates in each direction of Canadian trade. Second, the differences between the contiguity estimates after trade-volume effects are accounted for are reduced; i.e., trade-volume effects increase the difference

Table 3a Parameter Inferences from Sectoral Canadian Trade, 1997–2007. (1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

GOODS

AGRIC

FUELS

FOOD

LETHR

TXTLE

APPRL

WOOD

FRNTR

PAPER

−0.018 (0.056) 0.032 (0.071) −0.051 (0.102) 0.005 (0.038)

13.953 (1954.179) −0.211 (0.055)∗∗ 14.164 (1954.195) −0.372 (0.277)

−0.123 (0.006)∗∗ 0.713 (1.736) −0.836 (1.737) −0.077 (0.028)∗

−0.131 (0.003)∗∗ −0.110 (0.006)∗∗ −0.021 (0.007)∗ −0.125 (0.003)∗∗

−0.136 (0.004)∗∗ −0.122 (0.009)∗∗ −0.014 (0.010) −0.133 (0.003)∗∗

−0.126 (0.006)∗∗ −0.088 (0.025)∗∗ −0.038 (0.026) −0.113 (0.008)∗∗

−0.004 (0.133) 0.239 (0.315) −0.243 (0.334) 0.103 (0.156)

−0.135 (0.004)∗∗ −0.068 (0.020)∗∗ −0.067 (0.021)∗ −0.119 (0.006)∗∗

−0.124 (0.005)∗∗ 0.396 (0.670) −0.520 (0.672) −0.076 (0.018)∗∗

1.965 (0.925)∗ −0.583 (0.443) −0.121 (1.043) −10.466 (1384.252)

−0.387 (0.289) −2.079 (3.662) 2.482 (0.418)∗∗ 0.609 (0.187)∗

−1.183 (0.245)∗∗ −0.387 (0.080)∗∗ 2.826 (0.323)∗∗ 0.551 (0.104)∗∗

0.017 (0.282) 0.004 (0.071) 3.246 (0.295)∗∗ 0.535 (0.107)∗∗

2.512 (0.437)∗∗ 1.154 (0.529)∗ 3.976 (0.369)∗∗ 0.896 (0.211)∗∗

1.085 (0.310)∗∗ 0.632 (0.267)∗ 2.984 (0.460)∗∗ 0.512 (0.137)∗∗

0.476 (0.203)∗ 1.630 (2.424) 2.839 (0.537)∗∗ 0.678 (0.193)∗∗

2.414 (0.509)∗∗ 2.969 (3.217)

1.402 (0.439)∗ 27.130 (2.720)∗∗

0.972 (0.530)+ 20.362 (3.739)∗∗

1.263 (0.545)∗ 6.847 (4.729)

2.739 (0.603)∗∗ 20.496 (4.119)∗∗

2.580 (0.484)∗∗ 6.501 (3.058)∗

A. Scale parameters(s = 6.13) 0can −0.123 (0.004)∗∗ 0usa −0.061 (0.025)∗ −0.062 0can − 0usa (0.027)∗ −0.102 0symm (0.006)∗∗ B. Contiguity estimates CONTIG_PR_ST CONTIG_PR_ST(1 + s0usa ) CONTIG_ST_PR CONTIG_ST_PR(1 + s0can )

C. Tests of uniformity a3 − a2 ber + bbrdr

0.281 (0.301) 0.175 (0.216) 1.467 (0.292)∗∗ 0.357 (0.098)∗∗

1.828 (0.500)∗∗ 16.234 (2.499)∗∗

0.703 (0.412)+ 0.843 (0.689) 4.121 (0.450)∗∗ 3.662 (1.727)∗

0.347 (0.710) −3.184 (4.113)

2.842 (3.213) −37.484 (20.496)+

0.923 (0.276)∗∗ 2.275 (2.153) 4.818 (0.745)∗∗ 4.702 (4.538)

0.505 (0.730) −9.460 (6.161)

Notes: This table reports quantitative implications based on our theory. In panel A, we recover scale parameters for Canada and for US. Panel B, offers quantitative implications for the effects of contiguity on trade between Canadian provinces and US states. Finally, in panel C, we report tests for money neutrality. See text for further details. Standard errors, constructed with the Delta method, are reported in in parentheses. + p < 0.10. ∗ p < .05. ∗∗ p < .01.

186

J. Anderson, M. Vesselovsky, Y. Yotov / Journal of International Economics 100 (2016) 174–193

Table 3b Parameter Inferences from Sectoral Canadian Trade, 1997–2007. (1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

PRNTG

METL1

METL2

MCHNS

VHCLS

ELCTR

MNRLS

PETRL

CHMCL

MISCL

−0.115 (0.007)∗∗ 0.024 (0.133) −0.139 (0.136) −0.087 (0.014)∗∗

−0.119 (0.004)∗∗ −0.089 (0.009)∗∗ −0.030 (0.011)∗ −0.111 (0.004)∗∗

−0.116 (0.005)∗∗ −0.111 (0.006)∗∗ −0.005 (0.009) −0.114 (0.004)∗∗

−0.136 (0.001)∗∗ −0.114 (0.004)∗∗ −0.021 (0.005)∗∗ −0.132 (0.002)∗∗

−0.138 (0.002)∗∗ −0.091 (0.013)∗∗ −0.047 (0.013)∗∗ −0.130 (0.004)∗∗

−0.112 (0.006)∗∗ −0.048 (0.028)+ −0.064 (0.031)∗ −0.092 (0.007)∗∗

0.395 (1.296) 0.017 (0.132) 0.378 (1.379) 0.066 (0.129)

−0.134 (0.003)∗∗ −0.080 (0.022)∗∗ −0.054 (0.023)∗ −0.120 (0.007)∗∗

−0.137 (0.002)∗∗ −0.059 (0.031)+ −0.078 (0.031)∗ −0.129 (0.004)∗∗

2.488 (0.578)∗∗ 10.453 (21.617) 2.709 (0.458)∗∗ 0.576 (0.153)∗∗

2.627 (0.614)∗∗ 3.018 (2.574) 2.424 (0.418)∗∗ 0.716 (0.193)∗∗

0.320 (0.281) 0.145 (0.136) 1.852 (0.403)∗∗ 0.497 (0.139)∗∗

0.619 (0.256)∗ 0.198 (0.095)∗ 2.160 (0.266)∗∗ 0.625 (0.135)∗∗

1.343 (0.210)∗∗ 0.403 (0.089)∗∗ 0.348 (0.262) 0.059 (0.046)

1.071 (0.330)∗∗ 0.472 (0.197)∗ 1.555 (0.270)∗∗ 0.235 (0.052)∗∗

1.843 (0.402)∗∗ 1.301 (0.533)∗ 3.104 (0.463)∗∗ 0.979 (0.239)∗∗

0.292 (0.919) 0.322 (1.132) −0.022 (0.592) −0.076 (1.897)

1.350 (0.376)∗∗ 0.689 (0.326)∗ 2.508 (0.583)∗∗ 0.446 (0.123)∗∗

0.766 (0.512) 0.490 (0.409) 2.414 (0.290)∗∗ 0.393 (0.067)∗∗

4.561 (0.758)∗∗ 12.007 (4.897)∗

1.864 (0.597)∗ 3.263 (3.610)

1.204 (0.413)∗ 18.331 (2.229)∗∗

0.252 (0.468) 22.579 (2.221)∗∗

1.921 (0.351)∗∗ 37.356 (1.630)∗∗

1.889 (0.270)∗∗ 17.367 (2.008)∗∗

1.819 (0.627)∗ 14.635 (3.334)∗∗

−0.622 (1.201) −9.681 (4.916)∗

2.147 (0.472)∗∗ 16.617 (2.910)∗∗

3.142 (0.482)∗∗ 21.964 (2.800)∗∗

A. Scale parameters(s = 6.13) −0.128 0can (0.004)∗∗ 0usa 0.522 (1.316) 0can − 0usa −0.651 (1.317) 0symm −0.096 (0.019)∗∗ B. Contiguity estimates CONTIG_PR_ST CONTIG_PR_ST(1 + s0usa ) CONTIG_ST_PR CONTIG_ST_PR(1 + s0can )

C. Tests of uniformity a3 − a2 ber + bbrdr

Notes: This table reports quantitative implications based on our theory. In panel A, we recover scale parameters for Canada and for US. Panel B, offers quantitative implications for the effects of contiguity on trade between Canadian provinces and US states. Finally, in panel C, we report tests for money neutrality. See text for further details. Standard errors, constructed with the Delta method, are reported in in parentheses. + p < 0.10. ∗ p < .05. ∗∗ p < .01.

xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx Table 3c Parameter Inferences from Sectoral Canadian Trade, 1997–2007. (1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

SRVCS

TRNSP

CMNCN

WHLSL

FNNCE

BUSNS

EDCTN

HELTH

ACMDN

OTHER

−0.065 (0.026)∗ −0.048 (0.033) −0.016 (0.055) −0.057 (0.011)∗∗

−0.065 (0.028)∗ 0.304 (0.627) −0.369 (0.651) −0.016 (0.031)

−0.099 (0.010)∗∗ 0.184 (0.274) −0.283 (0.280) −0.053 (0.027)+

−0.112 (0.007)∗∗ −1.988 (10.200) 1.876 (10.205) −0.094 (0.014)∗∗

−0.106 (0.009)∗∗ 0.370 (0.721) −0.477 (0.727) −0.068 (0.019)∗∗

−0.062 (0.036)+ 1.342 (7.985) −1.403 (8.017) −0.023 (0.042)

−0.089 (0.009)∗∗ −0.381 (0.089)∗∗ 0.292 (0.095)∗ −0.050 (0.034)

0.080 (0.118) −0.041 (0.030) 0.121 (0.136) 0.000 (0.028)

−0.074 (0.021)∗∗ 1.483 (7.963) −1.557 (7.979) 0.003 (0.042)

0.598 (0.356)+ 0.421 (0.346) 0.777 (0.354)∗ 0.469 (0.319)

0.352 (0.374) 1.008 (2.319) 0.145 (0.370) 0.087 (0.245)

−0.006 (0.348) −0.012 (0.736) −0.077 (0.399) −0.030 (0.155)

1.442 (0.532)∗∗ −16.132 (85.368) −0.464 (0.511) −0.145 (0.144)

1.123 (0.363)∗∗ 3.672 (5.993) −0.000 (0.401) −0.000 (0.139)

0.700 (0.692) 6.460 (40.112) −0.086 (0.681) −0.054 (0.406)

1.180 (0.733) −1.573 (0.554)∗ −0.888 (0.639) −0.405 (0.256)

1.369 (0.970) −8.111 (1.981)∗∗

1.872 (0.636)∗ −3.538 (2.528)

2.937 (0.917)∗ 4.964 (3.005)+

2.649 (1.968) −9.153 (4.022)∗

8.087 (1.526)∗∗ −14.419 (4.547)∗

A. Scale parameters(s = 6.13) −0.093 0can (0.013)∗∗ 0usa 0.233 (0.441) 0can − 0usa −0.327 (0.452) 0symm −0.053 (0.022)∗ B. Contiguity estimates CONTIG_PR_ST CONTIG_PR_ST(1 + s0usa ) CONTIG_ST_PR CONTIG_ST_PR(1 + s0can )

C. Tests of uniformity a3 − a2 ber + bbrdr

0.954 (0.405)∗ 2.320 (3.462) 0.247 (0.357) 0.106 (0.170)

2.234 (0.998)∗ −0.596 (2.514)

0.222 (0.746) 1.201 (2.379)

4.505 (1.418)∗ 3.051 (3.075)

0.248 (0.352) 0.186 (0.302) 1.079 (0.383)∗∗ 1.610 (1.286)

−0.885 (0.754) −4.795 (3.484)

1.282 (0.476)∗∗ 12.938 (66.899) 0.549 (0.432) 0.301 (0.303)

2.383 (1.241)+ −6.667 (3.523)+

Notes: This table reports quantitative implications based on our theory. In panel A, we recover scale parameters for Canada and for US. Panel B, offers quantitative implications for the effects of contiguity on trade between Canadian provinces and US states. Finally, in panel C, we report tests for money neutrality. See text for further details. Standard errors, constructed with the Delta method, are reported in in parentheses. + p < 0.10. ∗ p < .05. ∗∗ p < .01.

J. Anderson, M. Vesselovsky, Y. Yotov / Journal of International Economics 100 (2016) 174–193

187

Table 4 PPML Panel Gravity Estimates, 1997–2007, Goods/Aggregate.

INTERNAL_DIST DIST_USA_CAN DIST_CAN_USA CONTIG_PR_ST CONTIG_ST_PR ER_CA BRDR_USA_CAN

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Main

Lagged

PairFEs

pre-2002

post-2002

IntraProv

Aggr

−0.728 (0.038)∗∗ −2.994 (0.307)∗∗ −1.166 (0.277)∗∗ 0.281 (0.301) 1.467 (0.292)∗∗ −0.002 (0.215) 16.235 (2.327)∗∗

−0.728 (0.038)∗∗ −2.994 (0.307)∗∗ −1.167 (0.277)∗∗ 0.280 (0.300) 1.466 (0.292)∗∗ −0.143 (0.214) 16.245 (2.322)∗∗

−0.728 (0.038)∗∗ −2.994 (0.307)∗∗ −1.166 (0.277)∗∗ 0.280 (0.301) 1.467 (0.292)∗∗ 0.043 (0.205) 16.235 (2.335)∗∗

−0.717 (0.036)∗∗ −3.177 (0.280)∗∗ −1.349 (0.258)∗∗ 0.371 (0.265) 1.234 (0.275)∗∗ 1.561 (0.303)∗∗ 19.038 (2.134)∗∗

−0.736 (0.040)∗∗ −2.821 (0.326)∗∗ −1.040 (0.293)∗∗ 0.245 (0.318) 1.689 (0.341)∗∗ −0.518 (0.188)∗∗ 13.736 (2.494)∗∗

−0.980 (0.044)∗∗ −3.220 (0.409)∗∗ −0.884 (0.424)∗ 0.676 (0.373)+ 0.980 (0.319)∗∗ 0.568 (0.259)∗ 11.038 (2.692)∗∗

1014

1014

1014

507

507

−0.738 (0.085)∗∗ −2.994 (0.307)∗∗ −1.165 (0.280)∗∗ 0.281 (0.301) 1.466 (0.292)∗∗ −0.001 (0.215) 16.093 (2.512)∗∗ −0.007 (0.056) 1014

−0.123 (0.004)∗∗ −0.061 (0.025)∗

−0.123 (0.004)∗∗ −0.061 (0.025)∗

−0.123 (0.004)∗∗ −0.061 (0.025)∗

−0.126 (0.003)∗∗ −0.076 (0.017)∗∗

−0.121 (0.005)∗∗ −0.048 (0.033)

−0.123 (0.004)∗∗ −0.061 (0.025)∗

−0.114 (0.006)∗∗ 0.017 (0.088)

INTRAPROV_DIST N B. Parameter inferences 0can 0usa

1014

Notes: This table reports sensitivity experiments for aggregate trade and aggregate Goods trade. Panel A of the table reports PPML gravity estimates and Panel B reports quantitative implications. Column (1) reproduces the findings for aggregate goods from the first column of Table 1a. The estimates in column (2) are obtained with lagged exchange rate. Bilateral fixed effects are used to obtain the ER estimates in column (3). Columns (4) and (5) allow for time-varying ER effects for the periods pre-2002 and post-2002, respectively. The specification in column (6) allows for differential effects of inter-provincial and intra-provincial distance. Finally, column (7) reports results obtained with aggregate data. Standard errors in Panel A are clustered by pair and are in parentheses. Standard errors in Panel B are obtained with the Delta method. See text for further details. + p < 0.10. ∗ p < .05. ∗∗ p < .01.

xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx Table 5 PPML Panel Gravity Estimates, 1997–2007, Services. (1)

INTERNAL_DIST DIST_USA_CAN DIST_CAN_USA CONTIG_PR_ST CONTIG_ST_PR ER_CA BRDR_USA_CAN

(2)

(3)

(4)

(5)

(6)

Main

Lagged

PairFEs

pre-2002

post-2002

IntraProv

−1.162 (0.055)∗∗ −2.712 (0.520)∗∗ −0.478 (0.527) 0.954 (0.405)∗ 0.247 (0.357) −0.055 (0.171) −0.541 (2.477)

−1.162 (0.055)∗∗ −2.712 (0.520)∗∗ −0.478 (0.527) 0.954 (0.405)∗ 0.248 (0.357) −0.022 (0.179) −0.540 (2.476)

−1.162 (0.055)∗∗ −2.712 (0.520)∗∗ −0.478 (0.527) 0.954 (0.405)∗ 0.248 (0.357) −0.037 (0.177) −0.540 (2.476)

−1.167 (0.054)∗∗ −2.864 (0.520)∗∗ −0.507 (0.525) 0.928 (0.402)∗ 0.264 (0.358) 0.090 (0.316) 0.701 (2.445)

−1.158 (0.055)∗∗ −2.598 (0.520)∗∗ −0.455 (0.530) 0.974 (0.408)∗ 0.235 (0.357) −0.004 (0.169) −1.487 (2.500)

1014

1014

1014

507

507

−0.583 (0.110)∗∗ −2.634 (0.400)∗∗ −0.629 (0.372)+ 0.878 (0.291)∗∗ 0.341 (0.286) −0.065 (0.179) 7.668 (3.127)∗ 0.419 (0.077)∗∗ 1014

−0.093 (0.013)∗∗ 0.233 (0.441)

−0.093 (0.013)∗∗ 0.233 (0.441)

−0.093 (0.013)∗∗ 0.233 (0.441)

−0.097 (0.012)∗∗ 0.212 (0.392)

−0.090 (0.014)∗∗ 0.252 (0.487)

−0.127 (0.009)∗∗ 0.234 (0.441)

INTRAPROV_DIST N B. Parameter inferences 0can 0usa

Notes: This table reports sensitivity experiments for Services trade. Panel A of the table reports PPML gravity estimates and Panel B reports quantitative implications. Column (1) reproduces the findings for aggregate services from the first column of Table 1c. The estimates in column (2) are obtained with lagged exchange rate. Bilateral fixed effects are used to obtain the ER estimates in column (3). Columns (4) and (5) allow for time-varying ER effects for the periods pre-2002 and post-2002, respectively. Finally, the specification in column (6) allows for differential effects of inter-provincial and intra-provincial distance. Standard errors in Panel A are clustered by pair and are in parentheses. Standard errors in Panel B are obtained with the Delta method. See text for further details. + p < 0.10. ∗ p < .05. ∗∗ p < .01.

188

J. Anderson, M. Vesselovsky, Y. Yotov / Journal of International Economics 100 (2016) 174–193

Table 6a Sectoral PPML Panel Gravity Estimates, 1997–2007.

A. Gravity estimates INTERNAL_DIST DIST_USA_CAN DIST_CAN_USA DIST_MEX_CAN DIST_CAN_MEX CONTIG_PR_ST CONTIG_ST_PR ER_CA_USA ER_CA_MEX BRDR_USA_CAN BRDR_MEX_CAN

B. Scale parameters 0can usa 0usa

can

0can

mex

0mex

can

N

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

FOOD

LETHR

TXTLE

APPRL

WOOD

FRNTR

PAPER

PRNTG

−0.620 (0.029)∗∗ −2.502 (0.408)∗∗ −0.155 (0.223) −15.401 (7.394)∗ −1.954 (2.642) −0.417 (0.287) 2.497 (0.419)∗∗ −0.650 (0.269)∗ −0.926 (0.412)∗ 3.725 (3.290) 122.844 (63.642)+

−0.670 (0.036)∗∗ −3.378 (0.319)∗∗ −2.114 (0.257)∗∗ −22.082 (10.029)∗ 1.409 (4.114) −1.160 (0.230)∗∗ 2.862 (0.336)∗∗ 0.214 (0.141) −0.404 (0.472) 26.500 (2.741)∗∗ 151.297 (87.583)+

−0.453 (0.034)∗∗ −2.679 (0.302)∗∗ −1.921 (0.423)∗∗ −11.129 (4.655)∗ −1.676 (1.533) −0.021 (0.283) 3.321 (0.308)∗∗ −0.731 (0.373)∗ 0.988 (1.387) 20.370 (3.785)∗∗ 90.123 (39.535)∗

−0.530 (0.036)∗∗ −2.225 (0.396)∗∗ −1.350 (0.426)∗∗ −11.736 (6.371)+ −0.258 (1.795) 2.453 (0.436)∗∗ 4.095 (0.392)∗∗ −1.409 (0.511)∗∗ 3.072 (25.174) 7.735 (4.504)+ 96.981 (51.517)+

−0.816 (0.037)∗∗ −0.836 (0.708) −0.331 (0.270) −12.260 (2.119)∗∗ −5.711 (1.770)∗∗ 0.923 (0.276)∗∗ 4.818 (0.745)∗∗ 1.663 (0.262)∗∗ −6.972 (0.946)∗∗ −11.125 (5.982)+ 120.604 (21.882)∗∗

−0.632 (0.044)∗∗ −3.479 (0.621)∗∗ −1.362 (0.265)∗∗ −49.209 (13.010)∗∗ −16.859 (3.418)∗∗ 0.985 (0.334)∗∗ 3.140 (0.513)∗∗ 0.202 (0.310) −231.834 (47.970)∗∗ 19.283 (4.587)∗∗ 399.101 (105.114)∗∗

−0.662 (0.030)∗∗ −2.749 (0.373)∗∗ −0.225 (0.232) −18.526 (4.111)∗∗ −3.889 (3.027) 0.470 (0.200)∗ 2.849 (0.536)∗∗ −0.087 (0.134) −1.562 (0.678)∗ 6.645 (3.027)* 163.421 (40.973)∗∗

−1.021 (0.032)∗∗ −4.798 (0.524)∗∗ −0.248 (0.466) −42.887 (14.793)∗∗ −2.787 (1.143)∗ 2.485 (0.578)∗∗ 2.711 (0.458)∗∗ 0.237 (0.182) −8.581 (2.148)∗∗ 11.768 (4.952)∗ 341.234 (119.311)∗∗

−0.123 (0.007)∗∗ 0.489 (0.942) −0.157 (0.003)∗∗ −0.111 (0.070) 1114

−0.131 (0.003)∗∗ −0.111 (0.006)∗∗ −0.158 (0.002)∗∗ −0.241 (0.227) 1038

−0.136 (0.004)∗∗ −0.125 (0.008)∗∗ −0.156 (0.003)∗∗ −0.119 (0.040)∗ 962

−0.124 (0.007)∗∗ −0.099 (0.018)∗∗ −0.156 (0.004)∗∗ 0.172 (2.337) 849

−0.004 (0.133) 0.239 (0.315) −0.152 (0.002)∗∗ −0.140 (0.007)∗∗ 1081

−0.133 (0.005)∗∗ −0.087 (0.015)∗∗ −0.161 (0.001)∗∗ −0.157 (0.001)∗∗ 933

−0.124 (0.006)∗∗ 0.317 (0.491) −0.157 (0.001)∗∗ −0.135 (0.022)∗∗ 1015

−0.128 (0.004)∗∗ 0.507 (1.258) −0.159 (0.001)∗∗ −0.103 (0.025)∗∗ 1061

Notes: This table reports PPML estimates of the scale effects and of the effects of exchange rate fluctuations on Canadian goods trade with US and Mexico. Column (1) includes estimates for aggregate manufacturing and the numbers in the next nine columns are for individual sectors. Panel A reports gravity estimates. All estimates are obtained with time-varying, directional (importer and exporter) fixed effects and the years included in our sample are 1997, 1999, 2001, 2003, 2005 and 2007. The dependent variable is nominal exports. Standard errors are clustered by pair and are in parentheses. Panel B reports estimates of the scale parameters that are recovered from the indexes in Panel A as described in the text. Standard errors for the scale parameters are obtained with the Delta method. + p < 0.10. ∗ p < .05. ∗∗ p < .01.

between the effects of directional contiguity. Third, many of the contiguity estimates remain positive and significant even after removing volume effects, and exhibit directional symmetries. Test of Uniformity. The test of passthrough uniformity has two components, scale neutrality (0can − 0usa = 0) and net neutrality (bˆ er + bˆ brdr = 0), reported in Panel C of Tables 3a–3c. Net neutrality implies that the direct partial equilibrium effect of exchange rates on bilateral US–Canada sectoral balance of trade is zero. For 14 of 19 goods sectors, scale neutrality is rejected. Net neutrality is rejected in 14 cases as well. The coincidence of the two rejections, implying rejection of passthrough uniformity, obtains for 11 of the 19 goods sectors. For services, in contrast, scale neutrality and net neutrality are rejected in 5 of 9 sectors. Both are rejected for Business services, Health services, and Other services. In the case of Health services, the test statistic is based on an almost surely mis-specified trade cost function evidenced by a positive distance elasticity for Canada’s exports to the US. We conclude that passthrough uniformity can be rejected for a majority of goods sectors but not for services sectors.

5. Sensitivity experiments This section presents and discusses the results of several experiments that test the robustness of our main findings. First, we

offer two alternative specifications that address potential endogeneity concerns. Second, we estimate a specification that allows for time-varying scale and exchange rate effects. Third, we test for intranational scale effects. Fourth, we obtain estimates with aggregate data across all sectors in our sample. Finally, we obtain multilateral scale effects by introducing Mexico to our sample. 5.1. Endogeneity experiments Two alternative specifications test for potential endogeneity with respect to exchange rates. For brevity we present only the representative results for aggregate goods and services, reported in Tables 4 and 5, respectively. We report the base case results for goods in column (1) of Table 4 (from column 1 of Table 1a), and in column (1) of Table 5 we report the main estimates for services (from column 1 of Table 1c). The top panel of Tables 4 and 5, labeled ‘A. Gravity Estimates,’ present the gravity estimates from our experiments and in the bottom panel, labeled ‘B. Parameter Inferences,’ we recover the volume parameters. Begin with goods. First, in column (2) of Table 4, we use lagged exchange rate values to eliminate any simultaneity between trade and exchange rates. Overall, the results from column (2) are not statistically different from the main findings from column (1). Three properties stand out. (i) The estimates of the standard gravity variables in columns (1) and (2) are virtually identical. This suggests that

J. Anderson, M. Vesselovsky, Y. Yotov / Journal of International Economics 100 (2016) 174–193

189

Table 6b Sectoral PPML Panel Gravity Estimates, 1997–2007.

A. Gravity estimates INTERNAL_DIST DIST_USA_CAN DIST_CAN_USA DIST_MEX_CAN DIST_CAN_MEX CONTIG_PR_ST CONTIG_ST_PR ER_CA_USA ER_CA_MEX BRDR_USA_CAN BRDR_MEX_CAN

B. Scale parameters 0can usa 0usa

can

0can

mex

0mex

can

N

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

METL1

METL2

MCHNS

VHCLS

ELCTR

MNRLS

PETRL

CHMCL

MISCL

−0.741 (0.065)∗∗ −2.509 (0.264)∗∗ −0.636 (0.477) 0.070 (2.101) −0.157 (1.540) 2.590 (0.603)∗∗ 2.435 (0.415)∗∗ −2.002 (0.246)∗∗ −1.923 (0.567)∗∗ 5.219 (3.645) −17.085 (20.992)

−0.786 (0.029)∗∗ −2.860 (0.290)∗∗ −1.816 (0.203)∗∗ −25.002 (7.273)∗∗ −11.045 (2.209)∗∗ 0.288 (0.281) 1.901 (0.411)∗∗ 0.493 (0.271)+ −2.253 (0.497)∗∗ 17.615 (2.168)∗∗ 274.032 (60.741)∗∗

−0.742 (0.050)∗∗ −2.345 (0.293)∗∗ −2.519 (0.307)∗∗ −16.996 (8.274)∗ −12.120 (3.950)∗∗ 0.550 (0.261)∗ 2.272 (0.290)∗∗ 0.450 (0.447) −166.389 (55.600)∗∗ 21.414 (2.289)∗∗ 137.000 (67.176)∗

−0.659 (0.050)∗∗ −3.712 (0.416)∗∗ −2.819 (0.291)∗∗ −31.063 (10.709)∗∗ −13.416 (2.859)∗∗ 1.108 (0.264)∗∗ 0.653 (0.347)+ 0.138 (0.247) −1.100 (0.301)∗∗ 35.463 (2.466)∗∗ 347.025 (87.294)∗∗

−0.367 (0.049)∗∗ −2.097 (0.284)∗∗ −1.207 (0.240)∗∗ −16.121 (3.407)∗∗ −3.895 (1.109)∗∗ 0.881 (0.365)∗ 1.889 (0.324)∗∗ −0.500 (0.223)∗ −57.154 (15.688)∗∗ 14.582 (2.520)∗∗ 136.944 (27.561)∗∗

−1.029 (0.037)∗∗ −3.118 (0.410)∗∗ −1.594 (0.345)∗∗ −18.366 (6.723)∗∗ 0.062 (1.979) 1.773 (0.396)∗∗ 3.197 (0.475)∗∗ 2.515 (0.294)∗∗ 1.133 (1.272) 11.854 (3.424)∗∗ 125.064 (56.703)∗

−1.015 (0.056)∗∗ −0.288 (0.692) −0.923 (0.650) 3.554 (1.937)+ 1.444 (4.500) 0.295 (0.918) −0.027 (0.589) 0.214 (0.625) 3.315 (3.754) −9.914 (4.816)∗ −66.264 (39.651)+

−0.582 (0.038)∗∗ −3.242 (0.281)∗∗ −1.166 (0.290)∗∗ −19.656 (8.666)∗ 3.138 (1.641)+ 1.307 (0.374)∗∗ 2.528 (0.581)∗∗ −0.934 (0.240)∗∗ −1.392 (0.513)∗∗ 17.394 (2.793)∗∗ 116.343 (71.210)

−0.661 (0.037)∗∗ −3.977 (0.309)∗∗ −1.200 (0.297)∗∗ −21.169 (12.981) −1.483 (1.719) 0.682 (0.516) 2.503 (0.307)∗∗ −0.449 (0.233)+ −13.533 (24.367) 21.905 (2.934)∗∗ 171.344 (104.991)

−0.115 (0.007)∗∗ 0.027 (0.136) −1.879 (51.115) 0.609 (7.560) 971

−0.118 (0.004)∗∗ −0.093 (0.008)∗∗ −0.158 (0.001)∗∗ −0.152 (0.002)∗∗ 1093

−0.111 (0.008)∗∗ −0.115 (0.006)∗∗ −0.156 (0.004)∗∗ −0.153 (0.003)∗∗ 1047

−0.134 (0.002)∗∗ −0.125 (0.006)∗∗ −0.160 (0.001)∗∗ −0.155 (0.002)∗∗ 1041

−0.135 (0.005)∗∗ −0.114 (0.011)∗∗ −0.159 (0.001)∗∗ −0.148 (0.005)∗∗ 1010

−0.109 (0.007)∗∗ −0.058 (0.023)∗ −0.154 (0.003)∗∗ −2.855 (85.449) 987

0.411 (1.365) 0.016 (0.130) −0.210 (0.026)∗∗ −0.278 (0.358) 955

−0.134 (0.003)∗∗ −0.082 (0.021)∗∗ −0.158 (0.002)∗∗ −0.193 (0.016)∗∗ 1061

−0.136 (0.002)∗∗ −0.073 (0.023)∗ −0.158 (0.003)∗∗ −0.090 (0.084) 1031

Notes: This table reports PPML estimates of the scale effects and of the effects of exchange rate fluctuations on Canadian goods trade with US and Mexico. Panel A reports gravity estimates. All estimates are obtained with time-varying, directional (importer and exporter) fixed effects and the years included in our sample are 1997, 1999, 2001, 2003, 2005 and 2007. The dependent variable is nominal exports. Standard errors are clustered by pair and are in parentheses. Panel B reports estimates of the scale parameters that are recovered from the indexes in Panel A as described in the text. Standard errors for the scale parameters are obtained with the Delta method. + p < 0.10. ∗ p < .05. ∗∗ p < .01.

the ER effects are orthogonal to the effects of distance, border and contiguity. We exploit this property in the next experiment. (ii) The estimate of the relative ER effect is still not statistically significant, as it was in the main results. (iii) We do not find any statistically significant effects on the structural parameters in our model, which are reported in Panel B. Comparison between the numbers in columns (1)–(2) of Table 5 reveals that these results are confirmed for services as well. The second alternative specification applies the methods of Baier and Bergstrand (2007), who convincingly account for endogeneity of free trade agreements. Specifically, we include the full set of country-pair fixed effects in addition to the directional (exporter and imported) fixed effects. The intuition is that the bilateral fixed effects can successfully absorb the correlation between the trade policy variable and the unobservable error term in the gravity model in order to eliminate endogeneity. The introduction of country-pair fixed effects results in the following econometric specification:

obtain estimates of the standard gravity variables, which are needed to recover the structural parameters in our model, we apply a twostage procedure similar to the one from Anderson and Yotov (2011). In particular, first we estimate Eq. (23) to obtain the ER effects after addressing endogeneity, then we restrict the ER estimates in a constrained second-stage optimization, where the bilateral fixed effects are replaced with the standard set of gravity variables. Orthogonality between the ER effects and the standard gravity covariates validates this approach. Results for goods are reported in column (3) of Table 4. There are no statistically significant differences from estimates in column (1). In particular, (i) the standard gravity estimates are identical; (ii) the ER estimates are not statistically significant; and (iii) the structural parameters from columns (1) and (3) in panel B are identical to each other. The estimates for services from column (1) and column (3) of Table 5 are identical as well. 5.2. Time-varying scale effects

Xij,t

= exp[a˜ 0 + b˜ er ER CA EXP + xij + gi,t + hj,t ] + u˜ ij,t ,

(23)

where xij is the full set of bilateral fixed effects for any two trading partners in our sample, and all other variables are defined as before. All time-invariant standard gravity covariates (such as distance for example) will be absorbed by the bilateral fixed effects. To

In the next experiment, exchange rate effects can vary over time due to splitting the data in two periods: before and after 2002. Choosing 2002 to allow for time-varying ER effects has two advantages in addition to being the mid-year in our sample. First, the Canadian dollar depreciated steadily during the period 1997–2002, while it appreciated steadily between 2002 and 2007. These patterns provide

190

J. Anderson, M. Vesselovsky, Y. Yotov / Journal of International Economics 100 (2016) 174–193

an opportunity to look for asymmetric trade responses to ER changes. Second, splitting the time series at 2002 may pick up any changes in trade due to changes in border security after the 9/11 events. Pre-2002 and post-2002 estimates for goods are reported in columns (4) and (5) of Table 4, respectively. Similar estimates of the standard gravity covariates were obtain in the two periods. There are also two important differences between the estimates in columns (4) and (5). First, we obtain a large positive and statistically significant estimate on ER_CA_EXP for the period before 2002, but a negative and significant estimate on ER_CA_EXP for the period after 2002. Based on the definition of b˜ er , as a relative effect capturing the response of Canadian exports to US relative to Canadian imports from US, our estimates imply that when the CA dollar was depreciating, in the pre-2002 period, Canadian exports responded much more than Canadian imports. The trade response was asymmetric in the post-2002 period too, when the CA dollar was steadily appreciating. This time however, the response of Canadian imports was stronger than the response of Canadian exports. Hence, the negative estimates of b˜ er for the post-2002 period. The asymmetric passthrough of depreciation is stronger than the passthrough of appreciation, taking the scale elasticity as constant, a result consistent with those of Delatte and Lopez-Villavicencio (in press) based on price comparison data. The second difference between columns (4) and (5) is in the estimated scale parameters. There is an insignificant estimate of 0usa in the post-2002 period (column 5), although the estimate lies within 2 standard deviations of the pre-2002 estimate. The mechanical reason for the significance test result is that the standard error doubles in the post-2002 period. Interpreting this weak evidence as indicating a fall in the scale elasticity (in absolute value), the result is consistent with a thickening of the US border after 9/11. (The scale parameter inference is identified independently of exchange rate effects so it is not subject to passthrough asymmetry.) For services there are no significant differences between the pre-2002 and post-2002 estimates. This finding underlines the conclusion drawn from the main results that the scale and passthrough channels are weakly identified in the services data, and splitting the data does not help. The results are reported in columns (4) and (5) of Table 5. 5.3. Intra-national scale effects Scale effects might operate differentially at provincial borders, so that inter-provincial trade is subject to scale effects relative to intraprovincial trade. To test for sensitivity to differential treatment of intra- from inter-provincial trade we introduce an additional distance regressor to our main econometric model. INTRAPROV_DIST is equal to the logarithm of bilateral distance for trade within the Canadian provinces and territories, and equal to zero otherwise. By construction, the estimates on INTRAPROV_DIST should be interpreted as deviations from the corresponding estimates on INTERNAL_DIST and a significant estimate of the coefficient on INTRAPROV_DIST would point to potential intra-national scale effects, now capturing forces that act differentially on interprovincial vs. intra-provincial trade. Our results for aggregate goods appear in column (6) of Table 4. These estimates reveal that there is no difference between the effects of distance on intra-provincial vs. inter-provincial trade. The estimates for aggregate goods are representative of the estimates that we obtain for individual sectors, which we omit here for brevity but are available upon request. At the individual goods sector level, we find that the estimates on INTRAPROV_DIST are not statistically significant for 14 of the 19 sectors in our sample. For three sectors (including Fuels, Wood, and Printing), we obtain positive and significant estimates on INTRAPROV_DIST, suggesting that distance is a smaller impediment to intra-provincial trade, while we obtain negative estimates on INTRAPROV_DIST for Metals and for Paper, which

means that intra-provincial distance is a larger impediment to trade than inter-provincial distance in these sectors. Our findings are much different in the case of services, where, without any exception, we find that the estimates on intra-provincial distance are always significantly smaller than the estimates on interprovincial distance. The estimates for aggregate services, which we report in column (6) of Table 5, are representative of our findings for individual sectors, which are available by request. Overall, our interpretation of the differences between intra- and inter-provincial distance elasticities is the same as for the difference between internal and international distance elasticities: scale economies tend to be closer to exhaustion the more localized the market. This pattern is much more pronounced for services than for goods, no surprise given the localization of much of services activity. Another possible explanation for our findings for services is measurement error in our data. We study this possibility in our next experiment. 5.4. Aggregate data Arguably, aggregate trade flows are measured with less error than industry-level trade flows. To test for the possibility that our estimates may have been affected by measurement error of this type, in this experiment we obtain aggregate estimates. These results appear in column (7) of Table 4. Overall, the aggregate estimates are as expected and support our main findings, but we also observe some differences. All but two of our aggregate gravity estimates are within the bounds determined by the corresponding estimates for aggregate goods and aggregate services. The first exception is the aggregate estimate on DIST_USA_CAN, which is larger but not statistically different from the estimates on DIST_USA_CAN for aggregate goods and aggregate services. The second exception is the estimate on ER_CA, which becomes positive and significant. One possible explanation for this result is that more weight is attached to the earlier years in the sample, when, as we saw in the two-period experiment from Section 5.2, the estimate on ER_CA is positive and highly significant for the years before 2002. Measurement error is another possible explanation for these differences. Importantly, the estimates of the aggregate scale effects in each direction are also between the corresponding estimates for aggregate goods and aggregate services. Specifically, our gravity estimates imply a negative and statistically significant estimate of 0can = −0.114 and an insignificant estimate of 0usa = 0.017. The latter is consistent with the small, negative, and marginally significant estimate for aggregate goods and the insignificant positive estimate for aggregate services. 5.5. A multilateral setting A natural question is whether scale effects are observable in other Canadian bilateral trade relationships and how their size relates to those with the US. Data availability and expositional simplicity motivated a focus on Canadian provinces’ relationships with the US. Here we include Mexico. Mexico and Canada are each other’s third largest trade partner (the US being first and China second).35 Additional reasons for inclusion are that Mexican and Canadian legal systems are quite different, and Mexican institutions differ from parallel institutions in Canada and the US. We expect, thinking about the difficulties facing pair-specific investment, that the scale effects for trade between Canada and Mexico will be stronger compared to the corresponding indexes that we obtained in our main analysis for trade between Canada and the United States. Data availability allowed inclusion of each Mexican manufacturing sector for the

35 We thank an anonymous referee for this suggestion. The referee also suggested China as another alternative. However, we only were able to collect Chinese production data for the years post 2003. Thus, we could not include this country due to lack of sufficient data on internal trade.

J. Anderson, M. Vesselovsky, Y. Yotov / Journal of International Economics 100 (2016) 174–193

whole period of investigation but precluded agriculture, fuels and services.36 Guided by the multilateral theoretical specification (9), the econometric specification becomes: Xij,t =ea0 +a1 INTERNAL ea5 DIST

DIST+a2 DIST CAN USA+a3 DIST USA CAN+a4 DIST CAN MEX

MEX CAN+a6 CONTIG PR ST+a7 CONTIG ST PR+bbrdr1 USA CAN

ebbrdr2 MEX

CAN+ber1 ER CA USA+ber2 ER CA MEX+hj,t +gi,t

+ 4˜ ij,t .

∗

∗ (24)

The new covariates in Eq. (24) include: DIST_CAN_MEX and DIST_MEX_CAN, which account for the directional effects of bilateral distance between Canada and Mexico; MEX_CAN, which is an indicator border variable that takes a value of one Mexican exports to Canada, and it is set to zero otherwise; finally,werelabeled the exchange-rate covariate ER_CA from the original specification to ER_CA_USA, and we added a new ER covariate, ER CA MEX = MEX CAN × ln(rcan mex,t ), which captures the effects of exchange rate fluctuations between the Canadian dollar and the Mexican peso.37 (Note that while in principle the introduction of Mexico offers an opportunity to introduce the standard gravity variable ‘Common Official Language’ to specification (24), we cannot add an indicator variable for common official language between Canada and the United States to specification (24) because this dummy would be perfectly collinear with the corresponding border variables.) Estimation results from specification (24) for each of the manufacturing sectors in our sample are reported in Tables 6a and 6b. Panel A in each table presents the gravity estimates and panel B in each table presents the corresponding scale parameters. First note that the introduction of Mexico has no significant effects on the original estimates based on Canada–USdata: none of the Canada– USgravity estimates from Tables 6a and 6b are statistically different from the corresponding main estimates in Tables 1a and 1b. This, by construction, also implies that the original scale parameters for Canada–UStrade are also very similar to the new estimates that we obtain after the addition of Mexico. The rich fixed effects structure of the econometric specification is a natural explanation for the robustness of the bilateral results to the introduction of a new country in the sample. Turning to the new results for Canada’s trade with Mexico, the gravity estimates in panels A reveal that the effects of distance on trade between Canada and Mexico are, on average, significantly larger in absolute value than the corresponding distance elasticities for trade between Canada and the United States. The non-loglinear effect of distance at the cross-country level is well documented in the trade literature, cf. Eaton and Samuel (2002).38 More novel, the effects of distance on Canada’s trade with Mexico are highly asymmetric. The effects on Mexican exports to Canada are very large and always statistically significant at any conventional level (except for Primary Metal products), while the effects on Canadian exports to Mexico are always smaller in magnitude and statistically significant for only about half of the manufacturing sectors in our sample.

36 Data on international trade between the Canadian provinces and territories and Mexico are from the Trade Data Online web interface of Industry Canada. Production data for Mexico come from the UNIDO Industrial Statistics database. Finally, and consistent with the rest of our distances data, we used the population-weighted procedure of Mayer and Zignago (2006) to construct the distances between Mexico and each of the Canadian provinces and territories. 37 Similar to the bilateral case of trade between Canada and the United States, we cannot include the directional/import counterparts of MEX_CAN and ER_CA_MEX due to perfect collinearity. 38 In analysis available by request we demonstrate that the effects of internal distance and of distance between Canada and US in each direction of trade are essentially linear.

191

Table 7 Sector Definitions and Labels. Sector definition

Sector label

Agriculture Mineral fuels Food Leather, rubber and plastic products Textile products Hosiery, clothing and accessories Lumber and wood products Furniture, mattresses and lamps Wood pulp, paper and paper products Printing and publishing Primary metal products Fabricated metal products Machinery Motor vehicles, transportation equipment and parts Electrical, electronic, and communications products Non-metallic mineral products Petroleum and coal products Chemicals, pharmaceutical, and chemical products Miscellaneous manufactured products Transportation and storage services, including transportation margins Communication services Wholesale services, including wholesale margins Finance, insurance and real estate services Professional, scientific, technical, computer, administrative, support, and related services Education services Health care and social assistance services Accommodation services and meals Miscellaneous services

AGRIC FUELS FOOD LETHR TXTLE APPRL WOOD FRNTR PAPER PRNTG METL1 METL2 MCHNS VHCLS ELCTR MNRLS PETRL CHMCL MISCL TRNSP CMNCN WHLSL FNNCE BUSNS EDCTN HELTH ACMDN OTHER

The scale estimates from panels B of Tables 6a–6b reveal three distinct patterns. First, there are large scale effects for Canadian exports to Mexico. The estimates of 0can mex are significant at any level (except for Primary Metal products) and they are remarkably stable across sectors. Second, similar to the case of trade between Canada and US, we find that the scale effects are asymmetric for trade between Canada and Mexico. The scale elasticities on Mexican exports are almost always smaller in magnitude than the scale elasticities for Canadian exports, they vary more across sectors, and they are not always statistically significant. Third, comparisons between the scale elasticities for Canada–UStrade and the corresponding numbers for Canada–Mexicotrade reveal that the scale elasticities on trade between Canada and Mexico in each direction are larger than the corresponding scale elasticities for trade between Canada and the United States. This result is consistent with our expectations based on bigger legal system and institutional differences between Canada and Mexico than between Canada and the US. It also points to potential benefits from digging deeper into understanding the determinants of the scale elasticities. Overall, the sensitivity experiments of this section support the base findings from Sections 4.1 and 4.2. The main additions are: (i) evidence of differences in the response of goods trade flows to exchange rate appreciation as opposed to depreciation; and (ii) evidence for heterogenous scale effects across trading partners. 6. Conclusion We develop a structural model that can identify external economies or diseconomies of scale in the structural gravity model. This significant departure from constant returns points toward a richer model that includes infrastructure detail as a determinant of bilateral trade. Application to more disaggregated trade data would permit direct measures of trade volume to be used to directly identify scale elasticities, increasing the precision of estimation. Disaggregation would also permit examination of more detailed analysis of scale

192

J. Anderson, M. Vesselovsky, Y. Yotov / Journal of International Economics 100 (2016) 174–193

effects, such as allowing for multiple points of border entry, multiple modes of transport and infrastructure detail. We also develop a structural gravity model of exchange rate effects due to incomplete passthrough at the low frequencies appropriate to gravity. Though motivated to model Canada–US trade in a decade of substantial exchange rate appreciation and depreciation, the model applies more generally to incomplete passthrough of other exogenous trade costs shifts at the border, such as free trade agreements. Our methods can yield passthrough elasticity estimates or elasticity of substitution estimates when combined with other externally derived parameter estimates. Implications for future work include efforts to drill down into the parametric scale and passthrough elasticities. As to the scale elasticity, following clues in descriptions about what actually happens at borders can yield useful improvements in modeling trade costs. As for the passthrough elasticity, forging links in a chain connecting high frequency price comparisons to low frequency trade flow inferences about this parameter should yield insights about the magnitude of this key parameter.

electric equipment; aircraft and engines; locomotives and railway stock; ships and boats; snowmobiles. Electrical, electronic and communication products: appliances; household equipment; household furnaces; household refrigerators and freezers; household cooking equipment; TVs, VCRs etc.; telephone and related equipment; broadcasting equipment; electric motors; transformers; batteries; wiring materials; lighting fixtures; other electric equipment. Non-metallic mineral products: cement; concrete products; lime; brick; gypsum; stone; asbestos; glass; abrasive products. Petroleum and coal products: gasoline; diesel; fuel oils; tar and pitch; naptha; asphalt; other petroleum products. Chemicals, pharmaceuticals and chemical products: industrial chemicals; hydrocarbons; organic acids; fertilizers; pharmaceuticals; soaps, detergents and other cleaning products; explosives; paints; ammunition; insecticides; inks; other chemical products. Miscellaneous manufactured products: scientific and lab equipment; measuring and other scientific instruments; clocks and watches; photographic equipment; pearls and precious stones; toys and games; shades and blinds; recordings; musical instruments; miscellaneous end-use consumer products.

Appendix A. Sector description

A.2. Services

Table 7 includes brief definitions of the sectors in our sample and corresponding sector labels that we use throughout the text. Next, we offer more detailed description of the sectors in our sample.

Transportation and storage services: Air, water and rail passenger and freight transportation; Bus (including school), ambulance and truck transportation; Urban transit and taxi transportation; Pipeline transportation of natural gas and oil; Grain and other storage; Warehousing. Communication services: Radio, television broadcasting; Cable programming; Telephone and telecommunication; Postal and courier. Finance, insurance and real estate services: Paid charges to financial institutions; commissions and investment banking; Mutual funds, Other securities and royalties; Real estate commissions; Life and non-life insurance; Pension funds; Paid residential and non-residential rent and lodging. Professional services: Architect, engineering, scientific, accounting, legal, advertising and other professional services; software, computer lease, data processing and other information services; Investigation and security services; Other administrative and personal services. Education services: Elementary, Secondary, College and University fees and tuition. Other education fees. Health care and social assistance services: Private hospital, private residential care and other health and social services; Child care outside the home; Laboratory, physician and dental services; Other health practitioner services. Accommodation services and meals: Hotel, motel and other accommodation; Meals outside the home; Board paid. Wholesale services: Wholesale trade and wholesaling margins. Miscellaneous services: Beauty and other personal care services; Funeral services; Child care in the home; Private household services; Photographic, laundry and dry cleaning, services to building and dwellings; Automotive and other repair and maintenance; Rental of office, machinery, equipment, automobile and truck; Trade union and other membership organization dues and political parties contribution; Motion picture production, exhibition and distribution; Lottery, gambling and other recreation services.

A.1. Goods Agricultural products: unmilled wheat; corn, barley, oats and other grains, excluding imputed feed; live animals; other agricultural products (unprocessed milk, eggs, honey, vegetables, seeds, tobacco and wool). Mineral fuels: crude oil; natural gas, excluding liquified. Food products: meat, fish and dairy products (including processed milk); fruit and vegetable products; feeds; flour; breakfast cereal; sugar; cocoa; coffee, tea etc. Leather, rubber and plastic products: tires; other rubber products; plastic pipes; other plastics; footwear; gloves; handbags; other leather products. Textile products: yarns and fibers;fabrics; ropes, tents and threads; other textile products. Hosiery, clothing and accessories: hosiery; knitted clothing; furs; custom tailoring; other clothing. Lumber and wood products: lumber and timber; plywood and veneer; wood chips; prefabricated buildings; wood containers; caskets and coffins; other wood products. Furniture: household furniture; office furniture; mattresses; lamps; furniture parts; other furniture. Wood pulp, paper and paper products: wood pulp; newsprint; tissue; wrapping paper; paperboard; coated paper and paper products; paper bags; stationery; other paper products. Printing and publishing: newspapers; magazines; books; business forms; advertising; miscellaneous printing components. Primary metal products: ferro-alloys; iron and steel ingots; steel castings; bars and rods; flat iron and steel; railway construction materials; oil and gas pipe; other pipes and tubes; primary forms of aluminum copper, nickel, carbon, lead zinc etc.; precious metals excluding gold; scrap and waste; other primary metal products. Fabricated metal products: boilers; tanks; plates; iron and steel structural materials; metal doors and windows; stampings; containers; wire and cable; chains; utensils; wire products; hardware; machine tools; furnaces; cooking equipment; iron and steel forgings; valves; plumbing fixtures; gas and water meters; firearms; other fabricated metal equipment. Machinery: agricultural machinery; bearings; pumps; conveyors; elevators; fans; furnaces; industry-specific machinery for construction, oil and gas, logging metal working and other industries; power hand tools; refrigeration and air-conditioning equipment; scales; vending machines; computers; miscellaneous machinery. Motor vehicles and other transportation equipment: automobiles; trucks; buses; mobile homes; trailers; specialized vehicles; motor vehicle engines and parts; motor vehicle

References Agnosteva, Delina E., Anderson, James E., Yotov, Yoto V., 2014. Intra-national Trade Costs: Measurement and Aggregation. NBER Working Paper 19872. Anderson, James E., 1979. A theoretical foundation for the gravity equation. Am. Econ. Rev. 69, 106–116. Anderson, James E., 2011. The gravity model. Annu. Rev. Econ. 3, 133–160. Anderson, James.E., Eric, van Wincoop, 2003. Gravity with gravitas: a solution to the border puzzle. Am. Econ. Rev. 93 (1), 170–192. Anderson, James E., Milot, Catherine A., Yotov, Yoto V., 2014. “How much does geography deflect services trade? Canadian answers”. Int. Econ. Rev. 55, 791–818. Anderson, James E., Yotov, Yoto V., 2010. The changing incidence of geography. Am. Econ. Rev. 100 (1), 106–116.

J. Anderson, M. Vesselovsky, Y. Yotov / Journal of International Economics 100 (2016) 174–193 Anderson, James E., Yotov, Yoto V., 2011. .Specialization: Pro- and Anti-Globalizing, 1990-2002. NBER Working Paper 14423. Antweiler, Werner, Trefler, Daniel, 2002. Increasing returns and all that: a view from trade. Am. Econ. Rev. 92 (1), 93–119. Baier, Scott L., Bergstrand, Jeffrey H., 2007. Do free trade agreements actually increase members’ international trade?. J. Int. Econ. 71 (1), 72–95. Baldwin, Richard E., 2006. .The euro’s trade effect. Working Paper Series 594, European Central Bank. Chaney, Thomas, 2008. Distorted gravity: the intensive and extensive margins of international trade. American Economic Review 98 (4), 1707–1721. Chaney, Thomas, 2014. The network structure of international trade. American Economic Review 104 (11), 3600–3634. Cheng, I.-Hui, Wall, Howard J., 2005. “Controlling for heterogeneity in gravity models of trade and integration”. Federal Reserve Bank of St. Louis Review, (January) 49–63. Costinot, Arnaud, Rodriguez-Clare, Andres, 2014. “Trade Theory with Numbers: Quantifying the Consequences of Globalization” Handbook of International Economics, vol. 4; Gita Gopinath, Elhanan Helpman and Kenneth Rogoff, Editors. Elsevier North-Holland, Amsterdam. Delatte, Anne-Laure, Lopez-Villavicencio, Antonia, 2012. Asymmetric exchange rate passthrough: evidence from major countries. J. Macroecon. in press. Disdier, Anne-Célia, Head, Keith, 2008. The puzzling persistence of the distance effect on bilateral trade. Rev. Econ. Stat. 90 (1), 37–48. MIT Press. Eaton, J., Samuel, K., 2002. Technology, Geography and Trade. Econometrica 70, 1741–1779. Feenstra, Robert M., 2004. Advanced International Trade: Theory and Evidence. Princeton University Press, Princeton. Fieler, Ana Cecelia, 2011. Non-homotheticity and bilateral trade: evidence and a quantitative explanation. Econometrica 79 (4), 1069–1101. Genereux, P.A., Langen, B., 2002. The Derivation of Provincial (Inter-Regional) Trade Flows: The Canadian Experience. Manuscript. Statistics Canada Goldberg, Pinelopi.K., Knetter, Michael, 1997. Goods prices and exchange rates: what have we learned. J. Econ. Lit. 35, 1243–1273. Griffoli, T.M., 2006. Explaining the euro’s effect on trade? Interest rates in an augmented gravity equation. IHEID Working Papers 10-2006. The Graduate Institute of International Studies, Head, Keith, Mayer, Thierry, 2013. What separates us? Sources of resistance to globalization. Can. J. Econ. 46 (4), 1196–1231.

193

Head, Keith, Mayer, Thierry, 2014. “Gravity Equations: Workhorse, Toolkit, and Cookbook”. Handbook of International Economics vol. 4. Elsevier North-Holland, Amsterdam. Hillberry, Russell, Hummels, David, 2008. Trade responses to geographic frictions: a decomposition using micro-data. Eur. Econ. Rev. 52 (3), 527–550. Hummels, David, 2007. Transportation costs and international trade in the second era of globalization. J. Econ. Perspect. 21 (3), 131–154. Jones, Ronald, Kierzkowski, Henrik, Lurong, Chen, 2005. What does evidence tell us about fragmentation and outsourcing?. Int. Rev. Econ. Financ. 14 (3), 305–316. Manova, Kalina, Zhang, Zhiwei, 2012. Export prices across firms and destinations. Q. J. Econ. 127 (1), 379–436. Kim, M., Cho, G., Koo, W., 2003. “Asymmetric Pattern of Intra-industry Trade Between the United States and Canada”. Agribusiness and Applied Economics Report No. 452. Center for Agricultural Policy and Trade Studies, North Dakota State University, Krautheim, Sebastian, 2012. Heterogeneous firms, exporter networks and the effect of distance on international trade. J. Int. Econ. 87 (1), 27–35. Eslava, Marcela, Tybout, James, Jinkins, David, Krizan, C., Eaton, Jonathan, 2015. A search and learning model of export dynamics. Meeting Papers, 1535. Society for Economic Dynamics, Martínez-Zarzoso, I., Nowak-Lehmann, Felicitas, 2003. Augmented gravity model: an empirical application to mercosur-european union trade flows. J. Appl. Econ. 6 (2), 291–316. Mayer, T., Zignago, S., 2006. Notes on CEPII’S Distances Measures. CEPII Olivero, Maria Pia, Yotov, Yoto V., 2012. Dynamic gravity: endogenous country size and asset accumulation. Can. J. Econ. 45 (1), 64–92. Silva, Santos, Jo ao, M.C., Tenreyro, Silvana, 2006. The log of gravity. Rev. Econ. Stat. 88 (4), 641–658. Urbina, Ian, Bradsher, Keith, 2013. Linking factories to malls, middleman pushes low costs. New York Times Aug. 7 Yi, Kei-mu, 2010. Can multi-stage production explain the home bias in trade?. Am. Econ. Rev. 100 (1), 364–393. Yotov, Yoto V., 2012. A simple solution to the distance puzzle in international trade. Econ. Lett. 117 (3), 794–798.