The estimation of non-tariff trade costs in the empirical literature either uses ‘prices’ (domestic and foreign) or ‘quantities’ (trade flows), while a further sub-classification distinguishes between ‘direct’ and ‘indirect’ methods, depending on the respective explicit or implicit treatment of the non-tariff indicators (see Deardorff & Stern, 1998 ; Ferrantino, 2006 , for surveys). A cursory examination of the literature reveals that the quantity based method appears to be the more popular, in large part due to easy access to detailed public databases of trade ( Berden & Francois, 2015 ). Examining the price-approach, Bradford (2003 , 2005 ) calculates (rather than estimates) the implicit non-tariff impact on prices (indirect method), whilst Dean et al . (2009) and Cadot & Gourdon (2014) , employ an explicit non-tariff variable ( i.e. , direct method) to estimate its price rising effect.

Direct methods, in both price and quantity approaches, employ secondary data sources (both quantitative and qualitative) to construct a coverage ratio or dummy variable to indicate the degree of pervasiveness or the presence of a non-tariff restriction within the commodity of interest, which subsequently enters as an explanatory variable in an econometric model (price-dependent or gravity-type quantity-dependent). To this end, data may be taken from inventories of standards and regulations ( e.g ., UNCTAD TRAINS database); notifications to the WTO on the implementation of new trade regulations or complaints by traders ( e.g ., WTO World trade Reports). In the context of the EU-US trade relations, EP (2014) use the SPS and TBT notifications to the WTO to build a non-tariff variable that when interacted with the EU-US and US-EU) route dummy, allows for the direct estimation of bilateral AVEs of non-tariff costs. In their study, the sample includes a cross-section (year 2012) of OECD countries. Unfortunately, these methods neglect the relative importance of each measure in restricting trade, while countries which are more transparent appear as more restrictive ( Chen & Novy, 2012 ).

Arita et al . (2015) build a non-tariff variable that collects the incidence of SPS type measures between the EU and US in those cases where official concerns have been raised by US and EU exporters. A data sample consisting of three years (2010-2012) and a selection of countries ranging between 20 and 35 (depending on the sector) is constructed to estimate directly the non-tariff impact on bilateral trade. The choice of sectors and the direction for which the AVE is estimated is contingent upon the concern raised ( e.g ., The EU non-tariff AVE is estimated for red and white meat, maize and soy; whilst non-tariff AVEs for fruits and vegetables are estimated in both directions). In a similar fashion, Winchester et al . (2012) narrow the focus to target specific non-tariff costs based on exhaustive databases covering ( inter alia ) TBT and SPS, for specific countries and sectors. A disadvantage of these databases (both global or tailored) is the limited sectorial- and country coverage, whilst the data typically refers to only a single year which precludes the use of a panel database which, from a pure econometric perspective, helps to mitigate endogeneity problems. Another method of data extraction employs questionnaire responses on traders’ perceptions of market access. ECORYS (2009) interacts this non-tariff score variable with a route dummy ( i.e. from EU to US and vice versa ), which feeds into a standard cross-sectional gravity equation, to estimate bilateral EU-US non-tariff AVEs for agri-food. This approach is potentially open, however, to criticisms of limited sample size and response bias.

Two more direct quantity gap studies merit note. Owing to its sector (HS6 aggregation) and country coverage, Kee et al . (henceforth KNO) (2009) is recognised as the most comprehensive source of commodity specific non-tariff AVEs. The authors employ an aggregate import equation to estimate restrictiveness indexes from a dummy variable that accounts for the presence of non-tariff barriers (SPS and TBT) to trade plus domestic support. A second study by Li & Beghin (2012) conducts a meta-analysis (27 papers are considered) of direct quantity gap estimation methods to explain the variation of trade effects of health, safety and sanitary regulations and standards. The study considers differences in non-tariff measurement, data disaggregation and size, different estimation techniques and approaches to deal with zero trade values.

In contrast to the direct approach, which can isolate the trade restrictiveness resulting from specific, or groups of non-tariff restrictions, the indirect or implicit approach is better attuned to examining the collective trade restricting impact of all trade barriers which may otherwise be hidden ( Dean et al ., 2009 ). Thus, indirect methods start by acknowledging that trade barriers imposed by the importer country cause distortions in trade, reducing import quantities ( i.e. quantity-gap) and/or increasing import prices ( i.e. price-gap). Indirect quantity approaches infer the non-tariff impact on trade flows from ‘border-effects’ ( e.g . Chevassus-Lozza et al ., 2008 ), ‘fixed-effects’ ( Fontagné et al ., 2011 ), or from the depth of past trade agreements ( Egger et al ., 2015 ). As an alternative indirect approach, the ‘residual gravity approach’ infers the non-tariff impact on trade from the residuals by comparing the value of observed imports constrained by trade barriers, with the expected value of imports in the absence of said trade barriers predicted by the gravity equation ( Ferrantino, 2006 ).

The residual approach has been more extensively applied in services sectors ( e.g ., Park, 2002 ; Francois et al ., 2005 ; Guillin, 2013 ), as the direct method requires extensive databases on regulatory regimes on services which have not, hitherto, been available ( Jafari & Tarr, 2014 ). On the other hand, in the area of merchandise trade, this method has been applied to agri-food ( Philippidis & Sanjuán, 2007a , 2007b ), whilst the IMF (2002) have used this same approach to estimate the trade restricting effect of all non-tariff barriers in order to calculate ‘trade potentials’ for certain groups of countries. Following this same approach EP (2014) also calculate the trade potential between the EU and US after eliminating all possible trade barriers.

A comparison between both direct and indirect approach shows that the former provides a statistical assessment of the impact of non-tariff impediments to trade through examination of the dummy coefficient. Although this is not as straightforward in the indirect approach ( Dean et al ., 2009 ), it is still possible (see next section). A further observation when inferring trade costs ( i.e. , AVEs) using quantities, is the sensitivity of the AVE estimate to the value of the chosen import demand elasticity and/or elasticity of substitution. The direct price gap approach gets round this problem, as it allows the direct estimation of the impact of non-tariff impediments on prices. Finally, direct and indirect econometric approaches are susceptible to misspecification bias ( e.g ., omitted variable bias), whilst it has been suggested that the accuracy of the non-tariff calculation in the residuals-gravity approach is potentially even more contingent upon the estimation technique and the quality of the model specification On the other hand, AVEs derived from non-tariff dummy variables depend crucially on the quality of the measurement of the non-tariff impediment under consideration ( Ferrantino, 2006 ).

As stated in the introduction, an important aim of this research is to provide a platform upon which compatible measures of non-tariff trade costs may be implemented into modelling databases as a basis for conducting global trade impact analyses. To this end, the GTAP database is chosen with its broad, yet comprehensive, coverage of agri-food trade. Having taken the decision to employ this level of aggregation, we effectively rule out the use of non-tariff specific dummies, since at the GTAP sector concordance, at least one NTM will always be present, thereby resulting in a limited variability of observations ² .

Consequently, an indirect or residual quantity gap approach was favoured, whilst the approach adopted here is ‘specific’, in that it estimates non-tariff impediments to trade on a bilateral basis differentiating between intra- and extra-EU ( i.e. , EU-US) trade routes. To provide statistical rigour often lacking in the indirect method (see discussion above), confidence intervals for bilateral non-tariff trade cost equivalents were calculated by bootstrapping, whilst pairwise t-tests for means were applied to test for statistically different AVEs across bilateral routes.

In its simplest form, the gravity model posits that trade between two countries is a positive function of GDP ( i.e. , ‘mass’) and a negative function of trade costs ( i.e. , distance). Empirical applications have extended this basic premise to encompass ( inter alia ) preferential trade ( e.g ., Hayakawa & Yamashita, 2011 ), contiguity ( e.g ., Thoumi, 1989 ), common language and/or ex-colonial ties ( e.g ., Rose & van Wincoop, 2001 ), or even to cater for the effect of distance along different hemispheres as well as remoteness ( e.g ., Melitz, 2007 ). Other developments ( e.g ., Hallack, 2006 ) account for the so called ‘Linder’ (1961) hypothesis, which states that countries with similar per capita incomes have a greater tendency to engage in mutual trade. This is seen as a test of the monopolistic intra-industry hypothesis, whilst the polar opposite that differences in per capita incomes (which proxy for differing factor intensities) promote trade can be interpreted as support for the Hecksher-Ohlin (HO) hypothesis. This framework has also been extended to account for the role of infrastructure ( Limão & Venables, 2001 ; Donaubauer et al ., 2016 ) and logistics performance indicators ( Martí et al ., 2014 ).

The general theoretically-consistent gravity equation derived by Anderson & van Wincoop (2003 , 2004 ) is formulated as follows:

[1]

where X _ij are exports from country i to country j; Y _i and Y _j represent GDP, Y _w is world GDP, t _ij are trade costs i.e. t _ij = 1 + t _ij where is an ‘iceberg cost’ imposed by country j on imports originating from country i; and s is the elasticity of substitution between varieties ( i.e. countries). The price index variables _i and P _j , denominated as ‘multilateral resistance’ terms, are a function of bilateral trade barriers (t _ij ), and reflect the level of difficulty for country ‘i’ to engage in trade with country ‘j’, taking into account their bilateral trade barrier relative to the average trade barriers that both countries face with all their trading partners. Empirically, these unobserved terms are proxied with country specific dummies ( Anderson & van Wincoop, 2004 ). Furthermore, provided that these variables do not change over the time horizon of the data, country-fixed effects may also capture consumers’ preferences in the importing country or the number of varieties in the exporting country ( Disdier et al ., 2008 ).

In this study, a class of Poisson ³ gravity model was favoured ( Santos Silva & Tenreyro, 2006 , 2011 ) known as the Pseudo-Maximum Likelihood (PML) ⁴ estimator ( Gourieroux et al ., 1984 ). The model assumes that the observed volume of trade between countries i and j , X _ij follows a Poisson distribution with a conditional mean (µ _ij ) which is an exponential function of the explanatory variables z: µ _ij = exp (ß ’z _ij ) ⁵ .

Time series trade data were obtained from UN ComTrade database ( http://comtrade.un.org/ ), which has been widely used in the literature ( Burger et al ., 2009 ; Fadeyi et al ., 2014 ; Serrano et al ., 2015 ). Thus, annual HS6 bilateral import trade values for the years 2001 to 2011, and 78 countries was collected and reconciled with the GTAP nomenclature of 18 agri-food sectors (Table 1) using WITS concordances (World Integrated Trade Solutions by WTO [ http://wits.worldbank.org/wits ]). ⁶

Data on ad-valorem applied tariffs were taken from versions six ( Dimaranan, 2006 ), seven ( Narayanan & Walmsley, 2008 ), eight ( Narayanan et al ., 2012 ) and nine of the GTAP database corresponding to the years 2001, 2004, 2007 and 2011, respectively. Population and GDP were from the World Bank ( http://data.worldbank.org/data-catalog/world-development-indicators/wdi-2014 ), whilst data for cultural and geographical proximities was taken from CEPII ( Mayer & Zignago, 2011 ).

Based on a review of the literature, a comprehensive gravity specification is formulated, and a full description of the variables is presented in Table 2. The explanatory variables of infrastructure and logistics were finally discarded owing to their incomplete country and/or temporal coverage from the available sources of data ⁷ . The Poisson regression model with an exponential mean function is presented in Eq. [2] ⁸ , where the sub-index i and j refer to the exporter and importer, respectively, whilst t refers to the year:

In the indirect quantity gravity approach, discrepancies between actual ( AX _ij , i.e. variable Xij in Eqs. [1] and [2]) and predicted ( PX _ij , i.e. E[Xij] in Eq. [2]) values of trade are taken to be indicative of trade barriers, as the prediction by the gravity equation is assumed to reflect potential trade after controlling for observed trade frictions. Given that applied tariffs are included explicitly in the model, trade barriers implied by the residuals are considered to be due to non-tariff barriers. Thus, the trade cost t _ij in Eq. [1] (after controlling for observed trade frictions) is assumed to reflect the NTMs:

[3]

where is the AVE of all the NTMs imposed by country j on imports originating from country i .

Following Park (2002) and Francois et al . (2005) , the AVEs are calculated by trading partner averaged over all import routes. Thus, in this study, for each country j, actual and predicted imports were summed over all its trade partners: and (in our application N<=78). In order to quantify the magnitude of the trade barriers, we also followed the aforementioned authors by normalizing each observed-to-predicted trade ratio with a benchmark of the largest ratio of observed-to-predicted trade, which is interpreted as a ‘relative free-trade benchmark’ ratio ( AX _b / PX _b ) ⁹ .

Solving for the AVE of non-tariff measures ( τ _ij ^NTM ) imposed by country j:

[4]

In the current study, the interest lies in the specific non-tariff restrictions between the EU and US ( i.e. , intra-EU; EU to US; US to EU), which in turn, requires a modification of Eq. [4]. For example, for EU imports of US goods, the AVE becomes:

[5]

where AX _US→EU (PX _US→EU )adds up observed (predicted) imports coming from the US to any country in the EU:AX _US→EU =∑ ^k _j=1 AX _US→j and PX _US→EU =∑ ^k _j=1 PX _US→j (k=number of Member States in the EU). ¹⁰ Sectoral substitution elasticities (s) across importing sources in [4] and [5] were taken from the GTAP database ( Hertel et al ., 2007 ) and Kee et al . (2008) . ¹¹ In addition to the point estimates, confidence intervals for bilateral non-tariff trade cost equivalents were calculated. ¹²

Making comparisons with existing literature is complicated by the choice of sectorial and country aggregation, the estimation procedure, the quantification method and years of reference in the sample. Using WTO (World Integrated Trade Solutions-WITS) concordances between GTAP and HS6 codes, KNO (2009) arithmetic average AVEs for agricultural and food imports by GTAP sector were calculated (columns 3 and 6, Table 3) based on a sample of 61 countries which are common to both KNO and the current study. ¹³ The current study employed Eq. [4] to calculate country non-tariff AVEs which were later averaged for each GTAP sector (columns 2 and 5, Table 3).

Examining Table 3, our average estimate (second column) for both the agri-food and agricultural composites (25% and 30%, respectively) were very similar to the corresponding KNO estimates (31% and 25% - third column), whilst in the case of the aggregate food processing sector, the results were also broadly comparable (20% vs 34% - columns 5 and 6). On a sector-by-sector basis, KNO estimates for processed food products (except beverages and tobacco) were higher than those reported with our approach, while our estimates for agricultural products were higher in six out of ten sectors. In almost half of the sectors examined, the gap in mean AVE estimates between both approaches was less than 10 percentage points, with a minimum difference of 5 percentage points for fruits and vegetables and beverages and tobacco. In the case of food processing, the largest difference can be found in processed rice (110% versus our estimate of 28%). In agriculture, the main differences occured in plant based fibres (18% versus our estimate of 82%).

As a further basis of comparison, Table 4 compares our country specific estimates (columns 2 and 6) from Eq. (4) with two quantity gap studies (KNO, 2009; EP, 2014) and the price gap approach of Bradford (2003) . In the case of agriculture, our non-tariff AVE estimate for the US and EU-27 was 27% and 29%, respectively. This is broadly compatible with corresponding estimates of 22% and 27% in KNO and 33% and 44% in EP (2014). Across the same three quantity gap studies, the divergence was higher for processed food, where KNO estimates were placed around the midpoint between the lower values in the current study and upper values of EP (2014). Despite the use of a different methodology, EU non-tariff AVE estimates in Bradford (2003) converged closely with the three quantity gap studies, whilst in common with our results Bradford (2003) concludes that agricultural non-tariff restrictions are more prohibitive than those of food in both the EU and US. Remarkably, all four studies conclude that ‘behind the border’ trade costs in the EU are more prohibitive than those of the US, although this finding is not consistent with non-tariff estimates of ‘food processing’ reported by OECD (2011) (30.1% and 49.5%, for EU and US, respectively). ¹⁴

In a next step, Eq. [5] was employed to calculate non-tariff AVEs on intra-EU and trans-Atlantic trade in both directions (Table 5). The point AVE estimates reported here for (trade weighted) agri-food imports to the EU from the US (35%) and to the US from the EU (27%) are lower than those reported by ECORYS (2009) (56.8% and 73.3%, respectively). Apart from the different sectorial coverage and disaggregation, differences in results are also attributed to the choice of data, econometric estimator and modelling approach. Our estimates are more in line with those of EP (2014). Thus, examining the simple average agri-food estimates in our study, we estimated AVEs of 47% and 39% for the EU and US, respectively, compared with corresponding estimates in EP(2014) of 54% and 48%. ¹⁵ Encouragingly, across the three bilateral routes, AVE point estimates for the ‘agri-food’ aggregate moved within the range of 17% to 35% (Table 5), which is more in line with the value of 31% reported by KNO (see Table 3). In addition, our AVE point estimates for extra-EU imports were 38% (agriculture) and 32% (food), which compares favourably with the corresponding extra-EU estimates by KNO (Table 4) of 27% and 40%. A similar observation is true for the US’s imports.

Examining intra-EU trade for primary agricultural commodities (Table 5), the trade weighted AVE mean estimate was 23%. Moreover, individual sector estimates ranged between 15% and 19% in the broad (heterogeneous) sectors of ‘other crops’ and ‘other grains’, to 48% in ‘cattle and sheep’ and 63% in ‘plant based fibres’ (Table 5), whilst in the case of intra-EU processed food trade, the AVEs were of a lesser magnitude and more homogeneous in magnitude across sectors. On trans-Atlantic trade routes, the US was estimated to impose AVEs as high as 91%, 82% and 71% on imports of ‘cattle meat’, ‘plant based fibres’ and ‘dairy’, respectively. In turn, the EU AVE peaks were apparent on imports of ‘dairy’ (90%), ‘cattle meat’ (71%) and ‘pig/poultry meat’ (63%).

Comparing across the three bilateral routes, with the exceptions of ‘oilseeds’, ‘cattle’ ‘wool’, and ‘beverages and tobacco’, the AVE was found to be lowest on the intra-EU trade route (as expected). In addition, in those sectors where US non-tariff AVEs are deemed higher than their EU equivalents ( i.e. , ‘cattle meat’, ‘processed sugar’ and ‘processed rice’), with the exception of the latter, the result was found to be statistically significant. In the case of ‘vegetables and fruit’, despite the close AVE estimates of 35-36% in both directions, a t-test of the mean bootstrapped values revealed that the US AVE was higher; a result supported by the higher upper limit within the US confidence interval (see below). In the majority of sectors (14), the results appeared to indicate that EU AVEs are equally or more prohibitive than those of the US, which further supports the observation made at the end of the previous section (Table 4). On the other hand, pair-wise t-tests showed that this finding is only supported statistically in eight of those sectors ( i.e. , ‘other grains’, ‘oilseeds’, ‘plant based fibres’, ‘other crops’, ‘white meat’, ‘vegetable oils’, ‘dairy’, ‘beverages and tobacco’). Moreover, when testing for differences in the trade weighted AVE means on trans-Atlantic routes for agriculture, food and agri-food categories, no statistical difference was found. Interestingly, EP (2014) also reports this same finding.

The calculation of confidence intervals helped to capture the degree of certainty behind the AVE point estimates over a time period of the sample. Inspecting the results in Table 5, EU AVE point estimates, particularly on intra-EU trade, exhibited greater accuracy than those of the US. On closer inspection of the data, there were cases of small import trade flows (owing to high self-sufficiency) accompanied by significant import volatility (particularly on ( inter alia ) ‘other grains’ and ‘processed sugar’ trade from the EU to the US), whilst in other sectors ( e.g ., vegetable oils and fats trade from the EU to the US) the presence of a structural break was observed relating to an array of indeterminate sector specific events ( i.e. , weather shocks, agricultural policies, data aggregation).

Comparing between (trade weighted) non-tariff and tariff AVE’s on EU-US bilateral trade (Table 5), for the selection of agri-food products under consideration (with one exception ¹⁶ ), NTMs (as expected) were found to be more trade prohibitive on trans-Atlantic imports in both directions. For aggregate agricultural imports, the tariff and non-tariff AVEs for the EU (US) were 3% and 38%, (2% and 30%), respectively. In the case of food, the corresponding estimates were 13% and 32% (2% and 27%), respectively. This finding reinforces the view expressed in the introduction that non-tariff restrictions have replaced tariffs as the dominant form of (agri-food) trade protectionism.