Globalisation and agri-food trade

Abstract

1. Introduction

Globalisation has shaped international trade over the last 50 years, causing markets to become more integrated due to new transportation technologies, the enhancment of communication capabilities, the reduction of trade restrictions and the evolution of international value chains (Jacks, Meissner and Novy, 2011; Yotov, 2012). However, primary agricultural products generally have a low per-unit value, can be perishable, are bulkier than non-agricultural products and do not benefit from production fragmentation. Furthermore, while processed food products generally have higher unit values, they can have strict transportation requirements to protect consumer health. As distance is the most encompassing proxy for bilateral trade costs (Keith and Thierry, 2014; Borchert and Yotov, 2017), an important question is, has globalisation—that rendered countries ‘closer’ over time—reduced the trade-diminishing effect of distance on international agricultural and food markets? This paper investigates the impact of globalisation on bilateral agricultural and processed food trade over the years 1986–2019.

Early studies investigating the impact of globalisation through distance on bilateral trade flows found that the effect of distance on trade increased over time as opposed to decreasing as predicted based on globalisation (Coe et al., 2002; Brun et al., 2005; Disdier and Head, 2008). The rise in the distance coefficients over time is known as the distance puzzle (Buch, Kleinert and Toubal, 2004). In a structural gravity context, Yotov (2012) and Yotov et al. (2016) offered an elegant solution for properly measuring the impact of globalisation on distance: the impact of distance on international trade costs must be measured relative to the impact of distance on intra-national costs by isolating intra-national distance and including the home bias effect in the model. After controlling for relative intra-national trade costs and the home bias effect, the impact of distance on bilateral trade costs declines over time, resolving the distance puzzle. Yotov (2012) utilises World Bank’s World and United Nations Comtrade data for 93 countries over the period 1965-2005 and includes food manufacturing at the two-digit ISIC level for the years 1995–2000. The results indicate that without isolating intra-national costs the impact of distance on bilateral food trade rises by 6.5 per cent between 1990 and 2000; however, after controlling for relative intra-national costs the impact of distance declines by 7.9 per cent. Yotov et al. (2016) provide further evidence that the distance puzzle has been resolved by showing that the solution to the distance puzzle is robust to the home bias effect by including a border indicator variable. However, no study has provided a detailed analysis of the impacts of globalisation—measured by changes in the estimated coefficient of distance—on bilateral agricultural and processed agricultural and food trade between 1986 and 2019. This paper aims to fill this gap in the agricultural trade literature.

Countries favour agricultural industries through complex domestic consumer and producer support programmes, trade policies (licensing requirements, tariffs, quota, export restrictions, sanitary and phytosanitary restrictions, technical barriers to trade, etc.) and state trading enterprises. In addition, agricultural and food products have a relatively inelastic demand for food (Guimbard et al., 2012; Greenville, Kawasaki and Jouanjean, 2019). Agricultural products can be bulky, perishable and have relatively low unit values, which has resulted in fragmented and local markets in agri-food, particularly before the 1990s (Reardon et al., 2003). As a result, per-unit trade costs may be higher and less sensitive to the effects of globalisation on trade costs than non-agricultural products (Bonuedi, Kamasa and Opoku, 2020; Egger et al., 2021; Gaigne and Gouel, 2022; Karagulle, Emlinger and Grant, 2022). Furthermore, global value chains, and thus the benefits of globalisation, have likely been less salient in agri-food industries than in other manufacturing industries. Lower trade costs can alleviate food insecurity, and many countries impose subsidies and barriers to imports and exports to be less reliant on the international markets of staple food production and promote self-sufficiency. Therefore, it is important to provide a detailed analysis of globalisation in the agricultural and food sectors. This study considers both primary and processed food items to provide a deeper understanding of globalisation and the impact of distance on bilateral trade costs in agri-food trade over the last three decades.

The agricultural trade literature utilising structural gravity models to examine the impact of globalisation on agricultural trade is scant.¹ Related work includes Karagulle, Emlinger and Grant (2022), who examined the effects of globalisation on agricultural trade by studying the evolution and determinants of ad valorem equivalent trade costs of friction variables and policy variables over the period 2001–2018 on disaggregated agricultural and processed food sectors. Karagulle, Emlinger and Grant (2022) utilised both inter- and intra-national trade data from Comtrade and UN Statistics Division Comtrade and the Food and Agricultural Organization for SITC (Revisions 1) products ranging from 0111 to 4224. They estimated the evolution of ad valorem equivalent trade costs and generally found agricultural trade costs rise over their 17-year sample. They then estimated the average impact of distance on agricultural trade costs for all agricultural and processed food commodities and found, as expected, that larger distances led to higher trade costs. With substantial heterogeneity in the results that depend on the time frame, sector and region, the agricultural trade literature is far from a consensus on how globalisation has impacted agricultural trade. The current study complements this work by examining the evolution of distance in agricultural and processed agricultural and food production trade, the relative effect of international distance and the home bias effect over a 33-year sample.

This study contributes to the agricultural trade literature in four ways: by (i) investigating the ‘distance puzzle’ in an agri-food trade context; (ii) providing estimates of the impact of intra-national trade costs on intra-national trade (domestic sales); (iii) providing estimates for the home bias effect for primary agricultural commodities (ACs) and processed agricultural and food commodities (PCs); and (iv) correcting for substantial omitted variable bias in trade friction variables when international distance is measured relative to intra-national distance ( i.e. estimating international and intra-national distances separately) and the home bias effect is included in agri-food gravity-model estimations.

To investigate the impact of globalisation on distance in agri-food trade, we utilise a structural gravity model with friction and policy variables, year-specific estimates for distance, controls for intra-national distance and the border effect. The model also includes importer-year and exporter-year fixed effects to control for multilateral resistance terms, which account for all observed (e.g., GDP and industry size) and unobserved (e.g., remoteness of a country) factors. These terms explicitly capture all third-party interactions, trade diversion and creation effects, and a country’s ability to import and export agri-food products in a given year.

The results show that countries with greater internal distances hinder domestic sales. Furthermore, the home bias effect is substantially larger in ACs relative to PCs. Importantly, separating internal distance from international distance and including the home bias effect in the gravity model correct for downward bias in the coefficient estimates of contiguous borders and for upward bias in the coefficient estimate of common language, colonial ties and distance. The distance puzzle exists in a structural gravity model without controlling for the relative intra-national distance or the home bias effect when comparing the estimated coefficients for 2019–1986. However, the distance puzzle disappears when the model controls for the relative intra-national distance and the home bias effect as the percent changes in estimated coefficients for distance are statistically insignificant. By contrast, for more recent years, the distance puzzle does not arise as the negative impact of distance on agricultural trade is smaller in 2019 compared to 1997 and 2008 whether or not the regressors for relative intra-national trade costs and the home bias effect are included in the model. Examining each commodity individually reveals substantial heterogeneity in the evolution of the impact of distance on bilateral trade.

The next section details the gravity model. Section 3 discusses data and sources. Section 4 presents the results. And Section 5 concludes the paper.

2. Gravity model

Based on the seminal work of Anderson and van Wincoop (2003), the structural gravity equation is

where |$X_{ij,t}^{k}$| is the value of exports from exporter i to importer j (including intra-national trade where i = j) in year t for product k, |$T_{i,t}^{k}=\sum_{j}X_{ij,t}^{k}$| is the value of production of product k in country i, |$T_{j,t}^{k}=\sum_{i}X_{ij,t}^{k}$| is the value of consumption of product k in country j, |$T_{t}^{k}=\sum_{i}\sum_{j}X_{ij,t}^{k}$| is the aggregate value of trade for product k, |$\varphi_{ij,t}^{k}$| represents both time-invariant bilateral trade costs, such as distance, common language, colonial relationship, borders, etc., and time-dependent trade policies, such as the General Agreement on Tariffs and Trade (GATT), World Trade Organization (WTO), Free Trade Agreements (FTAs), etc., |$P_{i,t}^{k}$| is the outward multilateral resistance, |$P_{j,t}^{k}$| is the inward multilateral resistance and |$\sigma^{k}\gt1$| is the trade elasticity of substitution.

To understand the evolution of globalisation in bilateral agricultural trade over the 33-year period between 1986 and 2019, we allow for differential effects of distance for each year in the sample. As is standard in the trade literature (Santos Silva and Tenreyro, 2006; Yotov et al., 2016), we use the Poisson Pseudo Maximum Likelihood (PPML) estimator to account for both heteroscedasticity and sample selection biases. Therefore, we define bilateral trade costs and trade policies as

where βs are coefficients, Cont_ij is an indicator variable equal to 1 if i and j have a contiguous border and 0 otherwise, C Lang_ij is an indicator variable equal to 1 if i and j have a common language and 0 otherwise, Colony_ij is an indicator variable equal to 1 if i and j have a common coloniser and 0 otherwise, |$GATT/WTO_{ij,t}$| is an indicator variable equal to 1 if both i and j were/are GATT and WTO members, respectively, in year t and 0 otherwise, |$FTA_{ij,t}$| is an indicator variable equal to 1 if both i and j have a free-trade agreement in year t and 0 otherwise,²|$\text{log}D\left(t\right)_{ij}$| is the log of international distance interacted with year indicator variables and is equal to 0 for intra-national distance where i = j and |$\text{log}D_{ii}^{Intra}$| measures intra-national distances and is equal to 0 for international distance where i ≠ j, |$HB_{ii}$| is a border indicator variable that is equal to 1 for intra-national trade (sales to self) and 0 for international trade, |$I_i$|is an indicator variable that is equal to 1 for county i and 0 otherwise, |$N$|is the total number of countries in the sample. The GATT/WTO and FTA indicator variables control for all observed (tariffs and NTMs) and unobserved (streamlining of policy and regulations) policy implications of these agreements. The variable |$HB_{ii}I_i$| controls for country-specific home bias effects, as domestic sales are typically larger than international sales, and for the unobserved effects that impact domestic sales versus trade differently (Anderson and Yotov, 2010).

As such, by logging the right side of (1) and substituting (2) into (1), the structural gravity model to be estimated using PPML is expressed as

where |$\beta_{0,t}^{k}=T_{t}^{k}$| is the intercept equal to the aggregate trade value, |$\lambda_{i,t}^{k}=T_{i,t}^{k}+P_{i,t}^{k}$| and |$\lambda_{j,t}^{k}=T_{j,t}^{k}+P_{j,t}^{k}$| are exporter-time and importer-time fixed effects and |$\epsilon_{ij,t}$| is an error term (Yotov et al., 2016). The exporter- and importer-time fixed effects |$\lambda_{i,t}^{k}$| and |$\lambda_{j,t}^{k}$| represent the outward and inward multilateral resistance terms and control for all time-varying non-discriminatory export and import factors that are observable (e.g. market size, importers’ and exporters’ domestic policies and climate) and unobservable (e.g. technology and political differences) that impact a country’s openness to trade and ability to import and export agricultural products. Note that, with the objective of examining changes in the impact of distance on agri-food trade over time, we are not able to include country-pair fixed effects to control for unobserved heterogeneity in the policy variables |$GATT/WTO_{ij,t}$| and |$FTA_{ij,t}$|⁠. Therefore, these two policy variables are included in the regression for completeness and to avoid any omitted variables.

3. Data

Bilateral trade values (including intra-national trade) for 209 import countries and 210 export countries are collected from the International Trade and Production Database for Estimation (ITPD-E) release 2 database published by the US International Trade Commission (Borchert et al., 2021). ITPD-E includes zero trade flows and domestic sales data for each commodity which is calculated by subtracting total exports from the (gross) values of the total production. For ACs, the dataset includes 4.8 million observations, 19 commodities³ and the years 1986–2019. For PCs, the dataset covers 9.6 million observations, 20 PCs and the years 1988–2019. Table 1 lists the industry, codes and number of observations for each industry utilised in the analysis. The AC with the most observations is 26 ‘Other agricultural products, n.e.c.’ with 559,554, while 9 ‘Raw and refined sugar and sugar crops’ has the fewest observations with 21,611. The PC with the most observations is 55 ‘Textile articles except apparel’ with 739,452, while 43 ‘Prepared animal feeds’ has the fewest observations with 325,852. Note that ITPD-E drops all irrelevant zeros by retaining observations that are maintained when estimating a structural gravity model with the PPML estimator and exporter-time, importer-time and country-pair fixed effects. This process omits singleton observations for which bilateral trade values are perfectly predicted by exporter-time, importer-time and country-pair fixed effects. Also, not all countries are available for each year. For instance, only 170 countries are available for the year 1986.

From the Dynamic Gravity Dataset, the US International Trade Commission, we collect data for standard gravity friction and policy variables (e.g. distance, common language, common coloniser, contiguous border, WTO, GATT and FTAs). The Dynamic Gravity Dataset provides population-weighted distance between country pairs based on an average distance between major cities in each country. Importantly, the database uses the city-based approach to measure internal distance within a country, which reflects the distance domestically produced and consumed agricultural products travel within a country, on average.

4. Results

This section considers the average effects of trade friction variables by stacking all 19 ACs into one dataset and by stacking all 20 PCs into one dataset. Then, heterogeneity across commodities is considered by examining each agricultural and processed food commodity individually.

Four gravity models are estimated.⁴ First, to provide a baseline, model (1) estimates equation (3) as a standard structural gravity with friction variables by combining intra-national distance with international distance in one variable and excluding the border variable, HB_ii. Model (2) controls for the international distance relative to intra-national distance, but not the home market effect. Model (3) controls for both the relative intra-national distance and the border effect. Model (4) is the preferred model and controls for the relative intra-national distance and country-specific border effects. By definition, the distance puzzle arises if distance hinders the AC trade more in 2019 compared to 1986, 1997 and 2008 (intervals of 11 years for the 33 years the AC dataset spans) and hinders the PC trade more in 2019 compared to 1988, 1998 and 2008 (intervals of about 10 years for the 31 years the PC dataset spans).⁵ To ease the analysis and interpretation, we calculate the percent changes in distance coefficients between 2019 and 1986, 1997, and 2008 for ACs and between 2019 and 1988, 1998, and 2008 for PCs. By comparing these percent changes across models and different time periods, we uncover whether the distance puzzle exists in ACs and PCs and understand how globalisation has impacted trade costs through distance in bilateral AC and PC trade.

4.1. Average effects

Table 2 reports the estimated coefficients for ACs and PCs for the four gravity models and the percent changes in the distance coefficient estimates.⁶ The results for the friction variables (contiguous borders, common language and colonial relationship) are discussed first, the policy variables (GATT/WTO and FTAs) are briefly considered, followed by the presentation of the distance variables and finally, the evolution of globalisation is revealed by examining the percent changes in distance over time in each model.

The results for contiguous borders, common language and colonial relationship reveal that bias in the coefficient estimates is controlled for by isolating intra-national distance and controlling for the home bias effect. For example, for contiguous borders, for both ACs and PCs, the estimated coefficients reverse sign as they shift from negative and statistically significant to positive and statistically significant when moving from model (1) to models (2)–(4)⁷, which controls for downward bias in model (1). Based on model (4), countries that share a border increase trade by 43.764 per cent (=|$100\times\left(\text{exp}\left(0.363\right)-1\right)$|⁠, on average, for ACs and by 68.203 per cent (=|$100\times\left(\text{exp}\left(0.520\right)-1\right)$|⁠, on average, for PCs. Furthermore, the coefficient estimates for common language and colonial ties decline when comparing the coefficient estimates in model (1) to those in models (2)–(4), indicating upward bias in the coefficient estimates in model (1) for both ACs and PCs.⁸ The estimated coefficients show that common language increases trade by 36.615 per cent |$(=100\times\left(\text{exp}\left(0.312\right)-1\right)$| for ACs and 50.230 per cent |$(=100\times\left(\text{exp}\left(0.407\right)-1\right)$| for PCs in model (4). However, while the coefficient estimates for colonial ties are positive and statistically significant for models (1)–(4), they are statistically insignificant for both ACs and PCs in the preferred model (4). These results reveal an important finding that controlling for relative intra-national trade costs for both relative intra-national trade costs and the home market effect and for relative intra-national trade costs and country-specific home market effect reduces bias in the estimated coefficient of these friction variables in model (1).

The distance coefficient estimates show similar patterns to that of contiguous borders, as for each year the estimates become less negative (i.e. closer to zero) when comparing the results from model (1) to model (2) and from model (2) to model (3) for both ACs and PCs; however, for model (4) the estimated coefficients generally become slightly more negative from model (3) to model (4). The distance coefficients indicate that, for ACs, a 1 per cent increase in distance lowers trade by 1.403 per cent in 1986 and 2.090 per cent in 2019 under model (1), which declines to 1.063 per cent in 1986 and 1.010 per cent in 2019 in model (4). For PCs, a 1 per cent increase in distance lowers trade by 1.246 per cent in 1988 and 1.746 per cent in 2019 under model (1), which declines to 0.829 and 0.895 per cent, respectively, in model (4). These results further corroborate the findings of the non-distance friction variables in that isolating intra-national trade costs and controlling for the home market effect reduce downward bias in the distance coefficient in both commodity classes.

The estimated coefficients for relative intra-national trade costs, d_ii, in model (2) for both ACs and PCs are negative, suggesting that countries with larger distances between major cities have higher transport costs and other domestic frictions, which lowers domestic sales. Specifically, a 1 per cent increase in the internal distance between major cities leads to a decline in domestic sales by 0.801 per cent for ACs and by 0.486 per cent for PCs, on average, in model (2) and by 1.331 per cent for ACs and by 0.556 per cent for PCs in model (3). These results are in line with the findings of Yotov et al. (2016) who find that intra-national distance lowers domestic sales of aggregate goods by 0.488 per cent. Our results are the first estimates of the impact of isolating domestic distance on domestic sales in the agricultural trade literature.

The estimated coefficients of the border-effect indicator variables are 5.727 for ACs and 0.936 for PCs. The estimated coefficient for ACs indicates a strong home bias effect, with intra-national (domestic) sales being a massive 307 (=|$\text{exp}\left(5.727\right))$| times larger than international trade. However, the results suggest a weak home bias effect for PCs, with domestic sales being only 2.55 (=|$\text{exp}\left(0.936\right))$| times larger than international trade. While the home bias of 2.55 for PCs is similar in magnitude to the home bias of 5.5 for aggregate trade reported in Yotov et al. (2016), the AC home bias effect is two orders of magnitude larger. The out-sized home bias effect in ACs compared to aggregate trade in Yotov et al. (2016) and PCs could arise because many countries do not want to be dependent on foreign markets for their primary agricultural products, which results in larger domestic sales relative to international sales. Furthermore, as previously mentioned, many developing countries have domestic goals of being self-sufficient in staple commodities such as rice, wheat, corn and meat products  and, thus, have extensive export restrictions and/or domestic policies that favour domestic sales relative to exports.⁹ These policies that favour domestic sales over exports largely do not arise outside of primary agricultural goods. Another explanation could be that processed food producers (e.g. Coca-Cola and Starbucks) can expand their consumer base by selling in foreign markets, which can result in a smaller estimate for the home bias effect. These are the first home bias effects estimated for agricultural production and provide key insights into agricultural and processed food trade.

For completeness, we report the coefficient estimates for the GATT/WTO and FTA. However, given that the primary focus of the paper is on distance and understanding the distance puzzle in agri-food commodities, we cannot include country-pair fixed effects to control for all observed (i.e. distance) and unobserved (i.e. political ties) factors that influence bilateral trade links. In this case, the coefficient estimates on the GATT/WTO and FTA suffer from endogeneity bias from omitted variables and the coefficients cannot be interpreted reliably.

Globalisation should result in the impact of distance on bilateral trade declining over time as the world gets ‘flatter’, as coined in Friedman (2005). However, based on the standard structural gravity model (1), the estimated coefficient for distance declined from −1.403 in 1986 to −2.090 in 2019 for ACs, a 49.038 per cent increase in absolute value, and from −1.246 in 1988 to −1.746 in 2019 for PCs, a 40.150 per cent increase in absolute value, confirming the distance puzzle for both ACs and PCs. Controlling for relative intra-national trade costs in model (2) reduces the increase (in absolute value) in the estimated coefficients for distance, but the estimated coefficient for distance still increases by 19.934 per cent for ACs and by 17.947 per cent for PCs (though statistically significant only at the 14 per cent level for PCs). Furthermore, including the control for the home bias effect in model (3) reduces the rise (in absolute value) in the impact of distance on bilateral trade, but while the percent changes are still positive at 2.866 per cent for ACs and 15.263 per cent for PCs, they are statistically insignificant. Under the most flexible specification with country-specific border effects (model (4)), the percent change for ACs is negative (though statistically insignificant) as the distance impedes bilateral agricultural trade about 5 per cent less in 2019 than in 1986. And, for PCs, the percent change in the distance coefficients is smaller than in model (3), but the estimated coefficient for distance is still larger in 1988 than in 2019 (though statistically insignificant).

Moving from model (1) to model (4), the percent difference in the impact of distance on bilateral trade between 1986 and 2019 lessens (and is even negative for ACs in model (4)). However, the distance puzzle only explicitly arises in model (1) for ACs and PCs and in model (2) for ACs. Therefore, while the distance puzzle does not arise as the percent changes are not statistically different from zero, we can reject the hypothesis that globalisation has reduced the impact of distance on AC and PC trade over time. These results are contrary to the findings in Yotov et al. (2016), where isolating intra-national distance alone (equivalent to model (2) in this paper) resolves the distance puzzle for aggregate trade between 1986 and 2006.

Globalisation may not lower the impact of distance on AC and PC trade because the bulky and perishable nature of ACs and PCs makes transporting more difficult. Furthermore, international value chains are a key mechanism through which globalisation has impacted transport costs, which are less prominent in agricultural and processed food products than other manufacturing commodities. Therefore, these results suggest that the effects of globalisation of making the world ‘flatter’ are not fully realised in agri-food industries between 1986 and 2019.

However, the existence of the distance puzzle and the evolution of globalisation vary depending on the period studied. For instance, the percent change in the distance variable for 1997–2019 and 2008–2019 for ACs and 1998–2019 and 2008–2019 for PCs for models (1)–(4) declines (in absolute values), suggesting that the distance puzzle did not arise. Furthermore, because the percent changes are negative and statistically significant, globalisation resulted in lower trade costs, proxied by distance, for more recent time periods. We surmise that the distance puzzle did not arise from 1997 onward because of the defragmentation of local food markets that arose (particularly in developing countries) as globalisation drove the so-called ‘supermarket revolution’ in the mid-1990s, starting first with grains and then fresh products, and advanced in procurement logistics technology and inventory management (Reardon et al., 2003; Humphrey, 2007; Atkin, 2015).

4.2. Heterogeneous effects

The effects of globalisation on distance will likely have differential impacts on individual ACs and PCs, which is the focus of this subsection. We first briefly summarise the results for intra-national distance and home bias effects (not presented in the table or figures).¹⁰ Then, we provide a graphical analysis and detailed discussion of the percent changes in the distance coefficient estimates over the three intervals discussed in the previous subsection.

The estimated coefficients for the intra-national distance and the border effect reveal substantial heterogeneity. For example, the results show that a 1 per cent reduction in intra-national distance reduces domestic sales of ACs by between statistically zero for commodity 9 sugar and 91.48 per cent for commodity 2 rice¹¹ and reduces domestic sales of PCs by between statistically zero for commodities 39 oils/fats, 40 dairy, 47 noodles and 54 textiles and 85.32 per cent for commodity 49 spirits. Sugar may have a very low domestic distance effect because it is largely sold in international markets (corroborated by the statistically zero home bias effect discussed later), while the majority of rice production is consumed domestically, especially in Asian countries, leading to high domestic sales.

The home bias effect shows that domestic AC sales are favoured to international sales by between statistically zero for commodities 1 wheat, 3 corn, 6 soybeans and 20 other meat and 1.29 million times for commodity 24 tobacco leaves. For PCs, the home bias effect shows that domestic PC sales are favoured to international sales by between statistically zero for commodities 36 meat, 38 processed fruit/vegetable, 41 grain mill products, 42 starch, 43 animal feeds, 45 sugar, 46 chocolate, 52 soft drinks and 55 textile articles and |$18,881.01$| times for commodity 49 spirits. Wheat, corn, soybeans, vegetables and their processed counterparts likely do not have a home bias because these commodities typically have highly intensive production with relatively low domestic sales in areas with comparative advantages in land, capital and weather. A similar argument could be made for sugar, chocolate, soft drinks and textiles. The massive home bias effect for tobacco could arise because of the particularly stringent domestic and international regulations on tobacco products and smokers’ strong preference towards a cigarette brand (Organization, 2013; Johnson, Coleman and Schmitt, 2016). We surmise that the reason for the large effect on spirits is similar to that for tobacco.

Figure 1 reports the percent change in distance coefficients between 1986 and 2019 (Panel a), between 1997 and 2019 (Panel b) and between 2008 and 2019 (Panel c) for all ACs for model (1) ‘Baseline’ depicted as |$\triangle$|⁠, model (2) ‘Intra Dist’ depicted as +, model (3) ‘Intra Dist+Border’ depicted as × and model (4) ‘Intra Dist+Ctry Spec Home Bias’.¹² As seen in Panel a, the results show substantial heterogeneity in the magnitude of the percent changes in distance over three decades and the existence of the distance puzzle (i.e. the impact of distance on bilateral trade is larger in 2019 than in 1986) across commodities. While the distance puzzle arises under a standard structural gravity specification of the ‘Baseline’ model (1) in eight of the 11 commodities depicted, we do not observe the distance puzzle for commodities 16 nuts, 21 cocoa and 26 other agricultural. For the set of commodities with the distance puzzle (in all but case 20 other meat), the gravity models that explicitly capture intra-national distance and the home bias resolve the distance puzzle—although commodities 7 other oilseeds, 10 other sweet and 11 pulses/legume require the more flexible country-specific home bias of model (4). Thus, while the distance puzzle is pervasive for highly aggregated trade data, it is not necessarily present at a disaggregated level.

Fig. 1.

Effects of distance on trade by AC. (a) Percent change between 1986 and 2019. (b) Percent change between 1997 and 2019. (c) Percent Change between 2008 and 2019.

Open in new tabDownload slide

Figure 1b and c depict the percent change in the distance coefficient estimates between 1997 and 2019 and between 2008 and 2019 for ACs and all four models. As with the average results reported in Table 2, for all but one of the individual commodities (20 other meat) over these two time periods, the distance puzzle does not arise, as the estimated coefficient for distance in 2019 is smaller than that in 1997 and 2008 in the baseline model (1). Furthermore, when isolating intra-national trade costs and controlling for the home bias effect, the difference in coefficient estimates for distance between 1997 and 2019, and 2008 and 2019 is mostly similar in magnitude or less negative. However, an interesting and somewhat counterintuitive result is that, for the gravity model with intra-national distance and the home bias, the distance puzzle emerges for commodities 10 other sweeteners and 11 pulses/legumes.

Figure 2 depicts the percent change in the distance coefficients between 1988 and 2019 (Panel a), between 1998 and 2019 (Panel b) and between 2008 and 2019 (Panel c) for all PCs. As seen in Panel a, the distance puzzle arises for the majority of commodities under the baseline model (1), except for commodities 37 fish and 50 wine, again highlighting the heterogeneity in the distance puzzle at the individual commodity level. Furthermore, including controls for relative intra-national trade costs and the home bias effect leads to smaller percent changes between 1986 and 2019 for almost all PCs, except commodities 42 starch and 49 spirits. In the most flexible specification of model (4), the distance puzzle is resolved in the majority of commodities, although the distance puzzle remains in commodities 36 meat, 42 starch, 49 spirits, 54 textiles and 55 textile articles as the impact of distance on bilateral trade remains higher in 2019 compared to 1988. These results further confirm that the standard structural gravity model results in biased estimates of distance and measuring intra-national distance relative to international distance and controlling for home bias are keys to accurate estimates of distance.

Fig. 2.

Effects of distance on trade by industry. (a) Percent change between 1988 and 2019. (b) Percent change between 1998 and 2019. (c) Percent change between 2008 and 2019.

Open in new tabDownload slide

Figure 2b shows that, as with the average results reported in Table 2, the distance puzzle does not arise for all PCs over this time span, as the estimated coefficient for distance is closer to zero in 2019 than in 1998 in the baseline model (1). While substantial heterogeneity exists with controls for relative intra-national trade costs and the home bias effect, the main conclusion generally holds because the distance puzzle is not observed during this time frame. Figure 2c reveals that, for the 2008–2019 interval, not only does the distance puzzle not arise as trade costs are lower in 2019 than in 2008 for the majority of commodities, but controlling for relative intra-national trade costs and the home bias effect dampens the globalisation effect of distance as the relative gap between the coefficient estimates for distance in 2008 and 2019 is smaller under models (2) and (3) than model (1).

5. Conclusions

As globalisation evolved over the last 50 years, the world has become ‘flatter’ meaning that the trade-reducing effect of distance has diminished over time. However, using gravity models, many studies have found a counterintuitive result where distance hindered trade less several decades ago as the distance coefficient became more negative over time—yielding the distance puzzle. Yotov (2012) and Yotov et al. (2016) showed that controlling for intra-national trade costs relative to international trade costs and the home bias effect results in less negative estimates for distance over time, resolving the distance puzzle. However, trade in agricultural products can differ from trade in aggregate and manufacturing goods due to complex support programmes, trade policies, the bulky and perishable nature of agri-food commodities and global value chains are not as prominent in agri-food industries as in other non-food industries. This study examines how globalisation has impacted the effect of distance on agricultural and processed food trade for the 33 years from 1986 to 2019 using a structural gravity model. In doing so, the study also examines the impact of internal distance and the home bias effect in agricultural and processed food products.

This study estimates four structural gravity models: the first is a baseline model where distance coefficients are estimated for each year in the sample (1986 –2019 for ACs and 1988–2019 for PCs). The second model isolates intra-national trade costs from international trade costs. The third model builds on the second by also including a control for the home bias effect. The fourth model builds on the third model by allowing the home bias effect to vary by country. We uncover the average effect by stacking the trade data for all 26 ACs into one dataset and all 20 PCs into another dataset. We also examine the heterogeneous effects by estimating the results for each commodity separately.

The results show that controlling for international relative to intra-national distance and the home bias effect can have a substantial impact on trade friction variables. Furthermore, countries with larger distances between major cities have higher intra-national trade costs, which hinders domestic sales. The results also reveal a substantially larger home bias effect for ACs compared to PCs.

The results indicate that heterogeneity exists depending on the time frame considered. For example, the average effects show that the distance puzzle arises as the estimated coefficients for distance in 1986 for ACs and 1988 for PCs are less negative than those in 2019 under the baseline structural gravity model (which excludes the relative intra-national distance and the home bias effect). After controlling for the relative intra-national distance and the home bias effect, the distance puzzle does not arise as the percent change between 1986/1988 and 2019 is statistically zero. Thus, while the distance puzzle is not confirmed, we also reject the hypothesis that globalisation lowers trade costs (proxied through distance) over this three-decade period. However, the distance puzzle is generally not observed, and globalisation led to smaller estimated coefficients for ACs in 1997 and for PCs in 1998 than those in 2019. An important avenue for future research is to investigate why the impact of globalisation on agri-food trade stagnated starting in the 1990s.

Furthermore, substantial heterogeneity exists when examining individual commodities. Therefore, the evolution of globalisation in agricultural and processed food commodity trade is substantially more nuanced than for aggregate or manufacturing trade. These results are attributed to the unique features (e.g. domestic producer and consumer supports, international trade policies, bulky and perishable nature of the products and less prominent global value chains compared to other manufacturing commodities) of food products.

Acknowledgements

I thank Yoto V. Yotov and three anonymous reviewers for helpful comments in revising the manuscript. I thank the the editor Dr Salvatore Di Falco for coordinating the review.

Funding

This work was supported by the United States Department of Agricultur, National Institute of Food and Agriculture, Agricultural and Food Research Initiative Competitive Program, and Agriculture Economics and Rural Communities (grant no. 2022-67023-36382).

Footnotes

References

Table A1 presents results for AC and PC models that include time-varying intra-national distance and time-varying border effects.