Econometric specification and identification strategy

Our benchmark estimation model is the following product-level gravity equation, wherein we model bilateral trade costs as a CES function of the product and time-varying country-pair difference in maximum residue limits (MRL_{i jkt}) and tariffs (Tariff_{i jkt}):

lnX_{i jkt}=ψikt+λjkt+αi j+β1M RL_{i jkt}+β2ln(1+Tariff_{i jkt}) +εi jkt (4.2) where i is the exporting country, j is the importing country,k is the product and t is time. Our parsimonious specification includes a host of importer-product-time (ψikt), exporter-product-time (λ_jkt) and importer-exporter (α_{i j}) bilateral fixed effects. These fixed effects control for all country and product-specific (e.g., production and expenditure) and country-pair specific time-invariant effects (e.g., bilateral distance, common language, contiguity). In line with the structural gravity literature,ψiktandλjkt also control for multilateral resistance (Anderson and Van Wincoop, 2003).

Hence, in principle, our model can only identify the effect of variables that are country-pair varying over time. Since these fixed effects eliminate many confounding factors as possible, we are confident our estimation captures a pure trade cost effect. εi jkt is the error term which we cluster at the country pair-product level. We are primarily interested inβ1, which reflects the effect of differences in product-specific MRLs between countries on different measures of trade, prices and quality. The inclusion of country pair fixed effects (αi j) implies that identification ofβ1is achieved from changes in bilateral MRL differences over time.

4.4.2 Definitions of the different measures ofX_{i jkt}

The dependent variable in equation (4.2) varies depending on the specific research question. It represents for each importer-exporter-product-time the (i) extensive margin (ii) intensive margin (iii) product of both trade margins (iv) value of trade conditional on exports (v) import prices expressed as unit values (vi) quality and (vii) quality-adjusted prices. Here, we discuss these different measures.

Measures of the intensive and extensive trade margins

Using conventional gravity equations with total trade flows as the dependent variable, although now armed with solid micro-foundations (e.g., Anderson and Van Wincoop, 2003), may still be misleading as the extensive margin and intensive margins might respond differently to trade costs (Feenstra and Ma, 2014). Existing studies in the literature that have tried incorporating the two

70 Chapter 4. Trade, price and quality upgrading effects of agrifood standards margins have used mainly the Heckman two-step procedures. However, these suffer two limitations;

the incidental parameter problems of the first stage Probit equation in panel data contexts and the fact that the procedure only works well in bilateral trade equations when true exclusion restrictions exist (Helpman et al., 2008). Recently, several papers have also used a direct approach to decompose the impact of policies on the extensive and intensive trade margins. These include measures such as the number of products exported within a certain industry, counting categories that exceed a certain size or exports concentration indexes (see, e.g., Cadot et al., 2011; Persson and Wilhelmsson, 2016).

These simple counts, although transparent, are limited by the assumption that products have the same economic weight.

Following Feenstra and Kee (2008), we consider a theoretically-founded decomposition of overall trade into the extensive and intensive margins considering the economic weight of the products. This measure is very similar to a count of the exported varieties within a certain industry, but appropriately weights categories of goods by their overall importance in exports to an importing country. The extensive margin (E M_{i jkt}) is the fraction of all products k exported from country i to country j, where each product is weighted by the importance of that product in total exports toj in yeart. The intensive margin (I M_{i jkt}) is the bilateral trade flow fromito jrelative to the average world export to jin the same product category. The product of the two margins equals the ratio of exports from ito jrelative to country j’s total imports, i.e., it measures the relative export performance of each exporter in an importer-product-year. We move to the Appendix 4.7, the detailed description of the methodology to measure both the extensive margin and intensive trade margin. As a fourth measure of the dependent variable, we consider the absolute value of exports of productkfrom countryito

jin yeart.

Measures of price, quality and quality-adjusted price

The final bit of our analysis relates the differences in national standards to prices and quality of im-ports. Consumption scandals in the agrifood sector have prompted an increase in quality requirements of consumers and firms. Regulations such as the EU Food Law of 2002 makes it the responsibility of retailers to ensure that their suppliers from third countries meet EU food quality standards. As a result, besidesde jurepublic standards, retailers enforcede factomandatory standards as gatekeepers to filter products based on quality when dealing with geospatially dispersed producers. Exporters knowing the quality of their products will segregate and send different quality levels to different destinations. For example, in a World Bank report Jaffee et al. (2005) show that exporters in Kenya segregate low-quality produce from smallholders for less demanding destinations. For MRLs, products earlier targeted for a specific importer may end up in other markets with less stringent quality defini-tions, depending on the timing of chemical control preceding harvest. These differences in objective qualities of agrifood products (e.g., size, colour, production location), the presence or otherwise of pesticides and grading or certification schemes fit in the realm of vertical differentiation (Saitone and Sexton, 2010). Thus, agrifood products are not necessarily homogeneous but heterogeneous in quality.

Critical to this part of the analysis is how we measure unobservable “product quality”. It is standard in the agricultural trade literature to use prices (measured as unit values) to proxy quality

Chapter 4. Trade, price and quality upgrading effects of agrifood standards 71 (Fernandes et al., 2019; Bojnec and Fert˝o, 2017). For each HS6 digit productk, the bilateral trade data records the total nominal value of imports in US dollars from a given exporter, as well as the quantity in tonnes associated with these imports. Taking the ratio of trade values and trade quantities, we obtain so-called unit values, i.e.,p_{i jkt}=v_{i jkt}/q_{i jkt}.⁷⁸While unit values are available for a wide range of products and countries, they may not be precise proxies for quality. Prices may also reflect higher production costs, exchange rates or market power. Our approach follows Khandelwal et al. (2013) and recovers quality directly from observed trade data.⁷⁹ The intuition behind the Khandelwal et al. (2013) approach is simple: conditional on prices, varieties with higher quantities (market shares) are assigned higher quality.⁸⁰We assume quality is any attribute that raises consumer demand other than price (Khandelwal et al., 2013; Disdier et al., 2018). After estimating quality ˆq_{i jkt}, we obtain the quality-adjusted price component as the observed log prices less estimated quality, i.e., ln ˆp_{i jkt}=lnp_{i jkt}−ln ˆq_{i jkt}. That is the differences in product prices for the same level of quality.

See Appendix 4.7 for a detailed description of the quality estimation procedure. Applications of the Khandelwal et al. (2013) method in the agrifood sector include Curzi and Pacca (2015) and Movchan et al. (2019).

According to Feenstra and Romalis (2014), quality differences can explain some of the variations we observe in unit-values across countries. As an initial exploratory analysis to see how well our quality estimates correlate with observed unit values, we plot a graph of lnp_{i jkt}against ln ˆq_{i jkt}(Figure A4.2 in the appendix) that show that our estimated quality and unit values are indeed positively

Figure 4.1: Distribution of prices and estimated product quality of imports

0.0

78Information on unit values can be particularly noisy because the trade data may contain measurement errors at the disaggregated product level. This noise in the price data would also affect our quality estimates. To deal with potential outliers in the price and quality estimations, we screen the dataset and we exclude extreme unit values within the 1st and 99th percentiles. We also drop annual growth rates within the 1st and 99th percentiles. Finally, we drop estimated quality values within the 5th and 95th percentiles. This data cleaning procedure eliminates 3% of our observations.

79Whiles this method was originally applied at the firm-product-country-year level, subsequent applications have also been done at the product-country-year level, see e.g., Curzi and Pacca (2015), Breinlich et al. (2016). The limitation, however, is that different producers or firms may produce different qualities. Lack of farm/firm-level trade data implies that our quality estimates reflect the average quality of exports from a country.

80For instance, suppose bananas from Ecuador and Colombia are equally priced, but Colombia’s market share in destination market j is 20% and Ecuador’s is 10%, the quality estimate for Colombia will be higher. If bananas from Colombia were more expensive, then we would need to control for the price difference and this would reduce the quality estimate for Colombia.

72 Chapter 4. Trade, price and quality upgrading effects of agrifood standards correlated. As a second descriptive analysis, we plot the Epanechnikov kernel density estimates of our quality estimates and unit values for the first and last years of our panel.⁸¹The results presented in Figure (4.1) reveal that average quality and price of imports increased over the study period.

However, compared to prices, average quality did not change by much. The extent to which this is driven by cross-country and product differences in MRLs over time is one goal of this paper.