Food Quality and Product Export Performance: An Empirical Investigation of the EU Situation

Abstract

The relationship between product quality and export performance is investigated. Five EU countries (Italy, France, Spain, Germany, United Kingdom), 3 product categories (cheese, meat preparations, and wine), 3 export destinations (intra-European Union, extra-European Union, and world), and 2 periods (1995–1999 and 2000–2005) are analyzed. The regression results show that the connection between product quality and export performance depends on the product category (but not on the period) and differs (but not in all cases) according to the export destination. An implication arising from this analysis is that considering international marketing of quality food products in teaching and training programs could contribute to the enhancement of EU agribusiness competitiveness in increasingly liberalized and quality-conscious markets.

KEYWORDS:

• European Union
• export performance
• food quality
• general linear mixed model
• unit values

INTRODUCTION

Food quality has become an increasingly important topic during the last decades. In developed countries, driven by aging populations and growing diet-related health concerns, consumer demand now seems to shift toward higher quality, more natural, and healthier food (Regmi, 2001). Given today's globalized markets, it can be assumed that—in line with rising consumption—international trade of quality food products (QFPs) is increasing.

Yet, the concept of food quality is elusive. Overall, quality may be seen as an abstract construct, multidimensional in nature (Charters & Pettigrew, 2007). More specifically, in the literature, at least four different quality definitions are discussed. It is either referred to as excellence or superiority, as value, as conforming to specifications, or as meeting or exceeding customer expectations (Reeves & Bednar, 1994; Verdú Jover, Lloréns Montes, & del Mar Fuentes, 2004). The first approach defines quality as best in class, judged on some to-be-specified criteria. The value approach is an economic one (higher monetary value reflects higher quality), whereas the conforming to specifications is a technological one (a quality product is a product that fulfills some predefined technical standards, however low they may be). The fourth approach defines quality from the point of view of the final consumer: as long as she or he is happy with it, it is a quality product. Either way, in order to measure quality objectively, a generally accepted reference system is necessary, otherwise “the general term quality is very subjective and means very little” (Satin, 2002, para. 2). Ninni, Raimondi, and Zuppiroli (2006) state that the quality difference between competing products can be easily analyzed for measurable characteristics such as: reliability, durability, various indicators of performance, and health and safety. However, it becomes more subjective when it refers to intangible characteristics such as design, taste, and flavor. Here the boundaries of vertical and horizontal differentiation are blurred. Obviously, the intangible characteristics are of particular importance for food products. Therefore, for these goods, it may perhaps—despite Satin's assertion—be most useful to accept these having subjective qualities rather than one objective quality, meaning that different people can come to diverging quality assessments and, in the end, none of these is superior to another. ¹

Despite its increasing importance, and probably partly due to the difficulties of objectively defining quality, not much research has been done so far on the particularities of the international trade of QFPs. Thus, for example, it is unclear whether the nature of the international trade of QFPs is similar or potentially structurally different compared with the one of low- or average-quality food products. Because many QFPs may be more perishable, more sensitive to external influences (e.g., temperature, vibration, light), and often of higher (monetary) value, more demanding logistics and insurance issues may affect the international trade of these products, potentially resulting in different export patterns.

The objectives of the following analysis are threefold: first, to review the recent literature dealing with international trade in QFPs; second, to generate empirical insights into the export patterns (levels and destinations) of food products, depending on their quality level; and third, based on the obtained results, to derive conclusions of what this could potentially mean for agribusiness management and/or policy making.

This article's structure is as follows: After the introduction, Section 2 reviews the previous literature related to the topic. Section 3 describes the procedure of the empirical investigation. Section 4 discusses the obtained results. Section 5 concludes by summarizing and pointing out some implications that arise from the findings. The Appendix provides a technical treatment on the generalized least squares estimation method used in this study.

PREVIOUS STUDIES

The literature on the relationships between quality and trade performance is sparse. The few existing studies have in common that they investigate intercountry quality competition in the sense that they try to find out whether a country's exports of certain products have higher quality vis-à-vis other countries producing similar goods.

Aiginger (1997) suggested a method of how to use unit values (UV) ² in order to discriminate between price and quality competition in international markets using the case of German exports of industrial goods. Gelhar and Pick (2002) applied Aiginger's framework to U.S. food trade flows and found that almost 40% of U.S. food exports could be characterized as dominated by quality competition. For imports, the share amounts to 60%. However, the results for bilateral trade flows are much lower, which points to problems involved when using unit values and net trade figures of economic aggregates. Ninni et al. (2006) explored the role of quality of Italian food products in international markets. They regressed relative market shares of the Italian products in the import market of several different countries on a quality indicator based on UVs and other variables. The obtained results suggest that “the quality image of Italian goods offers protection for some traditional products, but that this protection is not strong enough to counteract price competition” (p. 2).

No previous studies (to the author's knowledge), however, have addressed the issue of intracountry product export performance, that is, whether a country tends to specialize in exporting QFPs or low- and/or average-quality products among the whole range of highly differentiated products it manufactures. In other words, within a nearly defined product category (e.g., cheese) is a country a high-quality exporter (which indicates that the country's high-quality products are internationally appreciated and sought after) or that the country is rather a low- or average-quality exporter (implying that the country's QFPs are more domestically preferred but that international demand for them is weak).

PROCEDURE

The general research approach is a deductive, empirical investigation. That is, international trade data are analyzed and conclusions are drawn from the findings.

The focus of the analysis is to estimate the relationship between food quality and product export performance (i.e., whether there is a positive or negative link between the two variables). Although this relationship is estimated within a larger regression model (i.e., one that considers a few additional control variables), it is not aimed at fully explaining export performance. Given that the used statistical models are not completely specified (i.e., other potential export performance affecting variables are left out), omitted variable error may be a problem in the obtained results. However, a serious bias of the estimated regression coefficients would only occur if the quality variable is significantly correlated with other, omitted export performance determinants. ³

Other factors that potentially determine product export performance are not easy to identify on theoretical grounds. Almost all previous studies have used companies, economic sectors, or entire countries as units of investigation when analyzing export performance. Here, the research question is whether within a larger food category, the export performance of similar products changes as a function of quality. Many companies produce a range of products (typically covering lower quality goods as well as premium products), thus company-specific factors apply to all of them. Other typical export performance determinants such as promotion programs, export enhancement policies (subsidies), or trade-regulating measures (quotas, tariffs, etc.) are usually sector specific (i.e., apply equally to an entire food category). Therefore, these factors do not vary across the products under consideration. However, there are a few potential export performance-influencing covariates that would be meaningful to consider in the regressions. These are indicators for logistics (transport and/or storage) complexity (e.g., perishability, sensitivity to vibrations), likely to result in higher marketing costs, and effective final consumer prices relative to incomes in the export markets. The export performance of a QFP, ceteris paribus, would be higher if logistics complexity is relatively lower and/or relative prices in the target market are lower. ⁴ Hence, product export performance may be seen as a function of real quality (product-specific superiority as perceived by the final consumer), related logistics complexity (as reflected in marketing costs), and consumer purchasing power in the target market. In the following, the indicator used for quality also comprises marketing costs and the target market will be roughly controlled for by considering different export destinations.

Hence, although omitted variable bias cannot fully ruled out, its magnitude is uncertain. In any case, here the primary interest lies in the signs of the estimated quality parameters and not in their size.

The operationalization of the variables under consideration is as follows. First, food quality (within a homogeneous product category) is measured by price (i.e., unit value as a price indicator). In other words, it is assumed that among similar products quality is positively related to price. Thus, the aforementioned value approach is adopted for defining quality. Second, trade performance is assessed by using relative export shares.

Measuring Quality

As price proxies, UVs (in € per kg) are used, obtained by dividing export values by export quantities. UVs are known to be imperfect price indicators (Holmes, 1973; King, 1993; Shiells, 1991). Their main problem is related to the fact that an observed change in the unit value may not necessarily be a result of an underlying price change but may simply reflect a change in the composition of the goods within the class of exports under consideration. Another problem relates to invoicing practices. The existence of a lag between the time of contract and the delivery of goods can result in differences between the contract value (i.e., the real price) and the UV calculated from customs declaration when a good is actually delivered in cases where the exchange rate changes in between. However, the magnitude of these measurement problems is not clear. For instance, whereas Shiells finds that for U.S. trade data import UVs are good import price proxy, Holmes shows for Canada that domestic UVs (national production values divided by output quantity) do not well represent industrial selling prices obtained by means of manufacturer surveys.

Despite these shortcomings, UVs have been suggested and used as an indicator of quality content (Aiginger, 1997; Gelhar & Pick, 2002). Because the UV is output per units of input (e.g., measured in kilograms), for homogenous and comparable goods the value can indicate differences in quality if unit production costs can be assumed to be equal across the considered countries. However, the UV will also reflect differences in costs and thus high UVs can indicate relative high product quality and/or relative high unit costs. In order to distinguish between these two cases, Aiginger suggested looking at the net trade position of the good (aggregate) under consideration. If within a country a good's UV is high and the good's net trade position is positive, cross-country UV differences must then be due to superior quality. However, if a good's UV is high but the corresponding net trade position is negative, this indicates a cost disadvantage. One problem with this approach is that it requires aggregate data (e.g., “cars,” “cheese,” or “ice cream”) or trade data on commodities (i.e., homogenous, undifferentiated goods) such as “meat of sheep, frozen,” “flat fish, fresh or chilled,” and so on in order to be able to calculate net trade positions. For highly disaggregated data of certain agricultural products (e.g., Gorgonzola cheese, Bordeaux red wine, etc.), no such trade balance can be calculated (because no other countries produce such goods and consequently no imports can exist). Thus, in general, the higher the level of disaggregation, the more accurate UVs may be as price proxies, but the more difficult it is to determine whether high UVs reflect high quality or high production costs. Yet, if absolute exports of products with high UVs are comparatively high, then this may nonetheless be an indicator of quality (due to an apparent international willingness to pay relative high prices for goods with many close substitutes).

UVs are calculated in relative terms (RUV) as symmetrized and normalized deviations from a category's mean unit value, as suggested by Laursen (1998). That is,

where k refers to the category (i.e., cheese, wine, or meat products; see later), c to the country, p to a particular product within k (e.g., Roquefort cheese), t to the period (1995–1999 or 2000–2005; see later), and n _ck is the number of products in a particular category for a certain country. Note that the range of RUV is [–100; 100], where positive (negative) values indicate above (below) average UVs.

Measuring Export Performance

Trade performance is assessed by relative export shares. The measure is a modified version of Balassa's index of revealed comparative advantage. It is defined as the deviation from the expected export share of a product within a product category, again symmetrized and normalized.

The index of revealed comparative advantage (RXA) was defined by Balassa (1965) as a measure for “the export performance of individual industries in a particular country …” (p. 105), which can be evaluated by “… comparing the relative shares of a country in the world exports of individual commodities […] where the data have to be made comparable through appropriate ‘normalization’” (p. 105). Thus, the original index was constructed in a form such as

where x _ct stands for the exports (in €) of the food processing sector of a country c in year t, and X _ct stands for total country exports of country c in year t. ∑x _ct designates total sector exports (considering C countries), whereas ∑X _ct refers to total aggregate exports in a particular year t. The RXA, therefore, is an index of the share of a country in the international market of a particular economic sector corrected (i.e., normalized) for the size of the country to which the sector belongs. The correction is necessary because larger countries can a priori be assumed to have larger market shares simply because of their size. Bowen (1983) showed that Balassa's RXA may also be interpreted as the deviation of actual exports, x _ct , from expected exports, E(x _ct ), where E(x _ct ) can be defined as , assuming that all countries engage in all economic activities in equiproportional shares.

The (theoretical) range of the RXA index is from [0; ∞[, with values ∈[0; 1[, indicating a comparative disadvantage and values ∈]1; ∞[ showing a comparative advantage of an economic sector of a particular country relative to the other sectors within this country. Thus, the RXA allows determining the strong and weak sectors within one country by ranking them by descending RXA scores. Unfortunately, however, the RXA may fail to reliably indicate comparative (or competitive) advantages within the same sector relative to other countries. The main problem in cross-country comparisons is the different frequency distribution of the RXA scores in each country. ⁵ In addition, De Benedictis and Tamberi (2001) show that the effective upper bound of the RXA for each country is different. This upper bound is equal to world total trade divided by total trade of country c (∑X _ct /X _ct ) and thus is in general relatively small for large and very high for small countries. Consequently, the index scores that are larger than one cannot directly be compared across differently sized countries (implying the aforementioned normalization for country size does not work perfectly in the standard notation of the RXA). Thus, because the conventional RXA range is inconvenient for interpretation, a symmetrization has been suggested by Laursen (1998), which yields a range for the RXA ∈[−1; +1] with values below (above) zero indicating below (above) average relative exports: ⁶

The main problem with using the RXA in the context of this research is again that it can only be applied to aggregate data or trade data on commodities (i.e., homogenous, undifferentiated goods) in order to be able to calculate world sector exports. However, for highly disaggregated data of certain agricultural products (e.g., Gorgonzola cheese, Bordeaux red wine, etc.), no such sums can be calculated because these products are only produced in one single country and thus country exports equal world sector exports. For this reason, a modified version of the RXA is used in the following, which is related to Bowen's (1983) interpretation of the Balassa index. Here the relative export performance (REP) is assessed by the share of the exports of a particular product p in category (k) exports as deviation from the expected export share (1/n _ck , where n _ck is the number of p in k for a particular country c), assuming that preferences for all products p are equal. That is,

where t refers to the period (1995–1999 or 2000–2005; see later). Note that the range of the REP is also [–100; 100], where positive (negative) values indicate above (below) average relative export performance.

It should be stressed that the REP, despite being derived from the RXA, does not compare export performance of the same products across different countries. Rather, it measures export performance of different products relative to other products within a certain product category and for one single country.

Data

The raw data were taken from Eurostat's COMEXT “EU25 Trade Since 1995 By CN8” database. These trade data are at the highest available level of disaggregation (eight-digit level).

The export data (€ values and kg quantities) are factored in four dimensions. First, reporters. The five largest EU countries (Germany [DE], Spain [ES], Great Britain [GB], France [FR], and Italy [IT]) were selected. Second, as destination area three different locations were chosen: within the European Union with 15 members (EU-15), outside the EU-15, and world (i.e., the sum of the former two). Third, three product categories (cheese, meat preparations, and wine) were selected. Each category contains a large number of individual products. The maximum number of products for each category is as follows: cheese 66, meat preparations 54, and wine 70. ⁷ However, the actual number of types included in the analysis depends on the individual country (see next paragraph). The categories were selected based on their importance to the EU food industry (meat, beverages, and dairy products are the three most important subsectors as measured by their shares in total sector value added; see Lienhardt, 2004). The fourth dimension is time. The years 1995 to 2005 were included in the analysis. Because the used trade data are volatile, the 11 years were averaged over two periods (1995–1999 and 2000–2005).

Overall, some 26,804 observations were used in the analysis. For the United Kingdom, the wine category was not analyzed because the country is not a significant wine producer. For Spain, only a few wine export quantities for the second period were available; thus only Period 1 exports could be included in the analysis. Despite these few missing values, the data set covers a whole target population rather than being of sample nature. This needs to be kept in mind when interpreting the statistical significance (i.e., deviations from zero larger than those to be expected due to sampling error) of the estimation results later in this article.

Data preparation involved the removal of reexport data and of outliers. ⁸ Unfortunately, the raw data also included export flows of products that could not have been produced in a certain country (e.g., Navarra red wines in German exports or Gorgonzola cheese in French exports). These reexports were removed from the original data set where possible. However, for many products (e.g., processed cheese, uncooked sausages, etc.), where the production is not restricted to a certain geographical area, no potentially existing reexports could be eliminated. Table 1 lists the actual number of products included in the empirical analysis for each category and country (the n _ck ).

TABLE 1 Number of Product Types Included in the Empirical Analysis for Each Product Category and Country (n _ck )

	DE	IT	FR	ES	GB
Cheese	27	25	31	13	15
Meat preparations	36	35	41	34	36
Wine	17	25	25	24	−

Estimation Methods

The relationship between export performance and food quality was estimated using regression analysis. First, ordinary least squares (OLS) estimators of the slope coefficients were obtained separately for each country and product category, controlling for a potential period effect by using a period dummy variable in all cases (except for wine in Spain). However, in no case did the dummy coefficient turned out to be significant at least at the 95% confidence level. Second, all data belonging to one country were pooled and feasible generalized least squares (GLS, which allows for heteroscedasticity across and correlation between different panels; see Cameron & Trivedi, 2005) was used to estimate the (nested) fixed effect of food quality on export performance, controlling for product category and time. A more detailed description of the GLS estimation procedure is provided in the Appendix. In the pooled estimations, the slope coefficients were allowed to vary across the product categories for all countries except for Germany. The justification for this is founded in the OLS results: when the slope coefficients for the different product categories displayed the same signs, only one GLS slope coefficient for the food-quality variable was estimated. Otherwise, the GLS slopes were allowed to vary.

RESULTS AND DISCUSSION

The estimation results are displayed in Figures 1 to 5 (OLS) and in Tables 2 to 6 (GLS). Overall, it emerges that the direction (and the significance) of the relationship between food product quality and export performance is not systematic but depends on the country, the product category, and the export destination. However, the direction (i.e., the sign of the slope coefficient) is in most cases independent of the used estimation method, yet the significance levels are not.

FIGURE 1 Germany: OLS (ordinary least squares) slope estimates* of deviations from expected exports (y) regressed on deviations from average unit values (x). Note. Bold regression lines indicate that the slope coefficient is statistically significant at the 95% confidence level. *Controlling for period (0 = 1995–1999; 1 = 2000–2005).

FIGURE 2 Italy: OLS (ordinary least squares) slope estimates* of deviations from expected exports (y) regressed on deviations from average unit values (x). Note. Bold regression lines indicate that the slope coefficient is statistically significant at the 95% confidence level. *Controlling for period (0 = 1995–1999; 1 = 2000–2005).

FIGURE 3 France: OLS (ordinary least squares) slope estimates* of deviations from expected exports (y) regressed on deviations from average unit values (x). Note. Bold regression lines indicate that the slope coefficient is statistically significant at the 95% confidence level. *Controlling for period (0 = 1995–1999; 1 = 2000–2005).

FIGURE 4 Spain: OLS (ordinary least squares) slope estimates* of deviations from expected exports (y) regressed on deviations from average unit values (x). Note. *Controlling for period (0 = 1995–1999; 1 = 2000–2005).

FIGURE 5 UK: OLS (ordinary least squares) slope estimates* of deviations from expected exports (y) regressed on deviations from average unit values (x). Note. *Controlling for period (0 = 1995–1999; 1 = 2000–2005).

TABLE 2 Germany: Pooled Regression Results (Feasible GLS Estimates) From a Linear Mixed Model With One Repeated Measurement

	Dependent variable: Relative (deviation from expected) exports
Regressors	Total	Intra-EU	Extra-EU
Tests of fixed effects (F values)
Intercept	18.3***	22.0***	22.5***
CATEGORY	2.32	1.54	2.16
PERIOD	7.94***	7.72***	2.24
Relative unit value	22.3***	10.1***	17.0***
Fixed-effect parameter estimates
Intercept	−7.69	−13.6	−11.4
CATEGORY = cheese	−17.1	−13.3	−22.4
CATEGORY = meat prep	−32.4**	−26.5*	−35.1**
CATEGORY = wine	0 ^a	0 ^a	0 ^a
PERIOD = 1 (1995–1999)	−5.68**	−6.12**	−4.19
PERIOD = 2 (2000–2005)	0 ^a	0 ^a	0 ^a
Relative unit value	−.771**	−.556***	−.872***
Variance/Covariance-matrix components †
	3,013.2***	3,043.2***	3,452.0***
	2,568.7***	2,611.7***	3,090.8***
σ12	2,632.8***	2,638.4***	2,965.2***
Model statistics
# of obs.	160	160	160
Restr. −2 log lik.	1,516.5	1,532.1	1,582.2
AIC	1,522.5	1,538.1	1,588.2
BIC	1,531.6	1,547.2	1,597.3

TABLE 3 Italy: Pooled Regression Results (Feasible GLS Estimates) From a Linear Mixed Model With One Repeated Measurement

	Dependent variable: Relative (deviation from expected) exports
Regressors	Total	Intra-EU	Extra-EU
Tests of fixed effects (F values)
Intercept	44.0***	43.0***	65.8***
CATEGORY	3.51**	3.70**	4.61**
PERIOD	.000	.018	.133
Relative unit value (CATEGORY)	.062	.058	.234
Fixed-effect parameter estimates
Intercept	−24.0**	−22.3**	−27.4**
CATEGORY = cheese	−10.8	−13.8	−16.8
CATEGORY = meat prep	−35.6**	−37.5***	−40.8***
CATEGORY = wine	0 ^a	0 ^a	0 ^a
PERIOD = 1 (1995–99)	.036	.275	−.649
PERIOD = 2 (2000–05)	0 ^a	0 ^a	0 ^a
Relative unit value (cheese)	.089	−.159	.191
Relative unit value (meat preparations)	−.036	.011	−.042
Relative unit value (wine)	−.074	−.074	−.160
Variance/Covariance-matrix components †
	2,922.7***	2,965.3***	2,690.2***
	2,796.6***	2,826.3***	2,693.0***
σ12	2,724.1	2,721.0***	2,563.1***
Model statistics
# of obs.	170	170	170
Restr. −2 log lik.	1,605.9	1,627.2	1,597.0
AIC	1,611.9	1,633.2	1,603.0
BIC	1,621.2	1,642.5	1,612.3

TABLE 4 France: Pooled Regression Results (Feasible GLS Estimates) From a Linear Mixed Model With One Repeated Measurement

	Dependent variable: Relative (deviation from expected) exports
Regressors	Total	Intra-EU	Extra-EU
Tests of fixed effects (F values)
Intercept	35.2***	39.3***	48.4***
CATEGORY	.259	.471	.318
PERIOD	.524	.094	.029
Relative unit value (CATEGORY)	2.04	.229	10.7***
Fixed-effect parameter estimates
Intercept	−30.2***	−30.5***	−31.0***
CATEGORY = cheese	−.931	−1.35	−10.2
CATEGORY = meat prep	−8.30	−11.2	−9.28
CATEGORY = wine	0 ^a	0 ^a	0 ^a
PERIOD = 1 (1995–1999)	1.23	.599	−.341
PERIOD = 2 (2000–2005)	0 ^a	0 ^a	0 ^a
Relative unit value (cheese)	−.057	−.091	.112
Relative unit value (meat preparations)	−.251*	−.007	−.803***
Relative unit value (wine)	.531	.253	.789**
Variance/Covariance-matrix components †
	2,816.9***	2,703.4***	2,616.3***
	2,758.2***	2,834.6***	2,781.2***
σ12	2,655.9***	2,592.6***	2,514.8***
Model statistics
# of obs.	194	194	194
Restr. −2 log lik.	1,833.0	1,858.8	1,859.9
AIC	1,839.0	1,864.8	1,865.9
BIC	1,848.7	1,874.5	1,875.6

TABLE 5 Spain: Pooled Regression Results (Feasible GLS Estimates) From a Linear Mixed Model With One Repeated Measurement

	Dependent variable: Relative (deviation from expected) exports
Regressors	Total	Intra-EU	Extra-EU
Tests of fixed effects (F values)
Intercept	25.6***	22.8***	39.8***
CATEGORY	1.97	4.17**	.644
PERIOD	1.69	.888	5.19**
Relative unit value (CATEGORY)	1.70	.347	2.82**
Fixed-effect parameter estimates
Intercept	−38.0***	−22.8**	−43.3***
CATEGORY = cheese	17.5	5.08	−2.41
CATEGORY = meat prep	−15.3	−30.2**	−15.6
CATEGORY = wine	0 ^a	0 ^a	0 ^a
PERIOD = 1 (1995–1999)	5.36	4.13	8.84**
PERIOD = 2 (2000–2005)	0 ^a	0 ^a	0 ^a
Relative unit value (cheese)	−.513	−.479	−.208
Relative unit value (meat preparations)	−.314	−.052	−.520**
Relative unit value (wine)	.646	.108	.735
Variance/Covariance-matrix components †
	2,785.1***	2,247.7***	3,014.2***
	2,965.0***	2,366.9***	3,161.5***
σ12	2,502.5***	1,885.5***	2,758.7***
Model statistics
# of obs.	118	118	118
Restr. −2 log lik.	1,176.0	1,165.8	1,176.0
AIC	1,182.0	1,171.8	1,182.0
BIC	1,190.1	1,179.9	1,190.1

TABLE 6 UK: Pooled Regression Results (Feasible GLS Estimates) From a Linear Mixed Model With One Repeated Measurement

	Dependent variable: Relative (deviation from expected) exports
Regressors	Total	Intra-EU	Extra-EU
Tests of fixed effects (F-values)
Intercept	16.4***	19.8***	25.1***
CATEGORY	.001	.055	.312
PERIOD	.773	.128	.012
Relative unit value (CATEGORY)	.404	.472	.584
Fixed-effect parameter estimates
Intercept	−33.9***	−35.6***	−32.0***
CATEGORY = cheese	−.529	−3.89	−8.11
CATEGORY = meat prep	0 ^a	0 ^a	0 ^a
PERIOD = 1 (1995–1999)	3.39	1.40	−.597
PERIOD = 2 (2000–2005)	0 ^a	0^a	0 ^a
Relative unit value (cheese)	.207	.275	.520
Relative unit value (meat preparations)	.148	.153	−.184
Variance/Covariance-matrix components †
	2,856.4***	3,135.0***	2,557.4***
	2,951.1***	3,049.1***	2,640.3***
σ12	2,526.7***	2,703.5***	1,855.7***
Model statistics
# of obs.	102	102	102
Restr. − 2 log lik.	1,008.7	1,013.4	1,032.0
AIC	1,014.7	1,019.4	1,038.0
BIC	1,022.4	1,027.1	1,045.7

In the theoretical better justified yet more complex GLS estimations, additional statistics need to be considered when interpreting the results. In particular, these are the estimates for the variance components and the maximum likelihood-related model fit statistics. These measures are all reported in Tables 2 to 6. In general, all variance components were estimated as statistically highly significant. Often, there seems to be a considerable difference between the error variance of the first and the second period, implying that heteroscedasticity may be an issue in the data. Significant residual autocorrelation is present in all cases. Thus, on statistical grounds, the use of feasible GLS seems to be justified. The overall fit of the estimated models is not very high as can be seen from the comparatively high values of the information criteria Restricted–2 log likelihood, Akaike information criteria (AIC) and Bayesian information criteria (BIC) (see Appendix for a theoretical discussion of these measures). However, it needs to be mentioned, and it can be seen from the scatter plots in Figures 1 to 5, that the here generally assumed linear relationship between relative export performance and quality may not hold in all cases. Often, a nonlinear (e.g., parable-shaped) relationship can be observed from the data with many countries exporting relatively few low- and high-quality products but many medium-quality ones. Yet, approximating the overall situation through a linear regression line may still yield a reliable indication of whether relatively more high-quality products than low-quality ones are exported (in which case the line's slope would be positive) or vice versa. Nonetheless, the overall fit of this line to the data can be expected to be poor.

Germany

In the OLS estimations (Figure 1), for all investigated product categories and export destinations, the relationship between product quality and export performance is negative. The slope coefficient is statistically significant (at least at the 95% confidence level) for total and intra-EU per capita cheese exports and in addition for wine exports to all three export destinations. Yet, in general, there does not seem to be a major difference between the intra- and extra-EU situations.

The findings differ slightly when using the pooled data estimator (GLS). As the results in Table 2 reveal, the relative export performance situation differs significantly between the two considered periods for intra-EU and total exports but not for extra-EU exports. As for product category-specific differences, the export performance situation does not generally differ significantly across the three included product categories, at least not between wine and cheese. (The mean export performance between the reference category wine and meat preparations differs, however, significantly across all export destinations.) Finally, and most important, food quality, as indicated by the relative unit value, is significantly negatively related to relative export performance across all export destinations.

Overall, then, it appears that Germany exports relatively more lower and medium-quality food products than high-quality ones in all three considered product categories.

Italy

Looking at the scatter plots (Figure 2), the Italian situation is almost completely different compared with the one of Germany. The estimated slope coefficients are positive for cheese, close to zero for meat preparations, and negative for wine to all export destinations. They are statistically significant (at least at the 95% confidence level) for extra-EU cheese exports and for wine to all export destinations.

The pooled data estimates reveal that, similar to Germany, in particular the wine and meat preparations, export situation differ structurally. However, in contrast to the German situation, no period effect can be observed. As for the quality indicator, given that its sign differs across the three analyzed product categories, the coefficients have been estimated separately for each category. However, none of the estimates differs significantly from zero (see Table 3).

Overall, the Italian export situation is less clear than the German one. Due to the diversities across the three products, the pooled data analysis does not yield significant estimates for the quality–export performance relationship. Nevertheless, when looking at the data, it appears that Italy is a significant exporter of high-quality cheese and meat preparations but not of high-quality wines.

France

Figure 3 shows that the situation in France is the exact mirror image to the one in Italy: a negative quality–export performance relationship for cheese and meat preparations and a positive one for wine. For intra-EU exports of meat preparations there seems to exist no clear relationship because it can be seen that mostly medium-quality products are exported and about equally few low- and high-quality ones. Only the slope coefficients for extra-EU exports of meat preparations and wine exports to all destinations are statistically significant (95% confidence level).

The GLS estimates show that no significant differences across the considered product categories and periods appear to exist. Similar to the Italian situation, the slope coefficients of the quality variable were estimated category specific. The results indicate that there is a significant controlled quality–export performance association for total and extra-EU exports of meat preparations (negative) and extra-EU exports of wine (positive; see Table 4).

Overall, France seems therefore to be a quality exporter of wine only. In the other product categories, export performance is highest for low- and average-quality products.

Spain

The Spanish situation (Figure 4) is the most heterogeneous among the countries considered. There appears to be a difference in the direction of the export performance–quality relationship between the intra- and extra-EU export situations in particular for cheese but also to a smaller extent for meat preparations and wine. For cheese, the relationship is negative for the total and intra-EU situation but positive for extra-EU exports. For meat preparations, the situation is almost vice versa: a positive relationship for the total and intra-EU situation and a neutral (or only slightly positive one) for extra-EU exports. For wine, the slope coefficients are positive except for the intra-EU situation. However, none of the OLS quality coefficients has been estimated as being statistically significant at the 95% confidence level.

The pooled data estimates show that there is a category effect for intra-EU exports and a period effect for extra-EU exports. Within the European Union, the exports of meat preparations differ significantly from the ones of wine. Outside the European Union, for some reasons, relative exports have been higher in the 1995–1999 period than in the 2000–2005 one. Similar to Italy and France, the quality coefficients were estimated category specific. The results show that the only significant effect (negative) is for extra-EU exports of meat preparations. The reason this is not reflected in Figure 4 is that by pooling the two periods into one data set, the regression line running through the combined data (almost) has a zero slope. If a regression line is produced separately for the two periods, the one for the second period has a positive slope whereas the one for the first period has a negative one. This also explains why there is a highly significant period effect for extra-EU exports (see Table 5).

Overall, the Spanish situation seems to be rather balanced. When looking at the data, it becomes obvious that high export performance is fairly evenly distributed across the entire quality spectrum. Hence, Spain exports cheeses, meat preparations, and wines of all quality categories to about an equal extent. However, it also appears that high-quality cheeses and wines tend to be exported rather to extra-EU destinations.

United Kingdom

The UK situation (Figure 5) is similar to the Spanish one. For cheese, the estimated slope coefficients are negative for intra-EU exports and positive for extra-EU ones, resulting in a (almost) zero slope coefficient for total exports. For meat preparation, the situation is vice versa: a positive relationship for intra-EU exports and a negative one for extra-EU exports, with a resulting slightly positive slope coefficient for total exports. However, none of the coefficients turned out to be statistically significant (95% confidence level; Figure 5).

The GLS results show that, given the underlying heterogeneous export situation, no parameter estimate is statistically significant (see Table 6).

Overall, and similar to Spain, when looking at the scatter plots, it appears that the United Kingdom exports products across the entire quality spectrum. However, the UK emphasis seems to rest on medium-quality products irrespective of the export destination.

CONCLUSIONS

In summary, this article has investigated the relationship between product quality (as indicated by UV) and export performance, both measured in relative terms. The estimation results show that the connection between product quality and export performance depends on the product category (but not on the period) and differs (but not in all cases) according to the export destination. Although the signs of the estimated slope coefficients are stable, the obtained statistical significance levels for these parameters depend on the estimation method (OLS or GLS).

The estimation results reveal that Italy displays strong export performance in high-quality cheese and meat preparations but not in high-quality wines. In France, the situation is exactly vice versa: higher relative export performance for lower and medium-quality cheeses and meat preparations but strongest export performance for high-quality wines. In both Spain and the United Kingdom, relative export performance appears to be equally distributed across the entire quality spectrum. Finally, in Germany, there is a negative quality export performance connection for all three product categories, implying that high-quality products play only a minor role in the country's exports.

The findings from this analysis suggest that although there does not seem to be a systematic relationship between food quality and export performance, high-quality products clearly play a considerable role in the food exports of at least some countries (in this study, Italy, France). This underlines the point that it is justified, and even necessary, to give special attention to this sort of trade. One possibility to do this, for instance, could be the introduction of specialized university courses in international marketing of quality food products. Today, a few other highly specialized postgraduate management courses already exist that may serve as a model (e.g., Essec Business School's Master's of Business Administration in Luxury Brand Management or the University of South Australia's Master's of Wine Marketing). In addition, specialized short courses (or other training programs) could be designed for the target group of already practicing agribusiness professionals. Given that the issue seems to be of relevance for several EU countries, an integrated, pan-European teaching/training program and thus the pooling of the expertises from different countries may perhaps be most useful. In any case, such specialized capacity-building initiatives for current and/or future agribusiness leaders can clearly be expected to contribute positively to the enhancement of the competitiveness of EU food companies, which operate in increasingly liberalized and quality-conscious markets.

Future research may investigate why some countries in some product categories perform better in exporting high-quality products than others. One possible reason could be that exports simply reflect production proficiency, that is, that countries which produce relative more high-quality food products also export more of these. However, this hypothesis would need empirical confirmation.

CONTRIBUTOR

Christian Fischer is Associate Professor of agricultural economics and management. He holds a doctorate in agricultural economics, Giessen University, Germany; specialized master's degree in agribusiness management (MSMAI), Lyons Graduate School of Management (EM Lyon) & Ecole Nationale Supérieure Agronomique Montpellier (ENSAM); a graduate certificate in international economics, The University of Adelaide, Australia; and a master of science degree in food economics, Giessen University, Germany.

The author gratefully acknowledges financial support from the H. Wilhelm Schaumann Stiftung, Hamburg, Germany.

Notes

Source: Author's calculations from Eurostat data.

Note. DE = Germany, IT = Italy, FR = France, ES = Spain, GB = Great Britain.

Source: Author's estimations from Eurostat data.

Note. *** (**, *) statistically significantly different from zero at least at 99% (95%, 90%) confidence level; significant levels are based on panel-robust standard errors.

^a Reference category.

†Restricted maximum likelihood estimates.

Source: Author's estimations from Eurostat data.

Note. *** (**) statistically significantly different from zero at least at 99% (95%) confidence level; significant levels are based on panel-robust standard errors.

^a Reference category.

†Restricted maximum likelihood estimates.

Source: Author's estimations from Eurostat data.

Note. *** (**) statistically significantly different from zero at least at 99% (95%) confidence level; significant levels are based on panel-robust standard errors.

^a Reference category.

†Restricted maximum likelihood estimates.

Source: Author's estimations from Eurostat data.

Note. *** (**) statistically significantly different from zero at least at 99% (95%) confidence level; significant levels are based on panel-robust standard errors.

^a Reference category.

†Restricted maximum likelihood estimates.

Source: Author's estimations from Eurostat data.

Note. *** statistically significantly different from zero at least at 99% confidence level; significant levels are based on panel-robust standard errors.

^a Reference category.

†Restricted maximum likelihood estimates.

In other words, de gustibus non est disputandum.

A unit value is the quotient of nominal sales to a physical unit of measure (Gelhar & Pick, 2002).

A multiple regression should operate on a complete list of causes. If omitted variables are correlated with included variables, then the estimated parameters will be biased. On the other hand, if omitted variables are orthogonal to all included variables, then parameter estimates will be unbiased, though the sampling error attached to the parameter estimates will be higher than it need be (Deegan, 1976).

Consider the example of high-quality cheese. A matured Parmesan hard cheese is easier to transport and store than an ultrafresh goat pyramid even though “objective” quality may be the same. Given the higher logistics ease, ceteris paribus, export performance of the former may be higher. Export performance of Parmigiano may also be higher because its main export markets may be more affluent countries than the ones to which the goat cheese is exported, thus making the former relatively less expensive and resulting in comparatively stronger demand.

See Hinloopen and Van Marrewijk's (2001) study on the empirical distribution of the Balassa index and the problem of the incomparability of RXA scores across countries.

However, it should be noted that despite symmetrization the problem of unequal distribution and different country-specific upper bounds remains.

The covered product codes are-cheese: 4061010–4069099, meat preparations: 16010010–16029099, and wine: 22041011–22042199. The exact description of the individual products can be found on the Internet through the Eurostat Comext database or can be obtained on request from Christian Fischer.

Only two product types (16022011 and 16022019: preparations of goose or duck liver) were removed in the meat preparations category. In almost all countries, these products displayed extremely high unit values and very low export performance, thus making them very influential outliers.

APPENDIX

Given the special nature of the data (i.e., a short unbalanced panel, or, more accurately, a nested cross section with one repeated measurement), a general linear model (GLM) was fitted for estimating the slope coefficient of product quality with regard to product export performance. Formally, the GLM (i.e., a fixed-effects specification that is assumed to be linear in parameters; Verbeke & Molenberghs, 1997) can be specified as

μ is the overall (grand) mean (a fixed parameter equivalent to the usual intercept), the fixed main-effect parameter for the two categorical predictors (i.e., “factors”) category (k = 1, …, K) and time period (t = 1, …, T _k ), the fixed parameters of l included (nonstochastic) covariates (l = 1, …, L), and the random error (disturbance) of . (For the research problem described in this paper, K = 3 and T _ck ∈ [1, 2], depending on the analyzed country and category, and L is in general larger than 1 because nested effects were considered).

Equation (A.1) can be simplified to , where denotes a (row) vector of m inputs (factors, covariates, and interactions) and the (column) vector of corresponding parameters (including the ). Rewritten in full matrix notation, Equation (A.1) becomes

where y is a response vector, X is a input matrix, β is a m × 1 parameter vector, and ϵ is a disturbance vector.

Given the panel structure of the data, where the different product types are treated as subjects, , and Var(y) = Var(ϵ) is assumed to be i.i.d. N(0, ), thus disturbances in the two panels may be contemporaneously correlated and potentially heteroscedastic (i.e., displaying nonconstant variance). As a consequence, in Equation (A.2) cannot efficiently be estimated by pooled ordinary least squares (OLS) regression. Instead, a method that takes into account autocorrelated and nonconstant residual errors needs to be used. One such method is feasible generalized least squares (FGLS), where

(see Cameron & Trivedi, 2005) and where implementation requires the consistent estimation of V, the variance—covariance matrix of disturbances ϵ. In order to consistently estimate V, one usually needs to make explicit assumptions about the underlying structure of its variance components, but it is possible (and was done in this case) to treat V as being completely unstructured, that is, .

The estimation of variance components can be done in different ways. Under the i.i.d. multivariate normal assumption for ϵ, maximum likelihood estimation methods are usually employed, with two possible options: maximum likelihood (ML) or restricted maximum likelihood (REML). A weakness of the ML method is that the estimates are biased in small samples (Cameron & Trivedi, 2005). Moreover, because REML does explicitly take into account the loss of the degrees of freedom involved in estimating the fixed effects, it is the recommended option in models containing many fixed-effect parameters (Verbeke & Molenberghs, 1997). The −2 times log-likelihood of REML is (Cameron & Trivedi, 2005)

where , |V| denotes the determinant of V, N the number of subjects, and p is the rank of X. The variance components of V can be computed by maximizing Equation (A.4); however, in general, there are no closed-form solutions. Therefore, Newton and scoring algorithms are usually used to find the solution numerically, starting with some initial value for residual error variance σ². Assuming i.i.d. N(0, σ²), residual sum of squares from OLS regression usually yields to be used as starting value.

Once V has been estimated, original data X and y are then accordingly transformed and OLS regressions are run on the adjusted data, yielding autocorrelation and heteroscedasticity-adjusted and respective panel-robust standard errors. For the significance tests of the included factors, Analysis of Variance (ANOVA) (i.e., “method of moments”-type) estimators, which equate quadratic sums of squares to their expectations and solve the resulting equations for the unknowns, are used. Baltagi, Song, and Jung (2001) showed that ANOVA methods perform well in estimating the regression coefficients in unbalanced nested error-component regression models. Given that the data set in this paper is unbalanced and that the main interest is in the significance of the remaining differences in the factor category (or marginal) means, Type III sums of squares are used (Hill & Lewicki, 2006).

In maximum likelihood-based FGLS regressions, conventionally only the final value of the (restricted) −2 times log-likelihood function and derived information criteria (such as the Akaike information criteria [AIC] and the Bayesian information criteria [BIC]) are calculated (see Cameron & Trivedi, 2005, for a discussion), on the basis of which appropriate (nested) models are selected. However, for assessing the overall fit of a model to the underlying data, theses statistics are less useful because they cannot be compared across different nonnested models.