Food Policy 43 (2013) 1–13

Contents lists available at ScienceDirect

Food Policy

journal homepage: www.elsevier .com/locate / foodpol

How much trustworthy and salubrious an organic jam should be?The impact of organic logo on the Italian jam market q

0306-9192/$ - see front matter � 2013 Elsevier Ltd. All rights reserved.http://dx.doi.org/10.1016/j.foodpol.2013.08.003

q I would like to thank Steinar Strøm for constant support and helpful commentsand Alessandro Sembenelli and participants of the EARIE Conference 2006 inAmsterdam for useful discussions. I also would like the editor and two anonymousreferees for the helpful comments and suggestions. I am grateful to the Italianbranch of Infoscan Resources Inc. for having provided the data. The usual disclaimerapplies.⇑ Tel.: +39 0577232688; fax: +39 0577 232661.

E-mail address: tiziano.razzolini@unisi.it1 See Council Regulation (EC, 1991) No 2092/91 of 24 June 1991, Council Regulation

(EC, 1999) No 1804/99 of 19 July 1999, Council Regulation (EC, 2007) No 834/2007 of28 June 2007 and the European Action Plan for Organic Food and Farming in 2004(EC-COM, 2004).

2 See De Magistris and Gracia (2008) for a recent article on the demand fofood in the South of Italy.De Magistris and Gracia (2008) and Torjusen et aprovide a review of the empirical analyses on the demand for organic food

3 See ISMEA (2007, 2012) for information on the recovery and growtdemand of organic products after 2004. We thank Enrico De Ruvo, ISinformation on the recent trends in consumption of organic jams.

Tiziano Razzolini ⇑University of Siena, Department of Economics and Statistics, Piazza S. Francesco 7, 53100 Siena, Italy

a r t i c l e i n f o

Article history:Received 5 November 2009Received in revised form 25 February 2013Accepted 5 August 2013

Keywords:Organic productsNutritional labelsProduct differentiationGEV models

a b s t r a c t

This paper studies the impact of organic logo in the Italian jam market. Using data on true sales in Italiansupermarkets in the 2002–2004 period we estimate the structural demand model developed by Berry(1994). This approach allows us to solve potential endogeneity problem in the estimation of the pricecoefficient in the demand equation, as well as other aspects related to multiple sources of differentiation.In a market where 62 percent of organic jams are diet this approach allows us to identify separately themarket power induced by the organic logo and by the low content of sugar. The estimation results fromvarious specifications of GEV class of models show that the organic attribute guarantees a degree of mar-ket power greater than the diet attribute. However, the protection from competition entailed by theorganic logo is not particularly strong, and the diet attribute does not guarantees additional protectionto organic jams. Most importantly, our findings show that consumers trust more big organic producersthus suggesting important policy implications on the supply and marketing side of the jam market.

� 2013 Elsevier Ltd. All rights reserved.

Introduction

The increased consumption of organic produce has been consid-erably supported by public interventions and by developments inthe supply and retailing chains. On the demand side of the market,Council regulations as well as national programs have promotedconsumers’ awareness of the characteristics of organic farming.1

On the supply side, economies of scale in the distribution togetherwith harmonization and standardization of organic produce have fa-vored their expansion from small specialized shops to mainstreamretailer channels (Torjusen et al., 2004). This phenomenon has raisednew challenges for organic produces, since it has exposed them tocompetition from cheaper conventional products in markets thatare often characterized by a high degree of differentiation.

This paper analyzes the impact of the organic logo and the com-petition between organic and conventional jams in the Italian mar-ket in 2002–2004 using real transaction data from supermarkets.

First, the Italian jam market represents an interesting case to beanalyzed with a structural model for the demand for organic prod-ucts. Although in the nineties the Italian market could not be con-sidered as mature as, for example, the Danish or British one, thedemand for organic products during 1990–1999 experienced aboom (Cremonini, 2007; Compagnoni, 2001; Torjusen et al.,2004).2 By the end of 2000, the market share of organic productssold by supermarkets exceeded the share sold in specialized organicshops (Pinton, 2001; Torjusen et al., 2004). All organic productsexperienced a reduction in demand between 2003 and 2004 due tothe general crisis (ISMEA, 2005, 2007), a stagnation that lasted untilthe end of 2007 (Cremonini, 2007; ISMEA, 2012). After 2008 con-sumption of organic production has been rising continuously until2012. During 2009–2012 the annual variations in the consumptionof organic jams have been 14, 6.7, 8.6 and 14.6 percent respectively.3

According to Rete Rurale Nazionale (2012) organic jams represented8.2 percent of total organic sales in 2011. The recent growth in theconsumption of organic jams and organic produce in general seemsto be driven by the increasing importance of private labels. The firstItalian supermarket started to offer organic products under its pri-vate labels in 1999, and all the other larger retailers started a fewyears later (Cremonini, 2007; Santucci and Pignataro, 2002). Private

r organicl. (2004)in Italy.h in theMEA, for

7 Carson et al. (1996), Cummings et al. (1995), Frykblom (1997) find relevant

2 T. Razzolini / Food Policy 43 (2013) 1–13

label jams, organic and non-organic, gained a larger market share,particularly in the latest 2000s.

Second, Italian jam market is highly differentiated: in additionto the organic nature of the product, jams are also characterizedby the diet feature, i.e. the low content of sugar. 62 percent of or-ganic jams in our sample are also diet, and are provided almostexclusively by the organic producers. This is an interesting phe-nomenon, since it might be a result of differentiation strategiesimplemented by Italian manufacturers. Given the increased levelof competition, such an analysis of differentiation strategies be-comes even more crucial. These differentiation strategies attemptto make organic product more attractive, since consumers of or-ganic products might be more concerned about diet and weight.4

To the extent that additional information on nutritional characteris-tics increases the willingness to buy organic food, there are econo-mies of scope in the labeling of such information, and there mightbe room for further regulatory interventions aiming at promotingthe demand for organic products and sustaining organic farming.5

Our findings from real market data indicate that the organiclogo ensures a low level of market power which is, nevertheless,higher than the market power induced by the diet attribute. Thelatter attribute thus seems to be an ancillary characteristic which,when combined with the organic one, does not provide additionalprotection from competition.

Most importantly, our estimates identify a relevant effect of thenumber of products sold by each manufacturer on consumers de-mand and suggest that buyers trust more manufacturers with awider variety of organic jams. In a market characterized by animbalance between small manufacturers/farmers and large retail-ers, this result indicates strong disadvantages for small suppliersand sheds light on the role of large retailers in the future develop-ment of organic farming. On the one hand, larger scales in the dis-tribution have lowered the costs and thus reduced the pricedifferential between organic and conventional goods, which is con-sidered as one of the major factors discouraging the consumptionof organic produce. On the other hand, these findings contributeto the debate on whether retailer-led interests may contrast withthe original organic values and on who is going to benefit fromthe success of organic produce (Vogl et al., 2005). The alignmentof process of production to the needs of mainstream distribution,combined with the importance of brand familiarity (Krystallisand Chryssohoidis, 2005) in this segment of the market may trans-form historical limits in the supply of organic produce, such as pooravailability and poor reliability in delivery, in an insurmountablebarrier for its diffusion among small scales of production.

The empirical analysis in this paper employs a structural de-mand model developed by Berry (1994) to provide a more detailedanalysis of the supply side of the market and in particular of thestructure of competition, the role of marketing, promotional dis-counts and the strategies of differentiation of the products.6

Although Berry’s (1994) approach has been successfully employedin the analyses of non-durable goods since the work of Nevo(2001), to our knowledge this is the first study to use this methodol-ogy in the analysis of organic products employing real market data.As recently discussed by Millock et al. (2002) and Wier et al. (2008),true market data are in principle to be preferred to information

4 See for instance Chen (2009), Gil et al. (2000) and Schifferstein and Oude Ophuis(1998). Hughner et al. (2007) and Yiridoe et al. (2005) review the concerns abouthealth issues as well as the differences in lifestyles affecting the demand for organicproducts.

5 The performance of small organic manufacturers certainly affects the welfare offarmers since jams can be easily and directly processed by the farmers themselves.See Santucci and Pignataro (2002) for information on the percentages of producerswith processing activities on the farms in Italy.

6 As noted by Padel and Foster (2005) the effect of promotional discounts byretailers needs to be investigated.

elicited from postulated behaviors.7 However, as noted by Grunert(2005), analysis with real data may also lead to estimating a willing-ness to pay which is ‘‘a great deal lower than those derived fromhypothetical methods’’.8 This result is most probably due to twofacts. First, respondents to surveys may overstate their willingnessto pay (Hansen and Sorensen, 1993). Second, when using true aggre-gate market data researchers are not able to observe many factors,such as perceived quality and attitudes that are instead availablein survey data. This creates an endogeneity problem, explained inSection ‘Identification and estimation procedure’, which is wellknown in the field of new empirical industrial organization anddue to which the estimated price coefficient may be biased towardszero.

The main advantage of the empirical methodology employed inthis paper is its ability to address the endogeneity issue. In partic-ular, we use the relevant variation in the number of jams offered inthe market during the period under analysis as our identificationstrategy. At the same time, this approach, conditional on a correctchoice of the demand specification model, allows us to take intoaccount unobserved consumers’ heterogeneity due to differencesin attitudes towards organic and diet goods that cannot be mea-sured with aggregate sales data. The unobserved attitudes towardsmultiple sources of differentiation and their implications for com-petition are taken into account by choosing the best specificationfrom the broad Generalize Extreme Value (GEV) class of models.

Finally, estimating a structural demand model for the specificcase of organic jam market has additional advantages. First, prob-lems due to the unobserved quality or visual aspects are less rele-vant.9 These visual aspects, typically unobserved in aggregates salesdata, might, in principle, create additional problems of endogeneity,as they are likely to be correlated with price. Since jams are a pro-cessed and labeled product, there exists no trade-off between sen-sory defects and willingness to pay for the organic nature of theproduct. Second, the problem of seasonality in consumption is lessimportant, since jams represent a natural way of preserving fruitsover time.

The remainder of the paper is organized as follows. The nextsection provides background information of the Italian jam marketand a description of the dataset. Section ‘Demand model’ describesthe different model specifications used to construct unobservedindividual probabilities and links the observed aggregate marketshares to the discrete choice probabilities. Section ‘Identificationand estimation procedure’ illustrates the identification strategyand the estimation procedure. Estimation results are discussed inSections ‘Empirical results and Conclusion’ concludes and suggestspolicy implications.

Data description

The national aggregate data employed in this study are kindlyprovided by the Information Resources Inc. (IRI). The data set con-tains monthly scanner information on values, quantities and prices(before and after promotional discount) sold by the top five

differences between the estimates from revealed and stated preferences. See alsoTrain (2009) for a discussion on the differences between revealed and statedpreferences in the context of random utility models. Verhoef and Franses (2002) findthat, for forecasting purposes, stated preference variables have poor predictive powercompared to models using revealed preferences.

8 Bähr et al. (2004), Bellows et al. (2008), Makatouni (2002), Onyango et al. (2007),and, Padel and Foster (2005) investigate the difference between actual and postulatedbehaviors in the demand of organic produce.

9 Thompson and Kidwell (1998), Huang (1996) and Yue et al. (2009) analyse theimpact of cosmetic damage on consumers’ demand. Kuhar and Juvancic (2010) arguethat visual attractiveness is one of the factors explaining the purchasing of organicproducts.

Table 1Number of jams classified according to organic ad diet features. Source: Author’s computation from IRI data.

Diet Non-diet

Panel (a) Number of types of jam and percentagesOrganic 25 Organic jams represent 76% of diet, Diet are 62% of organic 15 Organic are 8% of non-diet non-diet are 38% of organicNon-organic 7 Non-organic are 24% of diet, diet are 5%% of non-organic 175 Non-organic are 92% of non-diet, non-diet are 95% of non-organic

Diet Non diet Total

Panel (b) Sales percentage of organic and diet jamsOrganic 11% 2% 13%Non-organic 3% 84% 87%

Panel (c) Average discounted price according to organic and diet featuresOrganic 8.74 6.71 7.98Non-organic 7.27 4.89 5

T. Razzolini / Food Policy 43 (2013) 1–13 3

traditional producers and the by five organic manufacturers insupermarkets in Italy from May 2002 to May 2004.10

The sample consists of 331 jams with an annual total volumeand an annual total value of sales equal to 15,872,471 kilogramsand to 73,080,148 Euros, respectively. The dataset contains alsothe volumes sold by private labels, but they are excluded fromthe sample, since for these jams, we do not have information ontheir nutritional characteristics.

Organic jams constitute 15% of total volumes and show anincreasing level of consumption. Panel a in Table 1 shows that or-ganic jams represent 76% of the diet jams. Panel b reports the salespercentages of jams classified according to the organic and dietfeatures. It indicates that organic jams represent 13% of total sales,while sales of traditional jams (i.e. non-organic and non-diet) con-stitutes 84% of the market. Fig. 4 plots the number of organic jamsover the period under analysis. As can be seen from this figure,there is an initial increase in the number of organic jams which,however, declines after December 2003. This trend follows closelythe dynamics of all other organic products In Italy in the same per-iod. This dynamic suggests that organic products were more penal-ized than the conventional ones, and may reveal that for Italianconsumers the price premia required by organic products mayhave been perceived, in the period under analysis, as a luxury.

Panel c in Table 1 describes average discounted price classifiedby organic and diet features. On average, organic and diet jamshave the highest discounted price; the second highest is repre-sented by the diet and not organic jams. The cheapest jams arenon-diet and non-organic ones.

Table 2 reports jams’ characteristics and descriptive statistics oftheir market shares classified according to these attributes. Thedummy variable G4 is equal to one for a jar equal or heavier than400 g. The variable mono equals one if the jam is sold in a mono-dose format. The variable fruit takes value one if the jam is soldin the ‘‘ketchup-bottled’’ package. These packages are labeled as‘‘fruity’’ by jam manufacturers but do not contain higher contentof fruit. Manufacturers advertise these jams as ‘‘fruity’’ in orderto attract kids and their mothers and thus to expand the consump-tion of their products into the snack market. The variable Promo isequal to one if the jam is sold under promotional discount.

Table 3 reports estimates from a hedonic price regression. Incolumn 1, the diet attribute coefficient has a positive and signifi-cant impact on price, and this effect is three times larger thanthe coefficient of the organic dummy.11 In column 2, we introducethe interaction between organic and diet attribute which captures

10 The requirement to obtain an organic logo makes a conversion from traditionalproduction to the organic one not particularly convenient. Such conversion takes 3years and requires stating that explicitly on the label.

11 Note that we do not interpret these coefficients as premia that consumers arewilling to pay for these attributes. Manufacturers’ pricing behavior may bias theestimates of the demand parameters. The estimation procedure adopted belowaccounts for this problem.

the effect of the two healthy characteristics. Fruit jams have a nega-tive impact on price level, whereas mono-does jams have a positiveand significant coefficient. Finally, the coefficient for jam jars equalor heavier than 400 g is negative and significant.

Demand model

This section provides theoretical framework for estimation pro-cedures below. First, we outline briefly discrete choice models usedto construct individual probabilities. We then present a descriptionof aggregate market shares under the assumption that they areequal to individual probabilities.

Discrete choice models

We assume that individuals’ choices follow the theory of ran-dom utility models (RUM) and employ the discrete choice frame-work based on the Generalized Extreme Value class of modelsdeveloped by McFadden et al. (1978). According to this framework,each consumer chooses an alternative that yields the highest util-ity. The utility from consumption of product j for individual i attime t is defined as a follows12:

Uðpj; xj; nj; hÞ ¼ Uj ¼ dj þ ej; j ¼ 1; . . . ; J; ð1Þ

with

dj ¼ Xjbþ nj ð2Þ

where Xj are observable product attributes (e.g. prices pj) and nj isan unobservable product attribute. The vector h includes coeffi-cients b and parameters characterizing the distribution of the errorterm.

The individual utility is equal to the sum of the mean utility dj

and the random component ej. The mean utility is the average eval-uation of the utility of product j. The error term ej represents devi-ation from this common evaluation and captures consumers’heterogeneity in tastes. Since we do not observe individual choicesand, as a consequence, do not know any characteristics of the con-sumers, the error term is the unique component that captures dif-ferences in attitudes towards different jams. Therefore,assumptions on error term distribution have a fundamental rolein determining the choice, and different distributions of the errorterm lead to different implications in terms of market power andpatterns of substitution across different jams.

If the error term is assumed to be type I extreme value distrib-uted, the function for probability will have the well-known logitclosed form (see, for instance, Train 2009):

Pj ¼expðXjbÞPnk expðXkbÞ

¼ expðdjÞPnk expðdkÞ þ expðd0Þ

ð3Þ

12 For simplicity’s sake we omit the index i and the time index t.

Table 2Variables definition and mean and standard deviations of market shares.

Variables definition

Organic Jams composed of organic fruitDiet Jams characterized by low content of sugarFruit Jams in Ketchup-bottled packageMono Jams in mono-dose packagesG4 Jams jars equal or heavier than 400 gPromo Jams sold on promotional discount

Mean Standard deviation

Average market share 0.00032 0.0006Average price 5.80 2.48Average market share of organic jams 0.0002 0.0003Average market share of non-organic jams 0.00035 0.0006Average market share of diet jams 0.0003 0.0003Average market share of non-diet jams 0.00032 0.0006Average market share of fruit jams 0.0004 0.0006Average market share of non-fruit jams 0.0002 0.0005Average market share of mono–jams 0.0002 0.0002Average market share of non–mono jams 0.0003 0.0006

Table 3Estimates from hedonic price regression.

Dependent variable: price (1) (2)

Organic .935*** .445***

(.084) (.098)Diet 2.781*** 1.797***

(.087) (.135)Organic * diet _ 1.620***

(.171)Fruit �.888*** �.999***

(.068) (.069)Mono 1.364*** 1.226***

(.095) (.095)G4 �1.763*** �1.861***

(.079) (.078)Constant 4.035*** 4.189***

(.182) (.181)Fruit flavor dummies Yes Yes

Observations 4891 4891R-squared 0.56 0.57

⁄ 10% Significance level.⁄⁄ 5% Significance level.*** 1% Significance level.

4 T. Razzolini / Food Policy 43 (2013) 1–13

where Pj is the probability of buying product j, and Xj are the char-acteristics of the product j. d0 is the mean evaluation of the outsideoption, i.e. the decision not to buy.

The simple multinomial logit specification is subject to theIndependence of irrelevant alternatives (IIA) property.13 In orderto have more reasonable substitution patterns, we need to introducecorrelation between choices belonging to the same group. This cor-relation arise because jams belonging to the same group (definedby organic or diet attribute) may share other known by a consumerbut unobserved to an econometrician characteristics (and thus notincluded in the Xs but that affect the distribution of the error termej in Eqs. (1) and (2)). We can account for such similarity in productsdue to common unobserved characteristics by assuming that theindividual choices are well represented by Nested Multinomial Logit

13 This means that the ratio of two choice probabilities is independent fromcharacteristics of any third alternative. An implication of this property is that anincrease, for instance, in the price of organic jam should induce an increase in theprobabilities of other jams that is proportional to their respective market shares(since the ratio of two choice probabilities must remain the same). Intuitively, theincrease in the price of organic jam should induce previous consumers of that organicjam to switch to another organic jam.

models. Fig. 2 shows the decisional tree of a two-levels Nested Logitmodel, where nests are defined by organic or diet attributes. Then,the probability of buying a product j in the nest g has the followingclosed form:

Pjg ¼exp djg

1�rg

� �PSðgÞ

k¼1 exp dkg

1�rg

� � expðIgÞPGp¼1 expðIpÞ þ expðd0Þ

ð4Þ

where djg is the mean evaluation of product j and has the specificationgiven by equation (2); Ig are the inclusive values of group g14: and thecoefficient rg represents correlation in the unobserved portion of theutility for jams belonging to the same group, which properties will beexplained below in this section. The two-level nested logit differenti-ates jams according to a unique attribute (organic/non-organic or,alternatively, diet/non-diet).15 If, for instance, jams are divided intogroups according to the organic (diet) attribute, we assume that thereis no correlation in the unobserved part of the utility among jams thatshare the excluded diet (organic) attribute. We can differentiate jamsaccording to both diet and organic characteristics by introducing an-other dimension of differentiation (i.e. an additional nest) and byadopting a three-level Nested logit model. Fig. 3 shows the decisionaltree of a three-level Nested Logit model with the diet attribute defin-ing groups at the top and the organic attribute defining subgroups atthe bottom. The distribution of the error term assumed by this modelleads to the following closed form (see Train, 2009):

Pjsg ¼exp djsg

1�rs

� �PJðsðgÞÞ

k¼1 exp dksg

1�rs

� � expð IsðgÞ1�rgÞPSðgÞ

l¼1 exp IðlðgÞÞ1�rg

� �� expðIgÞPG

p¼1 expðIpÞ þ expðd0Þð5Þ

where Pjsg is the probability of buying a product j in the subgroup sof group g, djsg is mean evaluation of product j and has the specifi-cation given in equation (2), Is(g) and Ig are the inclusive values ofsubgroups and groups, respectively.

The coefficients rs and rg, together with b from equation (1), areour parameters of interest. rs (rg) represents the degree of correla-

14 We do not report the formula for inclusive values. Detailed description of theinclusive values as well as of the GEV class of models can be found in standardtextbooks on discrete choice, see, for instance, Train (2009).

15 We have tried to include also fruit flavour as a dimension of differentiation.However, empirical results (not reported) suggest that IIA property holds along thisdimension. This can be interpreted in terms of consumers’ ‘‘love of variety’’ fordifferent fruit flavors.

Fig. 1. Two-level nested logit model.

Fig. 2. Three level Nested logit model.

17 This property can be easily checked by taking the ratio of the probability of O, NDand a NO, ND jams as defined in equation (5). This property occurs because thechanges in the probabilities of O, ND and NO, ND jams due to the variation in theobservable attributes of O, D jams are proportional to their own probabilities PO,ND

and PNO,ND.

T. Razzolini / Food Policy 43 (2013) 1–13 5

tion in the unobserved part of the utility among the alternatives inthe same subgroup (group). When rs (rg) is equal to one there isperfect substitution among alternatives in the same subgroup(group). This has important implications in terms of substitutionpatterns among different jams. Let us assume that subgroups aredefined according to the organic attribute as in Fig. 2. If any ofthe Xj characteristics of an organic jam j worsens, the higher thecorrelation coefficients rs the more likely is that previous consum-ers of jam j will substitute towards another organic jam. This hap-pens because organic jams have some unobserved attributes incommon, the correlation of which is captured by the rs coefficient.These jams are perfect substitutes when rs is equal to one;whereas there is no correlation when the value of rs is zero. Thesame holds for the coefficient rg,16 that is, for correlation in theunobserved part of the utility among jams belonging to the samegroup. The values of these coefficients have thus important meaningin terms of substitution patterns among jams that share one of theattributes defining groups or subgroups. Intuitively, since higher cor-relation levels increase the probability of buying a jam in the samegroup (subgroup), coefficients rs and rg have a clear interpretationin terms of insulation from competition for other jams belongingto different nests.

Although the structure of the error term in NML models inducescorrelation among the alternatives, these models are still subject tothe Independence from Irrelevant Alternatives property (IIA) with-in nests and, moreover, from the Independence of Irrelevant nests(INN) property (see, for instance, Train 2009). As an example of thelatter property, let us consider the order of nesting in Fig. 2. Achange in the attribute of organic and dietetic jam (O, D jam)affects in the same way the jams that are non-diet and organic

16 The correlation coefficient rg has the same meaning also in the two-level nestedlogit of Fig. 1 when the unique nests are groups.

(O, ND) and jams that are non-diet and non-organic (NO, ND).17

Since OD and O, ND jams do not belong to the same organic sub-group, correlation due to the sharing of the organic dimension isthus lost.

INN property arises because the hierarchical order of nests im-poses mutually exclusive dimensions of differentiation. This obvi-ously constrains the substitution pattern across jams thatpartially share one of the ‘‘healthy attributes’’ but belong to differ-ent groups. Furthermore, the three-level NML model needs to re-spect McFaddens’ conditions. These conditions require thatrg < rg, meaning that correlation among jams increases from thetop to the bottom of the tree in Fig. 2.18 McFaddens’ conditions thusimply that jams in the same subgroups are more correlated thanjams belonging to the same group. In the example of Fig. 2 wheregroups are defined according to diet attribute and the subgroupsaccording to organic nature, the characteristic that implicitly yieldshigher correlation in the unobserved part of utility is the organicone. Bresnahan et al. (1997) describes the constraints of NML interms of substitution patterns and show that different order in thenesting structure yield significantly different estimates in terms ofcorrelation coefficients without violating McFaddens’ conditions. Be-cause of the high sensitivity of estimates to the order of nesting, therisk of a misspecification is very high. In our case, a hierarchical or-der in the choice between organic and diet features may be too

18 For two-level NML, McFadden’s condition requires that rg lies in the unit interval.See Train (2009) for a clear explanation of McFadden’s conditions and for theinterpretation of correlation coefficients in terms of dissimilarity (or heterogeneity)parameters.

Fig. 3. PDGEV model.

20 Berry (1994) implements this approach including also random parameters.Unfortunately, we cannot add random parameters since market shares only at thenational level are observed. Therefore, we cannot exploit variation of demographicvariables across different regions.

21 In the literature, market shares are often defined as the averages of all individualchoice probabilities. Since we do not observe consumers’ choices, all individuals areconsidered as being identical and therefore:

uj ¼XN

i¼1

Pi;j ¼ Pj

22 For example without the outside option, market shares of jams would beunaffected by a general increase in prices.

23 Source INN, ‘‘Alimentazione e nutrizione in Italia’’, Roma 1992, computed fromISTAT data.

24 In the logit case, Berry inversion simply means taking the logarithm of Eq. (3) for

6 T. Razzolini / Food Policy 43 (2013) 1–13

restrictive given the abundance of products that share bothcharacteristics.

To overcome the undesired properties of the NML model, weconsider the specification introduced by Bresnahan et al. (1997)which is called Product Differentiation and that belongs to theGEV class of models (henceforth PD GEV). In this specification,products are classified in non-mutually exclusive groups, and theyare thus allowed to compete with all other products with whichthey share at least one dimension. In this way, the PD GEV accountsfor correlation in error terms among non-nested cross products.

As is shown in Fig. 3, organic and diet attributes define non-mutually exclusive groups, there is no hierarchical order and thetwo decisions are simultaneous. PD GEV models can in fact be seenas a weighted sum of the two-level nested logit models. The prob-ability of choosing an organic and diet jams in a PDGEV models isthus as follows:

PjOD¼ao

Pj2Oe

djqo

� �qo�1

edjqo þad

Pj2De

djqd

� �qd�1

edjqd

aoP

j2Oedjqo

� �qo

þP

j2NOedjqo

� �qo� �

þadP

j2DedjqD

� �qd

þP

j2NDedjqD

� �qd� �

þed0

ð6Þ

where qO and qD are the heterogeneity parameters and representthe degree of dissimilarity of jams belonging to the organic and dietgroup, respectively. For these parameters, a value equal to zero rep-resents a case of perfect insulation from competition.19 If both val-ues are equal to one the model degenerates to the simplemultinomial logit. The parameters aO and ad are respectively equalto (1 � qo)/(2 � qo � qd) and (1 � qd)/(2 � qo � qd). They representthe weights in the sum of each of the two-level nested logit, and theysum up to one. These parameters measure the portion of an alterna-tive allocated to the dimensions o and d, that is, how much of thechoice of alternative j is driven by unobserved attributes commonto organic or non-organic jams, and how much depends instead onunobserved attributes shared by jams belonging to the same dietor non-diet groups. The PDGEV specification thus estimates sepa-rately the degree of correlation (1 � qo and 1 � qd) associated withorganic and diet attribute without constraining one of these charac-teristics to be more relevant than the other one, as does the three-le-vel nested logit.

Aggregate demand equations

Ideally we would like to observe the individual choice and esti-mate equations 4, 5 and 6 by maximum likelihood. Unfortunately,

19 Dissimilarity must be interpreted as the opposite of correlation. The q-param-eters can be considered as the complement of degree of correlation rs, that isqk = (1 � rk), k = o,d. However, he reader must keep in mind that (1 � qk) = rk are notdirectly comparable to rs and rg of the two-level and the three-level nested logitsince the structure and assumptions of the NML models are different from the PDGEVones. We use this notation to be consistent with the literature.

we observe only aggregate market shares at the national level. Fol-lowing Berry (1994),20 we assume that the aggregate market sharesare equal to the probabilities of buying product j, that is /j = Pj.21

Therefore, the market shares /j will be defined on the basis of thediscrete choice model assumed at the individual level, that is definedas one of the closed form expressions in Eqs. (4)–(6).

As shown by Berry (1994), a simple logarithmic transformationconverts the mean evaluation parameters of each jam j into a func-tion of the aggregate market shares and the outside option marketshare. This transformation applied to Eq. (3) yields the followinglinear expression:

lnð/jÞ � lnð/0Þ ¼ dj ¼ Xjbþ nj ð7Þ

where /j is the aggregate market share of jam j (by assumptionequal to the probability Pj) and /0 is the aggregate market shareof the outside option. In order to have reasonable implications,22

it is necessary to define the outside option – i.e. the choice not tobuy any of the alternatives. This is defined as the difference betweenthe potential level of the consumption of jam (i.e. the consumptionthat would have occurred if all the population consumed jams)and the total level of actual realized consumption. The total marketvolume is computed as the average consumption of jams multipliedby the number of potential entrants. Average monthly consumptionof jams is equal to the annual average consumption (4.6 kg)23 di-vided by twelve. In the estimation, the mean evaluation parameterd0 of the outside option is normalized to zero.24

Berry inversion can be applied also to Eq. (4) for the two-levelnested logit (see Berry, 1994) to obtain:

djg ¼ Xjgbþ njg ¼ ln /jg � rg ln /jjg � ln /0 ð8Þ

where /jg is the aggregate market share of jam j, /j|g is the withingroup market share of jam j, and rg is the degree of correlationamong jams that belong to group g.

Parameters bs and qs can be estimated by rearranging Eq. (8)and obtaining the following linear expression:

ln /jg � ln /0 ¼ Xjgbþ rg ln /jjg þ njg ð9Þ

Assuming that jams are divided in two nests according to organic(diet) attribute, we can constrain the correlation parameter rg tobe the same inside the organic (diet) or non-organic (non-diet)group, or we can allow this coefficient to differ in the two nests.In the former case, we have a restricted estimation; while in the lat-ter case an unrestricted one.25

the aggregate market shares. The aggregate market share of jam j thus becomes :

u0 ¼ 1Xn

k

expðdkÞ þ expðd0Þ !,

25 See Bresnahan et al. (1997), Verboven (1996) for examples of unrestrictedestimation.

2829

3031

3233

Num

ber o

f org

anc

jam

s

May20

02

Septem

ber20

02

Februa

ry200

3

July2

003

Decem

ber20

03

April20

04

time

Source: Author’s computation from IRI data

Fig. 4. Number of organic jams in the sample over time.

T. Razzolini / Food Policy 43 (2013) 1–13 7

The transformation of Eq. (5) yields the following linear equa-tion for the three-level Nested logit case (see Verboven 1996):

djsg ¼ Xjsgbþ njsg ¼ ln /jsg � rs ln /jjsg � rg ln /sjg � ln /0 ð10Þ

where /jsg is the aggregate market share of jam j, /j|s is the withinsubgroup market share of jam j, /s|g is the share of subgroup s insidegroup g, rg is the degree of correlation among jams that belong togroup g, and rs is the degree of correlation among jams that belongto subgroup s.

As for the two-level NML, by rearranging Eq. (10) we obtain thefollowing expression:

ln /jsg � ln /0 ¼ Xjsgbþ njsg þ rs ln /jjsg þ rg ln /sjg0ð11Þ

Razzolini (2009) shows that also for PD GEV models it is possible tocompute the mean evaluation parameter dj as a function of marketshares, outside option market share and conditional market shares.The result is the following expression:

dj ¼ ln /j � ln /0 � ln ao/1�qojjM þ ad/

1�qdjjN

� �M ¼ o; no;N

¼ d;nd ð12Þ

By rearranging Eq. (12) we obtain the following non-linearequation:

ln /j � ln /0 ¼ ln ao/1�qojjO þ ad/

1�qdjjD

� �þ Xjbþ nj ð13Þ

26 Manufacturers usually stipulate long-term contracts with distributors and theydo not control promotional discounts at the retailer level.

27 For a similar interpretation see Bresnahan et al. (1997).

Identification and estimation procedure

Estimation of the demand parameters needs to take into ac-count the presence of endogeneity. Some factors observed by con-sumers and producers but not by an econometrician might raisethe demand for a product with higher unobserved quality. This,of course, influences positively market shares and prices creatinga spurious correlation. Failing to control for that will result inbiased and inconsistent estimates. Identification requires that thisproblem of endogeneity (see, for instance, Berry, 1994; Nevo, 2001)must be solved by the use of instrumental variables and that thenumber of instruments must be at least equal to the number ofendogenous regressors. In our case, the endogenous regressorsare prices and within group (and subgroup) markets shares, sincethe latter are determined by the inclusive values (which containprices). Theory suggests using the following instruments: the num-ber of rival products in the same cluster, the number of productssold by the same firm in the same cluster, the sum of the charac-teristics of a rival and own products in the same cluster, etc. Thesevariables are exogenous and affect price through the Nash–Ber-trand equilibrium condition (see Berry, 1994; Nevo, 2000, 2001;

Verboven, 1996). In the Appendix, we describe the supply side ofthe structural model and how the pricing decision is affected,through the variables included in /j and o /j/ o pj, by the numberand characteristics of jams sold by the same firms and by rivalmanufacturers.

As mentioned in the introduction, the dynamics in the numberof jams over time provides a good source of variation to identifychanges in prices due to different equilibrium conditions. We alsoconsider promotional effort as exogenous and we use it as a regres-sor in our demand equation. Most of promotional discounts onjams are decided by retailers (i.e. supermarkets) on the basis oftheir fidelization program. Promotional discount can be seen asan independent choice of retailers that is not related to unobservedfactors that may affect simultaneously consumers’ purchases andbuyers’ pricing decisions.26 Usually unobserved (by the econometri-cian) promotional effort has been considered in the literature as asignificant portion of the unobserved utility.

Equations (9) and (11) provide linear expressions of demandparameters and can be estimated by standard GMM procedure.Equation (13) is estimated with a non-linear GMM procedure.The estimation algorithm is explained in details in the Appendix.

Empirical results

Table 4 shows the GMM estimates of Eq. (8) when the discretechoice model at the individual level is assumed to be a two-levelnested logit with the organic attribute as the unique dimensionof differentiation. Columns 1–5 differ slightly in the use of regres-sors and instrumental variables. The price coefficient is alwaysnegative and significantly different from zero.

The dummy for organic attribute is negative and significant.Since the logarithm of the jams’ market share in the correspondingorganic (or non-organic) group captures the correlation in the errorterm (see Eq. (9)), the dummy for organic nature can be interpretedas a negative demand shifter27: a negative sign of this variablemeans that consumers interested in organic feature constitute asmall proportion of the overall jam buyers. The dummy for diet attri-bute is negative and significant in columns 1–2 but not significantlydifferent from zero in columns 3–5.

The correlation coefficient for organic jams, ro, and the correla-tion coefficient for traditional jams, rNo, are both significant buttake different values. rNo is very close to unity and represents acase of quasi-perfect correlation. Traditional (i.e. non-organic) jams

Table 4Estimates GMM estimates of Eq. (9). The underlying discrete choice model is an unrestricted two level nested logit according to organic attribute.

Dependent variable ln (/it) � ln (/0) (1) (2) (3) (4) (5)

Price �.037*** �.029** �.004 �.004* �.004*

(.014) (.013) (.003) (.002) (.002)rO .415** .396** .465** .456*** .340***

(.189) (.188) (.195) (.196) (.212)rNO .970*** .977*** .995*** .995*** .995***

(.010) (.008) (.002) (.002) (.002)Organic �5.781*** �6.013*** �6.716*** �6.78*** �7.302***

(1.247) (1.27) (1.77) (1.782) (1.928)Diet �.076* �.117** .014 0.58 .009

(.044) (.055) (.014) (0.559) (.017)Promo .473*** .467*** – – –

(.153) (.148)Promoorg – – 1.571*** 1.599*** 1.764***

(.590) (.593) (.640)Promonon-org – – .016** .0158*** .0150**

(006) (.006) (.006)Fruit �.066*** �.074*** .001 .0004*** .0004***

(.024) (.026) (003) (.0034) (.0034)Mono �.089*** �.150*** .008 .010*** .0098**

(.035) (.053) (005) (.005) (.0054)NO .111*** .125*** .082*** .083*** .091***

(.035) (.038) (029) (.029) (.031)NNO .0001 .002* �.0002* �.0001 �.0001

(.0006) (.0008) (.0001) (.0002) (.0002)G4 �.128*** �.117*** �.005 – –

(.046) (.042) (.007)Constant �2.975*** �3.506*** �2.676*** �2.683*** �2.681***

(.138) (.303) (.030) (.0273) (.028)

p-Value F test time dummies 0.00 0.000 0.00 0.00 0.00p-Value F test fruit flavor dummies 0.89 0.90 0.90 0.90 0.90p-Value Hansen J statistics 0.20 0.243 0.343 0.415 0.355

Note: Standard errors are in parenthesis.rO Correlation coefficient for alternatives in the organic subgroup.rNO Correlation coefficient for alternatives in the non-organic subgroup.NO = number of jams sold by the same firm in the organic group.NNO = number of jams sold by the same firm in the non-organic group.*** 1% Significance level.** 5% Significance level.* 10% Significance level.

28 When we exclude one of the ‘‘healthy’’ attributes from the dimensions ofdifferentiation of NML, the corresponding dummy becomes positive or in general notsignificantly different from zero (see for example diet attribute in the estimates ofTable 4 or organic attribute in the estimates of Table 5). This happens because theinclusion of the within group market shares /j|g captures the correlation in theunobserved part of utility so that the dummy simply represents a negative shift in theaggregate demand.

8 T. Razzolini / Food Policy 43 (2013) 1–13

are, thus, perceived as quasi-perfect substitutes. The correlationcoefficient for organic jams is instead much lower and ranges from0.40 to 0.45. Thus, correlation in the error term between organicjams exists but is not particularly strong. This fact has a straight-forward implication for substitution patterns: any change that de-creases the mean utility of an organic product induces consumersto substitute ‘‘less’’ towards other organic jams.

The results for variables fruit and mono are less clear-cut. Esti-mated coefficients for these coefficients are negative and signifi-cant in columns 1 and 2, but are positive and significant incolumns 4 and 5. Surprisingly, in column 1 and 2 the G4 variableis negative and significant, even if packages heavier than 400 rep-resent the majority of jams sold.

The dummy for promotional discount is restricted to have thesame value for both organic and non-organic jams in columns 1and 2 and is always positive and significant. In columns 3–5, thepromotional discount dummy is split for the organic and for thenon-organic groups. The coefficient on promotional discount ishigher for the group of organic jams than for the non-organic ones.Thus, promotional discount seems to have a differential impact forthe organic jams. Discounts are much more effective in encourag-ing consumption of organic jams than the demand for traditionalones. All the specifications in Table 6 include time dummies andflavor dummies. The former are always significant, while the latterare never significant.

It is worth noting, that the Hansen test and the C test for over-identifying restrictions prove to be useful guides for the choice ofthe instrumental variables. Although jams sold by the same firm

in the same group are successfully used as instruments in theabove mentioned literature, both tests reject the use of these vari-ables. Since a possible reason of the rejection of null hypothesis ofthe C test is the exclusion of these variables from the regression,we also include among the Xs the number of jams sold in the sameorganic and non-organic group by the same firm (NOrg NNorg). Inter-estingly, the effect of the number of jams sold by the same firm isalways positive and significant for firms involved in the productionof organic jams, while this is not always the case for firms produc-ing traditional jams. This fact suggests that there is a positiveexternality on consumers’ preferences, i.e. consumers trust morefirms that produce a wider range of organic jams.

Table 5 shows GMM estimates of Eq. (8) when the underlyingdiscrete choice model is assumed to be a two-level nested logitwith diet attribute as the unique dimension of differentiation. Pricecoefficients are, with the exception of column 2, negative and sig-nificant. The coefficient of the organic dummy is positive and sig-nificant, whereas the coefficient of the diet attribute is alwaysnegative and significant.28 Similarly to the results for organic andnon-organic groups, the estimates of the correlation coefficient for

Table 5GMM estimates of Eq. (9). The underlying discrete choice model is an unrestricted two level nested logit according to diet attribute.

Dependent variable ln (/it) � ln (/0) (1) (2) (3) (4)

Price �.003** .003 �0.003*** �0.003***

(.002) (.003) (0.000) (0.001)rD .896*** .838*** 0.856*** 0.858***

(.0108) (.022) (0.018) (0.018)rND .994*** .995*** 0.997*** 0.997***

(.001) (.001) (0.001) (0.001)Organic .0258*** .020** 0.033*** 0.037***

(.008) (.009) (0.007) (0.003)Diet �3.347*** �4.107*** �3.828*** �3.802***

(.140) (.2852) (0.234) (0.232)Promo 0.466*** – – –

(.0073)Promodiet – .438*** 0.396*** 0.388***

(.073) (0.009) (0.0635)Promonon-diet – .023*** 0.010*** 0.010***

(.005) (0.003) (0.003)Fruit �.005** – �0.0002 –

(.002) (0.0014)Mono �.0041 – 0.006** –

(.003) (0.003)ND .064*** .069*** 0.0626*** 0.063***

(.005) (.007) (0.006) (0.006)NND �.000 �.000 �0.000 0.000

(.000) (.000) (0.000) (0.000)G4 .009** .008*** – –

(.004) (.006)ROrg .033*** .035*** .034*** .034***

(.002) (.003) (.003) (.003)RNorg .002*** .00011 .005*** .006***

(.001) (.002) (.0009) (.000)Constant �2.818*** �2.788*** �2.790*** �2.780***

(.027) (.031) (0.022) (0.022)

p-Value F test time dummies 0.00 0.000 0.00 0.00p-Value F test fruit flavor dummies 0.99 0.99 0.99 0.99p-Value Hansen J statistics 0.169 0.733 0.680 0.639

Note: Standard errors are in parenthesis.rD Correlation coefficient for alternatives in the organic subgroup.rND Correlation coefficient for alternatives in the non-organic subgroup.ND = number of jams sold by the same firm in the diet group.NND = number of jams sold by the same firm in the diet of non-organic jams.RD,org = number of rival organic jams sold in the diet group.RND,org = number of rival non-organic jams sold in the non-diet group.⁄ 10% Significance level.** 5% Significance level.*** 1% Significance level.

T. Razzolini / Food Policy 43 (2013) 1–13 9

diet group, rD, are significantly smaller than the correlation coeffi-cients for the non-diet group, rND. The overidentifying restrictionstests indicate that also in this case the number of products sold bythe same firm in the same group (variables Ndiet Nnondiet) have tobe considered as additional regressors and not as instrumental vari-ables. As before, this suggests the existence of positive externality.Consumers prefer to buy jams produced by a firm that is more in-volved in the production of diet ones. Moreover, overidentifyingrestrictions also induce to include the number of rival organic jamsin the group (variables RD,Org RND,org) as an additional regressor.These variables have a positive and significant effect. However, whenwe exclude the organic attribute from the dimension of differentia-tion of the Nested multinomial logit, the number of rival organicjams is the unique measure of the availability of organic products,29

and the fact that these variables are significant suggests that theexclusion of the organic dimension may lead to a misspecificationof the model.

29 It is worth noting that in the specification with organic feature as a uniquedimension of differentiation the number of rival dietetic jams sold in the group wasnot significant.

The dummy for promotional activity is positive and significant.In columns 2–4, we split promotional dummy and the coefficienton promotional activity for diet jams becomes higher than thesame coefficient for the non-diet ones. Coefficients on dummiesfor high content of fruits and mono-dose jams have values closeto zero.

Table 6 shows the estimated coefficients from a restrictedthree-level nested logit model with the diet feature at the topand the organic one at the bottom.30 The reverse order of nestingyields estimated correlation coefficients that do not satisfy McFad-den’s inequality conditions.31

The estimates are robust to overidentifying restrictions test. Thedummy for organic product is always significant and positive. Thesign of the dummy for diet jams is not robust to different specifi-cations. Its coefficient is significant only in column 3, where thesign is negative. The specification in column 2 contains the numberof jams sold by the same firm in the group, NF, which has a smallpositive and significant effect on the mean utility. Heterogeneity

30 No unrestricted version satisfies McFadden’s conditions. Estimates are availableupon request.

31 These estimates are available upon request.

Table 6GMM estimates of Eq. (11). The underlying discrete choice model is a restricted three levels nested logit according to diet at the top and organic at the bottom.

Dependent variable ln (/it) � ln (/0) (1) (2) (3) (4)

Price �0.274** �0.308** �0.139** �0.579⁄

(0.131) (0.140) (0.069) (0.327)rs 0.703*** 0.583*** 0.790*** 0.609***

(0.144) (0.191) (0.104) (0.235)rg 0.645*** 0.511** 0.760*** 0.501⁄

(0.171) (0.220) (0.119) (0.290)Organic 0.295⁄ 0.324** 0.116⁄ 0.769⁄

(0.153) (0.150) (0.064) (0.463)Diet �0.300 0.252 �0.892** 0.772

(0.707) (0.937) (0.449) (1.42)Promo 0.711** 1.019** 0.547** 0.764

(0.360) (0.470) (0.273) (0.549)Mono �0.161 �0.064 �0.017 �0.600

(0.125) (0.064) (0.051) (0.419)Fruit 0.000 0.064 0.027 �0.123

(0.992) (0.040) (0.027) (0.113)G4 �0.950** �0.872** �0.424** �2.242⁄

(0.468) (0.402) (0.216) (1.305)Constant �3.08*** �4.95*** �3.491*** �1.104

(0.625) (1.08) (0.443) (1.85)NF – 0.006** – –

(0.003)

p-Value F test time dummies 0.00 0.000 0.00 0.00p-Value F test fruit flavor dummies 0.99 0.99 1 0.99p-Values Hansen J statistic 0.274 0.274 0.113 0.503

Note: Standard errors are in parenthesis.NF = number of the jams sold by the same firms inside the group.rg Correlation coefficient for alternatives in the group.rs Correlation coefficient for alternatives in the subgroup.*** 1% Significant.** 5% Significant.

32 These two results imply that consumers of a traditional jam will switch toanother traditional jam if the utility of their previous choice decreases.

33 Hall and Pelletier (2007) provide a recent discussion on Non-nested specificationtests in GMM models. Uncertainties on finite sample properties and on the optimalchoice of the weighting matrix raise doubts on the adoption of these tests.

10 T. Razzolini / Food Policy 43 (2013) 1–13

at the top, according to the diet dimension seems to be higher thanheterogeneity among jams in the subgroup determined by the or-ganic feature. This means that the attribute associated with thehigher level of correlation in the unobserved part of the utility isthe organic one.

The first column in Table 7 shows estimates from the restrictedPD GEV model. The heterogeneity parameters are both positive,but only the organic one is significantly different from zero. Thecoefficient equals to 0.29 and indicates that between products thatshare the same organic status there is a high level of correlation,but not perfect substitution, since the coefficient is far from zero.

The number of products sold by the same manufacturer (NM) inthe same group (organic or not) is positive and significant. Thisspecification, however, has two pitfalls: Hansen test rejects thevalidity of the instruments and the price coefficient is positiveand significant. The unrestricted PD GEV specification thus doesnot provide a flexible enough structure to fit the data.

The second column of table 7 contains the estimates from theunrestricted PD GEV model. Price coefficient is negative and signif-icantly different from zero at the 10% level. The coefficient of thedummy for promotional discount is positive and significant. Thedummies for organic and diet features are both negative and signif-icant, but the former is much more negative than the latter. Sincethese variables are interpreted as demand shifters, it suggests thatthe fraction of consumers of organic products is much lower thanthe fraction of consumers of diet jams, and both represent a smallpart of the overall population. It is worth noting, that this is one ofthe few specifications that yields significant estimates for the fla-vor dummies. In this specification, therefore, different flavors playa role in shifting demand. As above, the number of organic prod-ucts sold by the same firm (NOrg) is positive and significant onlyfor organic manufacturers. This result thus reveals the importanceof having a larger market share willing to gain a higher level ofconsumers’ trust.

The most important results provided by these specifications arerepresented by heterogeneity parameter estimates. Unrestrictedestimations give precious insights into the way heterogeneity dif-ferentiates along all possible realizations of organic and diet sta-tuses. The heterogeneity parameter for organic jams is positiveand significant. Its value is equal to 0.38, and suggests the exis-tence of insulation from competition (i.e. correlation due to the or-ganic attribute), even if not complete as in the case of perfectsubstitution. The heterogeneity parameter for diet jams is equalto 0.97, and is very close to unity. Thus, the insulation from com-petition given by the diet feature is almost non-existent. The het-erogeneity parameter is negative for the non-organic group andpositive for the non-diet one, but both parameters are not signifi-cantly different from zero.32 Therefore, the hypothesis of perfectsubstitution inside the non-organic and non-diet groups cannot berejected.

Although we cannot perform a test to identify the correct spec-ification,33 the unrestricted PD GEV model seems to be the one thatfits the data the best and yields the lowest value of the GMM objec-tive function with a p-value of Hansen statistics of 0.828. Further-more, the results from the unrestricted PD GEV model giveadditional and more accurate information than the estimates ofthe restricted PD GEV and restricted three-level nested logit specifi-cations. Both three level NML and PD GEV models indicate that theorganic characteristic is more relevant than the diet one. The PDGEV estimation, in particular, reveals that having a diet attributeper se does not protect from competition, whereas non-diet jamsare perceived as perfect substitutes.

Table 7GMM estimates of Eq. (13). The underlying discrete choice model is a restrictedPDGEV in column 1 and unrestricted PD GEV in column 2.

Dependent variable ln (/it) � ln (/0) (1) Restricted (2) Unrestricted

Price 0.234*** �0.046*

(0.026) (0.0247)qR

O0.294*** –(0.012)

qRD

0.034 –(0.15)

qO – 0.384**

(0.140)qNO – �0.442

(0.29)qD – 0.975***

(0.021)qND – 0.069

(0.12)Organic �0.99*** �4.60***

(0.21) (0.24)Diet �0.88*** �1.78***

(0.24) (0.42)Promo 0.95*** 0.51***

(0.074) (0.058)Mono �0.99*** �0.085

(0.21) (0.046)Fruit �0.88*** �0.050

(0.24) (0.043)NM 0.003** �

(0.001)NOrg – 0.136***

(0.026)NNorg – 0.001

(0.001)G4 0.594*** �0.144***

(0.048) (0.036)Constant �6.69*** �3.10***

(0.49) (0.45)

p-Value F test time dummies 0.0000 0.0000p-Value F test type dummies 0.0000 0.0000p-Values Hansen J statistic 0.034 0.828

Note: Standard errors are in parenthesis.NM = number of organic jams sold by the same manufacturer (restrictedspecification).NOrg = number of organic jams sold by the same firm (unrestricted specification).NNorg = number of non-organic y the same firm (unrestricted specification).qR

O;qRD: Restricted heterogeneity parameters (organic and non-organic, diet and

non-diet).qO, qNO, qD, qND: Unrestricted heterogeneity parameters (organic, non-organic, diet,non-diet).*** 1% Significance level.** 5% Significance level.* 10% Significance level.

T. Razzolini / Food Policy 43 (2013) 1–13 11

Overall, organic and non-organic statuses guarantee insulationfrom competition although they differ in the measure of inducedmarket power. A result which is confirmed both by two-levelNML model with the organic feature as a unique dimension of dif-ferentiation and by the PD GEV model is that the degree of insula-tion provided by organic feature is not complete. In contrast,traditional jams are perceived as perfect substitutes betweenthemselves.

Conclusion

This paper has analyzed the demand for organic jams in the Ital-ian market over a period of initial diffusion followed by a slow-down and has highlighted some barriers to the consumption ofthese products. Our results indicate that, although the organic logoguarantees a positive insulation from competition, the magnitudeof this effect is not particularly high, at least in the period underobservation. This result probably reflects the fact that the Italianmarket of organic products, although rapidly growing during the

late 1990s and the beginning of 2000s, was not as mature as otherEuropean markets, and that Italian consumers in a period of crisis,were still perceiving the price premia of organic jams as excessive.The existence of a price barrier has been confirmed by the successof several promotional discounts, experimented by retailers in2002–2004, that stimulated the demand for organic jams. Otheradditional characteristics, such as the diet attribute offered bythe majority of organic manufacturers were proved not to be suc-cessful in providing additional protection from competition withconventional products.

Our findings also indicate that the number of organic productssold by the same manufacturers is a key feature of the demandfor organic jams. This result, obtained from the analysis of realdata, has relevant policy implications, since it reveals that the ac-cess to the market through mainstream retail channels may be lim-ited for small organic manufacturers (Torjusen et al., 2004; Wieret al., 2008). Small organic producers may fail to meet the require-ments of supermarkets, such as prompt availability of products,quick delivery and, especially, the large volumes necessary to re-duce prices. Moreover, the typical scale of productions in organicfarming may not be sufficient to cover the slotting fees necessaryto obtain shelf space in the supermarkets. All these limits togetherwith the evident role of variety and brand familiarity in the organicsegment of the market are consistent with the recent sales trendsof organic products and the recent offer by the main Italian retail-ers of organic products under their own private labels. As discussedby Jonas and Roosen (2005) and Grunert (2005), private labels havegreater chances to gain consumers’ trust using the supermarketbrand and to reduce prices thanks to the larger economies of scalein the distribution and to retailers’ ability to buy at lower prices,especially in agro-food industry. However, this imbalance in thesupplier-retailer relation which usually benefits larger suppliersin the entire agro-food industry may be even more penalizing forsmall scale production in organic farming and in the stages of foodprocessing that can be done by the farmers themselves.

European Institutions (EC-COM, 2004; EP et al., 2005) have at-tempted to promote such organizational structures as farmers’cooperatives and consortiums designed to ease group marketingstrategies and access to big distributional channels for small pro-ducers. However, additional interventions are needed in Italy tostimulate grouping of organic producers in order to increase theirbrand and bargaining power as well the ability to provide a greatervariety of products as requested by the market.

Appendix A. Supply side of Berry (1994) model

Let us assume that we have F firms. Each of these firms mayproduce some subset Ff of the j = 1, . . . , J different jams. Assumingconstant marginal costs the profit function of firm f is:

pf ¼Xj2Ff

ðpj �mcjÞN/j � Cf ð1aÞ

where /j is the market share of jam j, N is the total number of con-sumers (i.e. size of the market), mcj are marginal costs, Cf is firm ffixed cost.

Assuming the existence of a Bertrand–Nash equilibrium inprices, the following first order condition for price pj must holdfor any jam j sold by firm f:

@pf

@pj¼ 0() /j þ

Xr2Ff

ðpr �mcrÞ@/r

@pj¼ 0 ð2aÞ

From equation (2a) it is clear that firm f internalizes the effect of achange in price pj on the market shares /r of other jams r (r – j) itproduces. As Eqs. (3)–(6) in the main text show, the market share

12 T. Razzolini / Food Policy 43 (2013) 1–13

of any jam j, is affected by prices of all other jams, included the onesproduced by the same firm.

We can derive from the J first order conditions in Eq. (2a), theprice–cost margin. By defining Sjr = �o /r/ o pj, j, r, = 1, . . . , J,

X�jr ¼1 if 9f : fr; jg � F0 otherwise

�ð3aÞ

and X a J � J matrix with Xjr ¼ X�jr � Sjr , the system of J first orderconditions can be written in vector notation as follows:

/ ¼ Xðp�mcÞ ð4aÞ

where /, p and mc are J � 1 vectors representing market shares,prices and marginal costs, respectively. Price cost margins can thusbe defined as:

p�mc ¼ X�1/ ð5aÞ

In order to provide an intuitive explanation of the effect of correla-tion parameter on mark-up, we show, as in Berry (1994) Eq. (33),the price cost-margin of a firm producing a unique product whenthe discrete choice at the individual level is a two-level nested logit(Eq. (4)):

pj �mc ¼ð1� rgÞ=bp

½1� rg/jjg � ð1� rgÞ/j�ð6aÞ

where bp is the price coefficient. This equation shows that whencorrelation is equal to zero, mark-up of jam j depends only on itsmarket share /j; where correlation coefficient rg equals one, thenprice–cost margin depends uniquely on the within group share,/j|g. A simple derivative of Eq. (6a) with respect to rg, show thatthe mark-up of a jam depends positively on the correlation in theunobserved part of utility among jams belonging to the same group.

Estimation algorithm

To estimate Eq. (13), we apply a GMM estimation suggested byBerry et al.(1995), Bresnahan et al. (1997), Nevo (2000, 2001). Theunique difference with these procedures is that we do not need torecover mean utilities iteratively: Eq. (12) provides in fact a closedform expression of mean utilities. Following the literature, we esti-mate PD GEV demand parameters and heterogeneity parametersexploiting their different nature. The vector of parameters b entersin Eq. (13) in a linear fashion while heterogeneity parameters enternon-linearly.

The procedure can be described as follows:

Step 1: For an initial guess of heterogeneity parameters qs, Eq.(12) gives the vector of mean evaluations d that equalizes pre-dicted market shares with observed market shares.Step 2: Residuals are defined as a difference between d and thevalue predicted by the ‘‘linear’’ regressors. We interact theseresiduals with instruments and construct the GMM objectivefunction as follows:

Jð~hÞ ¼ ngnð~hÞ0S�1gnð~hÞ ð7aÞ

where ~h contains linear b and non-linear q-parameters, and S is aconsistent estimate of E½gig

0i� ¼ E½eiziz0i� obtained with 2SLS estima-

tion. The objective function is thus minimized with respect to thelinear parameters b by exploiting the linearity of these parametersin the first order conditions.

Step 3: We repeat steps 1 and 2 iteratively by implementing aNelder–Mead non-linear procedure that search for the valuesof the heterogeneity parameters that minimize the value ofthe GMM objective function specified in step 2. As suggestedby Nevo (2000) Nelder–Mead simplex procedure is more robust

than the usual standard non-linear method and is less sensitiveto the choice of starting values.

References

Bähr, M., Botschen, M., Laberenz, H., Naspetti, S., Thelen, E., Zanoli, R., 2004. In:Raffaele, Zanoli (Ed.), The European Consumer and Organic Food, vol. 4. Schoolof Business and Management, University of Wales, Aberystwyth.

Bellows, A.C., Onyango, B., Diamond, A., Hallman, W.K., 2008. ‘‘UnderstandingConsumer Interest in Organics: Production Values vs. Purchasing Behavior’’.Journal of Agricultural & Food Industrial Organization 6 (1), 1–28.

Berry, S., 1994. Estimating Discrete Choice Models of Product Differentiation. RANDJournal of Economics 25, 242–262.

Berry, S., Levinsohn, J., Pakes, A., 1995. Automobile prices in market equilibrium.Econometrica, 841–890.

Bresnahan, T., Stern, S., Trajtenberg, M., 1997. Market segmentation and the sourcesof rents from innovation: personal computers in the late 1980s. RAND Journal ofEconomics 28, S17–S44.

Carson, R., Flores, N.E., Martin, K.M., Wright, J.L., 1996. Contingent valuation andrevealed preference methodologies: comparing the estimates for quasi-publicgoods. Land Economics 72, 80–99.

Chen, M.-F., 2009. Attitude toward organic foods among Taiwanese as related tohealth consciousness, environmental attitudes, and the mediating effects of ahealthy lifestyle. British Food Journal 111 (2), 165–178.

Cremonini, P., 2007. Organic and Fair Trade Marketing in Italy. In: Simon, Wright,Diane, McCrea (Eds.), The Handbook of Organic and Fair Trade Food Marketing.Blackwell Publishing, Oxford, pp. 228–238.

Compagnoni, A., 2001. Organic food and farming: towards partnership and action inEurope. Conference Proceeding.

Cummings, R., Harrison, G., Rutstrom, E., 1995. Home-grown values andhypothetical surveys: is the dichotomous choice approach incentivecompatible. American Economic Review 85, 260–266.

De Magistris, T., Gracia, A., 2008. The decision to buy organic food products inSouthern Italy. British Food Journal 110 (9), 929–947.

EC, 1991. Council Regulation (EEC) No 2092/91 of 24 June 1991 on organicproduction of agricultural products and indications referring thereto onagricultural products and foodstuffs. Official Journal of the EuropeanCommunities, L198 (22.07.91), 1–15.

EC, 1999. Council Regulation (EC) No 1804/1999 of 19 July 1999 supplementingRegulation (EEC) No. 2092/91 on organic crop production of agriculturalproducts and indications referring thereto on agricultural products andfoodstuffs to include livestock production. Official Journal of the EuropeanCommunities L222 (24.08.1999), 1–28.

EC-COM, 2004. European Action Plan for Organic Food and Farming. Commission ofthe European Communities [COM(2004) 415 final]. Brussels.

EC, 2007. Council Regulation (EC) No 834/2007 of 28 June 2007 on organicproduction and labelling of organic products and repealing Regulation (EEC) No2092/91. Official Journal of the European Union, L 189 (20.07.2007), 1–23.

EP, 2005. European Parliament resolution on the European Action Plan for OrganicFood and Farming, Official Journal 320 E, 15/12/2005, pp. 0242–0247.

Frykblom, P., 1997. Hypothetical question modes and real willingness to pay.Journal of Environmental Economics and Management 34, 275–287.

Gil, J.M., Gracia, A., Sanchez, M., 2000. Market segmentation and willingness to payfor organic products in Spain. International. Food and AgribusinessManagement Review 3, 207–226.

Grunert, K.G., 2005. Food quality and safety: consumer perception and demand.European Review of Agricultural Economics 32 (3), 369–391.

Hall, A., Pelletier, D., 2007. Non-Nested Testing in Models Estimated via GeneralizedMethod of Moments. Working Paper Series No 11, North Carolina StateUniversity, Department of Economics.

Hansen, J.K., Sorensen, H.C., 1993. The Importance of Price for the Sale of EcologicalProducts. MAPP working paper no. 13, Aarhus School of Business, Aarhus.

Huang, C.L., 1996. Consumer Preferences and Attitudes Toward Organically GrownProduce. European Review of Agricultural Economics 23, 331–342.

Hughner, R.S., McDonagh, P.-., Prothero, A., Shultz, C.J., Stanton, J., 2007. Who areorganic food consumers? A compilation and review of why people purchaseorganic food. Journal of Consumer Behaviour 6, 94–110.

ISMEA, 2005. L’evoluzione del mercato delle produzioni biologiche: l’andamentodell’offerta, July.

ISMEA, 2007. Il mercato dei prodotti biologici: tendenze generali e nelle principalifiliere, December.

ISMEA, 2012. Report: Prodotti Biologici ISMEA, August.Jonas, A., Roosen, J., 2005. Private labels for premium products – the example of

organic food. International Journal of Retail & Distribution Management 33 (8),636–653.

Krystallis, A., Chryssohoidis, G., 2005. Consumer’s willingness to pay for organicfood: factors that affect it and variation per organic product type. British FoodJournal 107 (4/5), 320–323.

Kuhar, A., Juvancic, L., 2010. What determines purchasing behaviour for organic andinte-grated fruits and vegetables? Bulgarian Journal of Agricultural Science 16(2), 111–122.

Makatouni, A., 2002. What motivates consumers to buy organic food in the UK?Results from a qualitative study. British Food Journal 104 (3/4/5), 345–352.

T. Razzolini / Food Policy 43 (2013) 1–13 13

McFadden, D., 1978. Modelling the choice of residential Location. In: Karlquist, A.et al. (Eds.), Spatial Interaction Theory and Residential Location. North-Holland,Amsterdam, pp. 75–96.

Millock, K., Hansen, L.G., Wier, M., Mørch Andersen L., 2002. Willingness to pay fororganic foods: a comparison between survey data and panel data fromDenmark. In: 12-th Annual EAERE Conference, Monterey, June.

Nevo, Aviv., 2000. A Practitioner’s guide to estimation of random coefficients logitmodels of demand. Journal of Economics & Management Strategy 9 (4), 513–554.

Nevo, Aviv., 2001. Measuring market power in the ready-to-eat cereal industry.Econometrica 69 (2), 307–342.

Onyango, B., Hallman, W.K., Bellows, A.C., 2007. Purchasing organic food in US foodsystems: a study of attitudes and practice? British Food Journal 109, 399–411.

Padel, S., Foster, C., 2005. Exploring the gap between attitudes and behavior. BritishFood Journal 107 (8), 606–625.

Pinton, R., 2001. Marketing models: Supermarkets and Direct Supply - Synergie orOpposition. In: Conference Proceeding, 10–11 May 2001, Copenhagen,Denmark.

Razzolini, T., 2009. A note on the estimation of PD GEV models. Applied EconomicsLetters 16 (2), 193–197.

Rete Rurale Nazionale, 2012, Bioreport 2012, Rete Rurale Nazionale 2007–2013,Ministero delle Politiche Agricole Alimentari e Forestali, Roma, Italy.

Santucci, F.M., Pignataro, M., 2002. Organic Farming in Italy. Paper for the OECDWorkshop on organic agriculture. Washington DC, September 23–26, 2002.<http://orgprints.org/2159/1/Santucci_Pignataro_2002_Organic_farming_in_Italy.pdf>.

Schifferstein, H.N.J., Oude Ophuis, P.A.M., 1998. Health-related determinants oforganic food consumption in the Netherlands. Food Quality and Preference 9(3), 119–133.

Thompson, G.D., Kidwell, J., 1998. Explaining the choice of organic produce:cosmetic defects, prices, and consumer preferences. American Journal ofAgricultural Economics 80, 277–287.

Torjusen, H., Sangstad, L., O’Doherty Jensen, K., Kjaernes, U., 2004. Europeanconsumers’ conceptions of organic food: a review of available research. In:Professional Report No. 4, National Institute for Consumer’s Research, Oslo(SIFO), Norway.

Train, Kenneth, 2009. Discrete Choice Methods with Simulation. CambridgeUniversity Press.

Verboven, F., 1996. International price discrimination in the European car market.The RAND Journal of Economics 27 (2), 240–268.

Verhoef, P.C., Franses, P.H., 2002. On combining revealed and stated preferences toforecast costumer behavior: three case Studies. Econometric Institute Report,2002–04. Erasmus University Rotterdam.

Vogl, C., Kilcher, L., Schmidt, H., 2005. Are standards and regulations of organicfarming moving away from small farmers’ knowledge? Journal of SustainableAgriculture 26 (1), 5–26.

Wier, M., Jensen, K.O., Andersen, L.M., Millock, K., 2008. The character of demand inmature organic food markets: Great Britain and Denmark compared. FoodPolicy 33, 406–421.

Yiridoe, E.K., Bonti-Ankomahand, S., Martin, R.C., 2005. Comparison of consumerperceptions and preference toward organic versus conventionally producedfoods: a review and update of the literature. Renewable Agriculture and FoodSystems 20, 193–205.

Yue, C., Alfnes, F., Jensen, H.H., 2009. Discounting spotted apples: investigatingconsumers’ willingness to accept cosmetic damage in an organic product.Journal of Agricultural and Applied Economics 41, 29–46.