II. Literature Review

Wine consumption involves repeated purchases, as it usually happens with routine food products and beverages. The dynamics of purchases have been investigated as an exploratory behavior (Meixner and Knoll, Reference Meixner and Knoll2012), which explains why consumers buy the same brand or product or change them (variety) in repeated purchases (Barharrell and Denison, Reference Barharrell and Denison1995). Accordingly, a consumer's state of dependence occurs when purchases depend on past consumption history (Thomadsen and Seetharaman, Reference Thomadsen, Seetharaman and Verplanken2018). A positive state of dependence is detected as habits (automatic purchase process), inertial purchases (high probability of making the same purchase), loyalty (repurchase of the same brand), or as a positive effect of the last purchase on the current one. Conversely, a negative state of dependence reduces the probability of repurchasing the same product on the next occasion. This approach is named variety-seeking behavior (VSB).

The positive or negative state of dependence is usually linked to specific attributes, such as brand, flavor, size, and quality, rather than the product itself, while heterogeneity is found among consumers’ state of dependence (Thomadsen and Seetharaman, Reference Thomadsen, Seetharaman and Verplanken2018). Studies on food products such as ketchup and yogurt (Adamowicz and Swait, Reference Adamowicz and Swait2012) detected habit and variety-seeking segments.

The reasons behind the state of dependence have been largely investigated in the wine market as loyalty and VSB. Loyalty (Rungie et al., Reference Rungie, Laurent, Mtimet and Jarvis2006) reflects how strong the brand identity is in the mind of consumers who have acquired experience and formed their own habits.

Conversely, the VSB is historically linked to the optimal stimulation level, the arousal theory (McAlister, Reference McAlister1982; Venkatesan, Reference Venkatesan, Ward and Robinson1973), and the concept of satiation. McAlister (Reference McAlister1982) proposed a dynamic model for the attribute satiation, where the consumption history creates an attribute inventory that is depleted (the customer forgets the attribute) or replaced (repeated consumption) over time. More recently, Sevilla, Lu, and Kahn (Reference Sevilla, Lu and Kahn2019) built an extended model explaining that strategies mitigating the satiation depend on consumption features (quantity consumed, time, and frequency of consumption) and perception features such as categorization, attention, memory, and cognition.

A different approach is proposed in Yoon and Kim (Reference Yoon and Kim2018). Their study shows that consumers living in a constrained socio-economic status perceive their life as unsatisfactory. Accordingly, these consumers are inclined to compensate for this life status by choosing among a wide assortment of products, that is, they behave as variety-seeking purchasers (Yoon and Kim, Reference Yoon and Kim2018).

The conspicuous VSB literature allows us to draw a general profile of the variety-seeking customer revolving around the concept of sensation seeking. Generally speaking, VSB consumers prefer products connected to novelty and stimulation (Raju, Reference Raju1980), such as choosing new brands, avoiding routine behaviors while acting differently from time to time. In this framework, high-sensation seeker consumers tend to make impulsive choices and be more attentive to information. Inman (Reference Inman2001) suggests that consumers may change flavors more frequently than brands. The effect of satiation is observed especially in the food domain (Baltas, Kokkinaki, and Loukopoulou, Reference Baltas, Kokkinaki and Loukopoulou2017), where VSB is stimulated in products exhibiting hedonic attributes (Van Trijp, Hoyer, and Inman, Reference Van Trijp, Hoyer and Inman1996). Sensation seeking can be interpreted as a compensatory behavior in economically stuck consumers reacting to their unsatisfactory life conditions (Yoon and Kim, Reference Yoon and Kim2018).

When looking at extreme variety-seekers, referred to as switchers (Knox and Walker, Reference Knox and Walker2001), brand commitment lowers, and the desire to switch brands may be easily triggered by factors such as promotional prices or fancier packaging (Olsen et al., Reference Olsen, Atkin, Thach and Cuellar2015). On the other hand, it is important to consider context-related factors such as the wide product assortment offered in supermarkets, which can encourage switching behaviors (Diehl and Poynor, Reference Diehl and Poynor2010), or price discounting campaigns that attract customers and lock them in for their subsequent purchases (Thomadsen and Seetharaman, Reference Thomadsen, Seetharaman and Verplanken2018).

In the wine domain, Olsen et al. (Reference Olsen, Atkin, Thach and Cuellar2015) reported a literature review where VSB in wine is modeled using consumer features, product involvement, and wine knowledge. For instance, differences in VSB were found among age, income, and education level (Dodd, Pinkleton, and Gustafson, Reference Dodd, Pinkleton and Gustafson1996) or different risk-taking segments with older consumers likely to choose new wines (Mueller, Remaud, and Chabin, Reference Mueller, Remaud and Chabin2011). Personality traits can also affect wine and beer purchases (Gustavsen and Rickertsen, Reference Gustavsen and Rickertsen2019). VSB is often detected in innovators, early adopters, and risk-adverse consumers (Campbell and Goodstein, Reference Campbell and Goodstein2001). Personal involvement is higher in variety-seeking consumers since they are more aware of grape varietals (Madureira and Simoes de Sousa Nunes, Reference Madureira, de Sousa Nunes and J2013) or environmental aspects. Therefore, they switch the regions of their wine purchases (Dodd, Pinkleton, and Gustafson, Reference Dodd, Pinkleton and Gustafson1996), or they choose among sustainable wines (Kelley et al., Reference Kelley, Zelinskie, Centinari, Gardner, Govindasamy, Hyde, Rickard and Storchmann2017). The VSB may arise from an imperfect knowledge of wine consumers. Shoppers are often confused when selecting wines and can barely remember their last wine purchase (Hussain, Cholette, and Castaldi, Reference Hussain, Cholette and Castaldi2007). Following Dodd, Pinkleton, and Gustafson (Reference Dodd, Pinkleton and Gustafson1996), wine experts show a high VSB because of their better information level; conversely, Viot (Reference Viot2012) reports that subjective knowledge does not influence the number of wines chosen.

More recently, the VSB has been detected in consumers who prefer to taste different wines and therefore move across market segments because they know one or more wine attributes such as grape variety, origin, brand (Ellis and Caruana, Reference Ellis and Caruana2018) or, conversely, because they are not well informed about wines (Ellis and Thompson, Reference Ellis and Thompson2018). Moreover, a vast wine assortment encourages consumers to look for variety while eroding their loyalty to brands, especially in products such a wine having a strong hedonic component (Inman, Reference Inman2001; Ellis and Thompson, Reference Ellis and Thompson2018).

So far, the literature on VSB for wine has not paid attention to wine types. Specifically, the consumption of sparkling wines is affected by motivations that differ with respect to still wines. Charters et al. (Reference Charters, Velikova, Ritchie, Thach, Dodd, Fish, Herbst and Terblanche2011) thoroughly analyzed Champagne consumption, which carries symbolic motivations as these wines are often consumed on special occasions like celebrations, holidays, and in specific moments of the day, depending on the cultural context. Bubbles represent the distinctive trait of these wines, and they are often associated with positive emotions and dynamism. Conversely, the peculiarities of Champagne make its consumption differ with respect to “regular” sparkling wines that constitute a separate product category (Thach and Olsen, Reference Thach and Olsen2006).

The Italian literature on sparkling wines has increased recently due to an increase in the production of sparkling wines. Thiene et al. (Reference Thiene, Scarpa, Galletto and Boatto2013) pointed out consumer patterns and different willingness to pay (WTP) for Prosecco appellations by segmenting the wine market into groups using a latent class regression model on survey data. Onofri, Boatto, and Dal Bianco (Reference Onofri, Boatto and Dal Bianco2015) detected a segmentation among consumers showing different preferences for the two Prosecco denominations. Experimental research done by Vecchio et al. (Reference Vecchio, Lisanti, Caracciolo, Cembalo, Gambuti, Moio, Siani, Marotta, Nazzaro and Piombino2018) showed how consumers perceived the sparkling production process and brand differently (Charmat vs. Champenoise). Rossetto and Gastaldello (Reference Rossetto and Gastaldello2018) evaluated the different loyalty and VSB when purchasing sparkling wines in supermarkets. Using a model based on a Dirichlet distribution, they detected high loyalty toward Prosecco appellations and moderate VSB.

In the following, we analyze the dynamics of purchases of sparkling wines by using a different approach. We apply a multivariate HMM (Bartolucci, Farcomeni, and Pennoni, Reference Bartolucci, Farcomeni and Pennoni2013) to identify consumer types according to the dynamic patterns and detect possible effects of the customers’ socio-demographic characteristics. The proposed model accounts for a sequence of purchased brands and enables us to evaluate consumer loyalty while segmenting the purchase base, and it gives stores a method to manage long-term buying through suitable marketing strategies.

III. Data and Methods

B. Data

Data are provided by an international agency (Nielsen, Reference Nielsen2017) from a panel of 9,000 Italian households who recorded their purchases through a scan method over the years 2015–2016. This sample represents the Italian population regarding the area of residence, family size, income, age of the purchaser, and family life stage. The survey collected purchases every month; households may therefore perform multiple purchases in a two-year time horizon. We have a total of 22,362 purchases of sparkling wines made in non-specialized grocery stores (hypermarkets, supermarkets, minimarkets, and discounts) at our disposal. The data refer to the portion of the entire panel of 5,155 households that made at least one purchase in the reference period. Data contain detailed information on purchases: time and place of purchase, intrinsic and extrinsic wine attributes, such as the appellation, label, brand, producer location, and price (Bassi, Pennoni, and Rossetto, Reference Bassi, Pennoni and Rossetto2020).

Families make from 1 to 230 purchases in the reference period: 34.0% of households bought sparkling wine only once, 19.8% twice, 11.1% three times, only a small portion of the sample, 10%, made more than 10 purchases in the two years. Table 1 shows that the average expense per purchase in the whole period was approximately seven euros. The observed quantiles of price show five ordered levels of the categorical response.

Table 1

Descriptive Statistics of the Response Variables

The second response variable, namely the type of wine, exhibits that 33.5% of the observed purchases are of sweet wines and 35.3% of brut sparkling wines, while extra dry purchases are 19.2%. The share of no appellation sparkling wine purchases is 44.5%, while the Prosecco appellations represent 30% of the sampled purchases. High-price classic method appellations (Franciacorta, Trento, Oltrepo) represent 12% of the observed purchases.

Regarding the sample socio-economic characteristics, we used the categorical variables defined within Nielsen's methodology (Nielsen, Reference Nielsen2017) and reported in Table 2, along with some descriptive statistics. They include the Italian region of residence, the purchaser's age, the family income categorized as very low, below the average, above the average, and high according to the concentration curve of the per capita income, the number of family components, and the family life stages (defined according to the parents and children ages, see the footnote of Table 2). We observe homogeneous purchase distributions among all Italian regions; only Lombardy exhibits more purchases than elsewhere.

Table 2

Descriptive Statistics of the Covariates

Notes: Families are classified in life stages (age and children) as follows: “pre family” are singles or couples with no children, and the purchaser is younger than 35; “new” are families with children younger than 6; “maturing” are families that have children under 17 but not all of them are very young nor are any of them between 11 and 17; “established” are families with children between 11 and 17; “post” are families that are singles or couples (35–54) and do not have children under 18; “older couples” are purchasers that are over 55 and have no children under 18; finally, “older singles” who are 55 and older.

C. Methods

We employed the multivariate categorical response variables jointly to provide a dynamic overview of the purchase process over time. Homogeneous clusters of consumers are discovered according to the socio-economic features of the households. Specifically, we estimate a multivariate HMM (Bartolucci, Farcomeni, and Pennoni, Reference Bartolucci, Farcomeni and Pennoni2013, Reference Bartolucci, Farcomeni and Pennoni2014) with covariates influencing the latent model to investigate their effects on the purchase dynamics according to the responses shown in Table 2. The model provides the estimated probabilities of consumers being allocated in each segment at the beginning of the period (initial probabilities) and their movements across market segments over time (transition probabilities). It also allows us to disentangle the effects of socio-economic variables on these probabilities. The main assumption is that the latent process fully explains the observed consumer behavior, and the time-fixed and time-varying customer socio-demographic characteristics describe the dynamics of the underlying latent preferences.

The model is an extension of the latent class model employed with cross-sectional data (Clogg and Goodman, Reference Clogg and Goodman1984; Pennoni, Reference Pennoni2014). Recent inferential and computational developments of the HMM improved the model performance in analyzing time series and panel data in many fields of application (see Bartolucci, Farcomeni, and Pennoni (Reference Bartolucci, Farcomeni and Pennoni2013, ch. 3) for a literature review and some illustrative examples).

In the following, we consider the multivariate HMM based on a first-order Markov chain with homogeneous transition probabilities and covariates affecting the latent process while accounting for the unbalanced panel. The response variable is denoted as Y _ijt, that is, the customer's response i, i = 1, …, n, for the variable j, j = 1, …, r having c _j categories measured at time t, t = 1, …, T _i. It is worth noting that the number of purchases is customer-specific. The response vector is denoted as Y_it = (Y _i1t, …, Y _irt), and we assume that the distribution of the responses depends on a discrete latent variable assumed with a finite number of states ranging from 1, …, k, and denoted as U_i = (U _i1, …, U _iT). The components identified by this latent variable correspond to subpopulations of customers showing different purchasing behavior. We assume that the three responses on the same occasion, Y_it, are conditionally independent, given the latent process U _it, and the vectors of responses Y_i1, …., Y_iT are conditionally independent, given the latent process U_i. The conditional distribution of each response variable is denoted as

(1)$$P( {Y_{ijt\;} = y\;\vert \;U_{it} = u} ) = \phi _{\,jy\vert u}.$$

We account for the influence of socio-demographic features on the latent process by assuming that these affect customers’ latent preferences. Let x_it denote the vector of individual covariates available at the t time occasion for customer i. Each covariate contributes to the initial customer's allocation at the first purchase occasion across segments and to customers’ switch among segments after the first purchase. For this purpose, we model the initial and the transition probabilities as customer-specific through suitable multinomial logit models. Regarding the initial probabilities, we assume the following base-line category logit model:

(2)$$log\displaystyle{{\,p( {U_{it} = u} ) } \over {\,p( U_{it} = 1) }} = \beta _{0u} + {\boldsymbol x}^{\prime}{\boldsymbol \beta }_{1{\boldsymbol u}}, \quad u \ge 2, \;\;$$

where β _0u is an intercept specific for the latent state and ${\boldsymbol {\beta }^{\prime}}_{1{\boldsymbol u}}$ is a column vector of parameters measuring the effect of the covariates.

The transition probabilities of the HMM are modeled as

(3)$$log\displaystyle{{\,p( U_{it} = u{\rm \vert }U_{it} = \bar{u}) } \over {\,p( U_{it} = 1\vert U_{it-1} = \bar{u}) }} = \gamma _{0\bar{u}u} + {\boldsymbol x}^{\prime}{{\boldsymbol \gamma }_{1\bar{u}u}}\;t \ge 2, \quad\bar{u}\ne u, \;$$

where $\gamma _{0\bar{u}u}{\rm \;}$is the intercept and ${\boldsymbol \gamma }_{1\bar{u}u}$ is the vector of parameters representing the effects of the covariates on the transition probability from state $\bar{u}$ to state u.

Given a sample of independent customers, the log-likelihood of the model has been maximized through the Expectation-Maximization (EM) algorithm (Dempster, Laird, and Rubin, Reference Dempster, Laird and Rubin1977). The EM algorithm requires that the parameters be properly initialized, and additional details on how this is performed as well as on the steps of this algorithm can be found in Chapter 5 of Bartolucci, Farcomeni, and Pennoni (Reference Bartolucci, Farcomeni and Pennoni2013). The appropriate number of states that identify the customer segments is selected by using the Bayesian Information Criterion (BIC), which is a measure of the relative goodness of fit accounting for the sample size. Bacci, Pandolfi, and Pennoni (Reference Bacci, Pandolfi and Pennoni2014) compared the performance of different information criteria, and they showed that most criteria perform reasonably well in choosing the number of appropriate latent states; however, the BIC is preferred for its parsimony. Standard errors of the parameters are used as a measure of accuracy, especially for the covariate's effects. They are obtained as the squared root of the elements on the main diagonal of the inverse observed or expected information matrix computed at the maximum likelihood estimate (Bartolucci, Farcomeni, and Pennoni, Reference Bartolucci, Farcomeni and Pennoni2013). Suitable functions adapted from the R package LMest (Bartolucci, Pandolfi, and Pennoni, Reference Bartolucci, Pandolfi and Pennoni2017) have been used for estimating the proposed HMMs.

V. Discussion and Conclusions

The paper analyzes the dynamics of sparkling wine purchases over two years through a multivariate HMM, accounting for the sparsity of the purchases. The results point out homogenous latent profiles corresponding to groups of consumers sharing similar preferences in repeated sparkling wine purchases. The influence of the socio-economic characteristics, such as area of residence, age, income, family size, and family life stage, is assessed through estimated parameters on the initial and transition probabilities of the latent Markov chain.

The estimated parameters show significant effects of age, income, type of family, and area of residence. Older and high-income purchasers show loyalty to better quality sparkling wines, while young purchasers (young families) are more variety seekers. These results are in contrast with Bruwer and Li (Reference Bruwer and Li2007), who reported a high propensity to experiment with new wines in older consumers, while they are in line with Fountain and Lamb (Reference Fountain and Lamb2011) and Olsen et al. (Reference Olsen, Atkin, Thach and Cuellar2015), who showed that Millennials prefer different varietals. Regional segmentation reports differences in loyalty and in variety seekers. The residence area is likely to be associated with wine familiarity or knowledge, higher in northern Italian Regions, such as Veneto or Lombardy. Similarly, to Dodd, Pinkleton, and Gustafson (Reference Dodd, Pinkleton and Gustafson1996), the consumer purchases known brands to avoid searching for information or to lower the risk of making a wrong choice. Price has a strong effect on sparkling wine purchases, especially in high price positioned ones (Luxury and Sophisticated segments), while it produces variety-seeking tendencies, as also pointed out by some authors (Hussain, Cholette, and Castaldi, Reference Hussain, Cholette and Castaldi2007). The results concerning the appellation are in line with those presented in Rossetto and Gastaldello (Reference Rossetto and Gastaldello2018) for Prosecco and no appellation such as Ordinary Sweet or Brut groups; however, the HMM shows a lower market share of Prosecco segments and a higher VSB in its customers.

Compared to the existing models employed in the literature, such as loyalty and VSB, the proposed methodology based on the HMM works differently. For instance, Adamowicz and Swait (Reference Adamowicz and Swait2012) applied utility-maximizing on scanner data producing a market segmentation with respect to habits and VSB on two food products. Seetharaman (Reference Seetharaman2004) used a random utility model on scanner data to measure the sources of dependence. Recently, Wang and Liu (Reference Wang and Liu2020) adopted a random effect model on yogurt and soft drinks scanner data to investigate the effect of socio-economic covariates on the heterogeneous state of dependence. Ellis and Thompson (Reference Ellis and Thompson2018) evaluated the VSB in the wine market through the Van Trijp and Steenkamp scale (Reference Van Trijp and Steenkamp1992), while Olsen et al. (Reference Olsen, Atkin, Thach and Cuellar2015) adopted the Schwartz Value Inventory and the VARSEEK scale. The proposed HMM is based on latent variables having a discrete distribution, and it allows us great flexibility to model complex dependence structures such as those arising in wine or food markets having a wide assortment of sensory and hedonic attributes (e.g., cheeses). It permits to perform model-based clustering, identifying individuals with similar patterns of changes, allowing them to move between clusters across time.

The results provide strategic and managerial insights. First, the market is roughly divided into two main groups: consumers who repeat the purchase of low-priced sparkling wines (sweet or brut, having no appellation) and consumers who regularly buy Prosecco. Second, the loyalty is higher for Prosecco and Ordinary Brut than for Ordinary Sweet and high-priced sparkling wines. This result is associated with consumption occasions, usually higher for brut and Prosecco than for sweet and expensive wines, mostly consumed during celebrations or special events. Third, the transitions among states reveal the VSB of purchasers. Looking at Prosecco and ordinary brut consumers, we observe that Prosecco attracts people living in Veneto (they shift from Ordinary Brut to Prosecco), young consumers (less than 34 years of age), families with two or three components, and older families. On the other hand, consumers living in Sicily or Sardinia, between 45 and 64 years of age, with medium income after a certain period may leave the Prosecco group. Going to regional, income, and family life stage comparisons, we see a higher loyalty for Ordinary Sweet sparkling wines, and in the Sophisticated group, while living in northern regions, earning a low or high income, the age and family life stage may change the VSB of some purchasers. While income, and especially a low one, leads the consumer to be attracted by discounting strategies, the area of residence may influence the habits of consumers.

Implications for producers and sellers are referred to as loyalty and VSB and covariates effects. For instance, Prosecco consumers show an overall high loyalty due to heterogeneous purchase behaviors, which are pointed out by the HMM model used. Segmentation and purchasing results are useful for retailers to refine discount and assortment policies throughout the year. This analysis can be extended to food markets where items are frequently purchased throughout the year and show a wide assortment (sensory, hedonic attributes). When data becomes available, the HMM could evaluate consumers’ purchase patterns in switching among sensory attributes (e.g., flavors) and non-sensory ones (e.g., brand, price).

Data analysis was done at the category level (price range, type of wine, appellations), but the state of dependence among winery brands was not investigated. Also, the available panel data do not provide information on the wine knowledge of the purchasers or the customers’ values or attitudes, which are essential predictors in purchasing wine (Ellis and Thompson, Reference Ellis and Thompson2018). Furthermore, this analysis catches the sparkling wine purchase behavior in the large retail channel while excluding specialized wine shops and online channels. Thus, the paper does not evaluate high price market segments having a different brand assortment and price positioning.

Further research can be done on a sample that includes both sparkling and still wines. Sparkling is usually considered a different category; however, analysis on overall wine purchases may shed light on the associations among different types of wine while showing the wine portfolio of consumers. The analytical tool of HMMs may contribute to discovering relevant aspects of the consumer dynamics when applied with data obtained in other markets or from other marketing channels such as online sales.