Deployment of the Microeconomic Consumer Theory in the Artificial Neural Networks Modelling: Case of Organic Food Consumption

Abstract

Organic food consumption has become a significant trend in consumer behaviour, determined by various motives, among which the price does not play a major role, thus reflecting the Lancaster approach to the microeconomic consumer theory. Additionally, artificial neural networks (ANNs) have proven to have significant potential in providing accurate and efficient models for predicting consumer behaviour. Considering these two trends, this study aims to deploy the Lancaster approach in the emerging area of artificial intelligence. The paper aims to develop the ANN-based predictive model to investigate the relationship between organic food consumption, demographic characteristics, and health awareness attitudes. Survey research has been conducted on a sample of Croatian inhabitants, and ANN models have been used to assess the importance of various determinants for organic food consumption. A Three-layer Multilayer Perceptron Neural Networks (MLPNN) structure has been constructed and validated to select the optimal number of neurons and transfer functions. One layer is used as the first input, while the other two are hidden layers (the first covers the radially symmetrical input, sigmoid function; the second covers the output, softmax function). Three versions of the testing, training, and holdout data structures were used to develop ANNs. The highest accuracy was achieved with a 7-2-1 partition. The best ANN model was determined as the model that was showing the smallest percent of incorrect predictions in the holdout stage, the second-lowest cross-entropy error, the correct classification rate, and the area under the ROC curve. The research results show that the availability of healthy food shops and consumer awareness of these shops strongly impacts organic food consumption. Using the ANN methodology, this analysis confirmed the validity of the Lancaster approach, stating that the characteristics or attributes of goods are defined by the consumer and not by the product itself.

1. Introduction

In microeconomic theory, consumer preferences are one side of the optimisation problem, describing subjective views and habits; earnings and prices of goods and services are on the other side, forming a budget constraint. When subjective consumer desires and objective possibilities are confronted, as they are in the neoclassical microeconomic theory, the utility is maximised, or the expenditure is minimised [1]. However, consumers vary in their rational motives, significantly impacting the final purchasing decision. Hence, by combining the budgetary constraints with individual rationality motives, the final purchasing decisions are conducted when the budgetary constraints are the same [2,3]. The consumer is searching for a better product or service value, however, while the budgetary constraints are on the other side, real-life conditions drive sub-rational decision-making, which in turn decreases the consumer’s satisfaction [4,5].

This concept of the tradeoff between the budgetary constraints and individual rational habits can be investigated for various products and services, specifically those positioned as niche ones, therefore targeting one specific market segment. Healthy food could be considered one of these niche products as they are based on ecological and organic agriculture, which became increasingly popular as a concept in the early 21st century. Developed countries, such as Western European ones, e.g., Belgium, became pioneers in the organic food production trend [6].

Organic food consumers dedicate more time to their food selections due to their investigation of the details of the products they buy. Therefore, in search of the specific nutrients that they consider healthy, they buy products that offer the highest value in terms of food quality for their money. This kind of analysis is targeted at product characteristics, not the goods themselves, as it is described in the so-called Lancaster approach, a concept in which consumers demand traits and buy goods that have these traits [7]. In this paper, Lancaster’s theoretical concept has been applied to the empirical data, using artificial neural networks (ANN) to provide new insight into organic food consumption.

Organic food is not widely affordable, mainly due to the high prices of organic agriculture production. However, the price is not the only factor affecting an individual purchasing decision of organic food consumers since various factors, such as personal habits, play an important role in the food selection process, which can be impacted additionally by external factors, including the attitudes of close friends and family members [8]. Other motives are also relevant, such as the consumer’s ethnocentrism and the focus on the availability of the products [9]. However, the trend of organic food consumption is still not widely accepted, although it is on its way to having a pervasive presence, since according to the new European Union legislation and procedures, every EU country should have at least 25% organic agriculture by 2030 [10].

Croatia has been developing organic agriculture as it is a country of great agricultural abilities and resources, specialising in numerous organic food products. However, research about the organic food consumption of Croatian inhabitants is still scarce, focusing mainly on the monoculture or one specific population segment, using inferential statistical methods. The consumer perception of a single product, olive oil, was examined using Structural Equation Modelling (SEM), which revealed that olive oil’s benefit perception was strongly related to culture, tradition, pleasure, and healthy behaviour [11]. Student healthy food perception in Croatia was analysed using an ANOVA and a chi-squared test, and the authors found that the budget determines the healthy food usage frequency among the student population [12]. Cross-country research on food choice determinants using six types of conditioning motivations was implemented [13]. The research collected the responses from a questionnaire that was employed in 16 European countries, including the Republic of Croatia, and the data were analysed using SEM. The results indicated that health, emotions, environment, and convenience influence the consumers’ health food perception.

ANNs were developed at the end of the 19th and beginning of the 20th century, based on the interdisciplinary framework of physics, psychology, neurophysiology, mathematics, and computer science [14]. Nowadays, ANNs solve real-life problems that fall mainly into pattern classification, prediction, control, and optimisation [15]. Numerous types of research have indicated that considering the non-linear relationships is useful in various fields such as social research [16], and banking and finance [17,18].

The ANNs are mainly inspired by the human brain processes that include receiving, scanning, sending, analysing, and processing information. Linear regression could be considered a special form of the ANN since ANNs would be a simple linear regression without the softmax activation function. The softmax activation function has added non-linearity aspects to the neural network [19]. A multilayer perceptron (MLP) neural network function is a type of ANN that is widely used in various problem-solving efforts. They are based on a supervised procedure to minimise the prediction error and they contain three layers: input, hidden, and output [20].

ANNs have been used in various fields that are related to sensory and consumer science. The effects of input air temperature, feed flow rate, and maltodextrin ratio on the wall deposition flux and the moisture content of lactose-rich products were modelled using ANNs [21]. ANN has been combined with the particle swarm optimisation technique to determine the ideal amounts of lighting and sound in restaurants [22]. A particle swarm optimization-enhanced ANN that can forecast the coconut milk spray drying process has also been developed [23].

Several scientific papers are researching the effect of implementing ANNs for predicting the variables that are impacting the consumer awareness of food products (e.g., [24]). The review paper about the ANN applications for predicting consumer behaviour has revealed that ANNs have, so far, shown better accuracy than other methods due to their ability to capture the non-linear relationship between the variables [25].

ANNs have also been applied in the microeconomic field in several research studies. The authors of [26] used ANNs for investigating the microeconomic aspects of labour, consumption, and the accumulation of family assets, while the author of [27] used ANNs’ to perform microeconomic experiments regarding the economic agents’ behaviours. Other research has been developed to investigate ANN usage for the practical implications of microeconomic concepts, mainly in energetics [28], real estate market [29], and coastal area management [30].

Based on the above analysis, the following research gaps have been identified. First, motivations for consuming organic food in Croatia have been investigated mainly from the perspective of monoculture and affordability. In contrast, other motivations that are embedded in the Lancaster approach, such as availability and attitudes, have not been deeply investigated. Second, previous research has indicated that ANNs have a significant potential for predicting consumer behaviour, but they have been scarcely used in studies of predicting organic food consumption.

To shed some light on those research gaps, this research combines the microeconomic consumer theory and ANNs to provide insight into consumer behaviour in healthy food consumption, using data from the survey on the sample of Croatian citizens of the organic food usage. This research aimed to develop the ANN model that could be useful for predicting organic food, using variables that are easy to track, such as demographics, availability, and simple indicators of attitudes towards healthy eating. The Croatian inhabitants that are in social media groups that are interested in healthy eating were targeted to serve as the population for this research so to have a more diverse respondent base with an overall knowledge and interest in organic products [31,32]. Three ANNs were developed using a three-layer MLP network system, including an input layer with two hidden layers. Various partition rates for training, testing, and holdout have been implemented. The developed ANNs provided an insight into the relative importance of the individual rational habits to organic food usage among Croatian citizens, thus providing argumentation for the applicability of the Lancaster approach in this area.

2. Literature Review

In this section, a literature review is presented. Furthermore, the research gaps are explained, as well.

2.1. Deployment of Microeconomic Research with ANN Modelling

The linear models are mainly used to determine the relations between the dependent and independent variables. However, the potential bias might occur in the linear prediction models as a non-linearity relationship is often present. Economics has used ANNs for many purposes that mainly consider forecasting, econometrics, macroeconomics, etc. The microeconomics field was not involved in this research the majority of the time. However, there are indications that the usage of ANNs provides a more than satisfactory performance in model accuracy.

Several research papers have combined the microeconomic theory and ANNs, in various fields. The authors of [26] used ANNs to imitate economic subjects, and the researchers trained the model through a backpropagation algorithm. The same approach has been deployed in [27], regarding the economic agents’ (consumer, producer, etc.) behaviour.

Later applications of ANNs in microeconomics were specific to each industry. ANNs have also shown a relationship between the inputs and electricity production outputs when they are using data from the USA’s major electric utilities. The results show, with 97% accuracy, that electricity output can be calculated with the given inputs when using the ANN model [28]. When using the 2004 Household Budget Survey Data from Turkey, [29] used both ANNs and hedonic regression for the analysis due to the potential of having a non-linearity sample. The ANNs showed a better alternative for predicting house prices over hedonic regression. The author of [30] used ANNs in the economic management of coastal areas, which are affected by micro- and macroeconomic decisions. The research results had confirmed that ANN models can be used to support decision-making, especially for order prediction and performance evaluation in the management of coastal areas.

To summarise, non-linear relationships between the dependent and independent variables often occur in microeconomic problems. The presented research has indicated that a hybrid approach, such as those using ANNs, is often better for capturing the non-linear relationships than traditional statistical multivariate and econometric research are.

2.2. The Notion of Organic Food Consumption within the Lancaster Approach

Traditional demand theory assumes that the customers are driven mainly by the price when they are making their choices. However, numerous types of research indicate that consumer behaviour regarding healthy food choices is driven by motives other than the price [31,32], reflecting the Lancaster approach as one of the main opposing approaches to the traditional demand theory [7]. The Lancaster approach states that the characteristics or attributes of goods are defined by the consumer, and not by the product itself. It also indicates that goods (q = 1, …, m) are consumed due to their characteristics (z = 1, …, n), which is a fact that can be found for more than one good:

$$Z=\left[\begin{array}{ccc}{z}_{11}& \cdots & z_{1n}\\ ⋮& \ddots & ⋮\\ z_{m1}& \cdots & z_{mn}\end{array}\right]$$

(1)

The technological advances in the quality of a product increases the consumers’ utility [33]. Consumer demand and overall know-how for the organic food industry are what drive producers to change the ways of food production. Due to the utility increase as a result of customers choosing products for their attributes, such as is the way of organic food production, the consumer starts to change their habits. Therefore, the Lancaster approach can be easily transferred to field of organic food production.

Lancaster utility function approach is defined as the following:

$$\mathit{q}=f\left(Y,\mathit{p},\mathit{Z}\right)$$

(2)

where $\mathit{q}$ is a demand vector, $q^T=\left[\begin{array}{ccc}{q}_1 & \dots & q_m\end{array}\right]$ Y is the income level, price ${\mathit{p}}^T=\left[\begin{array}{ccc}{p}_1 & \dots & p_m\end{array}\right]$ are the prices of goods, and $\mathit{q}$ and $\mathit{Z}$ are a matrix of the attributes, i, in the product, j, respectively.

The utility maximisation model is:

$$U=\mathrm{U} \left(Z\right)$$

(3)

which is subject to

$$Z=g\left(\mathit{q}\right)$$

(4)

Since consumers buy goods, they pay for the goods even though they want its attributes. Therefore, the utility function should be transformed into a function of the goods:

$$\max U\left(\mathit{Z}\right)=U\left[{\mathit{Z}}^{-1}\left(\mathit{q}\right)\right]=V\left(q\right)$$

(5)

The budget constraint is:

$$pq=Y$$

(6)

Source: [33].

Alternatively, a model can also be transformed into an attribute view, but then budget constraint has to be redefined:

$$\max U\left(\mathit{Z}\right)$$

(7)

The budget constraint is:

$$p{q}^{-1}\left(Z\right)=Y$$

(8)

Lancaster concluded that consumers would act the same when the q values are changed into the z values. In other words, utility is expressed by the product attributes that are preferred by the consumers [33,34].

Consumer habits are often driven by social, psychological, and personal factors. The social factors that are affecting an individual’s habits are within their close environments, such as the attitudes of their peers, family, and friends, but they also include roles and status. Psychological factors include motivation, perception, learning, and attitudes and beliefs. Personal factors include age, income, occupation, and lifestyle [32]. Various actions and conditions can trigger these factors. Several studies indicate that labelling and branding play an important role in choosing organic products [35,36]. Consumers might be interested in the production and distribution process. More specifically, they might be interested in which type of seeds are the producers using, e.g., whether they are organic or not, or where is the farm’s location, and if the producer is certified with a label, e.g., a carbon footprint indicator. Research that was conducted in Germany revealed that the mix of the labelling of the product with the product characteristics, such as its origin, smell, and taste, with the combination of its price and quality, impacts the consumer’s buying intention the most, when compared to other factors [37].

To sum up, the above research indicates that in choosing healthy food, the consumer defines the characteristics or attributes of the goods, thus following the Lancaster approach [7], which is discussed in greater detail in the next section.

2.3. Determinants of Organic Food Consumption

The research about the determinants of organic food consumption reflects the Lancaster approach [7], indicating that the factors affecting an individual’s buying habits of organic food include price, demographic characteristics, retail store brand, and overall knowledge of organic products [38], therefore contradicting the traditional microeconomic approach that stresses that the price is the most important determinant in a buying decision. The authors of [39] provide a wide overview of consumer preferences across various food products in Germany, revealing that the consumers prefer to buy local organic food with higher prices over the lower-priced food from multichain stores. These research findings provide valid insights into consumer preferences, as more than one product has been included in the study and the survey was conducted in different regions across Germany. The authors of [40] revealed that consumers consider the ethical attributes to a greater extent than they do for the product price. The authors of [41] confirmed that the consumers are willing to pay more for an organic product than they would for a traditional one. Consumers value local products and brand authenticity more in that context and are willing to pay more for such products [42]. According to these research findings, consumers understand that they need to pay more to purchase an organic product and are willing to do so.

On the other side, there are differences in consumers’ willingness to pay more for organic products. Gender [43], age [44], and education [45] play a significant role in the consumer behaviour that is related to organic food consumption, although some researchers indicate that they are not completely relevant in that context [9]. Education is often the most important factor to consider when new products are distributed and it plays one of the most important roles in whether consumers purchase such a product [46], indicating that the higher the number of years of education, then the better the person’s general overview of healthy habits, a fact which was confirmed by the authors of [47] which considers the additional argument about the co-dependency of education, occupational class, and income with healthy food habits.

Behavioural habits determine which retail stores that consumers visit, which is an important element in their healthy food habits [44]. Retail stores could be divided into three groups: large retail chains (e.g., hypermarkets), small local shops, and niche market shops (in this context, healthy food shops). The shop’s availability of a specific product could significantly influence consumer behaviour. Although the consumers choose niche stores if the targeted product is unavailable in their local shops, they tend to decrease the purchasing of healthy food if it is not easily obtained [48].

The rate of organic product usage depicts the overall interest and know-how of healthy ways of living. For example, efforts to practice healthy eating impact the customer’s way of life because some of the consumer’s time might be used to prepare an organic meal using organic products, instead of them buying something on the “road”. Not only does this way of living decrease consumers’ daily spending, but the risk of certain acute diseases reduces, also [49].

Overall, the research indicates that numerous factors besides the price impact the consumer’s choice to buy organic food.

2.4. Usage of Machine Learning in Healthy Habits Research

ANNs have been so far used for several pieces of research on healthy eating habits. The authors of [50] show that the use of ANN modelling resulted in superior accuracy in predicting eating habits than more traditional approaches, such as SEM, did. The model, R², of the hybrid SEM-ANN model, increased the accuracy of the data by 27% when compared to the data of a model that used traditional SEM, showing that the non-linear approach has a significant potential for modelling eating behaviours when the researchers are using machine learning methods.

ANNs have been compared to decision trees, with mixed results when regarding which algorithm is more suitable [51]. Although the ANNs are not interpretable, they can be used only for prediction, and they do not provide the information that could indicate the relationship between the variables, they handle binary data very well, while having issues with categorical ones. On the other hand, decision trees can provide relevant insight into the variable relationship and can handle categorical data very well. Furthermore, the results of decision trees are easy to transfer to information systems as the set of rules that are created due to the tree form. However, ANNs are superior to decision trees in their prediction accuracy, which is relevant in microeconomic problems, and ANNs use all the inputs while the decision tree algorithm only uses relevant variables in its model. For these problems, the most relevant characteristic of the model is its ability to predict the outcome with high accuracy, even though visual representation and interpretability are important, thus leading to the conclusion that ANNs have significant potential in the development of prediction models in microeconomics.

However, several pieces of research have indicated no final solution to the comparison of ANNs and decision trees in generating the best predictions of organic food consumption. For example, [52] observed that the dietary quality that was based on food intake was subjected to a data mining process, including ANNs and decision trees, with mixed results, depending on the number of variables that were used. For the quality of a breakfast, which has a smaller number of food options, the ANNs method performed better, while for the main meal, with the higher number of food options, the decision tree method performed better. This result could be explained by the fact that decision trees handle more categories than ANNs. Therefore, one cannot conclude whether decision trees or ANNs are better if one does not fit the same dataset into both algorithms and check the accuracy of the models’ prediction.

The impact of sociodemographic factors on eating motivation using ANNs and neuron weight scores was investigated [53], defining eating motivation through three dimensions: health, emotional and economic. Variable age showed the strongest positive impact on health motivation, while living environment impacted emotional motivation. Gender showed the most positive and significant impact on economic motivation.

ANNs are also used to investigate other factors impacting food-related behaviour. Affective factors and normative cues may influence consumer-purchasing behaviour, driving unplanned and spontaneous buying, sometimes influencing consumers to act against their beliefs [54]. The impact of factors that are affecting consumer purchase intentions of organic food, including price, trust, and attitude, is investigated [24].

To summarise, ANNs have a significant potential in investigating the motivations behind consumer behaviour regarding organic food consumption due to their ability to capture the non-linear relationship between several variables and superior forecasting performance.

2.5. Research Gaps

The literature review has indicated the presence of the following research gaps. First, deploying ANNs has substantial potential in microeconomic research, especially considering the Lancaster approach. However, the ANNs’ forecasting ability strongly depends on the characteristics of the variables that are used as both the dependent and the independent in the model, with better forecasting performance in the models that use variables with smaller categories [51,52]. Second, healthy eating habits in the form of the purchase of organic products has been investigated mainly from product characteristics. In contrast, the contextual determinants, such as the type of retail store, health benefits awareness, and demographic characteristics, have not been sufficiently investigated, using the ANNs. Therefore, this research aims to model the determinants of organic food consumption, based on the Lancaster approach, using ANNs’ modelling, with the goal of developing a model for predicting organic food consumption using a simple set of indicators.

The next chapter presents the methodological approach that was developed for that purpose.

3. Methodology

3.1. Research Instrument

Table 1 presents the research instrument that includes the demographic variables, shopping habits in the preferred shop, and healthy eating importance and awareness.

Demographic characteristics are used in numerous research studies about organic food habits [53]; among them, gender [43], age [44], and education [45,46,47] are the ones that are most often researched since they have been often identified as the main factors determining organic food usage.

Shopping habits in terms of the customers’ preferred shop have been investigated, focusing on the large, small, and healthy food shops. Large shops are also known as hypermarkets/supermarkets, which denote a big area with lots of dietary, housing, hygiene, and similar products. These shops normally contain the cheapest options within the range of products they offer when compared to those in smaller shops [55]. In small shops, the selection of the products is similar, although sometimes they are more expensive and have a smaller range of choices. These shops are normally located in more convenient locations than the large shops are, such as residential areas and in the vicinity of schools or hospitals. Lastly, healthy food shops represent specialised, organic food-oriented shops which sell only a limited assortment of products, such as selected organic products that can be purchased from them. The number of such shops is smaller than the two above-mentioned ones, with more trusting customers that are driven by hedonic and utilitarian motives [56].

The dependent variable is the binary variable, measuring whether the respondents use organic food or not, in a dichotomous format.

Gender, age, and education modalities were defined using the typical categories [57,58]. The variables of healthy shop awareness and buying habits in large, small, or healthy food shops were measured using binomial variables [55,56]. Healthy eating importance was measured using a single statement with a five-point scale, following the recently proposed approach that was based on the research results indicating that when compared to multi-item measures, single-item measures can be a reasonable substitute because they frequently provide useful information [59].

We have opted for using the simplest possible form of the research instrument for the following reasons.

First, as previously stated, the features of the independent and dependent variables that are employed in the model significantly impact the capacity of ANNs to forecast, whereas when smaller category variables are utilised in models, there is a superior forecasting performance [51,52]. In addition, several authors have argued that artificial neural networks have some drawbacks [60], such as requiring an overlong training procedure, having a difficult topology selection, and having problematic input variables (attributes).

Second, since the research aims to develop a forecasting model for predicting consumer behaviour toward buying organic food, which could be easily deployed in practice, both dependent and independent variables were defined using the smallest possible number of categories and research items, following the recent research indicating that there is a preference towards the single items when compared to scales [59].

Third, another motivation for using the simplest possible form of the questionnaire is to avoid a non-response bias due to the reluctance of the respondents to fill in the lengthy surveys online, which not only increases the number of potential respondents who would give up on filling in the survey but also decreases the concentration of the participants, thus decreasing the quality of the collected answers [61].

Table 1. Research instrument.

Variable Group	Survey Question	Variable Name	Modalities	Source
Demographic characteristics	What is your gender?	Gender	Female, Male
	What is your age?	Age	18–25; 25–35; 35–45; 45–55; 55+	[58,59]
	What is the highest level of education at which you have graduated?	Education	Secondary school; College; Graduate; Post-graduate
Shopping habits and healthy food awareness	Do you buy food in a large shop?	Large_shop	0—does not buy food in the large shop; 1—buys food in a large shop
	Do you buy food in a small local shop?	Small_shop	0—does not buy food in a small local shop; 1—buys food in a small local shop	[56,57,60]
	Do you buy food in a healthy food shop?	Healthy_food_shop	0—does not buy food in a healthy food shop; 1—buys food in a healthy food shop
	Please estimate to what extent healthy eating is important for you.	Healthy_eat_importance	1—healthy eating is not important at all; 5—healthy eating is extremely important
	Are you familiar with the concept of healthy food shops?	Healthy_food_shops_awaren	0—not familiar with the healthy food shops in general; 1—familiar with the healthy food shops in general
Dependent variable	Do you use organic food in your diet?	Organic_food_usage	0—does not regularly use organic food 1—regularly uses organic food	[61]

Source: Author’s work, based on [56,57,58,59,60,61].

Table 1. Research instrument.

Variable Group	Survey Question	Variable Name	Modalities	Source
Demographic characteristics	What is your gender?	Gender	Female, Male
	What is your age?	Age	18–25; 25–35; 35–45; 45–55; 55+	[58,59]
	What is the highest level of education at which you have graduated?	Education	Secondary school; College; Graduate; Post-graduate
Shopping habits and healthy food awareness	Do you buy food in a large shop?	Large_shop	0—does not buy food in the large shop; 1—buys food in a large shop
	Do you buy food in a small local shop?	Small_shop	0—does not buy food in a small local shop; 1—buys food in a small local shop	[56,57,60]
	Do you buy food in a healthy food shop?	Healthy_food_shop	0—does not buy food in a healthy food shop; 1—buys food in a healthy food shop
	Please estimate to what extent healthy eating is important for you.	Healthy_eat_importance	1—healthy eating is not important at all; 5—healthy eating is extremely important
	Are you familiar with the concept of healthy food shops?	Healthy_food_shops_awaren	0—not familiar with the healthy food shops in general; 1—familiar with the healthy food shops in general
Dependent variable	Do you use organic food in your diet?	Organic_food_usage	0—does not regularly use organic food 1—regularly uses organic food	[61]

Source: Author’s work, based on [56,57,58,59,60,61].

3.2. Data Collection and Sample

In general, sampling is the process of choosing the population’s representatives to reveal the population’s characteristics. Several approaches to sampling are used; some approaches are random, convenient, and snowball sampling. The most commonly used method is random sampling due to its ability to control for random error, which refers to reducing the influence of variables that are unrelated to the study. However, the convenient sample is often used when the research is focused on minority groups [62], such as organic food consumers. A convenience sample is often used in the research of organic food consumption [63,64]. Several previous studies have indicated student samples’ applicability, even though the convenience sampling results may limit the findings’ generalizability. For example, similar dietary intake patterns were found in a convenience sample of teenagers in a Canadian city when they were compared to the national trends [65]. Moreover, the convenient sample method is often used with a distribution using social media by employing two different strategies: the push and the pull method. These strategies involve recruiting respondents that are actively looking to participate in paid surveys vs. recruiting respondents by using ads or posts on social media; the pull method results in a more diverse demographic profile of respondents [66].

This study investigated the factors influencing consumer organic food usage among the inhabitants of the Republic of Croatia, using a convenient sample, and it was aimed at Croatian residents who are members of Facebook groups that are interested in healthy eating. We have used the pull method to recruit respondents at our convenience by publishing posts on Facebook groups to attract a more diverse demographic respondent profile [66]. By using this approach, we have additionally ensured that all participants had a relevant history of participating in online debates about organic foods or engaging in online shopping behaviour, thereby certifying the content validity [67]. Overall, the Facebook groups accounted for 1230 members, among which 252 participants replied, thus yielding a 20.5% response ratio, which is comparable with similar research [68].

The online questionnaire was written in Croatian and distributed online between April and May 2021, as indicated in the research instrument section. The first part of the research comprised demographic questions, while the second part contained questions regarding organic food buying habits, attitudes, and usage.

The survey included the ethics statements in the beginning, which the participant needed to agree upon before entering the data into the survey. The participants conformed to the statements indicating that: (i) they were willingly participating in the survey, (ii) they provided the correct demographic data (age, education, workplace, work experience), and (iii) they understood that their answers would be treated anonymously and were solely for scientific research.

3.3. Artificial Neural Networks

This research used a three-layer neural MLP network system, including an input layer with two hidden layers. One hidden layer covered the radially symmetrical input layer, while the second hidden layer covered the output layer. These layers were important in the hidden layer output weighted-sum calculation and in calculating the index class for the input pattern [69]. The sigmoid function was used for initiating the neural networks function, and this function guarantees that the outputs will always be between zero and one and is used to classify ANNs [70,71]. Experimentation with various node combinations has been implemented to produce a well-built model.

Following the machine learning approach, the ANNs’ framework requires stages such as training, testing, and holdout to produce the best possible non-linear outcome [64]. Units are assigned randomly to three partitions of the dataset. First, the training part of the dataset was used for the initial modelling, in which the ANNs learn the relation between the target variable and its features. Second, a cross-validation technique used the testing part of the dataset to optimise the network architecture, parameters, and learning time. Finally, the holdout part of the dataset was used to validate the performance of the ANNs that have been designed using the training part of the dataset and fine-tuned using the testing part of the dataset.

Different partition rates for training, testing, and holdout have been implemented in the experiment, following the approach proposed by [70]: (i) ANN1 with 50% training, 30% testing, and 20% holdout; (ii) ANN2 with 60% training, 20% testing, and 20% holdout; (iii) ANN3 with 70% training, 20% testing, and 10% holdout.

IBM SPSS ver. 27 software was used for this research to construct the ANNs.

4. Research Results

Research results are presented below. First, we present the prevalence of organic food usage. Second, we outline the results of case processing using neural networks. Third, we discuss the classification efficiency of neural networks. Finally, the measurement of the variable importance is presented and discussed.

4.1. Organic Food Usage

The dependent variable, organic food usage, is shown in

Table 2. Dependent variable.

Variable Name	Modalities	N (%)
Organic_food_usage	0—does not regularly use organic food	64 (25.4%)
	1—regularly uses organic food	188 (74.6%)

Source: Author’s work, using SPSS.

. The results indicate that the majority of the respondents regularly use organic food (74.6%), while about one-quarter do not use organic food regularly (25.4%). Since the survey respondents were mainly active participants in social media healthy food groups, the sample consists of a high percentage of organic food-oriented individuals, which is higher than the percentage of such users in the whole Republic of Croatia.

Appendix A denotes the demographic characteristics. The following groups that were accounted for the most in the sample: female (78.2%), 25–35 age group (37.3%), and graduate degree (41.3%). Not surprisingly, the highest percentage of organic food consumers were among female respondents, in the 45–55 age group, and post-graduate degree holders. A Chi-squared test indicated that the observed difference is significant for gender at the 1% level (Chi-square = 12.353, p-value = 0.000), and for age at the 5% level (Chi-square = 11.149, p-value = 0.025).

Appendix B denotes the independent research variables measuring healthy eating awareness and healthy eating habits. The majority of respondents buy food from large shops (89.7%), followed by healthy food shops (54.0%), and then small shops (32.1%). Most responded that healthy eating is important or very important (85.7%). Most respondents are familiar with healthy food shops (73.0% of them). Among those that buy food from large shops, 72.1% also use organic food, while the ratio of consumers of organic food is even higher in the groups that buy food from small shops (86.4%) and healthy food shops (93.4%). A Chi-squared test indicated that the observed difference is significant for large shops at the 1% level (Chi-square = 7.106, p-value = 0.008), for small shops at the 1% level (Chi-square = 8.797, p-value = 0.003), and for healthy food shops at the 1% level (Chi-square = 54.992, p-value = 0.000).

4.2. Results of Case Processing

An activation sigmoid function was used for the hidden layer, while the softmax function was used for the output layer.

Table 3. Network information for three ANNs.

Layers	Partitions	# of Units	Activation Function	Variable Description
ANN1 (50%-30%-10%)
Input	136 (54.2%)	24	-	Independent: Demographic characteristics; Healthy awareness habits
Hidden	86 (34.3%)	4	Sigmoid
Output	29 (11.6%)	2	Softmax
ANN2 (60%-20%-20%)
Input	160 (64.0%)	23	-	Independent: Demographic characteristics; Healthy awareness habits
Hidden	49 (19.6%)	6	Sigmoid
Output	41 (16.4%)	2	Softmax
ANN3 (70%-20%-10%)
Input	170 (67.7%)	24	-	Independent: Demographic characteristics; Healthy awareness habits
Hidden	57 (22.7%)	6	Sigmoid
Output	24 (9.6%)	2	Softmax

Note: Dependent variable: Organic_food_usage; Source: Author’s work, using SPSS.

shows the number of nodes in the three model stages. Independent variables, demographic (age, education, and gender), and healthy food awareness habits variables are shown in the variable description part of it.

The following training, testing, and holdout ratios were used [50]:

• ANN1 with 50% training, 30% testing, and 20% holdout;
• ANN2 with 60% training, 20% testing, and 20% holdout;
• ANN3 with 70% training, 20% testing, and 10% holdout.

Figure 1 presents the architecture of ANNs, and it was prepared using specialist software.

The input layer consists of 24 nodes and one bias node, and the hidden layer consists of six nodes and one bias node. The output layer consists of two nodes.

According to [17,18], there are numerous ways of determining the model fit. Cross entropy error is the function that the neural networks minimise during the generation of the training and testing samples. The percentage of incorrect predictions indicates the ratio of the misclassified cases.

Table 4. Summary of the model fit for the three ANNs in training, testing, and holdout stages.

	Layer Description	ANN1 (50%-30%-10%)	ANN2 (60%-20%-20%)	ANN3 (70%-20%-10%)
Training	Cross Entropy Error	51.314	72.201	65.499
	Percent Incorrect Predictions	14.7%	20.0%	15.3%
	Training Time	0:00:00.02	0:00:00.02	0:00:00.04
Testing	Cross Entropy Error	30.996	15.257	25.436
	Percent Incorrect Predictions	16.3%	16.3%	19.3%
Holdout	Percent Incorrect Predictions	31.0%	12.2%	8.3%

Note: Dependent variable: Organic_food_usage; Source: Author’s work, using SPSS.

summarises model fit for three ANNs.

In the training stage, ANN1 showed the lowest cross-entropy error of 51.314, with the lowest percentage of incorrect predictions, at 14.7%. In the testing phase, ANN2 showed the lowest cross-entropy error value of 15.257, with the lowest percentage of incorrect predictions, at 16.3%. The holdout stage indicated that ANN3 was the one with the lowest percentage of incorrect predictions, at 8.3%.

Even though ANN1 showed the smallest cross-entropy error, it showed a high percentage of incorrect predictions, especially in the holdout stage, and this percentage was higher than it was in ANN2 or ANN3. The ANN3 model has been chosen as the winning model since it had the lowest percentage of incorrect predictions. Furthermore, in the training stage, the cross-entropy error of the ANN3 model was the second smallest. The testing stage cross-entropy error when combined with the smallest percentage of incorrect predictions in the holdout stage provides the criteria for the best model selection in this research.

4.3. Classifications Matrices

The classification matrices are used to gain an in-depth understanding of the predictions. The classification matrices present the % of the correct classifications for the two classes of the dependent variable.

Table 5. ANNs’ classification matrices.

Sample	Observed	Predicted ANN1			Predicted ANN2			Predicted ANN3
		1—Yes	0—No	% Correct	1—Yes	0—No	% Correct	1—Yes	0—No	% Correct
Training	1—Yes	94	11	89.5%	106	14	88.3%	117	9	92.9%
	0—No	9	22	71.0%	18	22	55.0%	17	27	61.4%
	Overall %	75.7%	24.3%	85.3%	77.5%	22.5%	80.0%	78.8%	21.2%	84.7%
Testing	1—Yes	56	7	88.9%	32	4	88.9%	39	4	90.7%
	0—No	7	16	69.6%	4	9	69.2%	7	7	50.0%
	Overall %	73.3%	26.7%	83.7%	73.5%	26.5%	83.7%	80.7%	19.3%	80.7%
Holdout	1—Yes	18	1	94.7%	28	2	93.3%	18	0	100.0%
	0—No	8	2	20.0%	3	8	72.7%	2	4	66.7%
	Overall %	89.7%	10.3%	69.0%	75.6%	24.4%	87.8%	83.3%	16.7%	91.7%

Note: Dependent variable: Organic_food_usage: 0—does not regularly use organic food; and 1—regularly uses organic food; Source: Author’s work, using SPSS.

represents the classification matrices for the three ANNs models. The values (0—No, and 1—Yes) denote the classes of the dependent variable: 0—does not regularly use organic food; 1—regularly uses organic food. For example, the ANN1 model correctly predicted in the training phase, percentage of those who regularly use organic food (89.5%).

In comparison, those customers that do not regularly use organic food are correctly classified to have a lower prevalence (71.0%). The correct percentage in the training phase showed the highest result of 85.3% for the ANN1 model. The ANN2 model was aligned with the ANN1 modes in the testing phase, showing the same percentage of 83.7%. The holdout stage indicated the highest correct value of 91.7% in the ANN3 model. It can be concluded that the ANN3 model has the highest overall accuracy in the holdout phase. However, the classification matrices indicate that ANN models predict organic food usage better. For example, ANN3 correctly predicted the presence of 100% of the organic food users in the holdout stage, while the percentage of the correct classifications for non-users of organic food is lower than that is (66.7%).

The box plot for the ANN3 model is presented in Figure 2. The first box plot indicates high predicted pseudo-probabilities for organic food users that are correctly classified. In contrast, the second box plot has low predicted pseudoprobabilities for organic food users that are incorrectly classified. On the other hand, the third and fourth box plots indicate moderate pseudo-probabilities for the non-users of organic food that are correctly and incorrectly classified. In other words, the diagrams indicate that the ANN3 model better predicts the behaviour of respondents who were actual users of organic food than those who were not users at all.

For further validation, an ROC curve has been implemented. It shows the classification performance for all possible cut-offs by a diagram of sensitivity vs. specificity. The area under the curve represents the sensitivity vs. specificity data for the research ANN models and two organic food usage categories in

Table 6. The area under the curve.

Dependent Variable Modalities	ANN1	ANN2	ANN3
1—Yes	0.876	0.839	0.856
0—No	0.876	0.839	0.856

Note: Dependent variable: Organic_food_usage: 0—does not use organic food; 1—uses organic food; Source: Author’s work, using SPSS.

, which shows the probability of a randomly picked conscript being accurately assessed or ranked if the person is more likely to use organic food. This interpretation is based on the non-parametric Mann–Whitney U statistics test used for the area under the curve computation. Furthermore, the maximum area under the curve = 0.876 (ANN1) was followed by ANN3 (area under the curve = 0.856), which demonstrated that the model has a good predictive standpoint for perceiving the organic food usage among the population Republic of Croatia. The maximum value for the area under the curve is equal to 1.000.

Figure 3 denotes the sensitivity and specificity of organic food usage (0—does not use organic food and 1—uses organic food) by the ANN3 model (the 70% training, 20% testing, and 10% holdout) created in the training and testing stages under the ROC curve. The 45-degree line connecting the bottom left of the axis to the top right of the axis represents a random guess of a category. The further the ANN model curve is away from the black line and the more to the left it is, then the better the classification is. According to the results we obtained in Table 6, in ANN3 the area under the curve is at 0.856 for the organic food category (yes) and this is the same for the no organic food usage (no) category. The ROC curve confirms that the ANN3 model has a moderately good fit for the research data.

Figure 4a,b are supporting sketches for assessing the performance of the classification model, presenting the gain and the lift of the ANN3 model. Gain is the proportion of all positive observations in the data to all positive observations up to a decile. The gain chart is a graph with the decile on the horizontal axis and the gain on the vertical axis. Lift is the proportion of positive observations up to the first decile that were observed using the model to positive observations up to that first decile that were anticipated to be observed using a random model. The relationship between the lift on the vertical axis and the matching decile on the horizontal axis is represented by a lift chart. Both charts confirm that the ANN3 model predicts the class of organic food users better than it does for non-users.

The variable importance is investigated in the next section to gain further insight into the factors affecting organic food usage.

4.4. Variable Importance

Table 7. Independent variable importance for the organic food consumption in the ANN models.

	ANN1 (50%-30%-10%)		ANN2 (60%-20%-20%)		ANN3 (70%-20%-10%)
Variables	Importance	Normalised Importance *	Importance	Normalised Importance *	Importance	Normalised Importance *
Healthy food shops	0.347	100.0% (1)	0.269	100.0% (1)	0.266	100.0% (1)
Awareness of healthy food shops	0.242	69.6% (2)	0.231	85.8% (2)	0.245	92.0% (2)
Small local shops	0.023	6.6% (8)	0.092	34.1% (5)	0.112	42.2% (3)
Education	0.064	18.5% (5)	0.058	21.5% (8)	0.111	41.7% (4)
Age	0.077	22.1% (4)	0.063	23.5% (6)	0.105	39.4% (5)
Gender	0.14	40.4% (3)	0.104	38.7% (4)	0.083	31.3% (6)
Importance of healthy eating	0.06	17.3% (6)	0.141	52.6% (3)	0.066	24.9% (7)
Large shopping chains	0.048	13.8% (7)	0.042	15.8% (7)	0.012	4.6% (8)

Note: * variable ranks according to normalised importance are in parenthesis; Source: Author’s work, using SPSS.

describes the importance and normalised importance that the independent variables have in the ANN models predicting organic food consumption. Importance is measured in absolute and percentage values, with the latter being the normalised importance. The variable importance denotes all the weights connecting the input nodes that pass through the hidden layer to the specific response. The closer the variable to one, then the higher the variable importance is.

As previously stated, the ANN3 model showed the best fit, so we have focused the discussion on variable importance to this model. The highest importance is the variable related to healthy food shops (100.0%), which measures whether respondents buy food in specialised food shops. The second highest importance is the variable related to the awareness of healthy food shops (92.0%). Three variables follow, with similar importance levels: small local shops (42.2%), education (41.7%), and age (39.4%).

The lowest level of importance are attributed to the variables related to gender (31.3%), importance of healthy eating (24.9%), and large shopping chains (4.6%). It can be concluded that the highest importance is found mostly among the variables reflecting the situational factors, shopping habits, and the availability of healthy food shops. Among the demographic variables, the educational level has the highest importance. In contrast, age and gender have lower importance than expected since previous research indicated that these factors strongly impact organic food consumption, e.g., [12].

These findings show that the research respondents are more likely to buy organic food in small or healthy food shops and have a good knowledge about the availability of healthy food shops in their neighbourhoods. These results align with the Lancaster approach since consumers show their interest for attributes and only indirectly for goods (Equation (8) instead of Equation (4)). In this way, the price of a good is an indirect price of a set of attributes, instead of a price of one good only. In this way, one good can satisfy more than one attribute, and the same can be found in more than one good. It is up to a consumer to find goods with the greatest attributes at the lowest possible cost. Namely, the consumer is the one that chooses their healthy habits and so they define the shops that they will purchase organic food in.

5. Discussion and Conclusions

Consuming organic food has emerged as a significant trend in consumer behaviour, which is driven by various factors, not the least of which is the Lancaster approach to microeconomic consumer theory. ANNs have also shown much promise for developing precise and effective models for forecasting consumer behaviour. This work tries to implement the Lancaster strategy in the developing field of artificial intelligence, while considering these two tendencies. This study aims to create an ANN-based predictive model to analyse the associations between attitudes toward health awareness, consumption of organic foods, and demographic factors.

This research combines microeconomic consumer theory and ANNs in the context of the healthy eating habits of Croatian inhabitants, measured as organic food consumption. The relevance of various determinants for organic food consumption has been evaluated using ANN models using survey research on a sample of Croatian citizens. An MLP-ANN was trained and tested by constructing an algorithm that included demographics (age, education, and gender) and healthy eating awareness and habits. The sigmoid activation function was used in the hidden layer, while the softmax function was used in the output layer. Its accuracy was tested by the ROC curve, cumulative gains, and lift charts. At the same time, the importance of the independent variables table showed that healthy food shops and awareness of healthy food shops affected the purchase of organic food use the most. The optimal number of neurons and transfer functions have been identified using a three-layer MLPNN structure which was built and validated. The first input comprises one layer, while the other two are hidden layers (the first covers the radially symmetrical input, the sigmoid function; the second covers the output, the softmax function). The testing, training, and holdout data structures were created in three different iterations, and the 7-2-1 partition produced the best accuracy.

The model with the lowest percentage of inaccurate predictions during the holdout stage, the second-lowest cross-entropy error, the correct classification rate, and the largest area under the ROC curve were selected as the best ANN model. The study’s findings indicate that consumer awareness of and access to healthy food outlets significantly impact the consumption of organic foods. This investigation, which used the ANN methodology, supported Lancaster’s claim that consumers, not the products themselves, define the qualities or attributes of commodities.

The main variables that impact the regular usage of organic food were identified. First, the research results indicated that buying food in healthy food shops, combined with the awareness of healthy food shops, are the strongest indicators of the regular usage of organic food, followed by a positive attitude regarding the importance of healthy eating. These results align with the Lancaster approach, where the consumer is the one that chooses their healthy habits and so they define the shops that they will purchase organic food at. Second, the research results indicate that demographic variables significantly impact the decision to buy organic food. Older female respondents were more positive about the regular usage of organic food, and healthy food users are primarily individuals above 30 years of age, specifically in the age group 35+, which aligns with [43,44]. Third, the research findings also show that healthy food consumers either buy food in small shops or healthy food shops, thus confirming the research results of [56].

These findings indicate the practical implications of our research. First, since the situational factor (availability of healthy food shops) had the strongest impact on buying organic food, the public authorities, especially in the area of public health, should investigate the options to subsidise the healthy food shops, especially those in small areas, such as suburbs and small towns. Second, since the attitudinal variable (understanding of the importance of healthy eating) and demographic variables (female and older age) had a strong impact on organic food consumption, public health authorities should work on the further promotion of educational programmes on healthy eating, especially among the younger male population [72]. Third, since the research results indicated that organic food availability is a strong prediction of it being bought, organic food producers should present their products in the shops where potential customers buy food to satisfy their demand, including in small and large shops.

The research has some limitations. First, the research base could have been expanded into a more diverse sample that includes less organic food-oriented individuals. The possibility of a large and more diverse sample might show general population responses and their impact on the dependent variable, organic food usage. Second, we have opted for using a limited number of demographics, behavioural, and attitude variables, which is beneficial for avoiding a non-response bias and improving the predictive performance of the ANN. Due to the simple form of the dependent variable (binary variable), the sigmoid activation function was used. Third, we have focused on only one country, Croatia, whose inhabitants are already quite aware of the importance of organic food due to the traditionally high-quality food and healthy diets, such as the Mediterranean diet is [11,12]. Fourth, in this paper, we have tested the Lancaster approach in the example of organic food. In contrast, a broader approach in terms of variable products and services should be used to ensure more general conclusions about the applicability of the ANN approach.

Based on the identified limitations, future research directions emerge. One of them is to expand the research framework and implement it in different demographic or geographic locations to add new variables and make it more appropriate for practical and academic implications. Furthermore, further research might expand the respondents’ base with a more general and younger population that are not so familiar with organic food. They could educate them on the healthy benefits of consuming organic food and research their answers while combining them with the results [72]. Finally, a wider sample of products, services, and countries should be covered to gain more general conclusions in testing the applicability of ANNs for researching the Lancaster approach.