Consumer Behavioural Patterns Towards Milk and Dairy Products: An Analytical Investigation of an Indian Dairy Supply Chain

The Indian dairy sector accounts for about 22% of global milk production and has the highest milk consumption with 20% of total world consumption (USDA, Report, Feb. 2020). The current demand for milk will be between two hundred to two hundred ten million tonnes and is expected to increase by an average of three per cent in the coming decades (DHAD, FAO, 2022). Milk production is expected to increase by three to five per cent in FY 2022-23. India’s dairy market will reach a value of INR 30,840 Billion at a CAGR of 14.8% by 2027 (Report, ICAR, 2019-2020).
 
In recent years, with the increasing demand for superior quality products and its brand’s value diversifying the customer expectations and satisfaction level in India (Diwakar et al., 2020)., the usage of milk has been shifting from traditional dairy products to value-added and sustainable milk products. It is estimated that the recent trends in consumption patterns and production levels have significant growth in the Indian dairy sector (Squicciarini et al., 2017) .
 
The earning population of India lies between the age group of fifteen to sixty years of age, which is a progressive indicator of the demographic pattern. In the past few years, health consciousness, high disposable income and willingness to spend in the younger population stimulated the retail sales of milk and dairy products (Laura, 2016). The consumption pattern of milk users varies from 55.4% in 1991 to 67% in 2019. It is observed that in India, a large amount of milk, about twenty-five to thirty per cent consumed by children below the age of fourteen. A paradigm shift in family setup reflects the changes in the standard of living and consumption patterns.
 
The Indian dairy industry comprises both formal and informal sectors for the large production of raw milk and dairy products. Among India’s top ten milk-producing states, Uttar Pradesh produced the highest amount of milk, about 30.52 million metric tons, followed by Rajasthan, 23,67 million tons in 2019 (Fig 1). However, milk production from an individual milch animal has been reduced over the past few years shown in figure (Table 1).




 
The motivation of the research is to understand the role of observed and unobserved demographic characteristics like household income, age group, employment, education level, food habits and others in the consumption pattern of milk users.
 
The conceptual framework shown in figure (Fig 2) is a natural progression of the phenomenon that describes the relationship between crucial indicators or constructs to be studied based on the proposed model of the study.
 


Related literature
 
Consumer behaviour pattern
 
It is evident that in present studies, the wide spread of epidemic diseases like COVID-19 and lockdown situations in various parts of the world changes the buying behaviour pattern of consumers related to health and hygiene products (Alejandro Acosta et al., 2021). Infectious diseases like the coronavirus mainly impact immunity and mental health. Naturopathy, yoga and Ayurveda as supportive intervention in the COVID-19 treatment of patients helps to improve immunity (Golechha, 2020).
 
Emerging trends in dairy markets
 
During the pandemic in India, digitalisation and localisation of Inbound logistics in the Dairy Supply Chain (DSC) mitigate the risks in dairy processing activities to ensure dairy milk collection, supply to the end-users with taking all precautionary measures, production of milk in case of shortage of staff to avoid spreading of infection. Distribution channels are working correctly to maintain an efficient supply chain (Alejandro Acosta, Steve Mc Corriston, 2021). 
 
Bovine milk plays a vital role in the human diet and possesses multiple health benefits to fight against cancer and cardiovascular and bacterial disease. Milk contains about 3.5 cents of Protein as an essential constituent of milk components (Surya et al., 2017).
 
Chemical constitution and complexities in milk
 
The pigmentation and colouration of milk and dairy products depend on factors related to primary and transformation processes. Carotenoids are chemical constituents which play a significant role in the colour of dairy products. The composition of carotenoids and retinol in milk depends on several dietary and non-dietary factors, like animal breed and feeding management (Gabriela Grigioni et al., 2018).
 
Buying intentions and rethinking strategies
 
Milk users pay a premium price when receiving health benefits from indigenous cow milk (Kumar, Rao and De, 2018). In the Milk Supply Chain (MSC), new technology facilitates techniques like robotic milking systems, making collecting and testing milk samples easier. Developments in analytical techniques have helped fractionate as well as characterise milk components (Priyadarshini, 2018). Knowledge of the biosynthesis of milk components has rapidly progressed over the past fifty years. Milk testing has also been transformed from a slow procedure done in the laboratory to rapid testing of multiple components (Lucey, Otter and Horne, 2017).
 
Buying theories to understand consumer pattern
 
The TPB (Theory of Planned behaviour ) in this context is pertinent to understanding the role of the decision-making ability of an individual. It also acts as an effective tool to choose various alternatives in a planned manner influenced by an individual’s past experiences and future expectations (Posthuma and Dworkin, 2000). This theory explains the cognitive factors like motivation, attitude and purchase intention of milk consumers (Wang and Kajungiro, 2019).

The study was conducted between July 2019- December 2019 in Rajasthan Technical University, Kota Rajasthan. Secondary Information retrieved from previous literature available in the dairy sector in the Indian and global context. Responses are gathered from the significant top ten milk-producing states of rural and urban populations in India. 472 responses are gathered from the finite population of milk users. An appropriate sample size is 384 (Bartlett, Kotrlik and Higgins, 2001). A systematic  stratified random sampling technique was used for data collection. The survey was conducted six months online during the COVID-19 pandemic outbreak. Responses are measured based on the previously tested multiple-choice questions five point rating scale. 
 
Instrument and Measurement: Based on previous literature Semi-structured questionnaire is designed to understand the purchase intention of milk users towards the quality of milk and dairy products. Fifteen items are retrieved from the demographic profile of the survey of milk consumers shown in Table (Table 2 and 3).




 
Statistical analysis
 
In this study, statistical analysis can be used to interpret the data gathered through surveys and studies, conceptual or statistical models and explore the relationship between the variables or set of items from which a sample can be drawn. Multiple Logistic Linear Regression (MLLRA) technique can confirm a connection between the target variable and two or more explanatory variables (Fung Jianing, 2013).
 
In the statistical approach, Logistic Regression, the target variable can be taken as the value 0 with a probability of failure (1-q) and value 1 with the chance of success (q). The objective of this technique is to predict an outcome of discrete variables.
 

log (p/1-p) =β₀+ β₁X₁ + β₂X₂+……+β_kX_k

 
 
This expression represents the relationship between the categorical explained variable (DV) and the predictor variables (IVs) to estimate the probability of certain events, where β₀, β₁, β₂……….,β_k. as a regression coefficient, computed from the dataset. When Y=1 and X₁, X₂,………, X_k are the predictor variables. The predictor variable log (p/1-p) is also a known explanatory variable in a linear equation. ⁷4.1 Generalised Linear Model Framework  in logistic Regression, the target variable is categorical, as. The factor function is used through R Software to convert the variables into factors (non-numeric to numeric) and run the logistic regression command. The glm function predicts binary outcomes from continuous explanatory variables to estimate the given data set.
 
These are generalised linear equations for given set of data file COM_train with explanatory variables are as follows:
 

glm (formula= Customer satisfaction ~1, family = Binomial (), data= COM_train)......(1)
 
glm(formula = Customer satisfaction ~ Frequency of the milk intake + Food habit +Price of the milk , family = Binomial (logit), data=COM_train)......(2)
 
glm (formula = Customer satisfaction ~ Frequency of the milk intake + Price of the milk, family = Binomial (logit), data = COM_train)......(3)

In this section, the strength of relationship among demographic variables measured in respect to gender male and female as respondents.
 
In contrast to western countries, consumers are more conscious about quality, convenience, and health concerns about dairy products and their value-addition (Silvia and Meiselman, 2010). There is a strong need to develop better awareness among consumers about the health benefits of milk, adulteration, harmonisation, and unfair trade practices (Reddy, 2016). From this study it has been observed that most of the respondents  prefer the AMUL brand, and have the largest milk share as a dairy company. In some areas of north India, the mother dairy is the leading brand and most prominent consumer base (Mangla Kumar Sachin et al.,  2019). Consumers prefer A₂ milk dairy farms at nearly 17.8%, and still an emerging market for dairy farms and milk processing units in India (Allison and Clarke, 2006). In dairy farming, sustainable supply chain practices and technological advancement of paramount importance for efficiency and productivity (Sinha and Mishra, 2023; Shruti and Latika, 2016).
 
 
Here, odd ratio is a measure of association between outcome and exposure; odds ratios are used to compute the relative odds of the occurrence of the outcome of interest like satisfaction or dissatisfaction in this model given exposure to the variable of interest like food habits, i.e. vegetarian or non-vegetarian, the odds ratio can also be used to determine whether a particular exposure is a risk factor for a specific outcome and to compare the magnitude of various risk factors.
 
Testing for logarithm function
 
Odds ratio in a linear regression model, the odd ratio logarithm function estimates the probability of success and failure to understand the presence of explanatory variables in response to predictor binomial variables.
 
Logistic regression provides the knowledge and strength of relationships among variables if it increases greater than 0.1 or decreases less than 0.1. We can say (odds for PV+1) (odds for PV), where PV is a predictor value, shown in the odd ratio table (Table 2).
Odds ratio: Satisfied: Dissatisfied=0.1/0.8=0.125/1
 
Conditional odds
 
For male =Satisfied/Dissatisfied=0.57/0.48=1.187.
If we consider satisfaction as success, then the probability of this event happening is more in males.
For Female = Satisfied/Dissatisfied= 0.69/0.51=1.352.
If we consider satisfaction a failure, then the probability of this event happening is less in females.
 

odds=p (1-p)
 

Where,
p= Probability of the event occurring.
So if p=0.1, the odds are equal to 0.1/0.9=0.111 (recurring).
 
Overall odds ratio
 
1.187/1.352=0.87
• From the overall odds ratio value of 0.87, it is concluded that males are more likely to be satisfied than female consumers.
 
Model testing
 
Wald test is used to compute the statistical significance of each beta (b) coefficient in the logistic regression model. 
A Wald test calculates the z statistic, which is: 
 

z = Var[ˆβ|X] = σ² (X'X) -1

 
Sigma (s2)  is the variance of the residuals and has to be estimated from the data and is unknown and X is the design matrix.
 
Likelihood - Ratio Test (LRT)
 
This test measures the ratio of the maximised value of the likelihood function for the entire model. (Li) over the maximised value of the likelihood function for the simple model (L₀).
 
(L_i) over the maximised value of the likelihood function for the simple model (L₀).
The likelihood-ratio statistic equals:
 

-2Log (L₀/L₁) = 2 [log (L₀) - log (L_i)] = 2 (L₀ - L₁)

 
The log transformation of the likelihood function yields a chi-square statistic that is considered in the case of backward stepwise elimination.
 
Results for model estimation
 
From the contingency table (Table 3), the contingency of transmission computes the conclusions as follows:
a. Probability of Food Habit if vegetarian: 336/446=0.07.
b. Probability of Food Habit if Non- vegetarian = 1-probability of Food Habit if vegetarian: 1-0.07 = 0.93 (i.e., 5/17).
c. Odd ratio of food habits non-veg. as opposed to non-vegetarian food habit=140/336=0.42:1
d. Likewise, the odds of non-veg opposed to veg. Food habits are the inverse=336/140=2.4:1.
 
Model Base for logarithm function is written as H₀: β = 0, where (Constant is zero).
Generalised linear model (glm) ~ LOGIT model 
 
In this study, customer satisfaction is the categorical dependent variable. Hence binomial function is used. 
 

glm (formula = CS ~ 1, family = Binomial(), data = COM_train)
 

A test statistic is used to build a model to identify the probability of success of all the predictor variables through backward stepwise transformation. 
 
The sample statistic of the output table (Table 4), drawn from the population ranges, lies between a min.-1.8528 and a maximum value of 0.6294. Inter- Quartile range one and 3Q is 0.6294 with a median of the same value, meaning there is a standardisation between observations scattered towards the central. The sample is approximately normally distributed. (Dispersion parameter for binomial family taken to be 1).
 


In the model base’s summary output table (Table 5), there is no predictor variable. Hence, null and residual deviance have the same values of 445.51 on 472 degrees of freedom. In the base model, no predictor variable is used. In the base model’s summary output of the ANOVA table, there is no predictor variable. Hence, null deviance and residual deviance have the same values. 

Base: log ( Odd ratio of food habit non-veg.-1)= -0.5447272


 
The intercept is <2e-16***, corresponding to the log odds for the target customer satisfaction variable. The logit of odds can be converted back to an odds ratio by taking the exponent of intercept exp( -0.5447272)= 0.58:1, 
Odd ratios 0.58 can also be converted back to probabilities

odds = p (1-p)
 

p=0.58/1.58= 0.3670886

 
The p-value of (Wald-X2) is 0.367, which is not significant and different from zero(0). In other words, we can say that the food habit of vegetarian consumers is 0.93 and non-vegetarian is 0.07 are no significant effect on customers satisfaction level about milk consumption.
 
Model: Fit
 
In this model, customers satisfaction =
 

β₀ + β₁ (FOMI) + β₂ (Food habit) + β₃ (PROM) + e

 
Where Frequency of milk intake (FOMI), Food Habits and price of milk (PROM) are included as predictor variables in the model.
 

H₀: β₀  = β₁ = β₂ = β₃ 0

 
The sample statistic from the population range lies between a min.-2.2359 and a maximum value of 0.8971. Inter- Quartile range 1 is 0.4139 and 3Q is 0.7027 with a median of 0.5421, meaning there is a standardisation between observations scattered towards the central. The sample is positively skewed and has approximately a normal distribution.
 
From the output table (Table 6), the p-value of (Pr (>|z|) of the explanatory variable PROM and FOMI is less than equal to 0.05 and the p-value of variable food habit is more than 0.05 in the logistic linear model fit. So, the H0 is rejected and suggests a significant effect of two variables on the dependent variable CS. The null deviance of model fit is 335.03 on 362 degrees of freedom and residual deviance is 319.87 on 359 degrees of freedom. The residual deviance is lower than the null deviance. Hence the deviance value is considered a good indicator and the model is valid but needs more transformation. 
 


Model: Fit 1
 
In this model, customers satisfaction (CS)  =

β₀ + β₁(FOMI) + +β₂ (PROM)+ E

 
Where frequency of milk intake (FOMI), Price of milk (PROM) are included as predictor variables in the model.
 

H₀: β₀ = β₁= β₂= 0

 
The sample statistic drawn from the population ranges lies between a min.-2.1929 and a maximum value of 0.8663. Inter- Quartile range 1 is 0.4351 and 3Q is 0.6960 with a median of 0.5526, meaning there is a standardisation between observations scattered towards the central. The sample is positively skewed and has approximately a normal distribution.
 
From the output table (Table 7), the p-value of (Pr (>|z|) of the explanatory variables PROM and FOMI is less than equal to 0.05 and all the predictor variables have significantly affected the target variables. Hence, the H0 is rejected and suggests a significant effect of predictor variables on the dependent variable CS. The null deviance of model fit is 338.91 on 364 degrees of freedom and residual deviance is 326.26 on 362 degrees of freedom. The residual deviance is lower than the null deviance. Hence, the deviance value is a good indicator and the model is valid and robust. 


 
From a given set of tain_data, Akaike Information Criterion (AIC) estimates the relative quality of statistical models and out-of-sample prediction error. AIC estimates the quality of each model relative to each other models. The threshold value for AIC is lower among all the tested models, considered the best model. The AIC of the base model is 447.51, the model fit is 335.03 and model fit1 is 332.26. So, from the above, all the statistics AIC of model fit1 is lower and considered the best fit for the data.
 
LRT (The likelihood ratio test) measures whether a model best fits a given data set if it demonstrates an improved model with fewer predictors to compare the existing model (fit1) with the base model. Moreover, the log difference between the current and base model is significantly different. The probability value of the model fit1 is less than five per cent in the given output table (Table 8). Hence the null hypothesis is rejected and provides evidence against a base model in favour of existing model fit1 for consideration as the goodness of fit and highly significant. 


 
In the case of linear regression, the proportion of variance is explained by the response variable and predictor variables termed R2. On the other hand, in logistic regression, the functional form of the equation contains a logarithmic function, So instead of pseudo-R2, in the case of Multiple Logistic linear regression model R2. McFadden’s R2 is used to define as 1- [ ln (LM) / ln (LO) ] where ln (LM) is the log-likelihood value for the fitted model and ln (LO) is the log-likelihood for the null model with only an intercept as the predictor. The measures range from zero to just under one, with a value close to zero indicating that the model possesses no predictive power. From the output table (Table 9), McFadden’s R2 is 0.03, ranging between zero to one and above 0.2 is considered satisfactory and variables can explain the proportion of the variance in the dependent variable.


 
From the output table (Table 10), Hosmer Lemeshow Statistic is a chi-square test statistic used to measure the predicted model is not significantly different in their observed values with a desirable outcome. It tests the goodness of fit if the probability value of the predicted model fit 1 in this study is not less than a five per cent level of significance and the null hypothesis fails to be rejected. Model fit 1 indicated that the p-value of Hosmer Lemeshow’s goodness of fit is 0.8072, which is greater than the 5% significance level and fails to reject the null hypothesis. This output would suggest no difference between observed and predicted values of model fit1 and adequately fitted with the data.


 
From the output table (Table 11 ), it is observed that the variable price of milk (PROM) strongly associates with other variables and is relevant in the logistic linear regression model fit1. 


 
Confusion Matrix or classification table statistic is used to measure how well the model predicts the target variable in a tabular form to represent the actual and predicted values to determine the model’s accuracy (Table 12). For the given logistic linear regression model fit 1, the recall matrix and sensitivity matrix for the actual and predicted values is 0.9918 with a cutoff of 0.3 on the train data, which is considered goodness of fit. 
From the logistic linear equation for model,
 

glm (formula = CS ~ PROM + Fscore, family = Binomial (logit), + data = COM_train) 
 

 
The output for manual transmission where probability is coded as 1. To understand the customer satisfaction level with observed variables, the price of milk is 0.47 and the F-score, i.e., frequency of milk intake is -0.17, which has a negative linear relationship (Table 13). Probability is being fitted about 95% of the manual transmission in model fit1 from the graph plotted (Fig 3).

On the basis of previous studies and present findings of the study, in India, consumers are generally price sensitive and focus on the quantity of the milk. They are consuming milk and dairy products for nutritive value  instead of focusing on  quality of milk products, fat content, and organic products.

All authors declare that they have no conflicts of interest.