1- Department of Social Medicine, Faculty of Medicine, Babol University of Medical Sciences and Health Services, Babol, Iran
2- Social Determinants of Health Research Center, Health Research Institute, Babol University of Medical Sciences, Babol, Iran
2- Social Determinants of Health Research Center, Health Research Institute, Babol University of Medical Sciences, Babol, Iran
Abstract: (3 Views)
Introduction and Objective: According to the National Document on Medicinal Plants and Traditional Medicine, ratified in 2013 (1392 SH) and established as a legally binding document for the government, medicinal plants were projected to capture a 20% share of the country’s pharmaceutical market by the year 2025 (1404 SH). As this planning horizon has now elapsed, a review of household income-expenditure baskets reveals that the contribution of these products to treatment remains below 2%. This failure may stem from an inadequate understanding of household health- and treatment-seeking behavior, as well as from the failure to employ appropriate stimuli. Prior studies indicate that the consumption of medicinal plants is a function of health-related variables and the level of individuals’ knowledge and belief in their effectiveness. Economic variables exert their influence indirectly, through changes in purchasing power. Given the interactive effects of such stimuli, their net impact remains ambiguous. For instance, rising income may increase consumption by enhancing purchasing power; conversely, medicinal plants may be perceived as inferior goods, in which case the relationship is inverse. Moreover, the relationship between medicinal plant consumption and the aforementioned stimuli may itself depend on the level of consumption, since the effect of the studied variables is unlikely to be uniform across high- and low-consumption households. The objective of this study is therefore to examine the relationship between pharmaceutical expenditures and the economic, social, and health-related variables of households.
Materials and Methods: The research data were drawn from the urban household income-expenditure database of the Statistical Center of Iran for the year 2025 (1404 SH). In total, data from more than 16,500 households were utilized. The variables affecting household consumption behavior were analyzed using the raw income-expenditure data of urban households provided by the Statistical Center of Iran. The methodology combines a neural network within the machine learning framework with quantile regression. The large volume of data, the need to remove outliers, and the identification of the most influential variables on medicinal plant consumption prompted the use of machine learning in the first stage. At this stage, following data extraction and cleaning, the data were randomly split into training (80%) and test (20%) sets. Machine training may be performed using various algorithms; in this study, a feedforward neural network was designed using the Keras library built on the TensorFlow backend. The general network architecture consisted of one input layer, three hidden layers, and one output layer. Once the objectives of the first stage were achieved, the causal relationship was examined using quantile regression. In the first stage, the variables of expenditure on herbal medicines, expenditure on physician-prescribed and non-prescribed drugs, household size, and household income were employed. After the first stage, the two most important variables were incorporated as predictor variables in the quantile regression model.
Findings: In the first stage, following training, the model was evaluated on the dataset using loss and mean absolute error (MAE) criteria. The descending trend of loss and error in both the training and validation curves indicates that the highest learning rate occurred within the range of zero to 20 epochs. The steep slope of the curve demonstrates that the model was capturing the principal patterns in the consumption data. Within the 40 to 50 epoch range, the curves flattened substantially, indicating convergence, whereupon further training with the existing data is unlikely to yield meaningful improvements in accuracy. The results of the first stage revealed that the economic and social status of households had a greater influence on medicinal plant consumption than their health status. Collectively, income and household size accounted for more than 70% of the relationship with consumption. Thus, economic and social variables exerted a stronger influence on the consumption of these plants than did health-related variables. Furthermore, the quantile regression estimates across various consumption percentiles indicated that the effect of these variables is not uniform across consumption levels, with the influence of income and household size being greater at higher percentiles.
Conclusion: Based on the obtained results, the market for medicinal plants is heterogeneous, and a uniform stimulus is not beneficial for the entire market. Segmenting the market into high-consumption and low-consumption groups enables marketers, manufacturing and distribution companies, and the government to utilize distinct tools for market development. While policies such as subsidizing these products would lead to increased consumption in the high-consumption market, improving insurance coverage and advertising (information dissemination) could be more impactful for low-consumption segments.
This study demonstrated that combining deep learning with causal functions can provide superior explanatory outcomes. Although deep learning models can be effective tools for analyzing complex relationships and offering more accurate predictions, they possess limited interpretability and causal analytical capabilities. Furthermore, their ability to measure the effect of explanatory variables on different levels of the response variable (medicinal plant consumption) is constrained. The estimation results indicated that both income and household size variables have a positive effect on consumption levels across all consumption percentiles. However, the magnitude of this impact is greater in high-consumption households than in low-consumption households. Consequently, the market for medicinal plants is heterogeneous, and uniform stimuli are not advantageous for the entire market. Differentiating the market into high- and low-consumption segments allows marketers, producers, distributors, and policymakers to employ varied strategies for market expansion. Specifically, while policies like product subsidies would stimulate consumption in the high-consumption segment, improved insurance coverage and public awareness campaigns (information dissemination) could prove effective for the low-consumption segment. Future studies could leverage richer datasets and employ more sophisticated models such as Recurrent Neural Networks (RNNs) or Reinforcement Learning (RL) models. This research represents a step towards utilizing advanced artificial intelligence tools to better understand human behaviors within the domains of health and economics.
Materials and Methods: The research data were drawn from the urban household income-expenditure database of the Statistical Center of Iran for the year 2025 (1404 SH). In total, data from more than 16,500 households were utilized. The variables affecting household consumption behavior were analyzed using the raw income-expenditure data of urban households provided by the Statistical Center of Iran. The methodology combines a neural network within the machine learning framework with quantile regression. The large volume of data, the need to remove outliers, and the identification of the most influential variables on medicinal plant consumption prompted the use of machine learning in the first stage. At this stage, following data extraction and cleaning, the data were randomly split into training (80%) and test (20%) sets. Machine training may be performed using various algorithms; in this study, a feedforward neural network was designed using the Keras library built on the TensorFlow backend. The general network architecture consisted of one input layer, three hidden layers, and one output layer. Once the objectives of the first stage were achieved, the causal relationship was examined using quantile regression. In the first stage, the variables of expenditure on herbal medicines, expenditure on physician-prescribed and non-prescribed drugs, household size, and household income were employed. After the first stage, the two most important variables were incorporated as predictor variables in the quantile regression model.
Findings: In the first stage, following training, the model was evaluated on the dataset using loss and mean absolute error (MAE) criteria. The descending trend of loss and error in both the training and validation curves indicates that the highest learning rate occurred within the range of zero to 20 epochs. The steep slope of the curve demonstrates that the model was capturing the principal patterns in the consumption data. Within the 40 to 50 epoch range, the curves flattened substantially, indicating convergence, whereupon further training with the existing data is unlikely to yield meaningful improvements in accuracy. The results of the first stage revealed that the economic and social status of households had a greater influence on medicinal plant consumption than their health status. Collectively, income and household size accounted for more than 70% of the relationship with consumption. Thus, economic and social variables exerted a stronger influence on the consumption of these plants than did health-related variables. Furthermore, the quantile regression estimates across various consumption percentiles indicated that the effect of these variables is not uniform across consumption levels, with the influence of income and household size being greater at higher percentiles.
Conclusion: Based on the obtained results, the market for medicinal plants is heterogeneous, and a uniform stimulus is not beneficial for the entire market. Segmenting the market into high-consumption and low-consumption groups enables marketers, manufacturing and distribution companies, and the government to utilize distinct tools for market development. While policies such as subsidizing these products would lead to increased consumption in the high-consumption market, improving insurance coverage and advertising (information dissemination) could be more impactful for low-consumption segments.
This study demonstrated that combining deep learning with causal functions can provide superior explanatory outcomes. Although deep learning models can be effective tools for analyzing complex relationships and offering more accurate predictions, they possess limited interpretability and causal analytical capabilities. Furthermore, their ability to measure the effect of explanatory variables on different levels of the response variable (medicinal plant consumption) is constrained. The estimation results indicated that both income and household size variables have a positive effect on consumption levels across all consumption percentiles. However, the magnitude of this impact is greater in high-consumption households than in low-consumption households. Consequently, the market for medicinal plants is heterogeneous, and uniform stimuli are not advantageous for the entire market. Differentiating the market into high- and low-consumption segments allows marketers, producers, distributors, and policymakers to employ varied strategies for market expansion. Specifically, while policies like product subsidies would stimulate consumption in the high-consumption segment, improved insurance coverage and public awareness campaigns (information dissemination) could prove effective for the low-consumption segment. Future studies could leverage richer datasets and employ more sophisticated models such as Recurrent Neural Networks (RNNs) or Reinforcement Learning (RL) models. This research represents a step towards utilizing advanced artificial intelligence tools to better understand human behaviors within the domains of health and economics.
Send email to the article author
| Rights and permissions | |
 |
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. |
