Eating to live or living to eat? Exploring the link between calorie satiation, Bennett’s law, and the evolution of food preferences

Abstract

As people become wealthy, the type of food they consume tends to fundamentally change. Bennett’s law states that the budget share of starchy food staples, such as cereals and rice, tends to decline as household income rises, while the budget share of other foods, such as meat and vegetables, tends to rise. We argue that an important factor behind this shift is the extent to which the caloric value of the household’s diet meets their energy needs. When the caloric value of the diet is too low, the demand for calories is unsatiated and household’s will ‘eat to live’ and prefer to consume more calorie-dense staple foods, relative to other foods. If the calorie intake exceeds daily energy needs, satiated consumers will ‘live to eat’ and prefer to dedicate additional income to consuming more non-starchy foods that satisfy other wants, such as taste, novelty, or social status. Using Sri Lankan household data, we employ a finite mixture model (FMM) to analyze how satiated and unsatiated groups of consumers alter their calorie intake as income rises. We find empirical evidence that supports this conjecture and suggests that calorie satiation plays an important role in enabling food preferences to evolve as income rises. Policy implications are also discussed.

1 Introduction

As people become wealthy, the type of food they eat tends to fundamentally change. Bennett’s law states that the budget share of starchy food staples (such as cereals and rice) tends to decline as household income rises, while the budget share of other foods (such as meat, sweets, and vegetables), tends to rise (Bennett 1941; Gouel and Guimbard 2019; Timmer et al. 1984). At low levels of income, hunger compels consumers to dedicate a major share of their food budget to consuming calorie-rich staples. This decline in the budget share of starchy food staples mimics the decline in the overall share of food expenditure that occurs on a more aggregate level, which is known as Engel’s law (Chai and Moneta 2010). Empirical evidence for Bennett’s law can be found in studies of the ‘nutrition transition’ that track structural shifts in the composition of food spending across many developing countries (e.g., Drewnowski and Popkin 1997; Law et al. 2020). A key focus of these studies is to understand what factors accelerate or inhibit the nutrition transition as income rises. These factors include geographic location, availability of imported foods, culture, and eating habits (Fourat et al. 2018; Landy 2009).

According to evolutionary consumer theory, the satiation of basic needs stimulates structural change in consumption patterns (Witt 2001; Lades 2013; Chai 2017;). Demand satiation describes a threshold level of consumption that occurs when per-capita quantity consumption of a good (with a fixed set of characteristics) ceases to rise beyond a particular level even as household income continues to grow. Slowdowns in per-capita demand reflect changes in consumer preferences that occur because the underlying needs of consumers that originally motivated consumption become satiated. The most prominent example of demand satiation is the case of food consumption used to satisfy the need for hunger (Ruprecht 2005; Manig and Moneta 2014).

Given the role that hunger likely plays in driving consumer preferences for starchy food at low-income levels, we argue that one potential contributing factor to Bennett’s law (and the nutritional transition) is the caloric energy needs of consumers and the biological systems that regulate feelings of hunger and satiety. Here it is worth noting that the main empirical focus of the nutrition transition literature has been thus far on measuring the income elasticity of starchy and non-starchy food (e.g., Drewnowski and Popkin 1997; Law et al. 2020). However, to date studies have not directly considered how the composition of the household food budget tends to change as the estimated calories consumed by the household increase. As will be discussed in Section 2, this is an interesting question because, while it is indeed true that richer households are able to consume more calories, there are number of factors that mediate the relationship between the household’s total calories consumed and its income. It is precisely this gap in the research that this paper seeks to address.

Based on these insights from evolutionary consumer theory, we empirically investigate the following hypothesis: If the caloric density of the household’s diet meets its daily energy needs, the household demand for calories is satiated and households cease to increase their calorie intake (Jensen and Miller 2010; Zhou and Yu 2014).¹ Reaching satiation enables consumers to dedicate further increases in income to spend to satisfy other needs. We hypothesize that if the calories derived from the diet are below the satiation threshold and do not meet daily energy needs, it is likely that Bennett’s law will not be observed. This is because consumers who do not attain enough calories to reach the subsistence threshold tend to experience adverse physical sensations of hunger, headaches, dizziness, and loss of energy (Warrilow et al. 2019). Bennett’s law is likely observed only if the calories derived from the diet are above the satiation threshold and do meet daily energy needs. That said, additional calories may be consumed for other reasons, for example, to satisfy other needs such as a demand for novelty which can be met by eating stimulating food² (Scitovksy 1976; Bianchi 2002; Chai 2012;).

A major challenge that we confront in empirically identifying these effects is that the precise level of calorie intake required for consumers to meet their energy needs is highly variable (Dasgupta 1995; Svedberg 2000), and influenced by range of complex factors such as age, sex, height, weight, health condition, physical activity, fitness level, and metabolic rate (Jensen and Miller 2010). Following Zhou and Yu (2014), we overcome this challenge by using a finite mixture model (FMM) to examine unobserved heterogeneity in the data by assuming the existence of two latent groups: unsatiated and satiated consumers. The FMM enables us to characterize these groups and study the empirical relationship between calories consumed and income separately for each of these groups.

Our results show that the unsatiated group of households in Sri Lanka tends to be relatively poor and spend most of their additional income on food. The estimated income elasticity of calories for this group is large: marginal increases income generates large increases in per-adult equivalent daily calories consumed. In contrast, the satiated group of households shows that they tend to be relatively wealthy, possess a relatively lower income elasticity of calories: marginal increases in income generate small or negligible increase in per-adult equivalent daily calories consumed, suggesting that calorie demand is satiated. The mean per-adult equivalent daily calories consumed for the unsatiated group was 1,351 calories, which is approximately half of the 2,897 calories mean for the satiated group. The share of spending on non-starchy food, such as meat and vegetables, was observed to be lower for the unsatiated group than for the satiated group. Taken together, this represents new evidence that calorie demand satiation plays a role in shifting household dietary patterns from starchy to non-starchy foods.

To further empirically verify the relationship between calories consumed and Bennett’s law, we employ an OLS regression to examine how the propensity of households to substitute away from starchy foods is correlated to the calorie density of the household diet. We hypothesize unsatiated households who consume relatively fewer calories have a lower propensity for households to substitute away from starchy food towards non-starchy food. Our results show a positive and significant correlation between the calorie density of household diets (the independent variable) and the budget share of non-starchy food (the dependent variable). This correlation suggests that as the overall density of the diet increases, there is a greater tendency for households to substitute away from starchy foods. It provides further indirect evidence about the role that satiation of demand for calories is linked to Bennett’s law.

The remainder of the paper is organized as follows. Section 2 discusses demand satiation, Bennett’s law, and the links to Engel curves. Section 3 describes the data and results of the analysis of Bennett’s law from the calorie dimension. Section 4 outlines the estimation techniques employed in this study, while Section 5 presents the empirical results from the estimations. Conclusions and policy implications are presented in Section 6.

2 Background

2.1 The concept of demand satiation

The concept of demand satiation is a focal point of evolutionary consumer theory that harnesses key biological concepts to develop new insights into long-run trends in household consumption patterns. Demand satiation refers to periods of stagnant demand growth where per-capita quantity consumption of a good (with a fixed set of characteristics) ceases to rise beyond a particular level even as household income continues to grow (Chai 2017; Lades 2013; Witt 2001). Case studies suggest that slowdowns in per-capita demand reflect changes in consumer preferences because the underlying needs that originally motivated demand growth become satiated. As a result of this satiation, further increases in household income are diverted to satisfy other needs that are not yet met (Witt 2001). This is known as the ‘escaping satiation hypothesis’. These are critical changes in the underlying motivation driving consumption that can lead to changes in the characteristics of goods and services and innovative activity (escaping satiation).

The classic example of demand satiation is the case of food consumption used to satisfy the need for hunger (Manig and Moneta 2014; Ruprecht 2005). Food consumption is innately linked to human biology since households consume food primarily to obtain nourishment and avoid hunger, which is a universal basic need shared by all humans. In cases where consumers fail to attain sufficient calories to reach the subsistence threshold, they tend to experience adverse physical symptoms such as hunger, headaches, dizziness, and loss of energy, as reported by Warrilow et al. (2019). Meeting the subsistence threshold leads to the disappearance of these physiological symptoms, which is linked to the homeostatic processes inherent in human physiology, as documented by Rolls (1986).

The universal importance of the need to avoid hunger and attain nourishment is reflected in the fact that food expenditure among the poorest tends to dominate household spending (Banerjee and Duflo 2007). This homogeneity in spending patterns among the poorest suggests that the need to avoid hunger and attain nourishment is innate and linked to basic homeostatic functioning of humans. As households³ become more affluent, it has been widely observed that the per-capita spending on food tends to decline, i.e., Engel’s law (Chai and Moneta 2010; Clements and Chen 1996). From an evolutionary economic perspective, it is conjectured that Engel’s law occurs because rising income enables capacity to consume sufficient calories and satisfy the basic need for to avoid hunger and attain nourishment (Chai and Moneta 2012; Manig 2010; Ruprecht 2005). As a result of reaching satiation levels of food expenditure, consumers will diversify their spending toward other wants and goods (Chai et al. 2015).

Manig and Moneta (2014) and Ruprecht (2005) have noted that when it comes to empirically studying the satiation of food expenditure that is associated with the reduction of hunger, it makes sense to focus on caloric value of food. Calorie intake is widely used to assess hunger and undernourishment (Jensen and Miller 2010). The calorie content of food ultimately serves to provide an accurate picture of how food effectively translates into reduced hunger. Manig and Moneta (2014) undertook a cross-sectional empirical investigation of contemporary Russian food spending patterns. They found that the average calories consumed per person in a household per week tends to flatten out at about 2,857 calories per day. After reaching this level, their results indicate that increasing income appears not to stimulate increases in average calories consumed.

These studies face a significant empirical challenge in identifying demand satiation by studying the calorie value of household diets. This is because there is a wide variation in calorie satiation levels observed across consumers (Manig and Moneta 2014; Ruprecht 2005), and the precise caloric intake required for consumers to meet their energy needs is highly variable (Dasgupta 1995; Svedberg 2000). Factors such as age, sex, height, weight, health status, level of physical activity, fitness level, and metabolic rate all play a role in determining the required caloric threshold for a consumer (Jensen and Miller 2010). This lack of information about these factors poses a challenge for estimating caloric demand and empirically identifying caloric demand satiation in food consumption patterns.

Two other phenomena that accentuate heterogeneity in the calorie density of household diets are the global obesity pandemic and the growth of food insecurity in some developing countries (Popkin 2011). This has led to an increasingly polarized distribution of calorie diets across households, with some households having access to high-calorie diets while others experience limited access to adequate nutrition. From the perspective of evolutionary consumer theory (Witt 2001), this polarization in the calorie density of the household diet is likely to occur as food becomes associated with satisfaction of other wants, such as the demand for novelty and avoidance of boredom (Chai 2012; Scitovksy 1976). The availability of a wide range of food options, combined with the marketing strategies of the food industry, can lead to the consumption of high-calorie diets by some households. In contrast, limited access to food options and financial constraints can result in households having to rely on lower-calorie diets, thus contributing to the bipolar distribution of diets across households. This increasing polarization in the calorie density of household diets may have significant implications for public health, particularly in relation to the growing prevalence of obesity and its associated health problems.

2.2 Bennett’s law

Bennett’s law states that the share of calories from starchy staples (such as cereals and rice) tends to decline as income rises (Bennett 1941; Gouel and Guimbard 2019; Timmer et al. 1984). At very low-income levels, consumers tend to dedicate a major share of their food spending to eating calorie-rich staples, such as rice or bread. As their income rises, the share of starchy staples tends to decline. Strong empirical evidence for Bennett’s law can be found in studies of nutrition transition that track structural shifts in food consumption patterns across many developing countries (e.g., Drewnowski and Popkin 1997, Law et al. 2020).

In terms of what drives Bennett’s law, scholars note that nutritional attributes of food become less important and other attributes of food play a more important role. This includes taste, appearance, status, and convenience (Behrman and Deolalikar 1987; Behrman et al. 1988; Jensen and Miller 2010; Ye and Taylor 1995; Yu et al. 2014). In line with the ‘escaping satiation hypothesis’ (Witt 2001), Tian and Yu (2015) found evidence that people become more interested in other attributes of food, such as taste and variety, rather than calories with income growth. This shift in focus may also explain why periods of fast economic growth are accompanied by a decline in per-capita calorie consumption (Deaton and Dreze 2009), for example, China during the 1980s (Meng et al. 2009). It is worth noting that several studies have found evidence that across the different categories of staples (e.g., rice, wheat, maize, etc.), rising income can also stimulate a substitution process. For example, several studies have found that household expenditure on wheat initially rises when household income rises (Deaton and Dreze 2009; Gandhi et al. 2004; Mason et al. 2015; Mottaleb et al. 2018).

Previous studies in the nutrition transition literature implicitly assume that there exists one uniform single transition pathway through which dietary patterns evolve as household income increases. In other words, scholars typically assume all households tend to change their dietary patterns in the same uniform way. However, considerable evidence has recently emerged that suggests there exists a high degree of heterogeneity in the how dietary patterns evolve as household income rises. For example, several studies show dietary transition paths in India are linked to local cultural and religious norms (Fourat et al. 2018: Landy 2009). This heterogeneity in transition paths raises the possibility that more than one transition pathway exists for a population of households. The importance of considering the possibility of multiple transition pathways is further highlighted by other studies that emphasize how urbanization, shifts in the labor market⁴ and the westernization of societies (Huang and Bouis 2001; Pingali 2007; Popkin 1999; Regmi and Dyck 2001) can influence the nutrition transition process.

A second implicit assumption in the nutrition transition literature is that as income rises, consumers tend to be less hungry (Kearney 2010; Kelly 2016). In other words, it is assumed that as consumers become wealthy, they are more likely to meet their daily energy needs as the calories consumed by the household increase. In reality, the relationship is likely to relatively complicated as minimum caloric intake necessary for a consumer varies based on a multitude of biological factors as noted in previously. To date, few studies have considered explicitly investigated the extent to which this assumption is valid by analyzing the calorie content of households at different income levels. Adopting such an approach can enable scholars to study how the substitution process between staples and non-staple foods may be conditional not only on income, but also on the calorie value of the household diet.

2.3 Calorie satiation and Bennett's law

We hypothesize that calorie satiation plays a role in enabling Bennett’s law to manifest in household dietary consumption patterns. If it is the case that the calorie value of the household’s diet does not meet their daily energy needs, it is unlikely that the budget share of starchy, calorie-dense staples will decline. As mentioned previously, the individuals who do not consume enough calories to reach satiation are likely to experience hunger and other adverse physical sensations of calorie deprivation, which motivates them to seek out and consume more food (Ruprecht 2005; Witt 2001). In such circumstances, their diet will concentrate on the cheapest source of calories, such as staples (Jensen and Miller 2010). We dub this the subsistence stage. These symptoms disappear when the satiation threshold is met (Warrilow et al. 2019). If the calories derived from the diet are above the calorie satiation threshold, meeting this threshold enables Bennett’s law to occur. In other words, once consumers exceed the subsistence threshold, the marginal utility of consuming additional calories declines, and the diet begins to shift towards foods that may be consumed for other purposes, such as taste, novelty and status (Behrman and Deolalikar 1987; Jensen and Miller 2010; Subramanian and Deaton 1996; Yu and Abler 2009; Zhou and Yu 2014). Based on this concept, we define the following two groups of consumers:

• a) Subsistence consumers: If the caloric intake does not meet energy needs, additional income will be used to increase the budget share dedicated to food staples which are a relatively cheap source of calories.

• b) Satiated consumers: If the caloric intake exceeds energy needs, additional income will trigger shifts in their budget share spending on non-starchy food types, which increases the variety of foods consumed.

We argue that a key outcome of consumers reaching the satiation threshold is that they begin to consume a wider variety of non-starchy food. In that sense, when examining the spending diversity of food consumption (Chai et al. 2015), reaching a level of food consumption where the calorie intake exceeds energy needs does not exclude the possibility that the calorie intake will continue to rise. It could be the case that satiated consumers who diversify their diet by expanding it into meat, vegetables and sweets may thereby trigger further increases in the overall calories consumed (Salois et al. 2012). However, it is worth noting that these alternative sources of calories are relatively expensive, compared to starchy staples. Whereas the subsistence group will seek to increase calories by consuming more (relatively cheap) staples, the satiation group may potentially increase calories by consuming more (relatively expensive) non-staple food. For a subset of consumers, reaching the satiation threshold where the calorie intake exceed energy needs may not lead to a slowdown in per-capita calories consumed. Rather, reaching the caloric threshold will trigger a qualitative change in the type of calories consumed by the household.

3 Data

The data are sourced from the Sri Lankan 2016 Household Income and Expenditure Survey (HIES) that was commissioned by the Department of Census and Statistics (2018). This survey adopted a two-stage stratified sample design to obtain a nationally representative sample of 21,756 households covering 25 districts⁵. The district is the primary domain used for stratification. The urban, rural, and estate sectors in each district of the country are the secondary domains. The urban sector refers to all areas governed by either a municipal council or an urban council. The estate sector comprises plantation areas that are larger than 20 acres and have no fewer than ten residential laborers. The rural sector includes all residential areas that are not part of the urban or estate sectors. The survey response rate was approximately 80% in each sector (Department of Census and Statistics 2018), which represents a very high response rate when compared to other consumer surveys. For example, the Living Costs and Food Survey in the United Kingdom reported 50.4% response rate in 2009 (To and McBride 2013).

The food consumption module of this survey includes both expenditure and quantity data on food and beverages consumed by the household. The data were collected from a nationally representative sample over a period of 12 months to capture seasonal variations in income, expenditure, and household consumption. The data include information on household income, expenditure on food and non-food items, as well as demographic characteristics. Additionally, the consumption of home-grown food and donated food was also recorded.

The food data was collected using a 1-week diary. According to Department of Census and Statistics (personal communication, May 31, 2022) at the commencement of the household’s reference week, an official representative provided the household with a diary sheet and completed the expenditure data related to food consumption of that household on the previous day, based on the information given by the household. The household was given specific instructions for completing their diaries for the next 6 days. The enumerator made another visit on day 3, to ensure each household had completed the diary sheet properly. At the end of 7 days, the enumerator visited the household to collect the diary sheet and to perform a brief interview to ensure the data entered were valid. The enumerator was instructed to go over each and every food expenditure sub-category, looking for any anomalies on the value of food based on the quantities the household had entered. If there were any disparities, the enumerator was responsible for revising the values, based on the general prices of such foods in the household’s region. The accuracy of these values was further verified at the data entry point, where the entry process was designed in such a way to avoid extraordinarily large or small values being entered for relevant food expenditure sub-categories. The diary data used in this study have been consistently identified as superior to recall data (Gibson and Kim 2007; Gibson and Rozelle 2002)⁶. As Sri Lanka has a high literacy rate of 93% (Central Bank of Sri Lanka 2021), the problems associated with the accuracy of the diary entries was also expected to be low.

For this study, we included more than 90% of the food items that are available in HIES data as these items are the most commonly consumed by many Sri Lankans and represent their typical eating patterns. Then, we aggregated these food items into 16 food sub-groups by following the food categories used in HIES⁷. Household daily food availability was derived by dividing weekly food quantity data by 7. To quantify the household daily caloric availability, the quantity of food each household consumed was mapped into calories. We used a food composition database compiled by the Biodiversity for Food and Nutrition (n.d.) in Sri Lanka to convert the quantity consumed of each food item by each household into its equivalent calories. These tables have been widely used in several papers (e.g., Lokuge Dona et al. 2018) as they cover a wide range of unique Sri Lankan food types that are not found in other recent international food composition databases. The total quantity of calories consumed was then aggregated at the household level. The total calories are computed as C = ${\sum }_{i}{w}_{i}{k}_{i}$ $,$ where ${w}_{i}$ is the quantity of item i and ${k}_{i}$ is the weight to calorie conversion factor for the same item⁸. The calories for a particular food group J, are derived for food items i ∈ J (Skoufias et al. 2011). For example, a household that reported consuming 250 g of chicken is assigned 272.5 calories (or 100 g of prawns is equivalent to 109 calories).

Finally, we computed the per-adult equivalent daily calorie consumption by dividing the actual daily caloric intake of each household by the appropriate FAO adult equivalent conversion factors as cited by Claro et al. (2010).⁹ These factors were developed based on information on food consumption and energy requirements from various countries, including developing countries. As a result, they provide a more accurate reflection of the dietary requirements of the Sri Lankan population, taking into account differences in food requirements, and nutritional adequacy of diets. Moreover, these conversion factors are widely used in several studies (Akerele et al. 2013; Akter et al. 2020; Azeem et al. 2018, 2019; Do et al. 2019; Eini-Zinab et al. 2021; El‐Shater et al. 2016; Karageorgou et al. 2018; Katungi et al. 2018; Piple et al. 2015; Rodrigues et al. 2020; Ryckman et al. 2021). These adult equivalent units enable us to take into account differences in the household composition and provide an expression of household food intake in proportion to the calorie needs of each individual based on a particular physical activity level (Weisell and Dop 2012). Due to the lack of information, other factors that influence the distribution of food among family members are ignored, such as social status, gender, age and parenting styles (Harris-Fry et al. 2017).

We considered all 21,756 households from the 2016 HIES sample. The data was cleaned by plotting observations and eliminating observations that reported extremely high household daily calorie consumption based on food groups following previous studies (Steffen and Yu 2018; Lovon and Mathiassen 2014; Wiesmann et al. 2009). The selected cut-off values for cereals, other staples, leafy vegetables, and fruit were 8,000, 7,500, 8,000, and 4,000 kcal, respectively. Observations with negative values and extraordinarily high monthly household income were removed. Additionally, observations with households larger than eight people were also removed. Moreover, we identified calculation errors in our dataset and excluded observations where the share of expenditure on starchy staples and non-starchy foods were both equal to zero. After this process, 20,082 households remained in the sample.

3.1 The calorie distribution and Bennett’s law

As a prelude to the main objective of this study, this section presents some stylized facts about calorie distribution and Bennett’s law. Figure 1 shows the kernel density estimates of the log of per-adult equivalent calorie consumption. The distribution is roughly normally distributed and shows no skew, indicating that where the majority of the data falls in the middle range of calorie intake, with fewer individuals consuming very high or very low amounts of calories.

Fig. 1

Per-adult equivalent daily calorie consumption distribution (2016)

Table 7 in Appendix 1 provides descriptive statistics of the sample. The mean per-adult equivalent daily calorie consumption was 1,654.67 kcal, which is considerably below the recommended average minimal daily calorie requirement of 2,030 kcal per person for Sri Lanka, determined by the Medical Research Institute of Sri Lanka (Department of Census and Statistics 2003). Approximately 75% of households were below and 25% of households were above the standard minimum daily requirements. The mean per-capita income for the sample was 13,229.78 Sri Lankan Rupees (SLR) per month, which is slightly lower compared to the census average per-capita household income of 14,473 SLR per month (Department of Census and Statistics 2018). Across Sri Lanka, households spent 39% of their total expenditure on food. This is quite typical for developing countries that spend 50% or more of their total available income on food (Banerjee and Duflo 2007; Qaim 2019). The mean per-capita expenditure was 12,438.35 SLR per month. The average household size in the sample was 3.97. This is slightly higher than the national average household size of 3.8 persons (Department of Census and Statistics 2018). About 74% of the households were headed by males whereas 26% were headed by females. The majority of the households in the sample were drawn from the rural sector (84%) and the rest were urban (16%). About 20% of households has access to social safety nets. Moreover, farming is a major occupation for about 16% of the sample while 84% are engaged in non-farming activities.

Table 8 in the Appendix 1 provides data on per-adult equivalent daily calorie intake, the average cost of 1,000 calories, and food diversity at the sectoral level and across income quartiles. It shows that per-adult equivalent daily calories are 1,654.67 on average and 1,514.65 and 1,912.10 in the two extreme quartiles, respectively. The average of 1,514.65 calories for households in the bottom quartile is significantly below the recommended average minimal daily calorie requirement for Sri Lanka (Department of Census and Statistics 2003), suggesting that much needs to be done to enhance the calorie consumption of poor households. Calorie consumption in urban households was slightly higher than their rural counterparts.

This data also enables us to calculate the average cost of calories consumed by the household. This is calculated by dividing the total amount that the household spent on each food category by its calorie value. The average cost of calories varied significantly by income quartiles and geographic location, as shown in Tables 8 and 9 in Appendix 1. Comparing the average cost of calories across poor and rich households in Table 8, rising income tends to stimulate change in the diet that results in household spending more per calorie consumed. This suggests that wealthy tend to consume higher-quality foods that are relatively more expensive. This is consistent with most of the studies conducted in various countries, which have shown that the average cost of calories among poorer households is cheaper relative to wealthier households (Abdulai and Aubert 2004; Hoang 2009; Skoufias et al. 2011). It is also worth noting that the average cost of calories is lower among rural households relative to urban households.

We also measure the diversity of the household diet using the Simpson index. This index can be written as I= 1-${\sum }_{i}^{n}{w}_{i}^{2}$ $,$ where ${w}_{i}$ is the calorie share of food group i. The Simpson index ranges from 0 to 1. The higher the index, the more diversified the diet. The Simpson index of the households in the poorer quartile is 0.56, while it is 0.61 for those in the wealthiest quartile. This indicates that wealthy households consume more diversified diets than poor households, which is consistent with previous studies (Thiele and Weiss 2003).

Figure 2 below shows the non-parametric estimates of log of per-adult equivalent daily calorie consumed and log of per-capita expenditure. This large positive slope at low incomes indicates some support for the notion that poor households spend more of their extra income on food and thus calorie consumption rises rapidly with income. At high income levels, calorie consumption becomes saturated, and consumers shift spending to other attributes of food, such as its taste and quality (Behrman and Deolalikar 1987; Jensen and Miller 2010; Salois et al. 2012; Ye and Taylor 1995; Zhou and Yu 2014).

Fig. 2

Non-parametric estimates of log of per-adult equivalent daily calorie consumed and log of per-capita expenditure

Table 9 in Appendix 1 reports the composition of household diets across the income distribution. Consistent with Bennett’s law (1941), the budget share of cereals declines as income increases. The budget share of high-quality protein-rich foods (meat, fish, milk, and milk products), other staples and fruit also grows as income increases. Columns 4–6 present the sources of calories consumed by households over the various food groups. Cereals, other staples, coconuts, and sugar make up about 57.46% of household calorie intake. Cereals are the major source of calories for all households, particularly those in the poorest quartile that obtains 18.63% of its calories from cereals while the wealthiest quartile obtains only 13.89% of its calories from cereals. However, the importance of cereals as a source of calories declines with income growth. Fruits, vegetables, and leafy vegetables contribute a substantial share of calories (12.82% for all households, 11.13% for the poorest quartile and 14.96% for the wealthiest). Moreover, some interesting patterns arise moving from the bottom 25% to the top 25% of the per-capita income in the sample; for example, other staples represent on average 14.35% of the total daily per-adult equivalent calorie intake, but this increases from 13.50% for the bottom to 15.68% for the top 25% of the population. On the contrary, the share of some foods (fats and oils, coconuts, and sugar) decreases from 36.72 to 33.18% from the bottom to the top. Also, the share of some other commodity groups does not vary as much (e.g., yams, vegetables, fish, dried fish, and eggs). These trends are also reported in Fig. 3.

Fig. 3

Food budget share and calorie share of starchy staples, non-starchy foods, and other foods by income quartiles Note: Foods included in the groups: starchy staples (cereals, other staples, yams, and pulses); non-starchy foods (vegetables, leafy vegetables, fruit, meat, fish, dried fish, eggs, and milk and milk products); other foods (coconuts, fats and oils, sugar, and confectionery and short-eats).

Columns 7–9 of Table 9 in Appendix 1 present the average cost of 1000 calories consumed across different food groups. These values show that cereals and pulses provide cheap calories to all households, particularly poor households. At the other extreme, meat, fish, dried fish, and eggs; fruit and vegetables; as well as milk and milk products are a relatively expensive sources of calories. Households in the poorest quartile achieve lower average cost per calorie relatively to wealthy households as the budget share dedicate to cereals is higher. Moreover, poorer households also consume less expensive items within all food groups except for pulses and eggs. Consequently, the average cost of calories consumed by poorer households is consistently lower than the average cost of calories consumed by wealthy household across a wide range of food groups. This suggests that wealthy households tend to possess a higher willingness to pay for other (non-calorie) attributes of food, such as taste and quality. These results are in line with previous studies, such as Abdulai and Aubert (2004), Manig and Moneta (2014) and Skoufias et al. (2011).

4 Empirical methodology

As mentioned above, a challenge to be confronted in this study is that the precise caloric intake required for consumers to meet their energy needs is highly variable across households (Dasgupta 1995; Svedberg 2000). Given the lack of information about these factors posed a challenge for estimating caloric demand and empirically identifying caloric demand satiation in food consumption patterns, following Zhou and Yu (2014), we overcame this challenge by adopting an FMM. Here the random variable is assumed to be drawn from a population that is a mixture of distinct subpopulations/classes (C) in proportions of $\pi$ 1,…..$\pi$C (Conway and Deb 2005). We hypothesize that there are two behavioral groups (subsistence consumers and satiated consumers) that exhibit unique transition patterns. The two groups are described as follows:

$${C}_{q,j}={w}_{q}\left({I}_{j},{X}_{j}\right)+{\varepsilon }_{q,j}$$

(1)

Cj is the per-adult equivalent daily calorie consumption and q = s,h, denotes the subsistence stage (s) and satiation group (h) for a household j. The per-capita income variable (I_j) in the model was selected as a predictor variable for these two latent classes since income is a key variable that is thought to increases in trigger calorie consumption in emerging economies (discussed above). The variable X_j, denotes a vector of other socio demographic factors, including household size, gender and education of household head, number of dependents and household location that influence per-adult equivalent daily calorie consumption. Furthermore, w_q (.,.) is a behavioral functions, and $\varepsilon$ _q,j is the error term. Each observation is treated as a combination of two different segments: subsistence consumers and satiation consumers. These are assigned by a probability $\pi$ _s and $\pi$ _h, with $\pi$ _s + $\pi$ _h = 1. Therefore, each observed per-adult equivalent daily calorie consumption is denoted as:

$${C}_{j}={{\pi }_{s}w}_{s} \left({I}_{j},{X}_{j}\right)+{{\pi }_{h}w}_{h}{\left({I}_{j},{X}_{j}\right)+{\pi }_{s}\varepsilon }_{s,j}+{{\pi }_{h}\varepsilon }_{h,j}$$

(2)

One can assume in Eq. (2) that $\pi$ _h is positively correlated with per-capita income. In contrast, $\pi$ _s declines as income increases. This means that as income rises, the probability that an individual behaves as if they are in the subsistence stage decreases. Alternatively, the probability of satiation stage behavior increases. Therefore, $\pi$ _s or $\pi$ _h could be a parameter representing the calorie consumption behavioral transition pattern. If $\pi$ _h ≤ $\pi$ _s, the subsistence stage dominates; and if $\pi$ _h ≥ $\pi$ _s , the satiation stage starts to dominate. Equation (2) is a FMM with two components. The observed data sample is considered a mixture of populations rather than a single one (Conway and Deb 2005). The FMM specifies a mixed probability density function as follows:

$$f \left(C\left|I, X,\theta \right.\right)={\pi }_{s}f\left(C\left|I, X,{\theta }_{s}\right.\right)+{\pi }_{h}f \left(C\left|I, X,{\theta }_{h}\right.\right)$$

(3)

where f represents the component density, which is considered to be a normal density function; the model then becomes a latent class regression model. The parameter vector is $\varphi =\left(\pi ,{\theta }^{^{\prime}}\right)$, where $\pi$ are mixing probabilities ${\pi }^{T}=\left({\pi }_{s}+{\pi }_{h}\right)$, ${\pi }_{q}>0$, $\sum {\pi }_{q}=1$, and ${\theta }_{q}={\left({\upbeta }_{q}^{^{\prime}},{\sigma }_{q}^{2}\right)}^{^{\prime}}$. The model assumes the prior probability of being a component was the same across all observations to estimate Eq. (3).

Once we have the FMM estimates of the two segments, we can compute the posterior probability of membership in each of the latent classes for each observation using the Bayesian principle, conditional on all observed covariates and outcomes. The posterior probability of one observation assigned to class q is defined as follows:

$$P\left(q\vert C,I,X,\varphi\right)=\frac{\pi_q\,f(C\vert I,X,\theta_q)}{\sum\pi_q\,f(C\vert I,X,\theta_q)}$$

(4)

We estimate the posterior probability of each component for each observation based on the estimation of Eq. (2). Such posterior probabilities decompose the sample into two components by assigning a probability. For instance, a posterior probability of 0.9 for component 1 implies that this observation is a mixture of 90% of the subsistence stage and 10% of the satiation stage. Following Zhou and Yu (2014), we categorize the sample into two components while fixing the threshold at $\pi$ _s ≥ 0.5 in this study.

The covariates included in the model were based on economic theory and previous empirical studies. Per-adult equivalent daily calorie consumption is used as the dependent variable following earlier studies of household calorie consumption (Dawson and Tiffin 1998; Subramanian and Deaton 1996; Tiffin and Dawson 2002). Demographic factors include household size, number of dependents, and household head’s gender and ethnicity. Household human capital is captured by the highest educational achievement of the household head. Economic factors that influence household consumption expenditure decisions include per-capita income. In terms of food expenditure, this was categorized into three broad categories (see Table 2 below):

• a) starchy staples (cereals, other staples, yams, and pulses);

• b) non-starchy foods (vegetables, leafy vegetables, fruits, meat, fish, dried fish, eggs, and milk and milk products);

• c) other foods (coconuts, fats and oils, sugar, and confectionery and short-eats).

Of these, food expenditure share on traditional starchy staples and food expenditure share on non-starchy foods were selected as covariates. In terms of demographic factors, the number of employed members in the household and sector of employment (whether household head is engaged in farming and non-farming) were also selected. We also include a variable indicating whether the household has access to social safety nets. Also, a dummy variable was included representing the household location. Table 7 in Appendix 1 provides a detailed description of dependent and independent variables.

5 Empirical results and discussion

The differences between latent classes are compared to assess whether the class characteristics are consistent with theoretical assumptions before evaluating the estimates of the FMM. Table 1 presents the summary statistics of the selected variables for each class based on the posterior probability after estimating Eq. (2). Class 1 covers 16,128 observations which make up 80% of the data sample, while class 2 contains 3,947 observations and 20% of the sample.

Table 1 Characterization of latent classes

There are substantial differences in per-adult equivalent daily calorie consumption, per-capita income, and per-capita expenditure between two distinct classes in the Sri Lankan context. As expected, class 1 is characterized by features pertaining to the subsistence-driven group, with low per-capita income, low per-adult equivalent daily calorie consumption, low food and non-food expenditure, and large household size with many dependents. Thus, class 1 could be identified as the subsistence consumers as they possess lower per-capita income and a lower level of per-adult equivalent daily calorie consumption; it relies more on traditional starchy staple-based cuisines and usually focuses more on the quantity of calories to meet their energy needs.

In contrast, class 2 is characterized by features related to the satiation-driven group with high per-capita income, high per-adult equivalent daily calorie consumption, high food and non-food expenditure, and small household size with fewer dependents. For example, the mean per-adult equivalent daily calorie consumption in class 2 is almost twice that of class 1. This implies individuals in class 2 have already reached the satiation stage. Also, the average per-capita income and household spending on food and non-food are clearly higher in class 2 than in class 1. Hence, households in class 2 may consume more high-value food products, which are expensive sources of calories. In contrast, households in class 1 may consume more starchy staples, which are the cheapest sources of calories. Moreover, as class 2 has relatively smaller household size and a smaller number of dependents compared to class 1, they are more likely to demand high-value foods and food eaten away from home on a more regular basis.¹⁰

In terms of choosing the number of classes, the FMM estimates are restricted to a two-component model for several reasons. First, the current literature on calorie consumption suggests that individuals follow two distinct calorie demand patterns as income rises (Zhou and Yu 2014). Second, our attempts to estimate the three-component model led to minuscule improvements for both the Akaike information criterion (AIC) and the Bayesian information criterion (BIC) to justify their further analysis. Therefore, our analysis estimated FMM using a two-component mixture: class 1 and class 2. Figure 4 illustrates the distributions of the predicted mean of per-adult equivalent daily calorie consumption for the two groups. The figure clearly shows that class 2 (satiated consumers) has greater calorie consumption variability compared to class 1 (subsistence consumers).

Fig. 4

Density plot of per-adult equivalent daily calorie consumption for the two classes (2016)

In addition, Table 2 presents an analysis of the distribution of calorie shares and average calorie prices for various food items across income distribution over two groups. The findings suggest that poor households, representing the bottom 25% of households, tend to consume a higher percentage of calories from starchy foods in both classes, with cereals being the primary source of calories. Specifically, poor households in class 1 and class 2 obtain 12.13% and 32.28% of their calories from cereals, respectively. However, the proportion of calories from cereals decreases with an increase in income for both classes. On the other hand, rich households, representing the top 25% of households, tend to obtain a higher percentage of calories from more expensive non-starchy foods, such as leafy vegetables, fruits, meat, and dairy products, than poor households in both classes. Furthermore, the results show that, on average, rich households pay a higher price per calorie compared to poor households in both starchy and non-starchy food categories. This trend is more pronounced for non-starchy foods. This suggests that rich households tend to possess a higher willingness to pay for other (non-calorie) attributes of food, such as taste and quality.

Table 2 Calorie consumption and the average cost of calories by two groups and income quartiles

5.1 FMM estimates for calorie intake

Table 3 reports results derived from both the traditional OLS model of calorie consumption and compares these to the FMM results.¹¹ In the OLS model that ignores the existence of the two latent groups, the parameter estimate for per-capita income and per-capita income squared reveals a nonlinear relationship that is consistent with existing studies (Gibson and Rozelle 2002; Skoufias 2003). Additionally, the results reveal that the maximum per-adult equivalent calorie consumption is reached at a per-capita income level of Rs.183,828/month (USD 1,263)¹². In the second OLS model, the results revealed that the coefficient for the interaction between income and budget share of staples was negative and significant. This indicates that as people get richer, they spend less on traditional starchy staples, which is consistent with Bennett’s law. In addition, the negative coefficient for the interaction term between income and budget share of non-starchy staples shows that the effect of income on calorie intake is weaker for households with higher budget shares of non-starchy foods. Further, the results show that the coefficients for the interaction terms between income squared and budget share of starchy staples and non-starchy food are both positive and statistically significant, indicating that as income increases, the positive effect on calorie consumption is stronger when households spend a higher proportion of their budget on these types of food.

Table 3 OLS and finite mixture models for calorie intake

Comparing OLS parameter estimates with the parameter estimates from the FMM model, a number of important differences can be observed. Considering the model estimates of FMM, two latent classes are identified in proportions of 0.71 (class 1 subsistence consumers) and 0.29 (class 2 satiated consumers), respectively. This implies that 71% of households in the sample resemble subsistence consumption, and 29% resemble class 2, satiated consumers. As expected, the parameter estimates indicate that subsistence consumers exhibit a significant and relatively stronger positive response to income growth. In other words, their marginal propensity to consume calories is relatively high. Specifically, the coefficient of per-capita income for subsistence consumers was 13.198, implying that a unit increase in per-capita income led to a 13.198 unit increase in per-adult equivalent calorie intake, holding other factors constant. Generally, the households in class 1, which typically comprise subsistence consumers, prioritize meeting their energy needs through consuming a greater quantity of calories as their income increases. In contrast, households in class 2 (satiated consumers) exhibit a relatively small positive coefficient of per-capita income (2.297) that is statistically insignificant. This indicates that their marginal propensity to consume calories is lower, relative to subsistence consumers (class 1). Further, a Wald chi-square test was conducted to compare the coefficients of income that are significantly different between class 1 (subsistence consumers) and class 2 (satiated consumers). The results of the test showed that the coefficients of income were significantly different between the two classes (chi² (1) = 10.85, p = 0.0010), indicating that the effect of income on calorie consumption varies depending on the level of satiation. This finding further confirms the validity of our results and provides valuable insights into the heterogeneity of household calorie consumption behavior.

As such, the FMM reveals notable differences in the income elasticity of calorie consumption across the groups, which is ignored in the traditional OLS analysis. Given additional income, households who are satiated more likely to shift additional spending into expensive sources of calories. Subsistence consumers, on the other hand, are more likely to increase their consumption of cereals. At the same time, the increase in budget share on traditional starchy staples and non-starchy foods has a positive and significant effect on calorie intake in both classes. However, in both cases, the calorie intake increase is substantially higher for satiated consumers (class 2) compared to subsistence consumers (class 1). These results suggest that satiated consumers are on a different transition path from traditional staples to a modern diet. It is worth noting that not all members of the satiated consumer are rich – some are relatively poor. This can be seen in Fig. 5 also provides a graphical summary of how the predicted mean of per-adult equivalent daily calorie consumption has changed over the income deciles across the two classes. This indicates that there are two distinct calorie demand transition pathways and individuals in class 2 have higher per-adult equivalent daily calorie consumption compared to class 1.

Fig. 5

Boxplots of distribution of per-adult equivalent daily calorie consumption for the two classes by per-capita income deciles (2016)

The FMM results also indicate education and living in an urban area had a positive and significant effect on calorie intake for subsistence consumers (class 1). The positive impact of education on calorie intake in class 1 is consistent with previous research that has demonstrated the role of education in enhancing nutrition knowledge and awareness, leading to improved dietary habits such as calorie consumption (Rathu Manannalage et al. 2022; Savari et al. 2020). Similarly, A high-calorie intake in class 1 urban households compared to rural households may reflect the greater availability and accessibility of food in urban areas, as well as a greater diversity of food choices.

The household size had a negative and significant effect on calorie intake for both classes, suggesting that per-adult equivalent daily calorie intake decreases in larger households. This observation is consistent with previous studies that have explored economies of scale in food consumption (Abdulai and Aubert 2004; Gibson and Rozelle 2002; Jayasinghe et al. 2016), which suggests that larger households may benefit from greater efficiency in food production and preparation, resulting in lower per-capita food expenditures and lower per-capita food consumption.

The results also indicated that farming had a positive and significant impact on energy intake for households in both classes compared to non-farming households. This suggests that households engaged in farming may have better access to nutritious foods, as well as the ability to grow their own food, which can lead to improved dietary diversity and energy intake. In terms of ethnicity, households headed by individuals of Sinhala ethnicity had a positive and significant effect on energy intake compared to other ethnicity category in class 2. Perhaps this is because Sinhala ethnic households may rely more on traditional diets, which have a high proportion of rice and other staple foods. However, in class 1, all considered ethnic groups had a negative and significant effect on energy intake. Moreover, being a beneficiary of a government subsidy program had a negative and significant effect on energy intake in class 1, indicating that households receiving such subsidies may be resource-poor and vulnerable. This finding suggests that government subsidy programs may not effectively address the nutritional needs of low-income households, particularly with regard to energy intake.

To summarize, the results in the previous section suggest that the extent to which the caloric value of the household’s diet meets their energy tends to the household’s nutritional transition pathway. Among subsistence consumers, the transition from starchy cereals to non-starchy foods is much slower. Income growth in this group is associated with a relatively higher increase in calories consumed and relatively small changes in the composition of the diet. Among satiated consumers, the reverse is true: income growth is predicted to generate relatively small changes in calories consumed, but relatively large changes in the composition of the household diet.

5.2 Calorie intake and the budget share of non-starchy food

As an alternative approach, in this section, we now switch to examining the factors that influence the composition of household diets. This enables us to show how the propensity of households to substitute away from starchy foods is correlated to the per-adult equivalent daily calorie consumption. To do this, we estimated a simple OLS regression model where the budget share of non-starchy food is the dependent variable. According to Bennett’s law this should be positively correlated to income, while the budget share of starchy food is negatively correlated to income. The OLS regression results are reported in Table 4. The second model is a robustness check in which the budget share on non-starchy food is replaced with the dietary diversity index. This is feasible as the diversity of food expenditure tends to increase as households switch away from cereals to other types of food. The Simpson index is used as a measurement of household dietary diversity.

Table 4 A model of the household food budget share on non-starchy foods

The results show that per-adult equivalent daily calorie intake has a positive and significant relationship with the budget share of non-starchy food, although the coefficient is relatively small. This suggests that as the overall calorie density of the household diet increases, there is a tendency for households to substitute away from starchy foods. The negative coefficient for the squared term of calorie intake indicates that once a household reaches the calorie satiation stage, the budget share of non-starchy food increases at a decreasing rate due to increases in household spending diversity which allows a diverse range of new goods to enter the consumption basket (Jackson 1984; Prais 1953). Moreover, the study shows that the maximum budget share of non-starchy food is achieved at a per-adult equivalent daily calorie intake of 4,165 kcal.

Per-capita income also has a strong positive relationship with the budget share of non-starchy food, as the coefficient of per-capita income is relatively large and statistically significant at 1%. This implies that households diversify their food spending across different food categories, leading to more diversified diets as they become more affluent. This is consistent with most of the previous studies that discuss the diversity of consumption patterns and their relation to income (Chai and Moneta 2010; Chai et al. 2015; Clements and Si 2018). The negative coefficient for the squared term of per-capita income indicates that at higher income levels, the marginal propensity to consume calories begins to fall due to increases in spending diversity across non-food categories (Theil and Finke 1983; Chai and Moneta 2012; Chai et al. 2015). Additionally, the study reveals that the maximum budget share for non-starchy food is reached at an per-capita income level of Rs.159,443/month (USD 1,095)¹³.

The level of education has a positive association with the dietary transition as more educated consumers are more concerned about nutritional balance and demand more diverse diets (Moon et al. 2002). The living in an urban area has a positive and significant effect on the shift in household dietary patterns away from starchy to non-starchy foods. This is supported by the fact that urban dwellers consume a more diverse diet that is particularly rich in animal proteins and fats compared to rural counterparts due to greater access to a wide variety non-starchy foods and markets (Levin et al. 1999). Also, the growing body of literature provides evidence that urbanization is one of the most important drivers for changing dietary patterns and nutrition (Huang and Bouis 2001; Kearney 2010; Pingali 2007; Popkin 1999; Regmi and Dyck 2001).

In addition, the demographic variable of household size has a significant positive impact on the budget share of non-starchy food, suggesting that larger households tend to allocate a greater proportion of their budget towards non-starchy food than smaller households. The possible explanation could be larger households may have more diverse dietary preferences and nutritional needs, which may require a greater variety of non-starchy foods. The results also indicate that as the number of dependents increases, the proportion of the household budget spent on non-starchy foods decreases. This could be because households with more dependents are more likely to purchase cheaper, more calorie-dense foods that are often high in starchy carbohydrates, rather than non-starchy foods like fruits and vegetables, which tend to be more expensive. Furthermore, household head’s ethnicity is also found to have a significant impact on the budget share of non-starchy food, with Sinhalese and Tamil households allocating significantly less of their budget towards non-starchy food compared to the other ethnicity category. In contrast, Muslim households allocate significantly more of their budget towards non-starchy food. These findings suggest that cultural factors and food preferences associated with each ethnicity may play a role in determining the share of the food budget allocated to non-starchy food. Lastly, being a beneficiary of a government transfer program has a negative and significant effect on the budget share of non-starchy food. This is possible because they are resource-poor and vulnerable households and allocate a substantial share of their food budget to starchy staples leading to less diversified and less nutritious diets.

Additionally, column II presents the results of the robustness check where the dietary diversity index is added into the specification as dependent variable. We can also derive the same conclusions in terms of coefficient signs and their significance in determining the relationship between calorie density of household food diets and dietary diversity index. Overall, these findings provide indirect evidence about the role that calorie intake and satiation play in driving the shift in household dietary patterns from starchy to non-starchy foods.

6 Conclusions and policy implications

This paper has developed a more detailed understanding of how household food diets tend to evolve as household income rises. Using Sri Lankan household data, we have uncovered evidence that shows how the caloric value of the household’s diet and the extent to which it meets their energy needs mediates how rapidly the household’s diet transitions from starchy food to non-starchy food. Using an FMM approach, we have presented evidence that there exist multiple nutrition pathways for different segments of the household population. Roughly 71% of Sri Lankan households were found to be in the subsistence group of consumers for whom the transition from starchy cereals to non-starchy foods is much slower as income rises. Income growth in this group is predicted to generate relatively large increases in calories consumed, and relatively small changes in the composition of the diet. Among satiated consumers (representing 29% of the population), the reverse is true: income growth is predicted to generate relatively small changes in calories consumed, but relatively large changes in the composition of the household diet. Taken together, these results confirm the existence of Bennett’s law. It also suggests that this law is deeply connected to biological mechanisms that regulate the feeling of hunger and satiety that motivate the consumption of starchy food.

How healthy are the identified nutrition pathways among subsistence and satiated consumers? Among the satiated consumers, rising income enables them to incorporate leafy vegetables and fresh fruits into their diets. This suggests satiated consumers may be more aware of the importance of these foods for maintaining a balanced diet and overall well-being. In contrast, subsistence consumers exhibit a greater reliance on affordable staples like cereals, yams, and pulses. These staple foods serve as the foundation of their diet due to their accessibility and lower cost. While these foods can provide necessary nutrients, they may not offer the same level of dietary diversity as the diets of satiated consumers. As a result, subsistence consumers may be more susceptible to deficiencies in certain vitamins and minerals. In terms of whether higher calorie consumption among subsistence consumers leads to higher rates of obesity, it is difficult to draw a conclusion as many of these consumers are involved in farm work and hence have higher daily caloric needs (see Table 3). Promoting the adoption of healthier food choices, particularly among subsistence consumers, is crucial to improve dietary diversity and overall nutrition levels.

This study has some limitations that suggest opportunities for further research. First, we were unable to use individual-level data on caloric consumption, as the food consumption data were only available at the household level. As a result, we used per-adult equivalent calorie intake as a proxy measure. While this measure is not perfect, it was the best available option given the data constraints. Secondly, the HIES survey data did not account for waste, losses, and non-food use within households. In addition, the survey did not provide detailed information about the values of food consumed outside the home, although it adequately captured the values of homegrown and freely received food. Consequently, our analyses did not consider these factors, and obtaining more qualitative and quantitative information on food consumption would improve the accuracy of calorie intake estimations using HIES data.

Furthermore, we investigated how the caloric value of the household’s diet and its ability to meet members’ energy needs affected the transition from starchy to non-starchy foods. However, we recognize that the dietary transition is complex and influenced by a variety of factors. Therefore, it is necessary to use multidimensional indicators to explore the nutrition transition from different perspectives, since it encompasses multiple dimensions and is not solely restricted to calorie intake.

In terms of policy implications, the study suggests that models of the nutrition transition that are used by policymakers to design welfare assistance schemes need to account for the existence of latent classes across households. In a nutshell, there is a complex range of factors that contribute to household dietary patterns. Our results suggest that policymakers should cast aside using a simple rule of thumb, such as the notion that all rich households have enough to eat. Instead, more attention needs to be paid to measuring the energy needs of households and the extent to which their diets are meeting these needs. This study also highlights that policies that raise the income of poor households will have quite a different impact on diets across these two groups. While poor households belonging to the subsistence group will likely use additional income to increase their calorie intake very effectively, the same is not true for poor households belonging to the satiated segment. It could be the case that an increase in the share of poor households belonging to the satiated segment could explain why the demand for calories is falling in many parts of the world (Deaton and Dreze 2009). We leave this question for further study. All in all, a better understanding of the behavior, location and size of these latent classes is important to ensure government policies that aim to improve the nutritional status of the most vulnerable households.

Appendices

Appendix 1

Appendix Tables 5, 6, 7, 8, 9, 10 and 11

Table 5 Household food groups

Table 6 List of variables

Table 7 Descriptive statistics

Table 8 Daily per-adult equivalent calorie intake, average cost of calories and dietary diversity by sector and per-capita income quartiles in the sample

Table 9 Food expenditure pattern, calorie consumption, average cost of calories

Table 10 Correlation matrix

Appendix 2

Table 11 FAO-based adult equivalent conversion factors