Sustainable diets are emerging as a core concept of global development dialogues and have become a cornerstone of achieving food security in an environmentally sustainable way⁽ Reference Burlingame and Dernini ^{1 )}. Dietary diversity is a key measure of sustainable diets and is currently being considered as one of the principal indicators in the post-2015 Sustainable Development Goals (SDG)⁽ Reference Webb ^{2 )}. A recent working paper on nutrition and environmental sustainability by the UN System Standing Committee on Nutrition⁽ Reference Remans, Ahmed and Herforth ^{3 )} provides ten principles to integrate nutrition and environment in food systems. The third principle proffered is to ‘maximize biological diversity at different levels of the food system, in the landscape, the markets, and diets’. The rationale behind this is diversity’s central role in ecological sustainability⁽ Reference Naeem, Duffy and Zavaleta ^{4 )}, ecosystem resilience⁽ Reference Steffen, Richardson and Rockstrom ^{5 )} and human nutrition⁽ Reference Naeem, Bunker and Hector ^{6 ,} Reference Loreau, Naeem and Inchausti ^{7 )}. UN Secretary General Ban Ki-Moon launched the ‘Zero Hunger Challenge,’ calling on all nations to work towards a future where everyone enjoys their right to food and all food systems are resilient. Since a nutritionally adequate diet is central to this right, a better understanding of the role of diversity in diets is a key to addressing this challenge.

Indicators for measuring dietary diversity⁽ Reference Ruel ^{8 , 9 )} have contributed to our understanding of how food variety is associated with nutrient adequacy⁽ Reference Foote, Murphy and Wilkens ^{10 –} Reference Moursi, Arimond and Dewey ^{12 )} and health⁽ Reference Arimond and Ruel ^{13 ,} Reference Rah, Akhter and Semba ^{14 )}, but the focus of these indicators has been on consumption rather than food provisioning. Diets are determined in large part by food availability and accessibility – factors which are influenced by geography, demography, economics and commerce⁽ Reference Kearney ^{15 )}. While the role of complex food systems has been recognized in determining consumer choices and diets⁽ Reference Sobal, Khan and Bisogni ^{16 ,} Reference Godfray, Crute and Haddad ^{17 )}, little is known about food diversity at different stages of these systems or how it is transmitted to dietary diversity at the household level.

These issues become critical for developing countries as policy makers navigate the dietary transition from smallholder farming-based food systems to industrialized agriculture and integration into global food trade. Urbanization, rising incomes, globalization, agricultural intensification and lifestyle changes affect food production and consumption with consequences for the healthfulness of diets⁽ Reference Godfray, Crute and Haddad ^{17 )}. The dietary transition results in the gradual replacement of diets high in localized and biodiverse varieties of legumes, fruits, nuts and wild caught animals and fish with simplified diets high in global commodities and processed foods^{( 18 )}, with important implications for both human health and environmental sustainability⁽ Reference Tilman and Clark ^{19 )}. Policy makers need metrics to understand the impact of agricultural programmes, trade policies, transportation infrastructure and market regulations on dietary diversity and which demographic groups are being impacted.

For food systems to be sustainable, the rate of utilization of natural resources employed in producing and distributing food cannot exceed the capacity of ecosystems to replenish those resources. Food security, of which nutrient adequacy is an essential condition, is the ultimate objective of a sustainable food system^{( 18 )}, but only recently have conservation policies begun to incorporate food security objectives^{( 20 )}. Developing effective policies to support sustainable food systems is hindered by a lack of metrics that are useful in understanding how diets can improve population health while conserving environmental resources⁽ Reference Johnston, Fanzo and Cogill ^{21 )}.

Previous research has demonstrated how the Functional Diversity indicator⁽ Reference Petchey and Gaston ^{22 )}, developed by ecologists to evaluate the impact of biodiversity in natural and managed systems⁽ Reference Flynn, Gogol-Prokurat and Nogeire ^{23 )} and to test the resilience of those systems⁽ Reference Laliberté, Wells and DeClerck ^{24 )}, can be adapted to describe the diversity of nutrients in cropping systems. In so doing, DeClerck⁽ Reference DeClerck, Fanzo and Palm ^{25 )} and Remans⁽ Reference Remans, Flynn and DeClerck ^{26 )} and co-workers provided evidence supportive of the proposition that agrobiodiversity can affect human nutrition, since the addition of key species belonging to distinct nutritionally functional groups on a farm increases the availability of nutrients. Their adapted indicator, called the Nutritional Functional Diversity (NFD) score, has advantages in linking agrobiodiversity to diets: (i) it can be applied at any scale to describe diversity in available nutrients from farm fields to markets to diets; and (ii) it reflects large nutritional differences in groups of foods that are not captured by a food variety score (a count of different foods consumed) while providing a continuous measure of nutritional variations in foods that are not captured by categorical measures such as a dietary diversity score (a count of the number of food groups consumed). A third advantage is that using a trait-based approach permits a methodological link between environmental and nutritional objectives, so that either agro-ecological traits, such as biodiversity of cropping systems, or nutritional traits, such as nutrient content of crops, could be studied using an analogous functional diversity metric in the same farm/household system.

While DeClerck et al.⁽ Reference DeClerck, Fanzo and Palm ^{25 )} and Remans et al.⁽ Reference Remans, Flynn and DeClerck ^{26 )} showed how NFD can be applied to measure the nutritional diversity of different cropping systems, these studies were limited in scope to thirty farms in Kenya and 170 farming households in three East African countries, respectively. Neither study applied the NFD score to foods consumed or examined the role of geography and markets in providing nutritional diversity.

To enable a better understanding of how diversity transmits through food systems, the present study expands the use of the NFD to a nationwide sample and evaluates it at different levels of the household food system. The specific objectives of the study are to: (i) estimate the relative contribution of home production and market purchases in providing nutritional diversity to agricultural households in Malawi; and (ii) examine how this food system provisioning varies by time, space and socio-economic conditions. By developing a methodology that uses the NFD with a commonly available data set, this research allows for subsequent policy making in a wide variety of countries that focuses on enhancing sustainable diets by linking household dietary diversity to agricultural and market sources.

Methods

Study population

The present study is based on data from the 2010–2011 Malawi Third Integrated Household Survey (IHS3), which was conducted by the National Statistics Office of Malawi with support from the World Bank and other donors. The survey was conducted nationwide between March 2010 and March 2011 with surveys conducted in every district throughout the year. The IHS3 used a stratified two-stage cluster sampling design with a total of 768 census enumeration areas, as the first-stage clusters, and 12 271 households in the overall sample. Because the aim of the present analysis is to understand the relative importance of market purchases and home produced foods in providing for dietary diversity, enumeration areas in which none of the households had any agricultural land were eliminated from the analysis. This reduced the total number of clusters in the final analytic sample to 741 and the final sample size to 11 814 households.

Food consumption module

The IHS3 follows the format of many of the World Bank’s Living Standards Measurement Study (LSMS) surveys and contains extensive demographic, agriculture, health, geographic and consumption data. The household food consumption module is an important module for the LSMS, both because it is essential for determining poverty thresholds and because it is the largest source of expenditure for many low-income households; thus, great care is taken in recording an exhaustive list of foods consumed. Enumerators recorded information on foods consumed in the previous week. This included cereals, tubers, plantains, nuts, legumes, vegetables, animal products, prepared foods from vendors, milk and dairy products, sugars and fats. Open-ended questions prompted respondents to specify any additional foods they may have consumed such as wild or traditional food sources not specifically mentioned. For each food consumed, respondents were asked whether the food was purchased, home produced or received as a gift. Foods received as a gift are included in the total NFD score, but are otherwise not examined here due their relatively low contribution to dietary diversity and lack of relevance to market and agricultural policy.

Nutritional Functional Diversity (NFD) score

The Functional Diversity metric was developed by Petchey and Gaston⁽ Reference Petchey and Gaston ^{22 )} to better evaluate the impact of biodiversity on the provisioning of ecosystem services. The metric uses a trait-based approach to quantify biodiversity in a way that does not give equal weight to species that fulfil a redundant function in the ecosystem (e.g. some bees, birds and moths all function as pollinators, even though they are different species). Their approach starts with a matrix of species’ traits that reflects the ecological contribution of each species to specific functions as a basis for measuring breadth of diversity within an ecological community or ecosystem. Ecological communities with a large number of species but a high degree of similarity in traits receive a relatively lower FD score, whereas communities with greater differentiation in traits receive relatively higher scores. For example, a system that produces rice, wheat and maize would have a lower FD score than one that produces maize, beans and squash, since the latter system maximizes trait differences for growth and resource use efficiency. DeClerck et al.⁽ Reference DeClerck, Fanzo and Palm ^{25 )} employed the Functional Diversity metric as specified by Petchey and Gaston⁽ Reference Petchey and Gaston ^{22 )}, but substituted nutrient content of foods for the species’ traits more typically used by ecologists and termed it Nutritional Functional Diversity. We calculated NFD as specified by DeClerck et al.⁽ Reference DeClerck, Fanzo and Palm ^{25 )}. Our methodological contribution is to apply this indicator at the household level to different levels of the household food system: home production, market purchases and overall consumption.

NFD is the extent of functional differences among foods available on a farm, in a market or consumed in one’s diet. These functional differences are based on the nutrient profiles of each food, i.e. the amounts of energy and seventeen different nutrients in a standard amount of each food. The NFD score is a relative measure, with higher scores indicating a more diverse diet. There are four main steps to calculating the NFD indicator as described below.

First, a food–nutrient matrix is created

In this matrix, each row is one of the foods in the IHS3 food consumption module and each of the columns is a nutrient, such that each cell of the matrix gives the nutrient content of each of the foods. After excluding processed foods (alcoholic beverages, soft drinks, meals eaten in restaurants) and those with negligible nutritional value (e.g. bottled water, salt and spices), eighty-seven foods were retained. The food–nutrient matrix was composed of energy and seventeen nutrients: protein, fat, carbohydrate, fibre, β-carotene, vitamin D, vitamin E, vitamin C, thiamin, riboflavin, niacin, folate, vitamin B₁₂, Ca, K, Fe and Zn. A food composition table for Malawi was not available so the nutritional values used came from composition tables for Tanzanian foods, which listed quantities of nutrients per 100 g of food products⁽ Reference Lukmanji and Hertzmark ^{27 )}. Nutrient contents for four foods not found in the Tanzanian food composition table were obtained from NutriBase (CyberSoft, Inc., Phoenix, AZ, USA). The nutrient values in the food matrix were then standardized in two ways: (i) they were divided by the RDA for an adult male^{( 28 )}; and (ii) these were then standardized to have mean=0 and sd=1.

Second, the food–nutrient matrix is converted into a food–food distance matrix

In this distance matrix, each of the rows and columns represents one of the eighty-seven foods, and each of the cells represents the ‘distance’ between a given food i and another food j. Distance is a measure of the difference between two foods, based on their nutrient composition. It is simply calculated as Euclidean distance is calculated in geometry. Specifically:

$$D_{{ij}} =\sqrt {(i_{1} {\minus}j_{1} )^{2} {\plus}(i_{2} {\minus}j_{2} )^{2} {\plus} \cdots {\plus}(i_{{18}} {\minus}j_{{18}} )^{2} } ,$$

where D _i,j is the distance between food i and food j, i ₁ is the amount of standardized nutrient 1 in food i and j ₁ is the amount of standardized nutrient 1 in food j. There are eighteen terms in the formula representing energy and the seventeen nutrients.

Third, the distance matrix is used to produce a cluster diagram, called a dendrogram

A cluster analysis was performed on the distance matrix to group the foods by nutrient similarities into clusters and to assess the distance between each cluster, as well as the distance between foods within a cluster. The group average method (also called the unweighted pair-group method using arithmetic averages, or UPGMA) was employed to do this⁽ Reference Podani and Schmera ^{29 )}. The food clustering and distance information was used to create the dendogram that graphically represents this information and can be used for calculating the NFD score.

Fourth, the dendogram is used to calculate the NFD score

Figure 1 illustrates how the NFD score is calculated from a dendrogram with a simplified example. In Fig. 1, there are five foods that have been clustered into groups. The horizontal lines are called branches and the vertical lines are called nodes. The longer the branches, the greater the functional diversity. Foods C and D are more similar to each other in their nutritional content than foods A and B are to each other, as can be seen by the longer branch lengths that connect A and B to their nodes. Therefore, foods C and D would represent more nutritional redundancy when both are included in a diet than foods A and B.

Fig. 1

A simplified illustration showing how the length of branches (horizontal lines shown in bold) from a dendrogram are summed and then divided by the potential NFD and multiplied by 100 to generate the NFD score (NFD, Nutritional Functional Diversity)

The potential NFD is calculated by summing the branch lengths, but not the nodes, of all foods with each branch only included once. In the example shown in Fig. 1, the total branch length is 14. A hypothetical household, Household X, only consumed four of the possible five foods (A, B, C and E) for a total branch length of 13. The total NFD score for Household X is 13/14×100=92·9. Household X purchased foods A, B and C with a total branch length of 10 for a market NFD score of 71·4 (10/14×100). Household X produced foods B and E for a home production NFD score of 57·1 (8/14×100).

Statistical methods

Mean NFD scores were calculated by geographic and temporal categories for total NFD, market NFD and home production NFD. One-way ANOVA tests were performed to test for differences among categories followed by t tests using the Sidak correction for multiple comparisons to test for differences between categorical levels. Sample characteristics were analysed by inclusion in the top four quintiles of total NFD and the bottom quintile. The Rao–Scott χ ² test was used to compare distributions of categorical variables by the two quintile groups and t tests were used to compare differences in means by the two groups.

A logistic regression model was conducted to provide odds ratios for the likelihood that a household’s total NFD score fell into the lowest quintile of national total NFD scores. The model included variables for demographic characteristics, region, access to markets, agricultural seasons, agricultural landholdings and access to agricultural extension services.

All data operations and statistical analyses were done using the statistical software package SAS version 9·3. Functional Diversity was calculated using PROC DISTANCE, PROC CLUSTER and PROC TREE. All data analyses were done using PROC SURVEYFREQ, PROC SURVEYMEANS, PROC SURVEYREG and PROC SURVEYLOGISTIC. All analyses were weighted using the IHS3 national sample weights and account for the clustered survey sample design in the calculation of standard errors.

Results

Table 1 shows that households in the bottom quintile of total NFD scores had a mean score for total NFD that is half of that of households in the top four quintiles (13·2 and 26·4, respectively). The difference was greater for NFD coming from market purchases, with the lowest quintile having a mean NFD score that is only 42 % of the mean for the top quintiles. NFD scores for home production did not show as much difference, but were still lower for the bottom quintile than the top four quintiles (6·4 and 8·2, respectively).

Table 1

Mean Nutritional Functional Diversity (NFD) scores for total consumption, market purchases and home produced foods by spatial and temporal factors; 2010–2011 Malawi Third Integrated Household Survey (IHS3)Footnote †

^a,b,cMean values within a column with unlike superscript letters were significantly different (P=0·05) using t test adjusted for multiple comparisons.

† All frequencies, means and t tests incorporated sampling weights and design characteristics in the calculation to provide nationally representative estimates.

Table 1 also presents mean NFD scores by geographic and temporal factors. The contributions of markets and home production to total NFD varied across agro-ecological zones, with households in the tropical highlands showing the greatest diversity from home production and the lowest diversity from markets, and households in the sub-humid zones having the least diversity from home production and the most from markets. Households in the southern region accessed the lowest diversity from home production.

As the distance from households to primary and secondary roads increased, households accessed less of their diversity from markets and more from home production for a lower mean total NFD. A similar association was found as the distance increased from households to population centres of greater than 20 000 residents. Those households closest to a population centre had the highest mean total NFD with decreasing mean total NFD as the distance increased. As the distance from a household to a population centre increased, households accessed less diversity from markets and more from home production, although with a decrease in mean home production NFD seen for households the furthest from a population centre.

Households located in villages within 5 km of a daily market had higher market NFD as well as total NFD scores than households in villages located further from daily markets. Households in villages within 5 km of a daily market had lower home production NFD than households in villages further from daily markets.

NFD varied across agricultural seasons. The growing season (coinciding with the rainy season), when farms are planted and tended, had the lowest mean total NFD score. The harvest season, characterized by cool weather and low rainfall, had the highest mean home production NFD score. The post-harvest season (coinciding with the dry season) showed the highest mean NFD score from market purchases and the lowest mean NFD score from home production.

Tables 2 and 3 demonstrate how households with low NFD scores, defined here as the lowest quintile of households based on their total NFD score, differ from households with higher NFD scores.

Table 2

Demographic, geographic and economic characteristics of households included in the study; 2010–2011 Malawi Third Integrated Household Survey (IHS3)

The sample n is the number of households surveyed without weighting, but the percentages were calculated using survey weights which accounts for the differences in distributions.

*P value for t test<0·05 comparing mean of the highest four quintiles with that of the lowest quintile.

**P value for t test<0·0001 comparing mean of the highest four quintiles with that of the lowest quintile.

† Differences between categorical distributions were tested using the Rao–Scott χ ² test. All results were significant at the P<0·001 level.

Table 3

Logistic regression results showing all terms included in the model predicting inclusion of households in the lowest quintile of total Nutritional Functional Diversity scores, 2010–2011 Malawi Third Integrated Household Survey (IHS3)Footnote †

† The model incorporated sampling weights and design characteristics to provide nationally representative estimates.

The frequencies show that there were greater percentages of households in the bottom quintile of NFD scores with a head of household who was older, female, less educated, and separated or divorced. These frequencies show that a greater percentage of households in the bottom quintile contained fewer persons than households in the top four quintiles. A greater percentage of households in the bottom quintile were located in villages without an agricultural extension office. There was no statistically significant difference in the distribution of households in the bottom quintile by either agro-ecological zone or region.

The mean total real annual expenditure per household as measured in Malawi Kwachas for households in the lowest quintile was less than half of that for households in the top four quintiles. Households in the lowest quintile had statistically significantly lower landholdings than households in the top four quintiles by an average of 0·3 acres (0·12 ha). Households in the lowest quintile were statistically significantly more likely to be further from a primary or secondary road, further from a population centre of greater than 20 000 residents and further from a daily market.

The results of a logistic regression model predicting inclusion in the bottom quintile of total NFD scores (Table 3) found that after adjustment for covariates, some of the differences that were significant in Table 2 were no longer significant. Households with a head younger than 30 years of age were less likely to fall into the lowest quintile by 21 % while households with heads older than 50 years of age were more likely to fall into the lowest quintile of total NFD by 35 %. Lack of a primary-school education increased the odds of being in the bottom quintile by 23·9 %. However, there was no significant effect found for households headed by a woman or for households headed by a married person. The logistic regression model found that each additional person in a household was associated with an increase in odds of being in the lowest quintile by 21·8 %. Each additional 1000 Kwachas of real annual household expenditure reduced the odds of being in the lowest quintile by 1·8 %. Compared with the Southern Region, households in the Central Region were 63 % more likely to fall into the lowest quintile while households in the Northern Region were similar to those in the Southern Region.

After adjustment for the covariates in the model, the distance to a daily market was no longer a significant predictor of being in the bottom quintile, but distance to the nearest primary or secondary road was associated with an increased risk of 1·2 % for each additional kilometre.

Households interviewed during the growing season were more than twice as likely to have total NFD scores falling in the bottom quintile (OR=2·2). Landholdings were not a significant predictor of being in the bottom quintile of total NFD, but the presence of an agricultural extension office in the village where the household was located reduced the risk by 17·7 %.