Download Predicting dietary intakes with simple food recall information: a case

European Journal of Clinical Nutrition (2003) 57, 1212–1221
& 2003 Nature Publishing Group All rights reserved 0954-3007/03 $25.00
www.nature.com/ejcn
ORIGINAL COMMUNICATION
Predicting dietary intakes with simple food recall
information: a case study from rural Mozambique
D Rose1* and D Tschirley2
1
Department of Community Health Sciences, Tulane University School of Public Health and Tropical Medicine, New Orleans, LA, USA;
and 2Department of Agricultural Economics, Michigan State University, East Lansing, MI, USA
Objective: Improving dietary status is an important development objective, but monitoring of progress in this area can be too
costly for many low-income countries. This paper demonstrates a simple, inexpensive technique for monitoring household diets
in Mozambique.
Design: Secondary analysis of data from an intensive field survey on household food consumption and agricultural practices,
known as the Nampula/Cabo Delgado Study (NCD).
Subjects: In total, 388 households in 16 villages from a stratified random sample of rural areas in Nampula and Cabo Delgado
provinces in northern Mozambique.
Methods: The NCD employed a quantitative 24-h food recall on two nonconsecutive days in each of the three different seasons.
A dietary intake prediction model was developed with linear regression techniques based on NCD nutrient intake data and easyto-collect variables, such as food group consumption and household size The model was used to predict the prevalence of low
intakes among subsamples from the field study using only easy-to-collect variables.
Results: Using empirical data for the harvest season from the original NCD study, 40% of the observations on households had
low-energy intakes, whereas rates of low intake for protein, vitamin A, and iron, were 14, 94, and 39, respectively. The model
developed here predicted that 42% would have low-energy intakes and that 12, 93, and 35% would have low-protein, vitamin
A, and iron intakes, respectively. Similarly, close predictions were found using an aggregate index of overall diet quality.
Conclusions: This work demonstrates the potential for using low-cost methods for monitoring dietary intake in Mozambique.
Sponsorship: Michigan State University and the Mozambican Ministry of Agriculture and Fisheries.
European Journal of Clinical Nutrition (2003) 57, 1212–1221. doi:10.1038/sj.ejcn.1601671
Keywords: Africa; Mozambique; household study; dietary assessment methods; diet quality index; nutrient intake
Introduction
A well-nourished population is not only important to a
country’s long-term development, but it is also a desirable
outcome objective in itself. Unfortunately, monitoring of
progress in meeting this objective can be expensive, since
large-scale quantitative surveys are time-consuming and
resource-intensive. Obtaining information on dietary intake
*Correspondence: Dr D Rose, Department of Community Health Sciences,
Tulane University School of Public Health and Tropical Medicine, 1440
Canal Street, Suite 2301, New Orleans, LA 70112, USA.
E-mail: [email protected]
Guarantor: D Rose.
Contributors: DR designed this study and performed the data analysis.
DT is co-director of MSU’s Mozambique Food Security Project, under
which both the original field work and this secondary data analysis
were conducted. DR was the lead author and DT was the second
author on the write-up of this paper.
Received 9 May 2002; revised 14 September 2002; accepted 16
September 2002
is especially difficult in Mozambique, as it is in other SubSaharan African countries, because of its large size, the lack
of adequate roads and transportation, the diversity of
languages spoken, and the relative inexperience in fielding
complex dietary questionnaires. Yet the information from
dietary surveys is essential for monitoring and program
design, as well as for geographical targeting to population
groups in need of emergency interventions.
Since a full and accurate assessment of dietary intake is a
costly and time-consuming activity, a number of attempts
have been made to develop relatively simple and inexpensive methods to do this. One such measure uses food variety
to predict dietary adequacy. In Mali, researchers weighed the
food intakes of household members and compared the
nutrients consumed in this food to simple measures of
dietary diversity (Hatly et al, 1998). These researchers found
that the number of different food groups consumed in a
3-day period was useful for distinguishing those with
Predicting dietary intakes in Mozambique
D Rose and D Tschirley
1213
inadequate diets from those with adequate ones. Unlike the
approach in Mali, analysts in Zambia developed a scoring
system that weighted foods differently depending on
the food group to which they belong. For example,
consumption of foods from the nutrient-rich meats group
received four points, whereas those from the cereals group
received two points. After adding up the points from all the
foods consumed in a 24-h period, household diets were
evaluated based on pre-established cut-points (FHANIS/CSO,
1998).
In rural Mozambique, the types of foods, their availability
and nutritional content, as well as the consumption patterns
and nutritional problems in the population are not the same
as those in Mali or Zambia. Neither are the constraints and
opportunities with regards to data collection efforts. A
simplified field tool would be an enormous advantage to
government agencies, donors, and national and international nongovernmental organisations involved in food
security interventions, since all of these organisations lack
the resources to regularly conduct intensive dietary surveys.
Even though there is difficulty in fielding intensive data
collection efforts, there are many local analysts in the
country who are quite practised at using statistical packages
on microcomputers. This human resource in data processing
implies that while field instruments should be kept simple
for data collection purposes, more sophisticated techniques
can be used to process the data, once obtained. Both the Mali
and Zambian tools described above begin with a simple
listing of all foods consumed yesterday. Processing of both
tools is quite simple as wellFeither a simple count of the
number of different foods or a simple summing of points.
Here we develop a tool that is simple to field, but is followed
by more complex processing techniques. This approach takes
full advantage of all the information collected in a simple
instrument as well as the human resources in Mozambique
capable of processing this information.
Figure 1 outlines the overall strategy for this approach.
Obtaining a detailed quantitative data set on household
dietary intake is a necessary first step. For our work, we rely
on a previous study conducted in northern Mozambique,
known as the Nampula/Cabo Delgado (NCD) study. Since
the NCD study collected quantitative information on food
consumption, it allows us to get estimates of household
nutrient intake in the study area. In the second phase,
statistical relations are explored between easy-to-collect
variables in the NCD database, variables similar to those
that could be collected in future efforts by various organisations, with these quantitative measures of household
nutrient intake. In this analysis, a dietary intake prediction
model is developed that allows prediction of a household’s
dietary intake level given some relatively simple information, such as the types of foods eaten by the household in a
24-h period or the number of members in the household.
The development of this modelFthat is, phase II in
Figure 1Fand its performance are the subjects of this paper.
In phases III and IV, the model can then be used with
Figure 1 Overview of strategy to get low-cost estimates for
monitoring dietary intakes.
information from a new survey to get predictions of household dietary intake in that area.
The techniques developed here are focused at the household level. Although malnutrition is a condition experienced
by individuals, the household is the entry point for most
interventions that seek to improve this condition. The
science of determining nutrient requirements at the household level is less developed than it is for individuals (FNB,
2000). However, it is more practical to monitor conditions at
the household level, since most agricultural and health
projects, as well as the surveys that monitor them, are
targeted at households. As is the case with the 24-h recall
instrument on which our approach is based, the targeting or
monitoring that can be done with this technique is designed
for populations of households, rather than for individual
ones.
Methods
Study area
The Nampula/Cabo Delgado (NCD) study was originally
designed to identify the impacts of various smallholder
cotton-growing schemes on household incomes and food
security in Mozambique (MAF and MSU, 1996, Strasberg,
1997). The study was conducted in Montepuez District in the
province of Cabo Delgado and in Monapo and Meconta
Districts of Nampula Province. These districts were purposively selected both because they encompassed the variety of
different cotton-growing schemes present and because they
are typical of the interior of northern Mozambique, where
maize- and cassava-based cropping systems predominate and
where cashew, as well as cotton, are often grown. Villages,
and households within them, were then selected using a
stratified random cluster design. In total, 388 households in
European Journal of Clinical Nutrition
Predicting dietary intakes in Mozambique
D Rose and D Tschirley
1214
16 villages were selected for the original study. An extensive
description of sampling procedures has been published
previously (MAF and MSU, 1996).
Dietary assessment
Household food consumption was measured using a 24-h
recall technique. Trained enumerators asked the person who
most often did the food preparation and cooking to identify
the type and quantity of each food ingredient prepared and
consumed during the previous 24 h. To facilitate recall of the
quantities used, the household respondent indicated the
amounts of each ingredient used with reference to her own
household utensils (eg, a cup, a can, a spoon). A quantitative
determination of the volume of the amounts indicated by
the household respondent was then made. To do this, the
utensil was filled to the level indicated by the respondent
with a food item of known density, such as dried maize, and
the weight of that amount of maize was then determined.
Dividing the weight of this maize by its density gave us the
volume of the particular food used by the respondent.
Volumetric determination was necessary, as opposed to
direct weighing, because households often did not have
extra amounts of the foods consumed in the previous day.
Interviews and weighings were always conducted in the
household, which facilitated this recall and volumetric
measurement. Information was solicited on foods eaten
other than the main meal items. Although it is possible that
some snacks between meals or foods eaten away from home
were not observed by the household respondent, snacking
and away-from-home food consumption were relatively
limited in the poor rural communities of Mozambique in
1995.
The weight of each food was then calculated by multiplying the volume of the food consumed by its density.
Through its work in Mozambique since the early 1990s, the
Michigan State University Food Security Project accumulated
a database on average densities of the most common food
items. Information on the densities of additional foods not
found in this database was obtained from research in a
neighbouring country (IFPRI, 1997). For most fruits and
vegetables, volumetric information was not used; weight
information was obtained directly by multiplying the
number consumed (eg, three tomatoes) by mean weights of
these items.
The nutrients consumed from each food were obtained
by multiplying the weight of the food consumed times
the nutrients per 100 g of that food. The latter information on the nutrient content of foods was obtained from
standard food composition tables used in Mozambique and
elsewhere in Africa (Leung et al, 1968; West, 1988; MISAU,
1991).
Two recall interviews, separated by about 1 week, were
conducted during each of three rounds: in May 1995
(‘harvest season’), in September 1995 (‘postharvest season’),
and in January 1996 (‘hungry season’). Nutrients for all foods
European Journal of Clinical Nutrition
listed in the 24-h period were summed for each interview day
and then averaged over the two interview days to get a mean
daily household intake for each round. This yielded 1140
observations on 388 households (94% of households had
dietary data in all three periods).
Household intakes were compared with international
standards to develop nutrient intake ratios (NIRs) for
each household. These ratios are similar to a nutrient
adequacy ratio (NAR) for an individual (Guthrie & Scheer,
1981), except that the numerator is a sum of all household
members’ intakes and the denominator is a sum of the
recommended allowances for each person in attendance
at household meals on the day preceding the recall. We find
the NIR to be useful for reporting nutrient intakes, since
it allows one to compare households with different age–
gender compositions. We purposefully avoid the previously
used ‘nutrient adequacy ratio’ terminology, in part because
ours is a household measure, and in part to avoid
misinterpretation regarding the nutrient adequacy of
household diets. Energy allowances were based on reference
weight data for Mozambique (James & Schofield, 1994)
and include the energy needed to maintain weight as well
as energy necessary for occupational and ‘socially desirable’
activities (FAO/WHO/UNU, 1985). Occupational activities
were assumed to be characteristic of a rural population
in a developing country, that is, requiring moderate to
heavy energy expenditures. NIRs were also calculated
for protein, vitamin A, and iron, with recommended
intakes obtained from appropriate sources (FAO/WHO/
UNU, 1985; FAO and WHO, 1988). These four nutrients
were highlighted because of their importance for public
health and nutrition in Mozambique. See Table 7 (of the
Appendix) for a detailed information on the recommendations used.
Mozambique Diet Quality Index (MDQI)
While information on intakes of specific nutrients is useful
for designing applied interventions to address specific
nutrition problems, policymakers often need simple summary measures of nutrition, so they can assess overall
progress in this area. We created the MDQI to summarise
dietary information related to the most important nutritional problems in Mozambique. Other than iodine deficiency, which cannot be assessed with our dietary
instruments, the key nutritional problems are protein and
energy malnutrition, vitamin A deficiency, and iron deficiency. Other nutritional deficiencies, such as niacin deficiency and vitamin C deficiency, have also been documented
in Mozambique, though not as frequently (GISMAV, 1998).
Thus, MDQI also gives weight to other important nutrients
by including a composite index, a mean intake ratio
composed of seven nutrients (MIR7)Fthiamin, riboflavin,
niacin, vitamin B6, folic acid, vitamin C, and calcium. Zinc
deficiency is common in developing countries and is likely
to be a problem in Mozambique, although it has not been
Predicting dietary intakes in Mozambique
D Rose and D Tschirley
1215
documented there. We have not included it in our index,
since our food composition databases do not have information on this nutrient.
The score on this MDQI ranges from 0 to 10 and is a sum of
each of the five component scores related to energy, protein,
vitamin A, iron, and the composite index MIR7. To compute
each component score, the NIR is first computed, then
truncated at 1.0 if the household consumed more than the
recommended amount, and then multiplied by 2. Truncation reflects the fact that excesses in consumption of one
nutrient do not make up for deficiencies in other nutrients.
Multiplying each of the ratios by 2 is simply a means of
converting the MDQI to a more convenient range of 0–10,
rather than 0–5. Based on the scores on this index, household diets were divided into three categories: acceptable, low
quality, and very low quality. Households that scored 7.5 or
greater on this index were considered to have acceptable
diets. Households that scored 6.0 or greater, but less than 7.5
points on this index were considered to have low-quality
diets. Those that scored less than 6.0 points were considered
to have very low-quality diets.
These cutoff points were based on a combination of
scientific judgement and practical policy concerns. The Food
and Nutrition Board (FNB, 1986) of the US National Research
Council outlined conditions for when the mean nutrient
requirement can be used as a cutoff point indicating
inadequate intakes for nutrients other than energy. Using a
typical assumption about the requirement distribution of a
nutrient (ie, the coefficient of variation is 0.15), it can be
shown that the mean nutrient requirement is 76.9% of the
recommendation for a safe level of intake. We use 75% as a
rough approximation to this figure, largely to facilitate
comparisons with other literature on this topic. This
corresponds to 7.5 on the 10-point MDQI scale as a cutoff
point for an acceptable diet. Based on the same rationale, the
intake of specific nutrients is described, elsewhere in this
paper, as low when the NIR is less than 75%. Practical
concerns about interventions that could be targeted to areas
of highest priority motivated our decision to split low
intakes on the MDQI into two categories, those that were
low (6.0–7.5) and those that were very low (o6.0). It can
never be known whether a given diet is truly low or
acceptable for a given household. However, the terminology
is useful for categorising groups of households based on
relative dietary quality.
Dietary intake prediction model
To begin developing a prediction model, variables
were considered that would be easy to collect in the
field, and which were also included in the NCD survey.
These variables included, for example, food items that
were consumed in the previous day or the age and
gender composition of the household. Note that food
consumption variables are identified here as easy-to-collect,
since reference is made to a simple listing of which foods were
eaten during the day. In contrast, quantitative information
on how much was consumed would be quite complex to
collect.
Since there are over 70 different food items in the original
NCD food consumption database, a number of different
food grouping systemsFincluding ones that contained 7,
11, 13, and 15 different food groupsFwere explored
to reduce this into a manageable number. The goal was to
find reasonably aggregated food groups, which would
be broad enough to encompass local foods from different
parts of the country. However, food groups that were
sufficiently disaggregated were needed so that nutrient
content would be relatively homogenous across foods within
the same group, a necessity for getting good predictions of
nutrient intakes. A system of 11 food groups was developed
which balanced these concerns. The complete list of food
groups and individual food items in each group is listed in
Table 1.
Linear regression was used to develop a prediction model
that would map food group consumption to nutrient
intakes. The household intake of a nutrient (expressed
as a percent of its recommendation, ie its NIR) was the
dependent variable and the consumption of foods and
other easy-to-collect variables were the independent variables. There were four main nutrients of interest: energy,
protein, vitamin A, and iron. There were also seven nutrients
that made up the summary measure of dietary variety,
referred to as MIR7 in the previous section. Thus, a total of
11 regression models, one for each nutrient, were estimated.
All models were estimated with ordinary least-squares
regression using all independent variables listed in Table 3
and the ‘regression’ command in SPSS. Other than for
vitamin A and calcium, which were estimated linearly, all
models were estimated with dependent variables in logarithmic form.
Several different expressions of the food consumption
variables were tested. One approach used a single variableF
the number of different food groups consumed in the
previous day. Another approach used a set of dichotomous
indicators of whether or not the household consumed a food
from each food group on the previous day. The final
approach used a set of variables expressing the number of
times per day the household consumed a food from each of
the 11 food groups. This latter approach performed best
among the different food variable alternatives.
We also experimented with a number of socioeconomic
variables, such as those related to household size (measured
in adult equivalents), land tenure, agricultural production
and agricultural sales, as well as seasonal indicators. Household size was a significant predictor in every nutrient intake
model, but none of the other socioeconomic variables
improved the prediction significantly enough to warrant
inclusion in the final models.
Statistical weights accompanying the NCD data set were
used for all descriptive statistics reported here (Tables 1, 4–6).
Regression analyses were performed unweighted.
European Journal of Clinical Nutrition
Predicting dietary intakes in Mozambique
D Rose and D Tschirley
1216
Table 1 Food items in each of the 11 food groups
Food group
Food items
Grains
Dried maize, maize flour, other maize products, sorghum, sorghum flour, fresh sorghum,
bread, rice, pasta, cookies
Cassava flour, dried cassava
Dried beans, dried peas
Dried peanuts, coconut, pumpkin seeds, sesame seeds, sunflower seeds, cashew nuts
Dried fish, fresh fish, beef, chicken, rat, bird, pigeon, snail, crustaceans, grasshopper, frog, milk, eggs
Pumpkin, dark leafy greens, red pepper leaves, cassava leaves, bean leaves, pumpkin leaves,
sweet potato leaves, cashew leaves, red peppers, mango
Papaya, lime, fresh cassava, fresh sweet potato (pale), tomato, fresh beans, fresh peas, fava beans
Mushrooms, onions, bananas, fresh maize, fresh yams, okra, apples, fresh peanuts
Sugar, sugar cane, honey
Oil
Beverages (including maize beer, cashew juice, cashew wine, tea, coffee), salt, candy
Tubers
Beans
Nuts and seeds
Animal products
Vitamin A-rich fruits and vegetables
Vitamin C-rich fruits and vegetables
Other fruits and vegetables
Sugars
Oils
Other foods
Results
Table 2 displays the mean nutrient intakes as a percent
of recommendations from the NCD study. Data were
collected during three different seasons. For the harvest
season, mean intake for energy was 93% of the recommendation, while for vitamin A it was only 28%. The second data
column presents mean intakes during the postharvest
season, when intakes for most nutrients were at their
highest. Intakes were at their lowest during the hungry
season, except for vitamin A, since this was the time when
households substituted leaves and other vegetables for staple
grains. Additional details on food and nutrient intakes in the
NCD study area have been published previously (Rose et al,
1999).
The preferred set of models derived from this work is
displayed in Table 3. Each column describes a model that
predicts the intake of a particular nutrient. The coefficient
estimates, derived from linear regression models, reflect
three main relations. First, they are affected by the nutrient
content of particular foods. For example, the largest
coefficient in the vitamin A model, 0.446, is on the vitamin
A-rich fruits and vegetable group. Consumption of beans,
nuts and seeds, which are good sources of protein, positively
affects the intake of protein. This can be seen by the sizable
positive coefficients on these foods.
Second, the coefficients in Table 3 may also reflect the
amount of food consumed at a given eating occasion. For
example, animal products are a rich source of protein, but
relatively small quantities are consumed at any one time in
Nampula and Cabo Delgado. Thus, the coefficient on animal
products in the protein equation is smaller than the
coefficients on some other food groups, such as nuts and
seeds or grains.
Third, the coefficients in Table 3 also reflect substitutions
between various food groups. For example, there is
a negative coefficient on the oils food group in the
protein model. Oils have no protein content and when
substituted for other foods that do have significant protein
European Journal of Clinical Nutrition
Table 2 Mean nutrient intakes in the Nampula/Cabo Delgado sample
by season
Mean intake as a percent of recommendation (standard errors in
parentheses)
Nutrient
Energy
Protein
Vitamin A
Iron
Harvest
seasona
(n=379)
92.7
150.2
28.1
103.4
(2.4)
(4.3)
(1.3)
(3.5)
Postharvest
season
(n=386)
102.0
151.6
19.5
155.2
(2.1)
(3.8)
(0.8)
(4.7)
Hungry
season
(n=375)
66.4
76.2
55.0
85.4
(1.8)
(2.6)
(3.0)
(2.9)
a
Harvest season data were collected in May 1995, postharvest season data
were collected in August 1995 and hungry season data were collected in
January 1996.
content, they could lower overall protein intake of households. This might occur if respondents substitute oils for the
amount of peanuts they use in the preparation of vegetable
dishes.
The coefficients in Table 3 form the basis of the dietary
intake prediction model. To apply this model in practice, one
would assemble a data set with information on the number
of times in a day foods from each of the 11 food groups were
consumed by each household. (This information could come
from a qualitative-type 24-h diet recall, which simply yielded
a listing of each food item consumed at each meal or snack
time in the previous day.) Then, to calculate the prediction
of vitamin A intake at the household level, one would
multiply the coefficient on each food group from the
vitamin A column in Table 3 by the number of times that
food group was consumed by the household. One would also
need to multiply the household size coefficient by the
number of adult equivalents in the household. The sum of
these products plus the intercept would yield a prediction for
the vitamin A NIR for that household for that day. In this
Predicting dietary intakes in Mozambique
D Rose and D Tschirley
1217
case the formula for a given household would be:
Pred VitA NIR ¼ 0:006 ngrains þ 0:090 nbeans
0:014 ntubers 0:033 nnuts
þ 0:084 nanimals þ 0:446 nvitAfv
þ 0:105 nvitCfv þ 0:050 notherfv
0:082 nsugars þ 0:018 noils
þ 0:096 notherfds 0:054 hhsize þ 0116;
where Pred_VitA_NIR is the predicted vitamin A NIR for
that household, nfoodgroup refers to the number of times
an item from a particular food group was consumed in
the previous day (eg, ngrains refers to the number of times
grain foods were consumed in the previous day), and hhsize
is the household size in adult equivalent units. This
description is intended to show only the mechanics of how
the dietary intake prediction model gets applied to household-level data. However, this tool is not designed for
targeting of resources to individual households. It should
be used only when considering groups or populations of
households.
In practice, the above calculations would be automated
with a computer program. This program would make a
similar calculation for each nutrient for each household
in the data set under consideration. In the present
study, these calculations were made in order to compare
the results from the prediction models with the actual
results of nutrient intake from the detailed quantitative
method used in the NCD survey. The results of
this comparison are presented for protein intake in the
harvest season in Table 4. There were 379 household
observations; 52 (14%) had low intakes of protein and 327
had acceptable intakes as determined by the quantitative
recall measurement technique implemented in that
study (see the far right column of Table 4). The dietary
intake prediction model predicted that from this sample,
46 (12%) would have low intakes and 333 would have
acceptable intakes (see the bottom row of this table). Of the
52 households with low intakes on the quantitative
measurement technique, 31 of them were also low using
the prediction model, so this model had a sensitivity rate of
60%. Of the 327 households with acceptable intakes using
the measurement technique, 312 were also acceptable with
the prediction model, for a specificity rate of 95%. For
vitamin A (not shown), the sensitivity of the prediction
model was much better, at 96%, but the specificity was
worse, at 49%. In all of these cases, we define a low intake to
be less than 75% of the recommendation. As discussed in the
Methods section, it can never be known whether a given diet
is truly low or acceptable for a specific household, but we
find this a useful threshold for comparing groups of households over time.
Information on the frequency of low intakes as
actually measured is compared with results obtained
for the prediction models for the four main nutrients in
Table 5. The first two columns display statistics for the
harvest season. For the most part, the predicted percent of
the sample with low intakes is fairly close to the results
Table 3 Dietary intake prediction model for selected nutrients
Dependent variablea
Independent variableb
Grains
Beans
Tubers
Nuts/seeds
Animal products
Vitamin A-rich fruits and vegetables
Vitamin C-rich fruits and vegetables
Other fruits and vegetables
Sugars
Oils
Other foods
Household size
Intercept
Adjusted R2
N
F
Energy
0.316**
0.298**
0.394**
0.240**
0.122**
0.050*
0.062**
0.100**
0.016
0.089
0.098
0.147**
0.739**
0.554
1140
118.68
Protein
Vitamin A
Coefficient estimates
0.289**
0.006
0.612**
0.090**
0.007
0.014
0.324**
0.033*
0.209**
0.084**
0.035
0.446**
0.071**
0.105**
0.100*
0.050*
0.071
0.082*
0.144*
0.018
0.146
0.096
0.145**
0.054**
0.457**
0.116**
Model statistics
0.646
0.565
1140
1140
174.16
124.14
Iron
0.201**
0.746**
0.492**
0.164**
0.119**
0.012
0.088**
0.129*
0.102
0.141
0.153
0.162**
0.545**
0.477
1140
87.46
a
Dependent variables are nutrient intake ratios, the sum of intakes of individuals in the household divided by the sum of their recommended allowances.
Food group variables refer to the number of times a food was consumed from each group per day. Household size is expressed in adult equivalents.
*
Coefficient estimate significant at Po0.05.
**
Coefficient estimate significant at Po0.001.
b
European Journal of Clinical Nutrition
Predicting dietary intakes in Mozambique
D Rose and D Tschirley
1218
derived from measurements of dietary intake. For example,
using the dietary intake prediction model, 42% of households are predicted to have low intakes of energy, whereas
40% had low intakes based on the full quantitative dietary
measurements. Measured and predicted low-intake rates for
the postharvest and hungry seasons are also displayed in
Table 5.
Predictions were also made on the MDQI scores for each
household. To do this, we first used the dietary intake
prediction model to make a prediction of each household’s
NIR for each of the 11 nutrients that form the MDQI, that is,
energy, protein, vitamin A, iron, and the seven nutrients that
make up the composite measure, known as MIR7. We then
calculated the MDQI for each household as described in the
Methods section, but used predicted NIRs rather than
observed values. As can be seen in Table 6, in the harvest
season this methodology predicted that 49% of households
consumed low or very low-quality diets, which is quite close
to the measured results of 52%. Predictions of the aggregate
percent of the population with either low or very low-quality
diets at different times of the yearFpostharvest or hungry
seasonsFwere also close to measured results. In the hungry
season, the model predicted that 81% of the sample would
have low or very low intakes, whereas 77 were actually
measured to have intakes falling in this category.
Discussion
This paper demonstrates an inexpensive method for
monitoring household diets in Mozambique. It uses a
previously conducted, intensive and quantitative study
of dietary intake to develop a prediction model. This
prediction model allows one to use simple, easy-to-collect
information on food group consumption to make assessments of overall dietary quality in a population. Comparisons of predictions using this technique with results
obtained from the quantitative measurements of dietary
intake in Nampula and Cabo Delgado provinces indicate a
model with a relatively robust set of coefficients. As shown in
Tables 5 and 6, it does reasonably well at predicting nutrient
intakes at vastly different times of the year, that is, at both
the low (hungry) and high (postharvest) points in terms of
consumption.
There are three plausible explanations for the reasonable
predictions yielded by the model. First, we extract a
maximum amount of information from a simple list
of foods. Rather than just a basic count of the number of
different food items consumed, as is done in some proxies
for diet quality, we categorise the foods into 11 homogenous
groups and count up how many times foods from each
group are consumed per day. This information allows us to
develop a set of coefficients that map food group consumption to nutrient intake. Second, our application of
this technique is on a relatively monotonous rural diet
with limited variety both in food selection and in recipes.
The entire Mozambique food composition table has
only 119 food items (MISAU, 1991), a small number
Table 4 Comparing the predictions of low intakes of protein in the
harvest season with those obtained from a quantitative measurement
method
Predictions
Measured results
Acceptable (X75% RDA)
Low (o75% RDA)
Totals
Acceptable
(X75% RDA)
Count
Row (%)
Col (%)
312
95
94
15
5
33
327
100
86
Low
(o75% RDA)
Count
Row (%)
Col (%)
21
40
6
31
60
67
52
100
14a
333
88
100
46
12a
100
379
100
100
Totals
Count
Row (%)
Col (%)
a
Bold-faced percentages, 14 and 12, represent the percent with low intakes as
determined by the quantative measurement method and by using the dietary
intake prediction model, respectively.
Table 5 Measured frequency of low intakes compared with predicted frequency using the dietary intake prediction model
Harvesta (n=379)
Nutrient
Postharvest (n=386)
Measured (% low)b
Predicted (% low)
Measured (% low)
40
14
94
39
42
12
93
35
25
10
99
14
Energy
Protein
Vitamin A
Iron
a
Predicted (% low)
28
15
100
16
Hungry (n=375)
Measured (% low)
Predicted (% low)
63
58
71
55
68
70
81
59
Harvest season data were collected in May 1995, postharvest season data were collected in August 1995 and Hungry season data were collected in January 1996.
A low intake refers to intakes less than 75% of the recommendation.
b
European Journal of Clinical Nutrition
Predicting dietary intakes in Mozambique
D Rose and D Tschirley
1219
Table 6 Measured and predicted results on the Mozambique Diet Quality Index (MDQI)a
Harvestb (n=379)
Percent of households with:
Acceptable diets
Low or very low
Low-quality diets
Very low-quality diets
Post-harvest (n=386)
Hungry (n=375)
Measured (%)
Predicted (%)
Measured(%)
Predicted (%)
Measured (%)
Predicted (%)
48
52
29
23
51
49
29
20
55
45
32
12
52
48
36
12
23
77
26
51
19
81
34
47
a
The MDQI is a composite measure based on the nutrient intakes of energy, protein, vitamin A, iron, and seven other nutrients. On a scale of 0–10, acceptable, low
and very low quality diets are those with 7.5 or more points, 6.0–7.5 points, or less than 6.0 points, respectively.
b
Harvest season data were collected in May 1995, postharvest season data were collected in August 1995 and hungry season data were collected in January
1996.
compared to the 1500 food items found in the South African
database (MRC, 1999) or the 6200 items in that of the
US (USDA, 1999). Previous research specific to the NCD
area has shown that, on average, only four different
food items are consumed per day (Rose et al, 1999). Third,
it should be emphasised that model development and
predictions were made on overlapping data sets. As discussed
below, the entire data set was used for model development,
whereas predictions were made with data from one season at
a time. Consumption does vary from one season to the next,
so these results are useful to study. However, a true test of the
success of the model would require comparison of predicted
and measured values using an entirely different data source
in northern Mozambique. Unfortunately, no such data set
exists.
In order to have the most representative prediction
model, we estimated our regressions pooling observations from three different times during the yearF
the harvest, postharvest, and hungry seasons. This allows
one to use the coefficients from this model to develop
estimates of nutrient intake for any time during the year
in which food consumption data can be collected.
The advantage of such a system is that dietary quality
can be monitored whenever it is feasible for the monitoring
agency, provided that subsequent monitoring surveys
are conducted at the same time of the year to ensure
comparability.
Although feasibility is an important concern, seasonal
issues are complex and their importance should not
be underestimated. The most convenient time to do
agricultural surveys is typically in the postharvest season,
which is usually the most food secure time of the
year. However, in the Nampula/Cabo Delgado sample,
even though this season was the best time of the year, close
to half of the households had low or very low-quality
diets. Thus, if this tool were to be used in the postharvest
season as a means for targeting resources or for monitoring
improvements, there would be no problem in finding
priority areas of need or in detecting changes over
time. Still, we have seen that household correlations
from one season to the next are not always strong. That is,
those that do worst in the postharvest season are not all
the same households as those with the lowest intakes
in the hungry season. Further research is needed on
the variability of household intakes across seasons
and its implications for designing realistic monitoring
systems.
Another concern with this approach is that the
simple dietary surveys needed to implement it are likely to
be based on just one day of data. As has been
shown previously, there is significant intraindividual
variation in intakes from one day to the next (FNB, 1986).
Thus, a distribution of intakes based on 1 day of data will
be more dispersed than a distribution based on averages of
intakes on 2 or more days from the same individuals, or in
this case households. We found this to be the case in the
NCD survey when we looked at the frequency of low intakes
based on 1 day of data as compared with 2 days of data, the
latter being what we report on in this paper. In practice, this
means that prevalence estimates of low intakes based on just
1 day of data will be higher than our results reported here for
NCD. This should not be a problem, as long as monitoring
agencies that begin collecting 1-day consumption data
continue to collect 1-day data in the future to ensure
comparability.
The results presented here suggest that simple food
recall instruments, when used in tandem with prediction
models developed from quantitative dietary surveys,
can provide reasonable assessments of household dietary
intake in Mozambique. It should be highlighted that
this technique is designed for monitoring purposes
at the population level, with the focus on detecting
global changes over time. This technique is likely to
have applicability in other developing countries both
because of the simple nature of the food recall data
and because of the widespread existence of microcomputers needed to develop and implement the
dietary intake prediction model. This prediction model
will need to be calibrated wherever it is to be used,
which implies that conducting or obtaining data
European Journal of Clinical Nutrition
Predicting dietary intakes in Mozambique
D Rose and D Tschirley
1220
from a detailed quantitative dietary survey is a
necessary first step in applying this technique in other
countries.
Acknowledgements
We would like to acknowledge the hard work of Paul
Strasberg, who coordinated the original field data
collection effort as part of his doctoral dissertation, as well
as all the Mozambican analysts and field workers, including
Jose Jaime Jeje, Ana Paula Santos, Higino Marrule, and Rui
Benfica who assisted us with this work.
References
Food and Agriculture Organisation and World Health Organisation
(FAO and WHO) (1988): Requirements of Vitamin A, Iron, Folate, and
Vitamin B12. FAO Food and Nutrition Series 23 Rome Food and
Agriculture Organisation.
Food and Agriculture Organisation/World Health Organisation/
United Nations University (FAO/WHO/UNU) (1985): Energy and
Protein Requirements: Report of a Joint FAO/WHO/UNU Expert
consultation. WHO Technical Report Series 724. Geneva: World
Health Organisation.
Food, Health and Nutrition Information System/Central Statistical
Office (FHANIS/CSO) (1998): FHANIS Urban Report: Monitoring of
the Household Food Security, Health, and Nutrition in Urban Areas.
Lusaka, Zambia: Central Statistical Office.
Food and Nutrition Board, Institut of Medicine (FNB) (2000): Dietary
Reference Intakes: Applications in Dietary Assessment. Washington:
National Academy Press.
Food and Nutrition Board, National Research Council (FNB) (1986):
Nutrient Adequacy Assessment Using Food Consumption Surveys.
Washington, DC: National Academy Press.
Grupo Inter-Sectonal de Mapeamento e Avaliação da Vulnerabilidade
(GISMAV) (1998): Avaliação da Vulnerabilidade em Moçambique,
1997/1998: Uma Análise Preliminar da Actual Vulnerabilidade da
Insegurança Alimentar e Nutricional. Maputo: Governo da República
de Moçambique.
Guthrie HA & Scheer JC (1981): Validity of a dietary score for
assessing nutrient adequacy. J. Am. Diet. Assoc. 78, 240–245.
Hatly A, Torheim LE & Oshaug A (1998) Food varietyFa
good indicator of nutritional adequacy of the diet? A case study
European Journal of Clinical Nutrition
from an urban area in Mali, West Africa. Eur. J Clin. Nutr. 52,
891–898.
International Food Policy Research Institute (IFPRI) (1997): Summary
of Conversion Factors and Densities; Major Crops in Malawi.
Washington, DC: IFPRI.
James WPT & Schofield EC (1994): Necessidades Humanas de Energia:
Um Manual Para Planejadores e Nutricionistas. Rio de Janeiro:
Institute Brasileiro de Geografia e Estatı́stica and Rome: Food
and Agriculture Organisation.
Leung WW, Busson F & Jardin C (1968): Food Composition Tables
for Use in Africa. Rome: Food and Agriculture Organisation
and Bethesda, MD: US Department of Health, Education, and
Welfare.
Medical Research Council (MRC) (1999): South African Food
Composition Database Version 1.2. Tygerberg: Medical Research
Council.
Ministério de Saúde, Repartição de Nutrição (MISAU) (1991):
Tabela de Composição de Alimentos. Maputo: Ministério de
Saúde.
Ministry of Agriculture and Fisheries and Michigan State University
(MAF and MSU) (1996): Smallholder Cash Cropping,
Food Cropping, and Food Security in Northern Mozambique:
Research methods. Working Paper No 22, Maputo: Ministry
of Agriculture and Fisheries and Michigan State University.
Rose D, Strasberg P, Jeje JJ & Tschirley D (1999): Household Food
Consumption in Northern Mozambique: A Case Study in Three
Northern Provinces. MAF/MSU Research Paper No 33. Maputo:
Ministry of Agriculture and Fisheries and Michigan State University.
Strasberg P (1997): Smallholder Cash Cropping, Food Cropping, and
Food Security in Northern Mozambique. Doctoral dissertation,
Department of Agricultural Economics, Michigan State University,
East Lansing.
US Department of Agriculture, Agricultural Research Service (USDA)
(1999): USDA Nutrient Database for Standard Reference, Release 13.
Nutrient Data Laboratory Home Page, http: //www.nal.usda.gov/
fnic/foodcomp.
West CE, Pepping F & Temalilwa CR (1988) The Composition of Foods
Commonly Eaten in East Africa. Wageningen: Wageningen Agricultural University.
Appendix
Nutrient reference standards (see Table 7).
Predicting dietary intakes in Mozambique
D Rose and D Tschirley
1221
Table 7
Males
Age
o1
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20–29
30–49
50–59
X60
Pregnant
Lactating
Females
Energya
Proteinb
Vitamin Ac
Ironc
Energya
Proteinb
785
1307
1456
1604
1729
1812
1910
1992
2056
2066
2088
2152
2180
2297
2397
2449
2528
2618
2987
2987
2987
2928
2928
2018
14.0
23.6
26.6
29.2
32.8
32.5
35.8
30.0
33.4
35.9
38.2
42.9
43.8
49.7
50.4
54.1
55.8
59.1
56.6
56.6
56.6
56.6
56.6
56.6
400
400
400
400
400
400
400
500
500
500
500
500
500
600
600
600
600
600
600
600
600
600
600
600
8
8
8
8
9
9
9
16
16
16
16
16
16
24
24
24
15
15
15
15
15
15
15
15
741
1107
1255
1397
1546
1698
1785
1771
1835
1810
1901
1914
1974
2029
2087
2143
2143
2150
2183
2183
2183
2186
2186
1834
þ285
þ500
13.3
19.1
23.4
26.5
30.2
31.8
35.5
29.1
33.2
36.5
41.6
44.6
44.0
48.5
50.2
55.5
51.7
52.1
49.7
49.7
49.7
49.7
49.7
49.7
þ7
þ 18
Vitaminc
Ironc
400
400
400
400
400
400
400
500
500
500
500
500
500
600
600
600
500
500
500
500
500
500
500
500
þ100
þ 250
8
8
8
8
9
9
9
16
16
16
16
16
16
27
27
27
27
27
27
27
29
29
13
13
29
29
a
These recommendations are based on reference weight data for Mozambique (James & Schofield 1994) and include energy needed to maintain weight as well as
energy necessary for occupational and ’socially desirable’ activities (FAO/WHO/UNU, 1985). Occupational activities are assumed to be characteristic of a rural
population in a developing country, that is, requiring moderate to heavy energy expenditures.
b
These levels are safe intakes (average requirement þ 2 s.d.) based on recommendations in FAO/WHO/UNU (1985) as applied to a Nigerian cassava-diet, that is
corrected for a reduced digestibility of 85%, and for reduced protein quality of 72% for ages 1–6 y, and 95% for ages 6–12 y (see Table 40 in FAO/WHO/UNU 1985).
Additional protein requirements for pregnancy and lactation are from the same source and assume a digestibility of 85% (see Table 50 in FAO/WHO/UNU, 1985).
Since protein recommendations are listed in grams of intake per kilograms of body weight, assumptions about weight were needed to calculate values in the above
table. We used reference weight data for Mozambique (James & Schofield, 1994).
c
Recommended levels of intake for vitamin A are in micrograms of retinol equivalents and for iron are in milligrams. These are safe levels, that is, average
requirements plus a safety factor, to meet the needs of most healthy people (FAO and WHO, 1988). Iron standards are based on the requirement to prevent anemia
from a low bioavailability diet (5%). For pregnancy and lactation, the requirement for menstruating women is assumed. For women over age 50, the iron standard is
reduced to 13 mg/day.
European Journal of Clinical Nutrition