You are on page 1of 16

Food Security (2023) 15:1383–1398

https://doi.org/10.1007/s12571-023-01388-y

ORIGINAL PAPER

Application of item response theory modelling to measure


an aggregate food security access score
Vonai Charamba1 · Lawrence N. Kazembe2 · Ndeyapo Nickanor3

Received: 15 February 2023 / Accepted: 24 July 2023 / Published online: 29 August 2023
© The Author(s) 2023

Abstract
Food security measurement is of paramount importance as it guides governance, policy formulation and intervention projects
targeting and monitoring and evaluation. The measurement of food insecurity has proven to be a difficult task owing to the
multi-dimensionality of the construct and different measurements have been developed to measure different dimensions of
food insecurity. However, it is difficult to apply the different measurements to a holistic food security measurement as their
classification might not agree. The current study proposes a composite food insecurity indicator by aggregating items from
Household Dietary Diversity Score (HDDS), Household Food Insecurity Access Scale (HFIAS) and Months of Inadequate
Household Food Provision (MIHFP) for Windhoek households’ data into a single measure using the Rasch Testlet Response
Model. The composite measure was internally validated against the HDDS, HFIAS and MIHFP and externally validated
against household income and the Lived Poverty Index (LPI) using Spearman's Rank Correlation and Cohen's Kappa. The
validation results suggest that the metric could be a promising measure of aggregate food insecurity worth further investiga-
tion and discussion. However, more research is needed in coming up with cut-points for categorizing households into food
insecurity statuses. In addition, the metric has only been computed and tested on a single urban sample in the Global South
and hence it is not generalizable to different setups. Other researchers who might want to use the index can try to estimate
and validate the index in different scenarios and suggest ways the index can be improved.

Keywords Aggregate food insecurity measurement · Food access · HFIAS · HDDS · IRT Testlet model · MIHFP

1 Introduction To be successful, targeting systems and policy formulation


must use indicators that are valid and reliable in identify-
Poverty reduction, food insecurity and undernutrition remain ing at-risk beneficiaries, progress monitoring and evalua-
high on the development agenda (HLPE, 2017; Leroy et al., tion and remain straightforward, simple and inexpensive to
2015) while targeting initiatives continue to be an important use (Chung, 1997). Historically, food availability (United
concern and have become more relevant in this era of climate Nations, 1975) has been the main aspect of food security
change hazards, ever-shrinking economies and aid budgets. measurement (Bertelli, 2020; Haysom & Tawodzera, 2018).
However, the conceptualization of food security has devel-
* Vonai Charamba oped significantly to incorporate the elements of individual
vcharamba@unam.na; charambavonai@gmail.com access to food, implying that food can be inadequate/inac-
Lawrence N. Kazembe cessible despite it being readily available in markets (Webb
lkazembe@unam.na et al., 2006). The construct was re-defined at the World Food
Ndeyapo Nickanor Summit of 1996 as “a situation that exists when all people at
nnickanor@unam.na all times have physical, social and economic access to suf-
ficient, safe and nutritious food that meets their dietary needs
1
Department of Animal Production, Agribusiness and preferences for an active healthy life” (FAO, 1996). In
and Economics, University of Namibia, Windhoek, Namibia
short, food security is measured as a household-level concept
2
Department of Computing, Mathematical and Statistical that refers to uncertain, insufficient or unacceptable avail-
Sciences, University of Namibia, Windhoek, Namibia
ability, access or utilization of food (Wunderlich & Norwood,
3
Faculty of Agriculture, Engineering and Natural Sciences, 2006). FAO (1996) defined four pillars of food security; (1)
University of Namibia, Windhoek, Namibia

13
Vol.:(0123456789)
1384 V. Charamba et al.

food availability (2) food access (3) food utilization and the variety of the diet by adding up the number of food
(4) food stability. However, this multidimensionality of the groups eaten by household members or individuals in the
attributes has posed challenges in its measurement. last 24 h before data collection. The MIHFP is a food secu-
Food security measurement is of paramount importance rity measure based on the total months in which households
as it allows comparisons between individuals, households had inadequate food provision (Bilinsky & Swindale, 2010).
and groups. It guides governance (Coates, 2013), policy The survey items for measuring the HFIAS, the HDDS and
formulations, project interventions and safety-net the MIHFP ate shown in Table 1.
programmes targeting, (Chung, 1997) and highlights changes In addition to coming up with an HFIAS score by adding
and progress towards achieving set targets as measurement up housing responses to the nine items in the scale, the score
drives diagnosis and response (Barrett, 2010). Given the top can be further categorized into the Household Food Insecurity
prioritization of food security by many governments and Access Prevalence (HFIAP) with four food insecurity classes
global development communities and its conceptualization as 1) Food Secure 2) Mildly Food Secure 3) Moderately Food
a challenge (McCordic et al., 2022), there is a need to improve Insecure and 4) Severely Food Insecure. There is no stand-
the understanding of food security measurement (Leroy et al., ard method for categorizing the HDDS and MIHFP into food
2015, Bartelli, 2020; Jones et al., 2013; Deitchler et al., insecurity classes. However, some scholars have attempted to
2010). However, there is a rigorous international debate on categorize the HDDS and MIHFP as shown in Table 2.
how best to quantify levels of food insecurity (Coates, 2013; Although the HFIAS, the HDDS and the MIHFP are
Haysom & Tawodzera, 2018; Headey et al., 2013; Leroy intended to measure the same dimension: food access, they
et al., 2015). Methodological concerns in the measurement might not be congruent in classifying households hence it
of food insecurity have been highlighted (Abuelhaj, 2007; might be challenging to incorporate all three measures dur-
Haysom & Tawodzera, 2018), and the measurement of ing safety-net targeting or progress monitoring and evalua-
the trait has proven to be a difficult task, owing to it being tion and policy formulation. It is argued that the HFIAS does
multidimensional (Barrett, 2010), resulting in different not capture important elements of household food security,
measures of food insecurity used interchangeably and at dietary diversity and coping strategies (Coates et al., 2006;
times classification by different indices is not in agreement. Kirkland et al., 2013) and thus should be used in conjunction
As a result, analysts have resorted to measuring the distinct with other primary indicators of dietary diversity and coping
facets of food security rather than holistically measuring the strategies (Kirkland et al., 2013), to adequately locate vul-
attribute. However, measurement tools designed to measure nerable households across geographical and cultural settings.
certain contexts can overlook important parameters in a Similarly, the National Research Council (2005) criticized
different context. the use of a similar scale, the Household Food Security Sur-
Several measures have been developed to measure house- vey Module (HFSSM) used in the Current Population Sur-
hold food access and these include; the Household Food vey, highlighting that it does not capture much information
Insecurity Access Scale (HFIAS), the Household Dietary relevant to capture frequency and duration of food insecurity
Diversity Score (HDDS) and Months of Inadequate House- levels. In addition, these measures were originally designed
hold Food Provisioning (MIHFP) developed and utilized for measurement in rural settings and might not be compat-
by the United States Agency for International Development ible with measurement in urban areas where households do
(USAID) Food and Nutrition Technical Assistance (FANTA) not directly rely on agriculture and it might be difficult to
Project. The HFIAS is a nine-item experience-based scale detect months of instability, diets might be influenced more
to measure the prevalence and severity of household food by ethnicity than food security status, coupled with lack of
insecurity in developing countries (Abuelhaj, 2007; ENN, safety-net interventions. Haysom and Tawodzera (2018)
2007), based on universal experiences; anxieties and uncer- urged a renewed focus on building food security metrics that
tainties households may have faced with limited access to apply to the unique characteristics of urban food systems. An
food. The nine HFIAS questionnaire items, arranged in order international workshop on simple tools for measuring house-
of the levels of severity of food insecurity they measure have hold food access and dietary diversity was held in Nairobi,
response codes 1 “never” if the household has not experi- 2007 and one of the objectives was to explore how tools can
enced the condition in the past month, 2 “rarely” for a house- be applied and “integrated” to improve the food security
hold that experienced the condition less than three times, 3 information systems (ENN, 2007).
“sometimes” for a household that had between three and 10 This study attempted to come up with a single measure
experiences and 4 “often” for a household with more than of household food access levels based on the responses to
10 experiences of the condition. The scale, ranging between items for measuring all three scales by considering the sec-
0 and 27, is a summation of the responses for a particular tions as subtests, defined by Wang and Wilson (2005) as
household. The HDDS is a proxy for the quality and quantity testlets, within one test intended to measure the food access
of food (Abuelhaj, 2007; Leroy et al., 2015), aimed to assess dimension. The three measures were selected for this study

13
Table 1  Survey questions for the HFIAS < HDDS and MIHFP used in the households’ survey
HFIAS Items HDDS Items MIHFP

I would like to start off by reading some statements that people Now I would like to ask you about the types of foods that In the past 12 months, were there months in which you did not
have made about their household’s food consumption over you or anyone else in your household ate yesterday. By have enough food to meet your household’s needs?
the past four weeks. For each statement, I would like you to yesterday we are referring to foods that were eaten during
tell me whether this happened rarely, sometimes, often or the day and/or at night.
never in the past four weeks.
Interviewer: hand show card to the respondent (READ list (Read the list of foods. Circle YES in the box if anyone in the Circle the months in which the household did not have
and circle only ONE answer for each question) household ate the food in question, and circle NO if no one enough food to meet the needs
in the household ate the food)
1. …, did you worry that your household would not have 1. Any pasta, bread, rice noodles, biscuits or foods made from 1. January
enough food? flour, millet, sorghum, maize, rice, wheat, or oats
2.… were you or any household member not able to eat the 2. Any potatoes, sweet potatoes, beetroots, carrots or any other 2. February
kinds of foods you preferred because of a lack of resources foods made from them
(money)?
3.…did you or any household member have to eat a limited 3. Any other vegetables 3. March
Application of Item response theory modelling to measure an aggregate

variety of foods due to a lack of resources (money)?


4…did you or any household member have to eat some foods 4. Any fruits 4. April
that you really did not want to eat because of a lack of
resources (money) to obtain other types of food?
5…did you or any household member have to eat a smaller 5. Any beef, pork, lamb, goat, rabbit, wild game, chicken, 5. May
meal than you felt you needed because there was not enough duck, other birds, chicken heads and feet, liver, kidney, heart,
food? or other organ meats/offal or products
6…, did you or any household member have to eat fewer meals 6. Eggs 6. June
in a day because there was not enough food?
7…was there ever no food to eat of any kind in your household 7. Any fresh fish, dried fish or shellfish 7. July
because of a lack of resources (money) to get food?
8…. did you or any household member go to sleep at night 8. Any foods made from beans, peas, lentils, or nuts 8. August
hungry because there was not enough food?
9…. did you or any household member go a whole day and 9. Any cheese, yoghurt, milk or other milk/dairy products 9. September
night without eating anything because there was not enough
food?
10. Any foods made with oil, fat, or butter 10. October
11. Any sugar or honey 11. November
12. Any other foods, such as condiments, coffee, tea 12. December

13
1385
1386

13
Table 2  Calculation and classification of households by individual indicators
Food security indicator Calculation guidelines Classification guidelines References

Household Food Insecurity Access Scale (HFIAS) Respondents are asked a set of 9 questions, eliciting percep- HFIAS See Coates et al. (2007) for compu-
Household Food Insecurity Access Prevalence (HFIAP) tions and experiences that represent generally increasing For the HFIAS, the responses are tation guidelines
levels of food insecurity, with a recall period of 4 weeks. summed to give a score from
In the case of affirmation response, the frequency of 0–27
occurrence is coded as 1 (rarely- once or twice), 2 some- HFIAP
times (3–10 times) and 3 often (> 10 times) 1 = Food Secure
2 = Mildly Food Insecure
3 = Moderately Food Insecure
4 = Severely Food Insecure
Household Dietary Diversity Score (HDDS) Respondents are asked a series of questions as to whether 0–3 Low Swindale and Bilinsky (2006)
any of the household members consumed the food groups 4–5 Medium
in the past 24 h. Cereals; roots and tubers; vegetables;
6–12 High
fruits; meat poultry and offal; eggs; fish and seafood;
pulses, legumes and nuts; milk and milk products; oil 0–3 Low Huluka and Wondimagegnhu (2019)
and fats; sugar and honey; miscellaneous. The responses 4–6 Medium
should be either “0” or “1.” Summing all the food groups 7–12 High
provides a household dietary diversity score ranging from
0 to 12.
Months of Adequate Food Provision (MAHFP) Respondents are asked to indicate the months that they 10–12 Food Secure Bilinsky and Swindale (2010)
encountered inadequate household food supply. Twelve 0–9 Food Insecure
months minus the total number of months during which a
10–12 Very high Mutea et al. (2019)
household was unable to meet its food needs.
7–9 High
4–6 Medium
0–3 Low
V. Charamba et al.
Application of Item response theory modelling to measure an aggregate 1387

because (1) they are normally collected in one questionnaire probability of endorsing a response option is monotonically
by institutions that adopt the FANTA measurement scales non-decreasing, that is, respondents with higher levels of
for food (in)security measurement and thus were readily the latent continuum have higher chances of endorsing more
available in the secondary data that was available for analy- severe items and item categories.
sis (2) the proposed metric incorporates the (chronic) dura-
tion of food insecurity (MIHFP) and acute food insecurity 1.2 Application of IRT in food security
through incorporation of the HFIAS and HDDS which cap- access measurement
ture access to dietary variety and perceptions about insuf-
ficient food quality and quantity into one measure. Although The application of IRT modelling to food security measure-
the proposed model might violate the comprehensibility- ment is not new. The World Food Programme (WFP) Voice
simplicity trade-off, that is, as much as it might portray the of the Hungry (VoH) uses the Rach IRT model for estimat-
food insecurity phenomenon, it might be too complicated, it ing the Food Insecurity Experience Scale. Nord (2014) illus-
is computed from household instead of individual-level data, trated the use of the one-parameter logistic (Rasch) model in
which requires less time and fewer resources to collect than food security modelling. The United States Department of
individual-level data (Jones et al., 2013). In addition, the Agriculture (USDA) uses the Rasch IRT model to measure
data is usually collected in the same survey. The household household food security for originally polytomous responses
level metric is appropriate for targeting food aid, screening dichotomized before estimation (Wunderlich & Norwood,
at-risk households and monitoring changes in food security 2006). A panel of experts reviewed the use of the Rasch
over time. model food insecurity measurement for dichotomous, origi-
nally polytomous responses and suggested the modification
1.1 The testlet item response models of the IRT model to handle the polytomous nature of the
data collected. Johnson (2004) illustrated the applicability of
Item response theory (IRT) models, are generalized linear IRT models in food security measurement, explaining vari-
mixed effects models (Johnson, 2004) and are a special case ous estimation methods, the appropriateness of IRT models
of latent variable models (Wunderlich & Norwood, 2006), on food security measures and the determination of cut-off
concerned with the measurement of inherently unobserv- points for the classification of individuals into food insecu-
able attributes such as food insecurity. They are stochastic rity categories. He concluded that the assumptions need to
mathematical models represented as item response function be met before the model can be used for inference and there
(IRF), which relates the respondent probability of observ- is a need to validate if the propensity measured by the IRT
ing a particular response to a categorically scored manifest model is related to the true food security score. Some schol-
variable (questionnaire item) to individual proficiency (𝜃 ), ars have applied the Rasch model to validate the Household
assumed to have a 0 mean Gaussian distribution, giving Dietary Diversity Score (Vellema et al., 2016; Charamba
the position of the respondent on the latent construct being et al., 2019) while some scholars applied the polytomous
measured (in this case food insecurity) and item discriminant IRT models to validate the HFIAS (Charamba et al., 2019;
parameter (a) and difficulty parameter (b). The latent vari- Deitchler et al., 2010; Bertelli, 2020). IRT methods have
able is regarded as fixed for a given respondent (household) an advantage in that they provide mechanisms to evaluate
and values for the item responses for that household are the contribution of an individual item to the latent measure
regarded as independent and represent different aspects of being estimated, by assessing the item parameter estimates
food security. The connection between 𝜃 and X ’s is assumed and item fit statistics (see Charamba et al., 2019) thus it is
probabilistic so that households with different proficiency not necessary to perform sensitivity analysis to assess indi-
(food insecurity) levels can still have different values for the vidual item contribution in the model.
observed X values. The parameter a indicates how much the
measurement error changes theta (𝜃) and large values result 1.3 The rationale for an aggregate food security
in low measurement error and the stronger the relationship access measure
between theta (𝜃) and the questionnaire item.
Standard IRT models such as the Rasch model, the two- Some scholars suggested disaggregated suites of indicators
parameter logistic (2PL) model, and the Generalized Partial to capture the complexity and diversity of the food insecu-
Credit Model (GPCM) for polytomous responses assume rity construct (Coates, 2013). However, the use of different
that the latent variable (𝜃) being measured is unidimensional measures to quantify the food access dimension makes it
and one person’s responses to different items are independ- difficult to compare results from different studies within
ent (local item independence) and responses of different the same spatial and temporal domains as different meas-
persons to the same item are independent (local person ures yield different estimates of food insecurity (Haysom
independence). In addition, IRT models assume that the & Tawodzera, 2018) and different metrics may classify

13
1388 V. Charamba et al.

the same household differently. For example, Mutea et al. insecurity levels. The Rasch IRT model and the area under
(2019) noted that the HDDS and the Food Consumption the curve (AUC) and the receiver operator curve (ROC)
Scale (FCS) revealed less food insecurity than the MAHFP, were applied to validate the HDDS with binary items while
Coping Strategy Index (CSI) and HFIAS. This might pose the Partial Credit Model (PCM) validated the polytomous
a challenge in household food insecurity classification for item response HFIAS. The Principal Components Analysis
reporting, beneficiary targeting in programme intervention (PCA) and the Mokken analysis were used to determine the
and policy formulation, especially since food security has dimensionality and monotonicity, that is, an increasing trend
been conceptualized as a development challenge (HLPE, of item severity is preferred, so that items and item response
2017; Leroy et al., 2015; McCordic et al., 2022). As a result, categories are arranged in such a way that a household with
there is a pressing need to develop decision-support mecha- a higher food insecurity levels has a higher probability of
nisms to support food security policy and research (Mock endorsing higher-order items or item categories. Their study
et al., 2013) and interventions. Several authors have sug- concluded that the HFIAS is unidimensional and items are
gested a re-think of food security measurement (e.g. Headey monotonic, with Loevinger coefficients > 0.5 implying that
& Ecker, 2013) with arguments that existing measures are they are monotonically arranged (see Charamba et al., 2019).
too varied. Maxwell et al. (2013), considered the possibility However, all items in the HDDS had Loevinger coefficients
of combining food security metrics or using them in a com- < 0.3, implying that they are not Guttman items, that is,
plementary way to yield a more multidimensional picture they are not monotonically arranged according to the level
of household food insecurity, arguing that the “holy grail” of food insecurity levels they measure. In addition, items 2
of food security measurement would be a single measure and 5 of the HFIAS were not monotonically placed among
that is valid, reliable and comparable over time and space, other items. The results from the ROC curve showed that
capturing different elements of insecurity. However, no sin- the HDDS is effective in separating Windhoek households
gle measure meets these criteria (Coates & Maxwell, 2012). according to their dietary quality and quantity. Vellema et al.
McCordic et al. (2022) combined the HDDS, HFIAS and (2016) validated the HDDS and HFIAS against household
MAHFP into a new index by using an adjusted geometric income and the Progress out of Poverty Index among other
mean and observed a correlation between food insecurity variables and Nickanor et al. (2017) compared household
and inconsistent resource access. They reiterated that how food security levels across households' socio-economic char-
scores are aggregated can significantly affect the index sta- acteristics and observed a relationship between food inse-
bility and such an index should provide means for normali- curity and household type and location, and LPI, with food
zation that ensures equal priority of measures being aggre- insecurity levels increasing with poverty levels. McCordic
gated. It is behind this background that the current study et al. (2022) observed a negative correlation between incon-
proposed an alternative aggregate food access measure by sistent access to water, electricity, cooking fuel and cash
combining the HFIAS, HDDS and MIHFP using the IRT income, with different correlation magnitudes observed
modelling technique. These indicators selected were selected across cities in the Global South. The contribution of the
because they have a low-respondent burden (Leroy et al., current study to food security measurement literature is that
2015) and are suitable for large surveys as they exclude it is probably the first initiative to try and bring together the
questions on household income and expenditure, consumed experience-based coping strategy and dietary diversity indi-
quantities, as well as quantities produced and distance to cators into one measure of food security using IRT methods.
markets which might be difficult to obtain. In addition, they
are usually combined in one survey by many institutions
measuring food insecurity based on the FANTA metrics. 2 Material and methods

1.4 The goodness of fit tests and construct validity 2.1 Research objectives

To be useful, food security indicators should be valid, well- The objectives of the study are to:
constructed, reliable, accurate and suitable for providing
analytical measurement for a given purpose (Johnson, 2004; 1. Create an aggregate metric for assessing food access, uti-
Leroy et al., 2015). An important step in the validation of lization and stability by aggregating the HDDS, HFIAS
food security measures is to assess the extent to which the and MAHFP score items using the Rasch testlet item
data are consistent with the assumption of the measurement response theory model.
model (Nord, 2014; Bertelli, 2020; Deitchler et al., 2010; 2. Internally and externally validate the proposed metric
Johnson, 2004). Charamba et al. (2019) conducted a validity against the indices whose items were aggregated and against
study on the applicability of the HDDS and HFIAS scales in common determinants and consequences of food insecurity.
measuring Windhoek urban households' (current data) food

13
Application of Item response theory modelling to measure an aggregate 1389

To address the stated objectives, the investigation first gives details on the distribution of PSUs, households and
created an aggregate index for Windhoek households’ food population per constituency. The second stage involved
access, utilization and stability by aggregating the HDDS, the systematic selection of 25 households from each of the
MIHFP and HFIAS items using the testlet item response 35 PSUs. The data on several demographic characteristics
model and validated the model against the Lived Poverty and socio-economic attributes including housing, income
Index (LPI) and household income levels. In addition, levels, access to water, cooking energy, electricity, access
the metric was correlated against the HDDS, HFIAP and to medical facilities and food access and consumption pat-
MIHFP categories (Mutea et al., 2019; Bilinsky & Swindale, terns were collected from household heads or their proxies
2010) to assess the individual contribution of the HDDS, using the AFSUN-HCP Household Food Security Baseline
HFIAS and MIHFP to the new measure. Survey questionnaire. Food access and consumption data
collected include the HFIAS, HDDS and MIHFP. The data
2.2 The data was collected using the tablet method.

The current study is based on secondary data analysis for


data collected through a face-to-face survey in Windhoek, 2.3 Estimation of the aggregate food access index
the capital city of Namibia. The data was collected from
urban households under a study funded by the Open Soci- A testlet is a bundle of items that share a common “stimulus”
ety Foundation for South Africa, conducted by the Depart- (Wang & Wilson, 2005), that is, items that are related to a
ment of Statistics and Population Studies of the University common sub-component of the latent variable being meas-
of Namibia (UNAM) in partnership with the African Food ured. Testlet models were developed to handle tests where
Security Urban Network (AFSUN), the Hungry Cities items can be grouped into subtests measuring a sub-category
Partnership (HCP) and Balsillie School of International of the stimuli being measured, as items in one subtest may
Affairs, aimed at assessing food security status for urban be related, probably violating the local item independence
households in the Global South. A total of 875 house- assumption as there is a possibility that items in a subtest are
holds were selected through a two-stage random sampling related. They include random parameters added to handle the
procedure. The first stage involved the random selection dependence that exists among items within the same bundle.
of 35 Primary Sampling Units (PSU) demarcated for the The Rasch Testlet Model (Wang & Wilson, 2005) is a model
2011 Population and Housing Census (Namibia Statistics for both dichotomous and polytomous items. Suppose X is the
Agency, 2013), using the probability proportional to size manifest variable that codes the response of a food security
(PPS) from all the 10 Windhoek constituencies, where the question as k, k = 1 … K, then the model used in the current
number of PSUs selected per constituency was by the num- study, generalized to allow items to have variant discrimination
ber of PSUs, number of households and population in the abilities and adjusted to cater for polytomous items has an IRF
constituency. Only one PSU was selected from Windhoek which relates the probability of endorsing a specific response
Rural constituency due to its proximity to Windhoek urban category given a specific proficiency level as in Eq. (1).
as the original study was mainly concerned with the meas-
urement of food security of urban households. Table 3

Table 3  Household survey sample selection


Constituency No. of PSUs Sampled No. of HH Sampled HH Population in Population in Population in
PSUs constituency sampled PSUs sampled HH

John Pandeni 63 2 3377 50 15121 559 130


Katutura Central 57 2 5096 50 24608 726 151
Katutura East 48 2 3756 50 18501 733 149
Khomasdal 136 3 10471 75 43,921 1128 247
Moses //Garoeb 158 6 13804 150 45564 1648 543
Samora Machel 195 5 13250 125 50110 1682 457
Tobias Hainyeko 155 5 12428 125 45912 1231 372
Windhoek East 101 3 7089 75 22712 617 254
Windhoek West 158 6 13837 150 53438 1814 520
Windhoek Rural 109 1 6330 25 22254 104 78
Total 1180 35 89438 875 342141 10424 2901

13
1390 V. Charamba et al.

( )
exp[ai 𝜃j − bik + 𝛾jd(i) ] 2.4 Categorization of the IRT food security measure
Pjik (Xi = k∕𝜃) = ( ) (1)
1 + exp[ai 𝜃j − bik + 𝛾jd(i) ]
The categorization of households into food security levels
where Pjik represents the probability of person j with ability similar to one of the expert judgement methods proposed by
𝜃j getting a score k on item i in a testlet d with step difficulty Johnson (2004) to categorize households under the HFSSM
bik , 𝜃j is the food insecurity level for household j and ai is into food secure, mildly food insecure, moderately food inse-
the discrimination parameter for item i . 𝛾jd(i) is the effect cure and severely food insecure categories were adopted in
of item i in testlet d on respondent j . The model exhibits a the current study. The method used assumed that a house-
monotonic non-decreasing relationship between the latent hold belonging to a particular food insecurity class has a fixed
variable and the probability of affirming an item response minimum probability of affirming a specific questionnaire
category. As 𝜃 increases, the probability of affirming or item. For example, experts may want to find a cut-point 𝜏2 so
endorsing a response category increases. As bik increases, that Pr(Xi7 = 1|hunger) = 0.95 or (some large probability)
the item category becomes “harder” and the probability of a where item 7 is the item that asks the individual if they were
person affirming a response category decreases for a house- ever hungry but could not( eat) over the past year, and then
hold with a specific value of 𝜃 . The discriminant parameter solve for the equation Pr7 𝜏2 = 0.95. In the current study,
a must be at least 0 and indicates the strength of the connec- the specific items demarcating households according to food
tion between the latent variable 𝜃 and the manifest question- security categories were determined according to Coates
naire item Xi . When a is near 0, the connection is weak and et al.'s (2007) categorization of households in HFIAP classes.
when a is 0, then there is no connection. When a is positive The current procedure is based on the existence of a measure
and large, the item is highly discriminating and there is a Yk , k = 1, 2, 3, 4 denoting the realization that a household is
strong connection between the item and the latent measure food secure, mildly food insecure, moderately food insecure
(food insecurity) implying less measurement error, and a and severely food insecure respectively. A perfect monotone
high ability to disaggregate households according to their relationship is assumed between the food insecurity category
food insecurity classes. The model parameters and hyper- and the propensity 𝜃 measured from the IRT model such that
parameters have the following prior distributions: there exists cut-points 𝜏1, 𝜏2, and 𝜏3 such that:
( )
𝜃 ∼ N(0,1); a ∼ LN(0.2,0.2); b ∼ N(0,1); 𝛾 ∼ 0, 𝜎𝛾2 ; 𝜎𝛾2 ∼ IG(1,1). ⎧1 if 𝜃 ≤ 𝜏1
⎪2 if 𝜏1 < 𝜃 ≤ 𝜏2
Yk = ⎨
where N, LN and IG denote the Normal, Lognormal and ⎪3 if 𝜏2 < 𝜃 ≤ 𝜏3
Inverse Gamma distributions respectively. The Lognormal ⎩4 if 𝜃 > 𝜏3
distribution was selected for the discriminant parameter ( a) and these points were used to categorize households into
to obtain non-negative values near 1 and to make the item- four food security categories. The relationship between Yk
characteristic curve monotonically non-decreasing. The test- and 𝜃 is assumed to be probabilistic, where 𝜃 depends on the
let model was chosen for the current study ahead of standard food security category Yk of household k . Let F(𝜃∕Y = y)
IRT for polytomous data based on the assumption that there denote the distribution of the propensity given the household
is a possibility of correlation that might exist for items in the food security status Y = y , the distributions are assumed to
same testlet (HFIAS, HDDS and MIHFP). All 9 items of the be stochastically ordered such that:
HFIAS scale (testlet 1), all 12 items of the HDDS (testlet 2)
and all the 12 items of the MAHFP (testlet 3) were used in F(t∕Y = 1) ≤ F(t∕Y = 2) ≤ F(t∕Y = 3) ≤ F(t∕Y = 4)
the current testlet model. The local item independence was
not tested for the current model, however, the magnitude of For all values t . The propensities are assumed to be nor-
the testlet variance of at least 0.25 imply significant testlet mally distributed within each food security category with
effects (Wang & Wilson, 2005). means 𝜇1 ≤ 𝜇2 ≤ 𝜇2 ≤ 𝜇4 and 𝜎1 = 𝜎2 = 𝜎3 = 𝜎4 .
The model parameters were estimated using the Bayes- According to the computation of the Household Food
ian estimation with Markov Chain Monte Carlo (MCMC) Insecurity Access Prevalence (HFIAP) categories (Coates
using the Gibbs Sampling methods in WinBUGS 1.4.3 et al., 2007), a household that responded with a 1 (never)
(Lunn et al., 2000). Convergence was checked by visualiz- or 2 (rarely) for item 1 of the HFIAS is assumed to be food
ing the trace and history plots of the four chains that were secure, while households responding the item with 3 (some-
initialized with different starting values as well as when times) and 4 (often) are considered food insecure. This can
the Brook-Gelman-Rubin (BGR, Gelman et al., 1995) be mathematically written as:
statistics becomes close to 1. The category characteris- ( )
𝑃 𝑟 food secure∕X1 ≤ 2 = 1 (2)
tic curves (CCC) were assessed to check if the combined
function is monotonically non-decreasing. and implicitly

13
Application of Item response theory modelling to measure an aggregate 1391

( )
𝑃 𝑟 food secure∕X1 > 2 = 0 result, there is an expected relationship between urban food
access and household income. The LPI is a barometer of
However, Eq. (2) implies that all items are Guttman items quality of life that measures the subjective experience of pov-
and all severely food insecure households have infinitely erty (Mattes et al., 2016) and is an indication of how often a
large propensities. However, the assumption was relaxed household goes without basic commodities, including food,
based on the findings from the study by Charamba et al. implying that a household that lacks basic commodities (food
(2019) where items 2 and 5 of the HFIAS scale were not included) is likely to be food insecure, making LPI a deter-
monotonically placed among other items of the scale and all minant and consequence of food insecurity. As a result, the
HDDS items were not Guttman-items, as the measure was new IRT index was externally validated against the house-
not designed to depict monotonicity as the food groups in hold income and LPI which captures most of the variables
the testlet are not weighted. It was assumed that households McCordic et al. (2022) validated their index against.
in each category have a probability of ascertaining a certain Income quartiles were computed by first ranking the
discriminatory item with a probability of 0.7 (expert judge- households from poorest to wealthiest based on household
ment). As a result, a response of 3 (sometimes) to item 1 income, and then grouping them into four income quartiles
“Worry that the household would not have enough food”, a and assigning 1 to the wealthiest and 4 to the poorest so that
response of 3 to item 3 “Eat a limited variety of foods” and each quartile had approximately 25% of the population. LPI
a response of 3 to item 6 “Eat fewer meals in a day” were was computed on a 5-point (0, absence of poverty -4, constant
set as
( boundaries
) to (calculate
) the cut-off(points
) such that absence of basic necessities) and then categorized into 4 classes
Pr13 𝜏1 = 0.7, Pr33 𝜏2 = 0.7, and Pr63 𝜏1 = 0.7, were according to the quartiles (< 1, 1–2, 2–3, 3–4) and these were
used for determination of 𝜏1, 𝜏2 an 𝜏3. considered to have a one-to-one correspondence with the food
security classes. The agreement between the household food
2.5 Internal and external validation insecurity levels measured by the Testlet model and other meas-
of the composite IRT food insecurity measure urement methods was assessed using Spearman’s rank correla-
tion coefficient (r) and Cohen's Kappa (𝜅) agreement.
The IRT food insecurity measure was checked for internal
index consistency by assessing the strength of the associa-
tion with the subscales that were combined to make up the 3 Results
aggregate measure, i.e., the HFIAS, the HDDS, and the
MIHFP using the Pearson’s Correlation coefficient (fol- From the category characteristic curve (CCC) in Fig. 1, the
lowing McCordic et al., 2022). A negative correlation will probability of endorsing an item ategory for the combined
mean that an increase in the subscale results in a decrease testlet data is monotonically non-decreasing. For example,
in the aggregate index while a low correlation will mean households with lower ability parameters (food insecurity
that the subscale did not play a major contribution (lower level) are more likely to endorse the first category (never) in
discrimination ability). The HFIAP was computed accord- the HFIAS testlet and “no” in the HDDS and MIHFP testlets
ing to Coates et al. (2007). As there is no standard way of and less likely to endorse category 4 (often). On the other
categorizing the HDDS and MIHFP, the categorization for hand, households high up on the food insecurity continuum
the HDDS in the current study was tailor-made for the cur- are more likely to endorse category 4. However, category 2
rent study following Oldewage-Theron and Kruder (2008), is less likely to be endorsed by households across all food
Swindale and Bilinsky (2006) and Huluka and Wondimageg- security classes as its distribution is flatter.
nhu (2019) and 1–3 food groups were categorized as low, Figure 2 shows the graphical presentation of the ability
4–6 groups as medium, 7–9 as high and 10–12 very high (IRT food insecurity) estimates for all the households in the
dietary quality. The months of inadequate household food study. The abilities seem to fluctuate between ±3 across all
supply were categorized by reversing Mutea et al. (2019) households, and there is no obvious pattern in the distribu-
categories for the MIHFP categories where households that tion of food insecurity levels according to the household
had inadequate supply for 9–12 months were considered to observation order.
be severely food insecure, 6–8 months as moderately food Figure 3 gives the distribution of the difficulty param-
insecure, 3–5 as mildly food insecure and households that eter estimates across items. The first 9 items are from the
experienced food shortages for 0–2 months as food secure. HFIAS scale, items 10 to 21 are from the HDDS section
For external validation, the IRT aggregate metric was while items 22 to 33 are from the MIHFP scale. Items in
correlated with the household income and Lived Poverty the HFIAS testlet are relatively difficult. Item 10, which
Index (LPI). As the surveyed households are urban holds, asked about the consumption of starch products such as
they purchase most of their food items for consumption, thus pasta, rice and maize, millet and sorghum flour meals,
household income is a determinant of food insecurity. As a was considered the easiest item (consumed by many) while

13
1392 V. Charamba et al.

Fig. 1  Item characteristics curves for the Windhoek Food Security Data

item 16 on the consumption of fish and item 17 on the to items in the HFIAS and the MIHFP testlets. For exam-
consumption of beans, nuts and lentils were considered ple, item number 12 on vegetable consumption and number
most difficulty (less consumed). Item 14 which collected 16 on fish and seafood consumption were considered less
data on meat consumption also peaked downwards, imply- discriminating.
ing that Windhoek households might have high meat con- From the results in Table 4, households are distributed
sumption patterns. according to their food insecurity levels from -3.54 (less
The discriminant parameter ( ai ) estimates for the items food insecure) to 3.03 (higher levels of food insecurity) with
are given in Fig. 4. The results show that the items in the an average of about 0. The first and second thresholds were
HDDS are less discriminating, that is, they contribute less to disordered as the first threshold is higher than the second
the aggregate households' food insecurity levels compared threshold while the third threshold was higher, implying

Fig. 2  Ability parameter esti-


mates for households

2
Ability param eter es tim ates

-2

0 250 500 750 1000


Household

13
Application of Item response theory modelling to measure an aggregate 1393

Fig. 3  Item Difficulty parameter estimates

that the transition from Category 2 to Category 3 was easier on the food insecurity continuum. In addition, items in the
than the transition from Category 1 to Category 2. However, MIHFP testlet were more difficult compared to items in the
threshold 3 was higher, implying that the last category of HDDS testlet, implying that there were months when some
the HFIAS testlet was endorsed by households higher up Windhoek urban households had inadequate food supply.

Fig. 4  Item Discriminant parameter estimates

13
1394 V. Charamba et al.

Table 4  Parameter estimates for the Windhoek food security data


Testlet Parameter Min Max Mean Std. Dev

Overall Abilitya -3.54 3.03 -0.01 1.09


Discriminant 0.35 2.67 1.41 0.57
Testlet1(HFIAS) Threshold 1 -0.33 0.66 0.03 0.36
Threshold 2 -1.03 1.32 -0.32 0.68
Threshold 3 0.87 2.02 1.20 0.37
Testlet 2 (HDDS) Difficulty 1 -3.27 4.49 -1.14 2.13
Testlet 3 (MIHFP) Difficulty 2 -0.18 1.14 0.76 0.37
a
Ability denotes the distribution of food insecurity levels in Windhoek

3.1 Calculation of cut‑off points for categorization according to Mukaka (2012) (r = −0.533) and increase with
of households into food insecurity classes the LPI, with a moderate positive correlation (r = 0.679),
that is, low-income households in extreme poverty had
The values that were imputed into Eq. (1) to calculate the θ val- higher food insecurity levels. The IRT behaved as expected
ues for cut-off points 𝜏1, 𝜏2, and 𝜏3, are shown in Table 5. The against the LPI and the income levels. Making the index
values were then used to categorize households according to externally valid.
estimated food insecurity levels and classes in Table 6 results.

3.2 Comparison with other food security/ 4 Discussion


insecurity measures
Poverty reduction, food insecurity and undernutrition remain
The results are shown in Table 6, together with the categorisa- high on the development agenda. The need for coming up
tion from the Household Food Insecurity Access Prevalence with good food insecurity measurements for project target-
(HFIAP: see Nickanor et al., 2017; Charamba et al., 2019). ing and governance cannot be underestimated. Targeting
The results in Table 6 show that there isn’t much dif- based on poor indicators can be costlier than not targeting
ference in the proportions of respondents categorized as at all. In practice, it is difficult to detect which measure is
food insecure according to the different indices and this is best for which situation (Maxwell et al., 2013), and often-
supported by high p-values from the Wilcoxon paired ranks times, indices classifications of households are not in agree-
test (p > 0.05). However, the IRT index estimated more ment. As such, having a single measure that captures all the
food-secure households than the other indices. information presented by different indicators might be the
Table 7 gives the correlation coefficients (in the upper best alternative in food security measurement. The objec-
triangle) and Cohen’s Kappa (κ) lower triangle) values to tive of the current study was to come up with an aggregate
measure the correlation and agreement between food secu- food insecurity indicator for improved targeting of safety-
rity measures and categories respectively, for the different net, policy intervention and governance by incorporating
measurement methods. The results showed that the IRT food the household experiential perceptions on food insecurity
insecurity measure had the highest correlations and highest (HFIAS), dietary quality measure (HDDS) and the duration
values of the Cohen’s Kappa agreement with the individual of food insecurity (MIHFP) into one food access measure.
food insecurity scores that were aggregated in its formation, The metric can be constructed using evidence-based data
thus the index is internally consistent. The IRT categories that can be collected timeously at a relatively low cost and
had a substantial agreement (𝜅 = 0.644) with the HFIAP analysis as they can be collected/are usually collected in one
categories and a fair agreement (𝜅 = 0.24) with the HDDS questionnaire, thus having limited effect on time and finan-
food insecurity classes. Food insecurity decreased with cial implications. In addition, the application of the testlet
an increase in income levels with a moderate correlation model requires limited time and resources as free software
and programmes are readily available.
Table 5  Calculation of cut-off points for categorization of households The category characteristic curves (CCC) for the com-
Probability Discriminant Difficulty Testlet Cut-off point bined data showed monotonicity, that is, lower-order response
categories were more likely to be endorsed by food-secure
0.7 a1 = 1.508 b13 = -1.1500 𝛾11 = -0.1338 𝜏1 = -0.45433 households while higher higher-order response categories
0.7 a3 = 1.731 b33 = -0.9284 𝛾11 = -0.1338 𝜏2 = -0.30512
were most likely to be endorsed by households that were more
0.7 a6 = 2.061 b63 = -0.5934 𝛾11 = -0.1338 𝜏3 = -0.04849
food insecure. Similar findings were observed by Charamba

13
Application of Item response theory modelling to measure an aggregate 1395

Table 6  Comparison of IRT and Model Food secure Mildly food Moderately food Severely food p-value
FANTA food security measures insecure insecure insecure

Testlet 22.1 3.5 7.3 67.1


HFIAP 16.4 3.4 13.2 67.1 0.992
HDDS 10.3 6.5 17.1 66.1 0.999
MIHFP* 10.1 3.4 19.1 67.3 0.996

et al. (2019) in their study to validate the use of the HFIAS compared to HFIAS and MIHFP items, which could be ema-
for the measurement of Windhoek households' food security nating from a similar diet across households.
levels. The curve for the second category was flat, implying The current index has been noted to be internally con-
that it was chosen by fewer respondents. This could imply sistent and externally valid with moderated correlations
that due to income disparities in Windhoek, there is a higher between the food insecurity status and LPI and income lev-
proportion of food secure and severely food insecure house- els, concurring with observations by McCordic et al. (2022)
holds than mildly and moderately food insecure. It could who observed significant moderate correlations between
also imply that the second category option was not neces- inconsistent access to water, medical facilities, electricity
sary for the Windhoek households. The difficulty param- and cash income against food insecurity index computed by
eter estimates for the HDDS testlet were lower, probably combining the HFIAS, HDDS and MAHFP via a mini-max
because households endorsed similar food groups. How- normalization and geometric means for aggregation. The
ever, the higher difficulty parameter for the MIHFP might lower Kappa values for HDDS and MIHFP (against rela-
mean that the households had several months of inadequate tively high correlations) may be because of the cut-off points
food provisioning. for categorization of the IRT, HDDS and MIHFP which is
The IRT food insecurity measure recorded the highest non-standard and might be worth further investigation. The
correlation and level of agreement with the FANTA indica- prevalence estimates that each indicator provides are func-
tors that were incorporated in its computation, suggesting tions of cut-off points used for the categorization of house-
that the index is internally consistent and the single compo- holds into different food security classes. Since there are no
nents measured by these indicators were a proper represen- universally acceptable cut-offs for most of these measures,
tation of the individual metrics being combined. In addi- these may be subject to further research and debate. The
tion, the metric behaved as expected, as it increased with cut-offs used in this paper are not intended for application
poverty levels, and MIHFP and decreased with increased in other contexts without location-specific consideration
dietary quality and income levels suggesting that poorer being considered. They are just meant to trigger investiga-
households had higher food insecurity levels. This is an tion and debate on the possibility of combining several met-
indication that the IRT measure can be further developed rics of food insecurity into one measure. Further studies are
into a food security measurement that combines access to required on algorithms to come up with cut-off points for
dietary variety and the months of inadequate household food different metrics, including the IRT metric itself. However,
provision and anxiety and uncertainty, insufficient quality the lower Kappa values might mean a different classifica-
and quantity food intake and their physical consequences tion of the same households by the indices. However, the
into one measure. The correlation coefficients and Cohen's food insecurity classes were compared with income and LPI
Kappa (κ) agreement measurements between the IRT food quartiles. It might not necessarily be the case that food inse-
security measures and the FANTA measurements were the curity classes in Windhoek have equal representation and
lowest for the HDDS. This is probably because the testlet this might have posed a challenge in the external validation
model considered the HDDS items to be less discriminating process used.

Table 7  Spearman’s rank IRT HFIAS HDDS MIHFP Income LPI


correlation and Cohen Kappa
agreement between metrics IRT 0.834a -0.579a 0.512a -0.533a 0.679a
HFIAS 0.644 -0.558a 0.499a -0.481a 0.661a
HDDS 0.240 0.215 0.351a 0.387a -0.584a
MIHFP 0.244 0.197 0.114 0.367a -0.617a
Income 0.440 0.340 0.140 0.428 -0.515a
LPI 0.514 0.388 0.213 0.434 0.680
a
Correlation significant at 0.01 level of significance

13
1396 V. Charamba et al.

The current study has been conducted on data collected insecurity levels and could have been dropped from the
from urban households. However, the applicability of the analysis. The two items have been noticed by Charamba
HFIAS, HDDS and MIHFP to urban households had been et al. (2019) to have low (0.21) and negative (-0.41) dis-
questioned (Haysom & Tawodzera, 2018) as they had been crimination ability respectively and also have the highest
originally designed for application in rural communities. It inlier-sensitive fit (infit) and outlier-sensitive fit (outfit)
is worthy of further investigation to assess the applicabil- statistics outside the acceptable range (Na et al., 2015;
ity of the current model to data collected from rural house- Nord, 2014), implying that they do not fit well in the scale
holds to which these tools had been specifically designed. for this data. Infit and outfit statistics, normally expressed
Considering the income disparity and dietary diversity for as the mean of summed squared standardized residuals
households in Windhoek, it might be worth further investi- (MNSQ), indicate the goodness of fit of a test item to the
gation to incorporate a variable that accounts for subgroups model and index being measured. Although the items had
of respondents that might exist within a population (see a minimal contribution to scale, they were not dropped
Wunderlich & Norwood, 2006). from the analysis because the IRT scale was meant to be
The validity of the IRT propensity as a measure of food an aggregate of the HFIAS, HDDS and MIHFP where
security narrows down to whether the IRT latent construct none of the items was dropped, and it was to be com-
is associated with the true food security (Johnson, 2004). pared against these individual scores for internal consist-
Johnson (2004) added that if the true food security is not ency checks. However, for estimation purposes, one might
related to the IRT construct, it doesn’t matter how well consider dropping off items with minimal contribution to
the item responses and the construct adhere to the spe- the latent measure. Item performance may not be universal
cific IRT model and thus can only be ascertained through and need not be generalized to all data sets. For example,
validity studies. A validity study suggested by Johnson items that performed poorly in the Windhoek households'
(2004) requires monitoring several individuals (house- data set might behave differently in a different set-up with
holds) over 12 months to determine their food security different food systems. The lack of statistical differences
status and then administer a food security tool, and then between the food security categories determined from the
conduct an analysis to determine the relationship between IRT index observed in Table 6 could have been because
the IRT propensity and the actual food security levels. the statistical power has been compromised by a small
Further studies incorporating such a validity study might sample size (only four categories). However, the problem
be worth undertaking to assess the validity of using the of a small sample size could have been addressed by the
combined measure in different contexts. The proposed employment of a non-parametric test for comparison.
measure does not cover all aspects of food access as those
go beyond what the HFIAS. HDDS and MIHFP measure,
but rather provide a proxy for its measurement. 5 Conclusion and recommendations
McCordic et al. (2022) suggested normalizing the
HFIAS, HDDS and MAHFP before aggregating them so Much of the demand for indicators of food insecurity is
that they maintain equal contribution, arguing that other to measure the impact of programmes, policies and to
methods would only support household ranking other than determine household groups that require interventions.
scoring. However, normalizing is not necessary for the The conceptual and measurement frameworks proposed in
current model neither is equal contribution assumed. The the current study have presented opportunities for a more
IRT model is advantageous as an estimation method in that holistic measure of food insecurity, as the proposed meas-
it provides mechanisms for assessing the contribution of ure has been externally validated against other measures
each item to each measure based on the item discrimina- of food security and household socio-economic status. In
tion and item goodness of fit parameters (see Charamba addition, most food security measurements have been orig-
et al., 2019). In addition, it supports scoring as the latent inally designed for measurement in rural settings, the com-
variable is measured on a continuous continuum. It will bined measure can provide an alternative for food insecu-
be difficult to perform sensitivity analysis and confusion rity measurement in urban areas where food systems are
matrices using the current model of food insecurity on a different and there is high-income disparity, especially in
continuous continuum and classification which is done Windhoek, and diets can be defined more by ethnicity that
later, is not part of the model algorithm. food security levels, coupled with lack of formal safety-
Although all the items in the HFIAS, HDDS and nets The study can help in the translation of research into
MIHFP were used to calculate the current metric, the policy as it avoids the use of single measures which might
items on vegetable and fish and seafood consumption had not be congruent. However, the measure had only been
lower discrimination ability, that is, they play a less signif- computed using a single urban sample in the Global South
icant role in separating households according to their food and hence is not generalisable unless it is tested in several

13
Application of Item response theory modelling to measure an aggregate 1397

different set-ups with different food systems, possibly both Bertelli, O. (2020). Food security measures in sub-Saharan Africa.
rural and urban settings. It is difficult to establish cut-off A validation of the LSMS-ISA scale. Journal of African Econo-
mies, 29(1), 90–120.
points for classifying households into different levels of Bilinsky, P., & Swindale, A. (2010). Months of adequate household
food insecurity. The empirical results in the current study food provisioning (MIHFP) for measurement of household food
do not specifically provide the thresholds or cut-offs for access: Indicator guide. Washington, DC: FANTA. Version 4.
any of the indicators discussed, but only to trigger further Charamba, V., Nickanor, N., & Kazembe, L. N. (2019). HCP discus-
sion paper no. 37: Validation of the HCP survey tool for meas-
research and debate on a composite food insecurity met- uring urban food insecurity: An item response theory approach.
ric. Further research and analysis over a much broader Coates, J. (2013). Build it back better: Deconstructing food security for
range of contexts will be required to empirically assess the improved measurement and action. Global Food Security, 2(3),
applicability of the proposed model in measuring house- 188–194.
Coates, J., & Maxwell, D. (2012). Reaching for the stars?: Univer-
hold food insecurity to substantiate the current findings sal measures of household food security. UNFAO international
and document if the proposed food insecurity construct scientific symposium on food security and nutrition security
applies to different contexts. Considering that data similar information: From valid measurement to effective decision mak-
to common metrics such as the HFSSM, the FCS had been ing. Rome.
Coates, J., Swindale, A., & Bilinsky, P. (2007). Household food inse-
incorporated into the scale, further research may consider curity access scale (HFIAS) for measurement of food access:
incorporating other metrics such as the CSI into the com- Indicator guide. Washington, DC: Food and Nutrition Techni-
putation of the food insecurity construct using the testlet cal Assistance Project, Academy for Educational Development.
model. For the IRT measure, cut-off points will require version 3.
Coates, J., Frongillo, E. A., Rogers, B. L., Webb, P., Wilde, P. E., &
suggestions from Johnson (2004). Houser, R. (2006). Commonalities in the experience of house-
hold food insecurity across cultures: What are measures miss-
Acknowledgements The researchers acknowledge the Open Society ing? The Journal of Nutrition, 136(5), 1438S–1448S.
Foundation for South Africa, the former Department of Statistics and Chung, K. (Ed.). (1997). Identifying the food insecure: The applica-
Population Studies (University of Namibia), the African Food Security tion of mixed-method approaches in India. The International
Urban Network (AFSUN), the Hungry Cities Partnership (HCP) and Food Policy Research Institute.
the Balsillie School of International Affairs (BSIA) for the secondary Deitchler, M., Ballard, T., Swindale, A., & Coates, J. (2010). Valida-
data that was used in this research. tion of a measure of household hunger for cross-cultural use.
Washington, DC: FANTA.
Funding Open access funding provided by University of Namibia. ENN. (2007). Simple tools for measuring household food access and
dietary diversity. Field Exchange, 31, 14. Retrieved April 2,
Data availability The data is available from the authors upon request. 2022, from www.​ennon​line.​net/​fex/​31/​simpl​etools
FAO. (1996). Rome Declaration on world food security. World Food
Declarations Summit. Rome: Food and Agriculture Organization.
Gelman, A., Carlin, J. B., Stern, H. S., & Rubin, D. B. (1995). Bayesian data
Conflict of interests The authors declare no conflicts of interest re- analysis. Chapman and Hall/CRC.
garding the publication of this paper. Haysom, G., & Tawodzera, G. (2018). Measurement drives diagnosis
and response: Gaps in transferring food security assessment to
Open Access This article is licensed under a Creative Commons Attri- the urban scale. Food Policy, 74, 117–125.
bution 4.0 International License, which permits use, sharing, adapta- Headey, D., & Ecker, O. (2013). Rethinking the measurement of food
tion, distribution and reproduction in any medium or format, as long security: From first principles to best practice. Food Security,
as you give appropriate credit to the original author(s) and the source, 5(3), 327–343.
provide a link to the Creative Commons licence, and indicate if changes HLPE. (2017). Nutrition and food systems. High-level panel of
were made. The images or other third party material in this article are experts on food security and nutrition. Rome: Committee on
included in the article's Creative Commons licence, unless indicated World Food Security.
otherwise in a credit line to the material. If material is not included in Huluka, A. T., & Wondimagegnhu, B. A. (2019). Determinants of
the article's Creative Commons licence and your intended use is not household dietary diversity in the Yayo biosphere reserve of
permitted by statutory regulation or exceeds the permitted use, you will Ethiopia: An empirical analysis using sustainable livelihood
need to obtain permission directly from the copyright holder. To view a framework. Cogent Food & Agriculture, 5(1), 1690829.
copy of this licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/. Johnson, M. S. (2004). Item response models and their use in meas-
uring food insecurity and hunger. Washington, DC: NAS Com-
mittee on National Statistics Workshop on the Measurement of
Food Insecurity and Hunger.
References Jones, A., Ngure, F., Pelto, G., & Young, S. (2013). What are we
assessing when we measure food security? A compendium and
Abuelhaj, T. (2007). Methodological concerns in the measurement of review of current metrics. Advances in Nutrition, 4, 481–506.
undernourishment, dietary diversity and household food insecu- Kirkland, T. M., Kemp, R. J., Hunter, L. M., & Twine, W. M. (2013).
rity. Retrieved July 23, 2018, from www.​ipcin​fo.​org Toward an improved understanding of food security: A methodo-
Barrett, C. B. (2010). Measuring food insecurity. Science, 327(5967), logical examination based in rural South Africa. Food, Culture &
825–828. Society, 16(1), 65–84.

13
1398 V. Charamba et al.

Leroy, J., Ruel, M., Frongillo, A., Harris, J., & Ballard, T. (2015). Vonai Charamba is a SeniorLec-
Measuring the food access dimension of food security: A critical turer (Biostatistician) at the Uni-
review and mapping of indicators. Food and Nutrition Bulletin, versity of Namibia. She studied
36, 167–195. for her PhD in Science (Statis-
Lunn, D. J., Thomas, A., Best, N., & Spiegelhalter, D. (2000). WinBUGS tics) at the University of Namibia
— a Bayesian modelling framework: Concepts, structure, and exten- and BSc and MSc in Science
sibility. Statistics and Computing, 10, 325–337. (Statistics) at the University of
Maxwell, D., Coates, J., & Vaitla, B. (2013). How do different indi- Zimbabwe. She is currently
cators of household food security compare? Empirical evidence teaching Biostatistics, Biometry
from Tigray (pp. 1–19). Feinstein International Cente. and Research Methods modules
Mattes, R., Dulani, B., & Gyimah-Boadi, E. (2016). Africa’s growth to undergraduate and postgradu-
dividend? Lived poverty drops across much of the continent. ate students in the School of
Afrobarometer. Agriculture and Fisheries Sciences and the School of Health Sciences
McCordic, C., Frayne, B., Sunu, N., & Williamson, C. (2022). The house- and Veterinary Medicine, at the University of Namibia. Her research
hold food security implications of disrupted access to basic services interests are in the application of Item Response Theory Modelling to
in five cities in the global south. Land, 11(5), 654. social research, Application of Generalized Linear Models to Agricul-
Mock, N., Morrow, N., & Papendieck, A. (2013). From complexity tural Research (Food security and sustainable agriculture), Bayesian
to food security decision-support: Novel methods of assessment Inference, Climate change and Social research. Vonai strongly advo-
and their role in enhancing the timeliness and relevance of food cates student-centred teaching and learning where students learn
and nutrition security information. Global Food Security, 2(1), through hands-on application. She believes learning is a cycle where
41–49. one continually learns from students and colleagues through informa-
Mukaka, M. M. (2012). A guide to appropriate use of correlation coef- tion dissemination. She has published some Agriculture, Climate
ficient in medical research. Malawi Medical Journal, 24(3), 69–71. change and Food Security work in peer-reviewed journals and had
Mutea, E., Bottazzi, P., Jacobi, J., Kiteme, B., Speranza, C. I., & Rist, S. several conference presentations.
(2019). Livelihoods and food security among rural households in the
north-western Mount Kenya region. Frontiers in Sustainable Food Lawrence N. Kazembe is a Pro-
Systems, 3, 98. fessor of Applied Statistics at the
Na, M., Gross, A. L., & West, K. P. (2015). Validation of the food University of Namibia. He previ-
access survey tool to assess household food insecurity in rural ously served as a Senior Biostat-
Bangladesh. BMC Public Health, 15, 1–10. istician at Malawi Liverpool
Namibia Statistics Agency. (2013). Namibia 2011 population and hous- We l l c o m e Tr u st C l i n i c a l
ing census: Main report. Windhoek, Namibia: Namibia Statistics Research Programme in Blan-
Agency. tyre, Malawi; consultant statisti-
National Research Council. (2005). Measuring food insecurity and cian with Malaria Alert Centre;
hunger: Phase 1 report. National Academies Press. and as a research fellow at Medi-
Nickanor, N., Kazembe, L., Crush, J., & Wagner, J. (2017). The super- cal Research Council of South
market revolution and food security in Namibia (rep., pp. i-87). Africa, Durban in the Malaria
Kingston, ON and Cape Town: African Food Security Urban Net- Lead Research Programme. He
work. Urban Food Security Series No. 26. has published extensively, with over 100 peer-reviewed publications in
Nord, M. (2014). Introduction to item response theory applied to food food security, malaria research and population health. In 2008, he was
security measurement: Basic concepts, parameters, and statis- recognized as a young scientist and in 2009, he was accorded a young
tics. Retrieved June 30, 2018, from http://w ​ ww.f​ ao.o​ rg/e​ conom
​ ic/​ research affiliate award by the Academy of Sciences in Developing
ess/​ess-​fs/​voices/​en Countries (TWAS), and a young biometrician medal from the Sub-
Oldewage-Theron, W. H., & Kruger, R. (2008). Food variety and die- Saharan Network of International Biometrics Society (SUSAN-IBS),
tary diversity as indicators of the dietary adequacy and health sta- in 2009. He has supervised over 10 PhDs and 50 MSc graduates. His
tus of an elderly population in Sharpeville, South Africa. Journal main research interests are in Bayesian statistical modelling and spatial
of Nutrition for the Elderly, 27(1–2), 101–133. analysis with applications in population health.
Swindale, A., & Bilinsky, P. (2006). Household dietary diversity score
(HDDS) for measurement of household food access: Indicator Ndeyapo Nickanor is an Associ-
guide. Food and Nutrition Technical Assistance Project, Academy ate Professor in the Department
for Educational Development. of Computing, Mathematical and
United Nations. (1975). Report of the World Food Conference. United Statistical Science in the School
Nations. of Science as well as the Execu-
Vellema, W., Desiere, S., & Haese, M. (2016). Verifying validity of tive Dean of the Faculty of Agri-
the household dietary diversity score: An application of Rasch culture, Engineering and Natural
modelling. Food and Nutrition Bulletin, 37, 27–41. Science at the University of
Wang, W. C., & Wilson, M. (2005). Exploring local item dependence Namibia. She is a research affili-
using a random effects facet model. Applied Psychological Meas- ate of the African Food Security
urement, 29(4), 296–318. Urban Network (AFSUN), which
Webb, P., Coates, J., Frongillo, A., Rogers, B., Swindale, A., & Bil- continues to extensively research
insky, P. (2006). Measuring household food insecurity: Why it’s food security in Southern Africa
so important and yet so difficult to do. Journal of Nutrition, 136, and the Hungry Cities Partner-
1404S-1408S. ship (HCP). She is also working with the African Centre for Cities at the
Wunderlich, G. S., & Norwood, J. L. (2006). Food insecurity and hun- University of Cape Town as part of the IDRC-funded Nourishing Spaces
ger in the United States. National Academies Press. Project. Her research interest continues to grow in the food security area.

13

You might also like