Data collection

This study utilized baseline data from a longitudinal study examining the effectiveness of the Arctic Char Distribution Project (AC/DP); a culturally sensitive food security intervention for pregnant women in Nunavik. The AC/DP was conducted in collaboration with the Nunavik Regional Board of Health and Social Services between September 2013 and April 2014 and was approved by the Laval University ethics committee. Written consent was obtained from all participants. Details about the AC/DP are described elsewhere.[31] A total of 130 pregnant women living in Nunavik participated in the study. The mean gestational weeks of participants at enrollment was 16.4 (SD = 6.5) with 40 participants in their first trimester and 90 after their 13^th week.

Food insecurity was measured at an individual level among participants using the HFIAS, slightly adapted for the Nunavik population based on feedback from local partners (see S1 File for the questionnaire). The “Worry” (did you worry that you would not have enough food?) and “Lack Resources” (Were you not able to eat the kinds of food you preferred because of a lack of resources) items were designed to measure mild food insecurity. (23) Items measuring “Limited Variety” (did you have to eat a limited variety of foods due to a lack of resources) and “Unwanted Food” (did you have to eat some foods that you really did not want to eat because of a lack of resources to obtain other food items you prefer?) were purported to address mild to moderate food insecurity. “Reduced Meal Size” (did you have to eat a smaller meal than you wanted because there was not enough food available for you?) and “Reduced Meal Frequency” (did you have to eat fewer meals in a day because there was not enough food?) items were designed to measure moderate to severe food insecurity. Any level of endorsement of the “House Empty” (was your house ever out of food of any kind because of a lack of resources to get food), “Sleep Hungry” (did you go to sleep at night hungry because there was not enough food), or “Whole Day Without Eating” (did you ever go a whole day and night without anything to eat because there was not enough food in your house?) items originally classified participants as severely food insecure. (32)

The HFIAS is a 9-item questionnaire that assesses the severity of household-level food access deficiencies; it is designed to track the effectiveness of food security interventions internationally.[32] The HFIAS has demonstrated acceptable test-retest reliability and convergent validity based on socioeconomic status, and internal consistency (α = 0.80 to 0.91) in Lebanon[21] and Tanzania.[20] In one study, participating in an urban agriculture program was associated with mild improvement in food insecurity measured by the HFIAS.[19]

Statistical analyses

Chronbach’s alpha describes the internal consistency of a scale.[12] This was computed as a preliminary assessment of the degree to which the HFIAS appears to measure the same latent construct. This was further supported by a principal component analysis conducted on the HFIAS data (see S1 Table and S1 Fig for a summary of those results).

Rasch rating scale model (RSM) is an item response theory (IRT) method designed to analyze data from measures using more than two ordinal response categories for each item, such as the HFIAS. Rasch models, including the RSM, allow for measurement of a person’s position along a continum of a latent variable, known as the person parameter, independently of the difficulty parameters, which describe the relative likelihood of various responses to each item.[33] This allows for a more accurate assessment of the construct validity of a rating scale than traditional scoring systems. Compared to other polytomous Rasch models, RSM is more appropriate for use in small samples.[34] Another strength of IRT models, including RSM, is that they allow one to determine the relative contribution of each item to information obtained from the scale along a continuum of the latent variable measured.[35] We conducted a one-parameter RSM analysis on the HFIAS data.

Difficulty parameters were estimated with STATA Data Analysis and Statistical Software[36] using Gaussian random effects method, which describes the level of food insecurity expected for each response while minimizing the difference between that and the observed levels.[37] Item information curves provide a visual representation of the difficulty parameters, based on the results of IRT analyses.[35] In health care research, they can demonstrate the level of severity each item addresses. If the scale demonstrates a high level of content validity, it should provide information about the entire spectrum of food security/insecurity.

Infit mean square values (MSV) for each item can be calculated with Rasch models to understand the degree to which the scale appears to be unidimensional.[38] We calculated the infit mean square values for the HFIAS using the Extended Rasch Modeling package for R.[39] Unidimensionality refers to the extent to which items on a scale measure a single latent construct, such as food security.[40] Infit MSV over 1.4 are considered to represent items that introduce too much statistical noise into the model and MSV below 0.6 indicate limited information is provided by the item due to high inter-item correlations. If the scale is unidimensional, infit MSV should fall within acceptable ranges. Two items were identified as potentially problematic due to cultural concerns. The RSM was run with these items removed sequentially. If the scale displays good discriminant validity, removing items should not improve its fit.

Akaike information criterion (AIC) and Bayesian information criterion [41] are indices of comparative model fit.[42] They indicate the relative amount of unmodeled variance provided by different statistical models. Lower AIC and BIC values indicate a preferred model. AIC and BIC indices were calculated for the RSM utilizing the original HFIAS scale and for the RSM of the HFIAS without the two potentially problematic items.

Person parameter (Theta) values, which are estimates of the latent trait experienced by each individual[43], in this case food insecurity, were calculated to provide a more accurate estimate of food security severity compared to the HFIAS original scoring procedure. Theta values are estimated independently of item parameters, which estimate the degree to which an individual’s responses on a scale are attributable to item characteristics, such as difficulty, rather than the latent variable.[44] Theta values cannot be calculated for the lowest and highest raw scores. These scores were imputed using Baysian estimation (see Baker & Kim 2004 Chapter 7 for description of this method[45]). Theta values are reported for demographic groups predicted to show differences in levels of food insecurity.

Summary of main findings

The HFIAS demonstrated unidimensionality and promising construct validity in this sample. However, the modified model without two items outperformed the original model. The House Empty item was designed to measure severe food insecurity, yet appears to capture mild to moderate food insecurity in this sample. This may be due to cultural factors regarding food sharing between households in Nunavik.[15] Individuals who do not have food in their houses can often access food from a friend or family member’s house. As such, this item does not necessarily tap into a severe lack of access to food in this sample.

The demographic variables most strongly associated with food security were age, having more than two fishers in the household, and having more than two hunters in the household. This substantiates the findings from other studies among Canadian Inuit populations suggesting that hunting and fishing, as means of providing families with country food and/or with supplementary income, are strong determinants of food security.[5, 7] As such, this further demonstrates the need to collect data on access to country food in future measurements of food security among Inuit populations. One unique finding was women who were married or in a domestic partnership were more food insecure than single women. This could be explained by crowding. [6]

Across all items, the scale was consistently able to differentiate between individuals experiencing each symptom “often” versus “sometimes” better than between those experiencing the symptoms “sometimes” versus “never”. This is unusual and might point towards ambiguity in understanding the response categories. It is often recommended that scales have an even number of categories, as participants frequently choose the middle option to avoid committing to a response.[46] The word “never” in questionnaires can also bias responses, as people often are reluctant to endorse absolutes, particularly when discussing high base-rate situations, such as the mild to moderate items on the HFIAS.

Comparisons with the literature

This is the first study to examine item-level psychometric properties of a food security questionnaire among an Indigenous population in Canada or the United States. The original scoring system of the HFIAS states that affirmative responses to the Worry and Lack Resources items indicate mild food insecurity.[32] The House Empty, Sleep Hungry, and Whole Day Without Eating items are theoretically purported to measure severe food insecurity while the remaining items primarily measure moderate food insecurity. Comparing this scoring system to the difficulty parameters displayed in the current sample in Fig 1, these items appear to function slightly different in the current sample. The Whole Day Without Eating item is associated with a greater difficulty parameter than the other items and thus seems to measure severe food insecurity. Likewise, the difficulty parameter for the Sleep Hungry item indicates that it somewhat taps into the upper level of food insecurity as well. The difficulty parameters for the remaining items appear to cluster together somewhat at an intermediate level of food insecurity, suggesting a degree of overlap and potentially redundancy. It is possible that the cultural differences in the streams of food acquisition affected participants’ response patterns on this scale. Fig 2 displays a model with two items identified as possibly culturally/linguistically inappropriate for this population removed, which increased the average level of information and reduced the overlap between the difficulty parameters for the items. This suggests a potentially stronger model, although removing these items limits our ability to utilize the original scoring system and necessitates the development of a new classification method for this population.

Strengths

This is the first study in Northern Canada to use a measure that theoretically captures different food sources and unique modes of access in the Inuit food system. Given the culture of Nunavik, and other Arctic regions’ mixed economy, it is particularly important to carefully validate food insecurity instruments locally. Currently, policy responses to food insecurity in Northern Canada are hampered by measurement debates [4], which reinforces the need to validate the tools employed.

This is one of a small number of studies that have investigated food security among pregnant women, a particularly important population given the potential for harm to the developing fetus.[22] Accurate assessment of food insecurity among pregnant women is instrumental in providing culturally appropriate and effective interventions.[29, 30]

This is one of the first studies to utilize a polytomous Rasch model in evaluating a food insecurity questionnaire. By using a polytomous model, we were able to uncover potential problems in their wording and/or format. Other studies that fit food insecurity data to Rasch models first reduced the data to dichotomous variables and thus only were able to distinguish between affirmative versus negative responses.[33] In doing so, they were not able to detect these differences. Furthermore, constructing an empirically-based scoring system using the person parameter estimates from the RSM has likely improved the accuracy with which the HFIAS can measure food insecurity and predict related health problems in this population and others.

Limitations

Some of the items in the HFIAS were overly long and utilized complex sentence structure. This may have compounded potential concerns about translation issues in Nunavik, where many people speak English as a second language and thus may have had difficulty understanding some questions in the HFIAS.

Because the population of pregnant women in Nunavik during the study period was considerably small, the sample size restricted the types of analyses that could be conducted. Since the sample was relatively demographically homogeneous, we were not able to parse it into groups in a manner that would allow us to perform differential item functioning analyses to accurately establish local independence and measurement invariance. This is recommended for future studies using a more demographically heterogeneous sample. However, these issues were balanced with the need to increase our understanding of the accurate measurement of food insecurity among pregnant women in Nunavik.

Food insecurity measures examine “deficiencies” in food access, availability, quality and/or utilization in a particular population. As all scales can be seen as imperfect measurements of latent variables, we can only assess the extent to which an instrument measures the components of a construct that it directly queries.[8] As such, we cannot assume that the absence of food insecurity, as measured by scales such as the HFIAS, is necessarily food security. This is especially of concern considering that the HFIAS only measures the access component of food insecurity,[26] when there have been concerns raised about the availability, supply, and utilization components of food insecurity in Canadian Inuit populations.[3] While understanding the deficiencies in access to food among such populations is indeed important, such tools are insufficient in understanding the full spectrum of food insecurity/security in these communities.

Future directions

To establish an accurate food insecurity prevalence rate for this population, it is recommended that future researchers develop a socially and biologically meaningful cutoff. This task should be driven by theory and research on food insecurity in indigenous populations and is beyond the scope of this article. A second recommendation is to develop a scoring program that can facilitate the use of the Theta values among clinicians and public health practitioners.