2.1. Study Design and Sampling
This analysis uses data from two multi-stage, cross-sectional, cluster surveys (2009 and 2012). The surveys were conducted in the same season and employed identical methods for sampling and participant recruitment. The baseline survey was conducted in September–December 2009, and was representative nationally and at the level of each of the three survey strata: the North (North, Far North, and Adamaoua administrative regions); the South (all seven remaining regions, with the exception of Yaoundé and Douala); and Yaoundé and Douala (the two major urban areas, comprising ~20% of total population) [
12]. Ninety clusters (30 per stratum) were selected using the 2005 census database (the most recent available at the time of the survey) and proportionate to population size sampling. The interim evaluation survey was conducted in October–November 2012, in the same 30 clusters in Yaoundé/Douala (15 per city) that were sampled in the 2009 survey.
The two surveys used identical methods to sample households within each cluster, but different individuals participated in each survey. Sampling within each cluster was conducted as follows: first, a random start point was identified by locating the approximate geographic center of the cluster, walking a straight line from the cluster center to the cluster perimeter in a randomly-selected direction, assigning a number to all households encountered along this line, and randomly selecting a household to begin sampling. Additional households were identified by the systematic sampling of adjacent households.
2.3. Participant Eligibility and Consent
Households were eligible to participate if there was at least one child of 12–59 months of age and one woman of reproductive age (15–49 years) who was the child’s primary caregiver. Children and women were eligible to participate if they had lived in the household for at least one month and did not have self-reported “severe fever”, diarrhea with dehydration, or another severe illness at recruitment or between recruitment and data collection (i.e., during the 72 h prior to data collection). If multiple eligible children were present in the household, one child was chosen at random to participate (by drawing from a hat). The index caregiver was selected as the primary caregiver of the eligible child. Caregivers were eligible to provide a breast milk sample if the breastfeeding child was at least one month of age, regardless of whether the lactating woman or breastfeeding child was selected to participate in the full survey.
Women provided informed oral consent for themselves and the child to participate, with permission from the head of the household where appropriate. The surveys were conducted in accordance with the Declaration of Helsinki. The 2009 survey was approved by the National Ethics Committee of Cameroon (Authorization No. 047/CNE/DNM/O9) and the Institutional Review Board of the University of California, Davis (Protocol #200917294). In 2012, the National Ethics Committee had suspended activity during a period of reorganization; thus, approval was obtained from the Cameroon Ministry of Public Health and the Institutional Review Board of the University of California, Davis (Protocol # 364876).
2.4. Data Collection
2.4.1. Household Characteristics and Fortified/Fortifiable Food Consumption
Bilingual (French and English) interviewers administered questionnaires to collect information on household demographic and socio-economic characteristics, including information on the household size and composition, educational level and occupation of the caregiver and head of the household, housing material, sources of drinking water and energy for light and cooking, and household possessions.
Information on the consumption of refined cooking oil was collected using a modified version of the Fortification Rapid Assessment Tool (FRAT), which was developed to quickly collect information on the consumption of fortified (or potentially fortifiable) foods [
14]. We retained the food frequency aspect of the FRAT questionnaire, but did not retain the use of a partial 24 h dietary recall to estimate the quantity of fortified foods consumed. In the present study, the interviewers administered a food frequency questionnaire (FFQ) which inquired about the consumption of refined oil in different preparations (e.g., oil in fried foods, oil in sauces, etc.). Respondents were asked how many days during the previous seven days they had consumed each food, and the number of times that they consumed the food on the previous day on which the food was consumed. To discriminate between refined (fortifiable) oil and unrefined oil (e.g., red palm oil), data were collected separately for each type of oil that the respondent consumed in the previous seven days. Women provided responses for themselves and for the child.
In 2009 only, 24 h dietary recall interviews with replicates in a ~10% subset (two non-consecutive days of data) were conducted to quantify the total nutrient intakes of women and children. Details of the methods of data collection and analysis are available elsewhere [
12,
15].
2.4.2. Blood Samples
Biological sample collections took place at a central location within each cluster. Containers for blood collection and storage were covered in foil to prevent exposure to light. Venous blood (5–7 mL) was collected by antecubital (or, for some children, metacarpal) venipuncture into tubes containing lithium heparin (Sarstedt, Nümbrecht, Germany). Blood samples were placed in a cooler with ice packs for <2 h until centrifugation to separate plasma (10 min at 2500× g). Plasma was aliquoted within an opaque, portable “hood” to minimize exposure to light and dust during sample handling, and storage vials were covered with aluminum foil.
In 2012 only, an additional 1–2 mL of blood was collected into tubes containing EDTA, for the measurement of hemoglobin and malaria infection in whole blood. For both surveys, hemoglobin was measured in venous blood using a portable photometer (Hemocue, Angelholm, Sweden); for a small number of children for whom sufficient venous blood could not be obtained for hemoglobin measurement, hemoglobin was measured in a capillary sample following a fingerprick. In 2009, current or recent malaria infection was assessed by measuring plasma histidine-rich protein (HRP-2) concentrations using a commercial CELISA kit (Cellabs, Sydney, Australia) [
16,
17]. In 2012, malaria was assessed in whole blood using a rapid diagnostic test (Malaria Ag Pf/Pan, SD Bioline, Standard Diagnostics, Gyeonggi-do, Korea). Individuals with positive rapid diagnostic test results were treated and referred to the nearest health clinic.
2.4.3. Breast Milk
In both surveys, breast milk samples were collected according to the casual sampling method [
13,
18]. The mother was asked to feed her child from the fuller breast. After exactly 30 s, the mother transferred the infant to the other breast, and manually expressed 5–10 mL of milk from the first breast into a plastic container covered in aluminum foil. The time of collection and time of the previous feed (from the breast from which milk was collected) were recorded.
The milk fat concentration was measured in triplicate in the field using the creamatocrit method [
19]. After swirling the milk to ensure a homogenous distribution of fat, the milk samples were drawn into nonheparinized glass microhematocrit tubes and centrifuged for 15 min at 1500×
g. The length of the lipid (‘cream’) layer and the total milk column were measured in duplicate using calipers to the nearest 0.1 mm. The median coefficients of variation (CVs) were 3.8% in 2009 and 3.1% in 2012 (mean = 4.9% for both). After removing an aliquot for milk fat analysis, the remaining milk sample was remixed and aliquoted into storage vials wrapped in foil.
2.4.4. Fortified Foods
During the interview, respondents were asked whether they had any refined oil in their home and, if so, whether they were willing to provide a sample for micronutrient analysis. Samples of ~10 g of oil were collected into sterile, plastic containers covered in foil for protection from light. The VA content of the oil samples was measured on the day of collection using a portable photometric instrument (iCheck Chroma, Bioanalyt, GmbH), according to the manufacturer’s instructions [
20]. This version of the instrument is intended for measurement of the common types of oil consumed in Yaoundé/Douala (palm and groundnut), but not cottonseed or soybean oil, so five household oil samples reported to be cottonseed or soybean oil were excluded from the analyses.
2.4.5. Sample Storage and Shipping
Aliquots of plasma, breast milk, and cooking oil were stored in a cooler with ice packs until the end of the day, when they were transferred to a freezer (with a backup generator) for storage at ≤−20 °C. Samples were shipped on dry ice to Germany for an analysis of the plasma proteins (2009 and 2012) at the VitA-Iron lab, and to the United States for an analysis of the breast milk retinol (2009 and 2012) and plasma HRP2 (2009 only) at UC Davis.
2.4.6. Laboratory Analyses
Plasma indicators of inflammation (C-reactive protein, CRP, and α
1-acid glycoprotein, AGP), and vitamin A status (retinol-binding protein, RBP) were measured using ELISA [
21]. In 2009, the interassay CVs were: RBP, 2.7%; CRP, 6.5%; and AGP, 3.5%. In 2012, the CVs of a plasma pool control sample on 11 plates were: RBP, 3.3%; CRP, 4.5%; and AGP, 5.5%.
Plasma HRP2 was detected using a commercial CELISA kit (Cellabs Pty, Ltd., Brookvale, Australia), following the manufacturer’s instructions, with positive and negative controls provided by the manufacturer. Samples with an absorbance value greater than the optical density of the negative control +0.05 U were considered positive for the P. falciparum antigen.
Breast milk retinol concentrations were measured under dim or yellow light by a reverse-phase HPLC system (Class VP; Shimadzu) [
13,
22]. After thawing at room temperature, milk samples were gently vortexed to homogenize them and were then saponified for 1 h at 60 °C in ethanolic KOH with pyrogallol. Retinal (O-ethyl) oxime was added as an internal standard prior to extraction with hexane [
23]. Samples were then dried under nitrogen, reconstituted in methanol, and injected onto a 3-μm C18 column using a mobile phase of 68% acetonitrile, 20% isopropanol, and 12% methanol by volume. Retinol was detected at 325 nm with a photo diode-array detector. Infant formula from the National Institute of Standards and Technology (NIST) was analyzed along with each batch of breast milk samples, according to the same procedures (NIST Standard Reference Material 1849, three to four NIST aliquots per batch). Retinol concentrations of the unknowns were calculated by comparing the ratio of retinol to retinal oxime in the NIST formula with that in the unknowns. The measured retinol concentration of the NIST formula was verified using NIST serum (SRM 968e). The within-day and between-day CVs of the NIST controls over the course of analysis were 3.5% and 8.9%, respectively, in 2009, and 3.6% and 13.4%, respectively, in 2012 (SAS proc varcomp).
2.5. Statistical Analysis
Data were analyzed using SAS 9.4 (SAS Institute Inc., Cary, NC, USA), with SAS survey analysis procedures. Weighting factors were applied to account for the respective population sizes of Yaoundé and Douala, and to adjust for the non-response within each cluster. For each survey, variables relating to socio-economic status were combined using factor analysis to create a score for socio-economic status. For the present analyses, the scores for the 2009 national survey were recalculated to represent Yaoundé/Douala only, for consistency with the later survey.
We calculated the frequency of the consumption of fortified oil in the past seven days by multiplying the number of days on which the food was consumed by the number of times per day that the food was consumed (on the most recent day on which the food was consumed) [
12]. Red palm oil and groundnut oil were excluded from the calculations of refined oil intake.
Details of the analysis of 24 h recall data have been reported elsewhere [
15]. Following the calculation of total nutrient intakes, the NCI method was used to estimate the usual intake distributions [
24]. We then simulated the effects of the fortification levels measured in this survey on dietary adequacy, as described previously [
15].
To report the mean RBP concentrations and prevalence of deficiency, RBP concentrations were adjusted for inflammation using regression analysis, employing a method adapted from Larson et al. [
25]. Separate linear regression models for women and children were developed to describe the relationship with CRP and AGP, including interactions or quadratic terms that were significant. These equations were then used to adjust individual values to concentrations equivalent to those in the absence of inflammation, defined as CRP and AGP concentrations of individuals at the 10th percentile of a group with CRP < 5 mg/L and AGP < 1 g/L. The reference CRP and AGP values derived from this dataset were 0.12 and 0.57 for children and 0.16 and 0.47 for women, respectively. However, as described below, differences in biomarker concentrations between the two surveys were examined using the unadjusted concentrations as the outcome variables, and controlling for CRP and AGP concentrations (as continuous variables).
We used previously-derived population-specific cutoffs to define VAD (inflammation-adjusted pRBP < 0.78 µmol/L for women and <0.83 µmol/L for children) and low VA status (pRBP < 1.17 µmol/L for women) [
11].
The change in micronutrient status and other indicators over time was examined using SAS survey regression procedures (proc surveyreg), with a binary variable representing pre- or post-fortification samples. Continuous outcome variables were examined for adherence to a normal distribution by the examination of histograms and Shapiro Wilke’s “W” [
26], and were transformed where necessary to achieve a normal distribution (W ≥ 0.97).
To control for potential confounding in the relationship between refined oil intake and VA status, we used logistic regression to create propensity scores for the frequent consumption of refined oil [
27]. We defined frequent consumption as ≥14 times/week, approximately the 75th percentile of consumption. Because “brand-name” cooking oil was more likely to be fortified than “bulk” oil, we also developed a separate score for the consumption of “brand-name” oil ≥ 7 times/week. Predictor variables included variables related to socio-economic status, including the type and location of residence; housing materials and the type of toilet; sources of lighting, water, and energy for cooking; occupation and employment status of the caregiver and head of the household; and caregiver education. The calculated propensity score for the total refined oil consumption was correlated with the frequency of refined oil intake (women: r
s = 0.31; children: r
s = 0.29,
P < 0.0001 for both), and the probability score for branded oil intake was correlated with branded oil intake (women: r
s = 0.31,
P < 0.0001; children: r
s = 0.35,
P < 0.0001), but not with RBP among women or children (
P > 0.14). The probability score for the total, but not branded-only, refined oil intake was marginally correlated with inflammation-adjusted RBP among children (r
s = 0.08,
P = 0.051), but not women (
P = 0.86).
For the adjusted analyses of difference in vitamin A biomarkers between surveys, variables were considered as potential covariates if they were correlated with either the outcome or with the refined oil consumption. Selected covariates were also included for theoretical reasons. Potential covariates included: age; residence in Yaoundé or Douala; household socioeconomic status (continuous score); CRP and AGP (both continuous); current or recent malaria; type of toilet used by the household (proxy for household sanitation and pathogen exposure); propensity to consume refined oil ≥14 times/week or branded refined oil ≥7 times/week; reported receipt of a vitamin A supplement in the previous six months, breastfeeding status, height-for-age Z-score, and weight-for-age Z-score (children); pregnancy or lactation, BMI (women); and receipt of postpartum vitamin A supplement, age of the breastfeeding infant, and milk fat content (breast milk vitamin A).
Square terms were evaluated where the relationship between the covariate and outcome variable did not appear linear, and selected interactions with CRP, AGP, and the child’s age were included. All possible covariates and selected interaction terms were then added to the regression model, and covariates were sequentially removed (beginning with interactions) if they were not significantly associated with the outcome (P > 0.05; P > 0.1 for interactions). If the removal of a covariate changed the regression coefficient of the “survey year” variable by more than 20% or caused the P value of the outcome to cross the threshold for statistical significance, the covariate was retained in the model. Age was retained in all models. Regression diagnostics, including measures of collinearity, normality of residuals, and leverage, were examined for “full models” (all covariates) and final models.
Finally, we conducted several plausibility analyses to assess whether any observed change over time was related to the consumption of fortified foods. First, we examined the relationships between micronutrient status indicators and the frequency of refined oil consumption post-fortification using Spearman correlations (using the SAS correlations procedure; Spearman, rather than Pearson, correlations were chosen because the frequency of oil intake was not normally distributed). Second, because we observed previously that the frequency of consumption of some foods was related to micronutrient status (likely because of the associations of each with underlying factors like socioeconomic status [
12]), to better assess whether the relationship between fortified food intake and micronutrient status differed pre- and post-fortification, we modeled this as the interaction between the survey year and frequency of consumption of fortified foods in regression models predicting concentrations of each VA biomarker. However, this was an exploratory analysis, because we did not base the sample size on that required to detect interactions.