Study design, setting and participants
This comparative cross-sectional study was carried out in Biriwa, a fishing community of 7086 people in the Mfantseman Municipality of the Central Region of Ghana [20]. Given its relatively large size, Biriwa was purposively selected from six fishing villages that were part of a larger pilot study conducted in the Central Region (the Invisible Fishers study) from May 2018 to August 2019 examining the impacts of various fisheries value chain and other behavior change interventions on anaemia among women of reproductive age (WRA) [21].
The inclusion criteria for participation in the study were, being an adult non-pregnant non-lactating WRA (18 to 49 years), living in Biriwa and being willing to participate in the study. A screening census was completed to list all adult women in the community according to their primary livelihood for the past two years; whether fish smoking involving burning of firewood (FSL women) or other livelihoods not involving burning of firewood and living in a household where no one smokes fish as a major economic activity (OL women). Due to concern about not achieving the estimated sample size, no restrictions were placed on the number of women who could be listed per household during the census. A total of 355 eligible FSL and OL women in 311 households were listed. Women participating in the Invisible Fishers study in Biriwa (N = 10) were ineligible to participate in this study. A sample size of 175 participants per group (i.e., FSL and OL) was estimated based on a 5% level of significance, 80% power, expected anaemia prevalence of 50% (using baseline anaemia prevalence for the Central Region from the Invisible Fishers study) and 35% (using prevalence of anaemia in the Upper West Region of Ghana) where fish smoking is expected to be only minimally practiced [1] among the FSL women and OL women, respectively, and 3% contingency to cater for incomplete surveys. As 194 eligible FSL women were listed from the census, we randomly selected 175 to participate in the study using the RAND function in excel. About 30% (n = 52) where from the same households (two or three [two instances] women from the same household). The remaining 19 eligible women were put on a waiting list and replaced women who could not complete the study due to reasons such as refusals, travel and relocation. The number of eligible OL women listed (n = 161) was lower than the estimated sample size so all of them were invited to participate in the study. About 19% (n = 30) where from the same household (two or three [one instance] from the same household).
Data collection and measurement of variables
Data collection took place from December 2018 to February 2019 and was completed with semi-structured questionnaires. Questionnaires were designed using the Kobotoolbox platform and loaded on Android tablets using the Open Data Kit (ODK) for data collection. Three research assistants with at least high school education were recruited and trained by the primary researcher to support the data collection activities. The questionnaire was pre-tested on women with similar characteristics in a neighboring community before it was administered to the actual study respondents. Face-to-face interviews were completed with participants in their preferred local language (Fante or Twi) or English at their homes or workplaces. Data were recorded with Android tablets by direct electronic data entry using the ODK.
The research assistants obtained information on household characteristics, as well as personal social demographic characteristics, reproductive history, health, recent diet and use of firewood from the selected participant in each household (see Additional file 1). Using the Urit12 HemoCue (URIT Medical Electronics Co., LTD, China) system, a lancet was used to prick the forefinger of each participant and a drop of blood was gently squeezed onto the sampling point of the system to obtain a digital reading of the haemoglobin concentration in the sample. Anaemia was defined as having haemoglobin concentration of less than 12 g/dl [22] based on one sample per participant. A one day 24-h recall method was used to record all foods and beverages (except water) consumed by the study participant in the past 24 h [23]. Wooden food models and household measures were used to help participants estimate quantities consumed. The frequency of consumption of different ASFs by the participant in the past week was captured with an abbreviated food frequency questionnaire listing seven categories of commonly consumed ASFs including fish and seafood, milk and milk products, livestock meats, eggs, poultry, organ meats and bush meats (see Additional file 1). This was a semi-quantitative questionnaire which required participants to specify the number of days in the past week they ate a particular ASF without specifying the portion size. From the 24-h recall data, we determined each participant’s dietary diversity score based on the 10 food groups used to compute the FAO’s Minimum Dietary Diversity for Women indicator [24] and computed recent total iron intakes for participants. The iron content of foods consumed in the past 24-h was determined using a food database (RIING food composition database, Nutrition Department, University of Ghana, unpublished). We estimated the bioavailability of the iron from the foods consumed using a previously published method [25]. For each eating event, 40% of the iron content of meat, fish and poultry (MFP) consumed was considered as heme iron and available and the remaining 60% non-heme while 100% of iron from non-animal sources consumed was considered non-heme. Bioavailability of non-heme iron was computed as 5, 10% or 15% of the total iron content of the food source depending on the quantities of MFP and vitamin C consumed in the same eating event or meal [25]. The proportion of the total estimated bioavailable iron intake from all foods contributed by ASFs consumed in the 24-h recall period was computed. Additionally, we calculated ASF diversity as the number of different categories of ASF (out of a total of seven) consumed by the participant in the past seven days. Completed questionnaires were reviewed for completeness at the end of each day. Participants with missing or incomplete responses were contacted the following day to complete the missing information.
Data analysis
Data were managed, cleaned and analyzed using the Statistical Software Package for Social Sciences (SPSS) version 22.0 (Chicago, USA) and SAS for Windows Release 9.4 (Cary, NC, USA). One duplicate record was identified and removed during data cleaning, otherwise all completed questionnaires were included in the analysis. Bivariate analyses using Student’s T-tests for continuous variables and Chi-Square tests for categorical variables were used to summarize differences in background and household characteristics of the FSL and OL women. Additionally, we compared group differences in mean ASF diversity (number of different ASFs consumed in the past seven days), total iron intakes, and mean percent contribution of ASFs to total iron intakes in the past 24-h.
Blood haemoglobin concentrations were compared between the two groups of women using a general linear model and ANCOVA for unadjusted and adjusted comparisons, respectively. The SAS PROC GLIMMIX procedure was used in both cases. Unadjusted and adjusted means with their 95% CIs were calculated. Anaemia prevalence was compared using a simple logistic regression model for an unadjusted comparison, and multiple logistic regression model for an adjusted comparison. The SAS PROC GLIMMIX procedure was used in both cases. A binary distribution and log-link function were specified in the SAS procedures so that relative risks between groups and their 95% CIs were calculated. Covariates for the ANCOVA and logistic regression models were selected by correlating anaemia with each covariate so that only those independent variables significantly associated with the outcome at alpha = 0.2 [26] were selected for the multiple logistic regression model. This conservative level of alpha was chosen to minimize the risk of type II error in variable selection [27]. In addition to covariates selected through correlation analysis (i.e., marital status, fever in the past two weeks, fish smoking livelihood, dewormed in past 3 months, ever been pregnant and age), other covariates including number of days spent smoking fish, fuel for cooking, ASF diversity, and access to a toilet facility were selected for the final ANCOVA and multiple logistic regression models if they were associated with haemoglobin concentration or anaemia, respectively.