Design
The 2014 Ghana Demographic and Health Surveys (GDHS) data was collected in line with a cross-sectional survey design [11].The survey employed two-stage probability sampling, stratified by urban and rural areas of the ten administrative regions of Ghana [11]. The sampling frame for the survey was based on the updated sampling frame from the 2010 Ghana Population and Housing Census [11]. The first stage of the probability sampling involved the selection of enumeration areas (EAs), which was stratified by place of residence [11]. A total of 427 EAs were selected, 216 in urban areas and 211 in rural areas [11]. The second stage of the probability sampling involved the systematic sampling of households [11]. The implementers of the survey undertook a household-listing operation in all the selected EAs from January through March 2014, and households to be included in the survey were systematically selected from the list [11]. Approximately 30 households were selected from each EA to constitute the total sample size of 12,831 households [11].
Data collection
Trained enumerators collected the data from early September to mid-December 2014 using paper-based questionnaires [11]. The selected sample size for the 2014 GDHS was 12,831 households, of which 12,010 were occupied [11]. Out of the occupied households, 11,835 were successfully reached, resulting in a 99% response rate [11]. In the GDHS, household heads provided information on their demographic characteristics and household characteristics such as household population and composition, housing structure, household assets, access to basic utilities, sources of drinking water, water treatment practices, access to sanitation facilities, and type of fuel used for cooking [11].
Study sample
The unit of analysis is households, and the dataset contains 11,835 households. The data were weighted using the household weight variable in the dataset. A sample of 513 households indicated that they cooked no food in the house, so they were excluded from further analyses. Therefore, the analytic weighted sample is 11,322.
Variables
Outcome variable:
In the dataset, the household head was asked to select the main source of fuel for cooking in the household from the following categories: electricity, LPG, natural gas, kerosene, coal, lignite, charcoal, wood, straw/shrubs/grass, and agricultural crop. Electricity, LPG, and natural gas was defined in our study as HP & CFFs and all other sources of fuels as non-climate friendly and health promoting. Additional file shows the proportion of households using each of the fuel types for cooking in Ghana (See Additional file 1).
Covariates:
Ten sociodemographic characteristics variables were selected as potential covariates in the study. Marital status was recoded as follows: never in union (as never married), married and living with partner (as currently married), widowed, divorced, and no longer living together/separated (as formally married). For the multivariable model, household wealth was dichotomized as follows: poorest, poorer, and middle (as poor) and richer and richest (as rich).
Household wealth index was already estimated and reported in the DHS data. This was created using household characteristics (source of drinking water, type of toilet, sharing of toilet facilities, main material for roof, walls and floors floor, and type of cooking fuel amongst others household characteristics) and household possessions and assets (ownership of television, radio, vehicle, bicycles, motorcycles, watch, agricultural land, farm animals/livestock, and bank account amongst others). DHS used a principal component analysis (PCA) to assign weights to each asset in each household and cumulative score were calculated from the assigned weights. Households were ranked according to the cumulative scores from the household assets. The cumulative percentage distribution of the wealth score was estimated and the wealth score values that corresponded to the four cut point values of the quintiles (20th, 40th, 60th, and 80th percentiles) were determined. Households with values less than or equal to the 20th percentile score were assigned poorest, those greater 20% but less than or equal to 40th % were assigned poorer, those greater than the 40th % and less than or equal to the 60th % score were assigned middle, greater than the 60th % and less than or equal to the 80th % score were assigned richer and the richest household were those with scores greater than the 80th percentile score. Wealth index was thus ranked into quintiles: poorest, poorer, middle, richer, and richest [12].
Analytic procedure
We used STATA version 14 for the data analyses. Summary statistics, chi-square test of independence, and Poisson regression were performed. The analysis accounted for sampling design (cluster, strata) and household weights using the ‘svy’ command in a default mode. The default ‘svy’ computes standard errors by using the linearized variance estimator called the first-order Taylor linearization. This process eliminated the incorrect estimation of the Standard Errors (SE) associated with the confidence intervals of the regression coefficients. After survey setting our data, we obtained prevalence ratios by performing Poisson regression. We achieved this by using the generalized linear model (glm) in STATA, setting the family to “Poisson” and the link to “log,” to avoid the potential overestimation of effects from reporting odds ratios when the prevalence of the outcome of interest is above 10%. The mathematical equation for Poisson regression is as follows: (see Formulas in the Supplementary Files)
The outcome variable has only two events (that is ‘0, 1’), where the value ‘1’ means a household was using clean cooking fuel and the value ‘0’ means that a household is not using clean cooking fuel.
Before performing the multivariable analysis, we check the assumption of multicollinearity and no violations were observed. The variance inflation factors of the covariates are all less than 2 (Table 1), which are far less than the cut of points of 10 [13].
Table 1 Collinearity statistics of the covariates
Variables
|
Tolerance
|
Variance Inflation Factor (VIF)
|
X1: Sex of HH
|
0.77
|
1.30
|
X2: Age of HH
|
0.89
|
1.12
|
X3: Marital status of HH
|
0.78
|
1.29
|
X4: The education level of HH
|
0.71
|
1.40
|
X5: Size of HH
|
0.86
|
1.16
|
X6: Number of women (15-49) in HH
|
0.85
|
1.78
|
X7: Number of men (15-49) in HH
|
0.89
|
1.12
|
X8: Household wealth
|
0.54
|
1.85
|
X9: Urban/rural residence
|
0.63
|
1.60
|
X10: Region
|
0.99
|
1.01
|