Beyond Household Socioeconomic Status: Multilevel Modeling of Supply-Side Determinants of LPG Consumption among 5,500 Households in Sub-Saharan Africa

Household transition to cleaner cooking fuels (e.g. liqueed petroleum gas (LPG)) has historically been understood as an “energy ladder” with clean energy access resulting from improvements in household socioeconomic status (SES). Recent studies have demonstrated the importance of supply-side determinants in increasing clean cooking, yet few large-scale studies have assessed their signicance quantitatively. As part of the CLEAN-Air(Africa) study, a population-based survey was conducted (N = 5,638) assessing cooking practices in peri-urban communities within Cameroon, Kenya and Ghana. Multilevel logistic and log-linear regression were used to assess socioeconomic and supply-side determinants of LPG usage (primary versus secondary fuel) and consumption (kilograms/capita/year), respectively. Supply-side factors (e.g. cylinder rell and transportation costs) and using single versus multi-burner stoves were better predictors of both the probability of primarily cooking with LPG and annual LPG consumption than household SES. These results suggest the need for policies promoting LPG access and stove equipment that meet household needs.


Introduction
Polluting fuels, including biomass (e.g. wood, charcoal), coal and kerosene, are used by approximately 3.8 billion individuals worldwide for cooking, heating and lighting. 1 Household air pollution (HAP) generated from incomplete combustion of these fuels results in ne particulate matter (PM 2.5 ) levels typically well above World Health Organization (WHO) guidelines. 2 Exposure to PM 2.5 in HAP is causally associated with many adverse health outcomes, including cardiopulmonary, and respiratory diseases. [3][4][5][6] Polluting fuel use also negatively impacts climate through deforestation from unsustainable harvesting of wood in certain locations, 7 and emissions of short-term climate-forcing pollutants; 25% of global anthropogenic black carbon emissions are produced through household biomass combustion. 7,8 Women, typically the primary cook, may travel long distances to gather polluting fuels in some settings, which negatively impacts their livelihoods. 9,10 In Sub-Saharan Africa (SSA), approximately 900 million people cook with polluting fuels. 11 Governments in SSA, including Cameroon, Ghana and Kenya, plan to expand population-level use of lique ed petroleum gas (LPG) as an interim solution for clean cooking to an aspirational target of between 35% and 58% over the next decade. [12][13][14] LPG, although a fossil fuel, does not emit black carbon, and has much lower PM 2.5 emissions than polluting fuels. 12,15,16 Using LPG for cooking can also decrease localized deforestation, and reduce time spent gathering and cooking with polluting fuels. 16,17 Historically, studies focused on determinants of clean cooking have emphasized the 'household energy ladder' model, by which improvements in socioeconomic status (SES) lead households to progressively transition to cleaner energy sources. [18][19][20] However, other aspects, including consistent fuel access, safety and convenience are also important facilitators of clean cooking fuel uptake. [21][22][23] Studies in India found that rural households provided with subsidies for LPG connections under the Pradhan Mantri Ujjwala Yojana (PMUY) programme continued to use polluting fuels alongside LPG ('stove stacking'), leading to less frequent LPG use compared with urban households. 24,25 Reduced LPG uptake likely resulted from poorer supply and access points in rural Indian villages. While studies conducted in Cameroon, 26,27 Tanzania 28 and Ethiopia 29 have also found supply-related issues to be important determinants of cooking fuel decisions, few large-scale studies in SSA have quantitatively assessed their effect on LPG consumption.
As partial use of clean cooking fuels (stacked with polluting fuels) does not provide su cient HAP exposure reductions to bene t health, 30 a comprehensive understanding of facilitators for sustained and exclusive LPG use is critical to maximizing health gains. This multinational modeling study of over 5,500 households in peri-urban communities of Mbalmayo, Cameroon; Eldoret, Kenya and Obuasi, Ghana, presents one of the largest assessments of supply-related impacts on LPG fuel usage in SSA. With urban populations in SSA predicted to double over the next 25 years, this analysis provides a timely examination across three rapidly developing communities.
A higher percentage of households primarily cooking with LPG contained a member with a university degree (22%) and were in the highest income quartile (23%), compared with households primarily using polluting cooking fuels (5% with university degree; 8% in highest income quartile) ( Table 1). In Eldoret and Mbalmayo, the proportion of households cooking primarily with polluting fuels that reported seasonal changes in income (72% and 75%, respectively) was 20-30% higher than those primarily cooking with LPG (42% and 58%, respectively). Among households primarily cooking with LPG, 59% had fewer than 5 family members, compared with 38% of those primarily cooking with polluting fuels (Table 1). *The sample of households in Eldoret, Kenya is smaller than the other two communities as random sampling was switched to purposeful sampling midway through data collection to ensure a higher sample of LPG households available for subsequent phases of the CLEAN-Air(Africa) study. Note: variables for some countries do not add up to total sample size due to missing data or elimination of variable categories for brevity.

Households Cooking with LPG
Over half (52%, n=1,458) of households cooking with LPG used it as a primary fuel; very few (4%, n=109) exclusively cooked with LPG and 44% (n=1,263) used LPG as a secondary fuel ( Table 2). In Obuasi, two-thirds of households reported using LPG as a primary fuel (67%; n=679), compared with one-third (37%; n=316) of households in Eldoret; in Mbalmayo LPG was used roughly equally as a primary and secondary fuel (48%; n=463). LPG was most frequently stacked with wood in Mbalmayo, and charcoal in Eldoret and Obuasi ( Figure 2). Nearly half (47%) of households primarily cooking with LPG said it was not always available for purchase ( In Eldoret, 72% of participants cooking exclusively with LPG were 10 minutes or less from a retailer compared with 47% and 36% of households using LPG as a primary or secondary fuel, respectively. Having an LPG retail point within walking distance was six times more common among households using LPG exclusively (61%) than those using LPG as a secondary fuel (11%).  The relationship between the proportion of households reporting LPG as a primary fuel and family size was modi ed by the number of LPG stove burners. Only 10% of primary LPG households with a large family size (≥7 members) used a single-burner stove, compared with 50% of the smallest households (1-2 members) (Figure 3). A monotonically increasing association existed between the proportion of households primarily cooking with LPG and proportion using multi-burner stoves among households with more than four members.  (Table S2). Demographics (R 2 marginal =0.11) and LPG supply-related factors (R 2 marginal =0.10) explained a higher proportion of model variability than SES (R 2 marginal =0.03) ( Table S2). Households with 1-2 members had more than twice the predicted probability (89% (95%CI: 79,95)) of primarily using LPG than households with 7-8 family members (39% (95%CI: 24,56)) ( Figure 4). Lower availability of LPG and higher re ll costs were associated with a lower predicted probability of primary use of the fuel in a monotonically decreasing manner ( Figure 4). Speci cally, 76% (95%CI: 60,87) of households reporting a re ll cost of <$0.86 USD/kg were predicted to use LPG as a primary fuel, compared with 65% (95%CI: 51,78), 58% (95%CI: 42,72) and 45% (95%CI: 28,62), of households reporting a cylinder re ll cost of $0.86-1.00USD/kg, $1.01-1.10USD/kg and >$1.10USD/kg, respectively.  Table S4 in Supplement).
Inability to afford the upfront costs of purchasing LPG stoves/equipment was the dominant reason (70%; n=1,889) reported for not currently cooking with LPG ( Figure 6). High re ll costs were cited as a barrier for LPG use by twice as many households previously cooking with LPG (37%) as those that had not (19%) ( Table S6). Concerns over LPG safety were reported by 18% (n=470) of households not currently using LPG; this concern was highest among households in Obuasi (30%; n=292); the proportion was twice as high as in Mbalmayo (14%; n=117) and ve times higher than in Eldoret (6%; n=61).

Discussion
By quantitatively modeling the impact of SES, fuel cost and accessibility on LPG usage, this multinational study demonstrates that supply-related issues, including inconsistent fuel supply and higher cylinder re ll and transportation costs are critical barriers to uptake and sustained use of LPG in peri-urban communities in SSA. Households indicating a consistent supply of LPG at retailers had over 20% higher predicted probability of using LPG as a primary fuel than those reporting inconsistent supply, irrespective of household SES (Table S4). Households reporting the lowest re ll costs also had a 30% higher probability of using LPG as a primary fuel ( Figure 4) and consumed 6.67 kg/capita/yr (95%CI: 5.16,8.60) more than households reporting the highest re ll costs ( Figure 5). High re ll costs were commonly reported by households cooking exclusively with polluting fuels, particularly in Obuasi (50%) and Mbalmayo (40%), possibly indicative that customers using typically smaller cylinders in Kenya (6 kg) compared with those used in Cameroon and Ghana (>12 kg) may be less sensitive to changes in the re ll price.
A negative dose-response relationship between higher LPG consumption and lower transportation cost and shorter time to obtain an LPG re ll ( Figure 5) matches with previous cooking fuel choice research conducted in rural communities in Ghana. [31][32][33] Thus, policies that improve proliferation of last-mile distributors (i.e. re lling stations) in peri-urban communities, which would decrease the collective population's travel time and associated costs, will likely pay dividends in terms of increasing populationlevel LPG usage. Other LPG supply chain enhancements, including increasing bulk storage and lling facilities, are important for expanding access 23 and have been identi ed as a priority by Cameroonian government. 13 Consumer nance mechanisms, including unconditional cash transfers, 34 micro nance and pay-as-yougo LPG, 35 which involve direct cylinder home deliveries, can help offset the nancial burden associated with transport and cost of re lls. PAYG LPG has shown promise in urban settings, 36 but will be more logistically challenging to implement in peri-urban areas due to higher transportation costs and enhanced distribution networks needed to ensure timely home deliveries. 13 The Pradhan Mantri Ujjwala Yojana (PMUY) programme achieved rapid expansion of LPG access but did not lead to higher usage among rural Indian households. 24,37 A study among 8,000 PMUY program bene ciaries proposed that long travel times from rural Indian villages to re ll points was a likely driver of 30% lower LPG consumption. 25 While this study proposed that fuel access is important at a village-level, we nd that accessibility may play a role at smaller scales in an African context; a 10-minute longer travel time to a retailer within a community may be a deterrent to LPG usage ( Figure 5). Indeed, a low intraclass correlation coe cient (0.01) in the consumption model emphasizes that LPG consumption was predominantly in uenced at a household-level (Table S3). These results contribute to growing evidence that accessibility/affordability of re lls are critical bottlenecks that should be priority targets for policymakers to expand LPG use.

Socioeconomic Status and Demographics
Younger households and smaller families were more likely to primarily use LPG and had higher consumption rates ( Figure 5), similar to ndings from a study of PMUY bene ciaries in India. 25 Moreover, using multi-burner LPG stoves greatly increased the probability of households with 5 members or greater primarily using LPG (Figure 3), likely because multi-burner stoves offer greater time and fuel savings to larger families. Using a double-burner stove was associated with substantially higher LPG consumption (21.92 kg/capita/yr (95%CI: 17.84,26.93)) compared with use of a single-burner stove (14.09 kg/capita/yr (95%CI: 11.53,17.21)). The ability to cook two meals simultaneously on double-burner stoves was an advantage of LPG over kerosene reported by households in Nairobi, with users associating the doubleburner stove with greater time and fuel savings. 36 Participants in another Kenyan study stated having "no need to stack" when using double-burner stoves. 38 Thus, increasing access to multi-burner LPG stoves seems a simple, yet potentially highly effective intervention to increase clean cooking.
In contrast to India, 24 years cooking with LPG was not signi cantly associated with consumption (Table  S5), potentially highlighting the importance of LPG fuel costs remaining price competitive in the long term to prevent reversion to polluting cooking fuels. This nding is further supported by households previously cooking with LPG being more likely than households with no prior LPG experience to cite high cylinder re ll costs as a reason for not cooking with LPG (Table S6). Additionally, the strongly negative association of family size with consumption and use as a primary fuel suggests that LPG usage is also dependent on its ability to prepare sizeable meals in a reasonable amount of time, 39 which may be more di cult for larger families.
Household income was not signi cantly associated with use of LPG as a primary fuel (Table S4), demonstrating that usage does not necessarily intensify in a linear manner with increasing SES, but may follow a complicated trajectory due to various supply-related and contextual factors. 40,41 LPG consumption statistics Self-reported frequency of stove usage during the past week was an insigni cant predictor of annual LPG consumption, possibly due to seasonal income uctuations reported by 65% of households (Table 2), which can lead to large within-household variation in consumption patterns. 37 However, fuel stacking was associated with 17% lower LPG consumption (2.3 kg/capita/year reduction).
Estimated median annual per capita consumption in Mbalmayo, Eldoret and Obuasi were relatively similar (40% higher, 13% higher, 8% lower, respectively) to national rates reported in the Global LPG Partnership National Feasibility Studies in 2017 (14.2 kg/capita/yr in Cameroon, 12.8 kg/capita/yr in Kenya and 25.0 kg/capita/yr in Ghana). [42][43][44] It is unclear if the differences result from population-level changes in consumption from 2017-2019 or bias in self-reporting of consumption (Table S7), which has occurred in previous studies. 45 Nonetheless, the self-reported consumption rates are about half that of households in more developed countries (e.g. Brazil, Indonesia, Peru) with more established LPG supply chains. 14 Safety Concerns LPG safety concerns were prevalent among households cooking with polluting fuels, particularly in Obuasi (30%) (Figure 6). A higher proportion in Obuasi compared with the other communities is likely attributed to the 'customer-controlled cylinder model' currently implemented in Ghana, which places customers in charge of cylinder maintenance, contributing to more frequent, serious LPG accidents. 42 The Ghana 'Sustainable Energy 4 All' plan called for a national switch to the 'branded cylinder recirculation model', which ensures proper re lling practices and correct disposal of cylinders, as LPG marketers are responsible for cylinder maintenance and safety. 46 Households Using Polluting Fuels There was high aspiration to use clean cooking fuels among those cooking exclusively with polluting fuels ( Figure 6). While high fuel costs were the dominant reason for not using LPG, overcoming the affordability barrier will not lead to universal use in these communities due to safety concerns and an inconsistent supply. Nearly half (47%) of participants currently not cooking with LPG have cooked with LPG previously. While unable to ascertain if households that formerly cooked with LPG used it as their primary fuel and whether they have completely stopped using it or routinely switch between fuels (e.g. due to periodic income uctuations), it is evident that LPG re ll costs play a key role in discontinuing use of LPG ( Figure 6). 47 Reversion to polluting cooking fuels has been reported in longitudinal studies, 48 with a prevalence as high as 35% (China). 49

Strengths and Limitations
This study is one of the largest modeling studies of determinants of LPG consumption in SSA. While calculation of self-reported annual LPG consumption via two different survey questions showed disagreement ( Figure S1), sensitivity analyses revealed that this discrepancy did not signi cantly impact study ndings (Table S8). As the direction of misclassi cation between the self-reported LPG consumption variables was the same among all households (consumption using self-reported number of annual re lls was always higher than that calculated via average cylinder lifetime), we expect that misclassi cation was non-differential and therefore biased toward a null nding. We recommend that future studies collecting self-reported data on LPG consumption phrase survey questions in terms of 'average cylinder lifetime' rather (or in addition to) number of annual cylinder re lls to protect against overreporting. As other studies have found low agreeableness between self-reported and objective measures, 45 we further recommend that absolute measures of LPG consumption reported in this study be interpreted with caution.
This study examined household energy decisions in a unique, peri-urban context. As the extent of fuel stacking and availability of free biomass typically varies between rural and peri-urban households, 27,33 and research has shown differences in LPG consumption in urban, rural and peri-urban settings, 42-44 the study results may not hold outside of peri-urban communities.

Conclusions
This study presents robust empirical evidence of the critical role of supply-side determinants and access to multi-burner stoves in increasing LPG consumption among peri-urban households in SSA. Supplyrelated factors may play a considerable role in LPG consumption even at small scales (e.g. 10-minute travel intervals) within a community. Energy policies targeting the LPG supply chain, including expanding distribution/retailing networks can help accelerate growth of the clean cooking market in Sub-Saharan Africa, leading to positive environmental and health impacts.

Methods
The Global Health Research Group on Clean Energy Access for the prevention of Non-communicable disease in Africa through clean Air (CLEAN-Air(Africa)) program consists of applied research and capacity building within SSA. 50 This study reports on ndings from the rst phase of CLEAN-Air(Africa), involving randomly administered surveys in three peri-urban communities: Mbalmayo, Cameroon; Obuasi, Ghana and Eldoret, Kenya. Mbalmayo is an agricultural town in central Cameroon with 60,000 residents that is an hour drive away from Yaoundé, the country's capital. Obuasi is a gold-mining community in southern Ashanti, Ghana with a population of almost 200,000 that is an hour drive away from Kumasi (capital city of the Ashanti region). Eldoret is a town surrounded by agricultural land, located at an elevation of over 2000 meters in Western Kenya and has a population of nearly 500,000; it is currently the fastest growing town in Kenya according to the 2019 National Census.
In each community, 2,000 households were surveyed to ensure a su cient sampling frame for comparative research between households cooking primarily with LPG and exclusively with polluting fuels in later phases of CLEAN-Air(Africa). The surveys included detailed questions on cooking behaviours, including WHO Harmonized survey questions on fuel use (https://www.who.int/airpollution/household/survey-harmonization/en/ ). Completion took approximately 20 minutes. Due to a switch from random to strati ed sampling (by primary cooking fuel type) midway through data collection in Eldoret (in order to ensure a su cient sample of households using LPG for future phases of CLEAN-Air(Africa)), population-based results are reported among a subset of 757 households for this location.
Surveys were administered to the main cook of the household using secure smartphone technology (Mobenzi Researcher in Cameroon/Kenya; Research Electronic Data Capture (REDCap) in Ghana) from April-August 2019. Mobenzi is a secure data collection software system (data encrypted at source) whereby data from prede ned surveys are collected by smartphone application and automatically uploaded via the phone's SIM card and synced to the Mobenzi cloud (or when the phone is connected to a wireless network if there is no mobile signal). 51 REDCap is an encrypted web application for creating and managing online surveys and databases; data are wirelessly imported directly to the database servers.
Two multilevel (households nested within communities) models were built: (1) use of LPG as a primary or secondary cooking fuel (logistic regression) and (2) self-reported quantity (kilograms) of LPG consumed per capita (log-linear regression). Results from logistic regression modeling are depicted as the averageadjusted predicted probability of using LPG as a primary cooking fuel. Findings from log-linear regression, are portrayed as the average-adjusted annual LPG per capita consumption (kilograms/capita/year). Model 1: LPG cooking fuel (primary/secondary) ik = β 0 + β 1 + b i + e ik (logistic regression) Model 2: ln(kg LPG/capita/year) ik = β 0 + β 1 + b i + e ik (log-linear regression) In Model 1, LPG cooking fuel (primary/secondary) ik represents whether the kth participant in community i uses LPG as a primary or secondary cooking fuel. In Model 2, ln(kg LPG/capita/year) ik represents the natural log transformed annual per capita consumption of kth participant in community i. In both models, β 0 is the overall intercept, β 1 represents xed effects, bi is the community-level random effect and εik is the residual error.
Annual per capita consumption of LPG was estimated in two ways: (1) multiplying self-reported cylinder size by number of annual re lls and dividing by the number of household members and (2) dividing 12 months by the self-reported average duration (months) of a cylinder re ll to obtain a second estimate of self-reported number of annual re lls and multiplying that by cylinder size then dividing by the number of household members. Sensitivity analyses examined the effects of using both metrics on the modeling results. The distribution of annual per capita LPG consumption (kg) was right-skewed so data was logtransformed before modeling. Principal components analysis (PCA) was run on owned assets to generate an alternative measure of household SES. 52 (add Vyas citation). The rst principal component was grouped into quartiles and tested as a predictor in both regression models.
Model t was assessed via a combination of the coe cient of determination (R 2 ), Akaike information criterion (AIC) and 10-fold cross validation (training and test datasets split at the community level to ensure more accurate evaluation of model performance), with consideration given to selection of parsimonious models. The cross-validated R 2 is reported for the linear model 53 and the area under the Receiver Operating Characteristic curve (AUC) is reported for the logistic model. 54 All data analysis was conducted in R (version 3.5.1). 55 Ethical approval was obtained from the University of Liverpool, United Kingdom and local Institutional Review Boards in each study country. Informed consent was obtained from all participants prior to conducting the study.