What happens after electrification? Exploring the evolution of appliance adoption in rural Kenya

Original research article What happens after electrification? Exploring the evolution of appliance adoption in rural Kenya Leticia Tomas Fillol a,* , Antti Pinomaa a , Nicol` o Stevanato b , Riccardo Mereu b , Samuli Honkapuro a a School of Energy Systems, LUT University, Lappeenranta, Finland b Department of Energy, Politecnico di Milano, Milano, Italy ARTICLE INFO Keywords: Appliance adoption Electricity access Income-generating activities mini-grids Demand growth ABSTRACT Design and dimensioning of off-grid power systems require a deep understanding of how electricity demand develops over time. However, a notable knowledge gap exists regarding how electricity loads evolve, particularly in remote areas of developing countries. This paper presents a granular analysis of appliance uptake through a case study in rural Kenya, and aims to provide real-world evidence to enhance demand forecasting in similar contexts. The study focuses on a mini-grid project and evaluates the evolution of appliance ownership during the early years following electrification. It examines differences between households with home businesses and standard households, i.e., those without income-generation activities at home, complemented by an exploration of the role of microfinance. The analysis is based on household-level survey data collected from the field. The findings highlight that establishing home businesses, access to microfinancing, and the prior presence of power systems are key drivers of faster appliance adoption. However, appliance adoption is gradual but slow, with many households owning only basic appliances even several years after gaining access to electricity. Households engaged in income-generating activities and utilizing microfinance are more likely to acquire welfare appliances at an earlier stage. 1. Introduction Decentralized electrification solutions are essential for achieving universal electricity access, particularly in rural under-served communities where grid extension is economically unfeasible. Off-grid minigrids (MGs) could provide grid-quality electricity to over 60 % of the African population currently lacking access to electricity [1]. Despite the potential of renewable-energy-powered MGs, gaps persist in optimizing system sizing to meet evolving energy demands. Appliance stock is a key driver of electricity consumption [2,3], but nevertheless, studies often use static categorizations of ownership, overlooking its evolution, which is a factor especially relevant to newly electrified households [4,5]. Limited data on changing electricity demand coupled with uncertainties and socio-economic complexities create challenges to accurate demand forecasting [6,7], affecting the performance and viability of off-grid systems [8]. This study explores the relationship between electrification duration and appliance ownership in a rural Kenyan village, focusing on the influence of the establishment of home businesses and access to microfinance. The findings can contribute to improving energy demand projections for off-grid communities. 1.1. Background The scientific literature highlights demand growth trends following electrification. A Tanzanian mini-grid study reported significant electricity purchase increases over 30 months [9]. While the study effectively quantifies demand growth after electrification, there is only limited examination of the factors driving the shift in demand. The importance of socio-economic factors, such as informal economies and income-generating activities, is emphasized by [10] and illustrated in [11], which provides specific evidence linking electricity access to economic activities that result in demand growth. Studies such as [12] model residential appliance uptake based on macroeconomic factors, and [3] examines appliance ownership in urban Ghana. However, while valuable at broader scales, these models do not address the socio-economic nuances of rural or off-grid contexts. * Corresponding author. E-mail address: [email protected] (L.T. Fillol). Contents lists available at ScienceDirect Energy Research & Social Science journal homepage: www.elsevier.com/locate/erss https://doi.org/10.1016/j.erss.2025.104015 Received 25 April 2024; Received in revised form 7 January 2025; Accepted 1 March 2025 Energy Research & Social Science 122 (2025) 104015 Available online 12 March 2025 2214-6296/© 2025 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license ( http://creativecommons.org/licenses/by/4.0/ ). Research in the Global South, like [13,14], examine appliance diffusion but often rely on generalized patterns, similar to [15], which treats SubSaharan Africa (SSA) as a single entity, obscuring localized trends critical for off-grid projects. A Rwandan study [16] highlights sociotechnical drivers, noting higher information and communication technology (ICT) device uptake in low-income households and gender disparities, though the absence of a temporal perspective limits long-term projections. In contrast, research on rural India [17] shows gradual appliance adoption over time, with slow uptake of power-intensive appliances, echoing findings from [16,18]. In terms of energy modeling, most studies rely either on static models or arbitrary assumptions. Nonetheless, there is growing interest in modeling demand evolution. Authors from [19] highlight load evolution as key for MG sizing but relies on trends that might not be wholly representative in many contexts. Similarly, [6] examines growing electricity demand in rural India using hypothetical appliance growth scenarios. However, the reliance on assumptions based trends limits the reliability of such studies as regards representation of dynamic realworld contexts, as noted by [10]. It can, moreover, be seen that most studies on appliance ownership found in the literature overlook the role of time. Additionally, socioeconomic factors like income-generating activities, microfinance, and presence of solar-home systems (SHS) are rarely considered, despite their potential significance. 1.2. Contribution and organization of the paper To address the above noted gaps, this study investigates the initial evolution of appliance adoption during the first six years following the commissioning of a rural MG in Kenya, as well as the key factors driving this process. We employ a simplified framework to classify varying levels of appliance ownership, which serve as proxies for different levels of electricity demand. The study also investigates the influence of homebased businesses on appliance ownership and the corresponding implications for electricity demand. Additionally, the work explores the strategies households employ to navigate the economic tradeoffs between electricity prices and appliance ownership and usage, providing a more comprehensive understanding of these dynamics. To achieve these goals, we combine econometric models with additional description analysis of survey results, which provides additional context and insights not captured by the econometric approach. This research is novel in two key respects. First, it provides a dynamic perspective on how appliance adoption evolves, and second, it introduces explanatory variables that are rarely considered in existing studies. The study contributes to scientific progress in the area of energy demand dynamics in SSA context, offering mini-grid developers insights into appliance adoption, which is information that can be employed for improving electricity demand projections during the first years of electrification. It also provides policymakers with evidence to develop strategies that enhance electricity benefits in newly electrified communities. This paper extends the work in [4] and presents updated results and more comprehensive analysis. The study addresses the following research questions: 1. What are the key factors contributing to appliance adoption in rural households? 2. How does appliance adoption over time differ between standard households and those with income-generating activities, and what are the implications for electricity demand? 3. How do households adapt to economic constraints while adopting and using appliances? The paper is structured as follows: The methodology employed in this study is explained in Section 2. Section 3 presents empirical results of various analyses, accompanied by key insights identified during the research. These findings are discussed further in Section 4. Finally, Section 5 provides the conclusions drawn along with practical recommendations for practitioners and developers. 2. Methodology Explanation of the research design applied in this work begins with a description of the case study, which is followed by a description of the survey method and the data collected, and then an account of the steps taken to ensure data reliability. Finally, the framework applied to categorize different levels of appliance ownership and the econometric methods employed for the data analysis are presented. The methodological approach adopted in this study is illustrated in Fig. 1. 2.1. Case study This study focuses on Faza, a community on Pate Island, Kenya, electrified by the diesel-powered MG. Built by the Rural Electrification and Renewable Energy Corporation (REREC) in 2016 and operational since 2017, the MG is managed by Kenya Power and Lighting Company (KPLC). It has a nominal output of 0.9 MVA from two generators (650 kVA and 500 kVA) and serves approximately 3500 connection points. 2.2. Survey method and data sample A bottom-up approach was employed to collect data at the individual household level through a structured questionnaire, detailed in supplementary material A. Data was collected in February 2023 through face-to-face and door-to-door interviews, with the responses filled into paper questionnaires before compilation into a laptop computer. The survey encompassed 192 respondents from the residential sector, with only one respondent per household answering the survey. The pre-planning for data collection was carried out with support from the local mini-grid operator through phone calls and virtual meetings to gain an understanding of the island's landscape. The data collection process took place over one week. Four enumerators, who were project team members, were trained on the questionnaire structure. Comprehensive details about the data collection process are provided in a previous publication [20], which specifically addresses this case. The publication includes detailed descriptions of the preparation, pre-analysis, data collection process, challenges faced, and the stakeholders involved. Household surveys have been used extensively to obtain individual dwelling information on electricity consumption or customers' behavior [7] and are still considered one of the most cost-effective approaches to understanding energy use in rural settings [21]. Other methods to examine energy consumption patterns include approaches based on DSO data or meter-based data [21]. Although the latter methods are comparatively more accurate, they also present disadvantages. For instance, dis-aggregated data is not possible when using DSO records, and meter-based data is more expensive, and impractical in some cases [21]. In this regard, given the absence of meters installed at the consumer end, face-to-face surveys were considered a more appropriate way to get input data at the household level as they enabled easier, more convenient, and cost and time-effective data collection. Nonetheless, the introduction of errors is a major problem that researchers encounter when utilizing household survey data [7,21], often due to interpretation and estimation errors made by the layman, the MG end-customer, thereby adding complexity to the analysis. Acknowledging these limitations and aiming to reduce bias derived from the data collection method, we employed a device-based bottomup approach presented in [21]. Such approach consists of gathering, through a structured questionnaire, detailed information on each appliance owned, including average daily usage, time slots when the device is normally used, and physical parameters. Other data collected include information on demographic and dwelling characteristics, monthly electricity expenditure, year of connection to the MG, use of L.T. Fillol et al. Energy Research & Social Science 122 (2025) 104015 2 microfinance, and the establishment of income-generating activities, as described in Table 1. The sample size for this study was calculated using the Bukhari method, as shown in Eq. 1. This method determines sample sizes for known populations by adjusting parameters, such as confidence level, margin of error, and sample proportion. Another frequently used approach is the Krejcie and Morgan method, which relies on the chisquare value instead of the z-score used in the Bukhari method. The Bukhari formula was chosen for this study due to its ability to optimize sampling effort while aligning with project budget and timeline constraints. Sample size = p(1−p)z2 e2 1+p(1−p)z Ne2 (1) where p is standard deviation, z is z-score, e is confidence-interval, and N represents population size. Using N as 3500, for the connection points to the MG, a conservative estimate of the energy consumption of the island, and p as 0.5, the sample size was calculated with a 90 % confidence level, and the resulting z-score was 1.65 with a margin error of 5 %. Notably, an important percentage of the interviewed sample utilizes electric appliances for income generation, we refer to these households as home businesses or households with income-generating activities in this study. Households with home businesses account for approximately 27 % of the surveyed households. Many of the commercial activities are related to juice, ice, or, in general, food commercialization, while other activities include carpentry, sewing, and digital enterprises like cyber cafes. In the sample, male-headed households are dominant, comprising 78 % of the total participants. The most common level of education is primary education (49.3 %), followed by secondary (25.8 %), unschooled (15.3 %), and tertiary education (9.6 %). Fig. 2 shows the percentage of households in our sample with the years of connection to the MG. About 19 % of the households in our sample received electricity from a nearby hospital before the MG was built. The households are excluded from Fig. 2; in other figures, such households can be distinguished by the years of electricity connection being indicated as seven or more (7+) years. Fig. 1. Schematic of methodological approach applied in the study. Table 1 Questionnaire data categories. Questionnaire Items Information Gathered Demographic characteristics Number of household members, age, number of children, number of adults, marital status, education, source of income, income Dwelling characteristics Number of rooms, location, ownership of SHS Appliances Appliances owned, power rating, purchasing year Energy consumption Monthly electricity expenditure, expenditure on other energy sources, typical time and frequency of appliance use Electricity Access Connection year, payment mechanisms Other Income-generating activities, use of microfinance Fig. 2. Connection year of survey participants. L.T. Fillol et al. Energy Research & Social Science 122 (2025) 104015 3 2.3. Sensitive data validation The main obstacles hindering the acquisition of accurate data include lack of knowledge, inability to recall, and reluctance to answer certain types of questions [22], especially for certain types of financial data, which are prone to be under-reported [23]. This very specific issue was noticed during the data collection process. Some of the participants were reluctant to provide honest feedback on money-related questions, for example, questions regarding income and electricity expenditure. On the other hand, the prevalent use of prepaid payments in rural areas is another concern [21]. Customers utilizing prepaid services do not have access to billing records, leading them to report their electricity consumption relying on memory-based estimates. Within our sampled population, 97 % of customers employ the prepaid method. However, only 23 % recharge monthly, with 43 % doing so weekly, 20 % biweekly, and the remainder engaging in daily or sporadic recharges. While ownership of appliances is objective information and easy to assess, the time of appliance usage and electricity expenses are more difficult to estimate. Therefore, in addition to employing a device-based bottom-up approach to collect data, and as a robust measure, we also performed an additional test to assess the reliability of self-reported electricity expenditure and appliance usage data. The first step was to calculate the expected electricity expenditure for each household based on self-reported appliance ownership, daily appliance usage, and electricity tariffs. The electricity prices used for this calculation were extracted from the KPLC website [24]. The value of the electricity price used in this calculation is 21.16 KES (2023)/kWh (0.13 USD/kWh). We considered each data point as reliable when the self-reported electricity expenditure fell between the range of ±50 % of the calculated electricity expenditure value. Otherwise, it was considered potentially unreliable. The decision to employ a ±50 % threshold aimed to strike an equilibrium between accommodating natural variations in monthly/seasonal energy consumption and fluctuations in electricity prices while ensuring the effective identification of data points that deviate significantly from the predicted values. We then created a reduced dataset that included only users whose self-reported electricity expenditure and self-reported daily appliance usage were consistent. This dataset was employed in the analysis that involved “monthly electricity expenditure” as well as in the evaluation of the tier framework, which is described next. 2.4. Appliance tier framework We employ an appliance tier framework to measure ownership levels in a straightforward manner, as illustrated in Fig. 3. As a reference point, we utilized the multi-tier framework (MTF) defined by the World Bank [25], which has been employed in previous studies for load modeling [26] and adapted for appliance diffusion evaluations [17] as well as estimation of load evolution [6]. In this study, we adapted the tier framework from [25], where tier 1 represents minimal appliance ownership and tier 5 the most advanced. Unlike its original use for assessing energy access, we apply the framework to quantify appliance diffusion over time in mini-grid-connected households, ensuring trends are relevant for electricity forecasting in similar contexts. Thus: •Tier 1 includes only lighting and mobile phones. The interquartile range (IQR) for electricity consumption is 14–24 kWh/month. •Tier 2 expands to include entertainment and office appliances such as a radio, television (TV), laptop (PC), or printer, as well as fans. With an IQR of 19–57 kWh/month. •Tier 3 introduces one medium energy-intensive appliance, such as a cooking device (e.g., blender, pressure cooker, electric kettle, or microwave) or a clothes iron, in addition to tier 2 appliances. The IQR is 24–118 kWh/month. •Tiers 4 to 5 include high energy-intensive appliances like refrigerators, freezers, water pumps, or heating systems. Tier 5 involves owning two or more high-power appliances or three or more medium-power appliances. Tier 4 households exhibit an IQR of 57–142 kWh/month, whereas the IQR of electricity consumption for tier 5 households is 95–142 kWh/month. Our tier definition includes only appliances observed in the sample (Fig. 3). For example, very high-power appliances like air conditioners, were rare. Instead, we account for equivalent energy consumption levels by combining highand medium-power appliances. A notable distinction from the standard MTF is observed in tier 5, which sets a minimum annual consumption of 3000 kWh/year for this tier, while the average in our tier 5 is 1600 kWh/year, with only 3 % of households spending over 5400 KES monthly (equivalent to 3000 kWh/year). Despite this, tier 5 still reflects the highest consumption levels in our framework, we consider that this adjustment in the higher tier ensures that our framework aligns with the real-world electricity usage observed in the empirical data and is suitable to reflect the reality of rural MGs in similar settings. To validate that the tier framework employed is consistent with different levels of electricity demand we first analyzed the appliance combinations by examining plots of the tier levels against energy expenditure, as shown in Fig. 4. Once confirmed the expected trend of higher consumption with higher tiers, analysis of variance (ANOVA) tests were conducted, as detailed in Section 2.5.1, which provided additional statistical confirmation. 2.5. Overview of econometric methods Three distinct statistical methods were employed to conduct a quantitative analysis of the data. Initially, ANOVA method was utilized Fig. 3. Appliance tier framework. Fig. 4. Electricity expenditure by appliance tier. L.T. Fillol et al. Energy Research & Social Science 122 (2025) 104015 4 to assess the appropriateness of the defined appliance tier framework in relation to electricity consumption levels. To investigate the determinants of appliance uptake, multiple linear regression (MLR) was iteratively performed using different sets of variables to ascertain variables with superior predictive power. Subsequently, ordered logistic regression (OLR) was applied to predict the appliance tier based on various factors. Finally, we employed Welch's t-tests to compare electricity expenditure among households with income-generating activities and standard households. The statistical analyses were carried out using Stata® software [27]. Based on previous research and the goals of the current study, relevant explanatory variables expected to influence the dependent variables were identified. Summary statistics of the independent variables are presented in Table 2. Variables employed can be continuous or categorical, and the latter can be either dummy or ordinal. The mean values for dummy variables in Table 2 represent the share of customers with the given characteristics. 2.5.1. Analysis of variance ANOVA is a statistical method used to analyze differences between the means of two or more groups. Another use case of ANOVA tests is comparing models to assess which model explains variability better [28]. In this study, we employed ANOVA for both the aforementioned purposes when evaluating the appliance tier framework. Model results are commonly evaluated through F and p-values. The F -value determines the significance of the variance between the means of the groups, whereas the p-value represents the likelihood of obtaining a result as extreme as the one observed. Typically, p-value below 5 % indicates statistical significance. Upon confirming statistical significance, Tukey's Honest Significant Difference (HSD) test was performed to find the means of the different groups (different tiers in this study) that differ significantly from each other. To evaluate the effectiveness of the appliance tier framework, we test the hypothesis that electricity expenditure increases with higher tier levels and that this categorization yields more robust results compared to using appliance count alone. This is done by comparing two ANOVA models. Both models have monthly electricity expenditure as a dependent variable and differ in terms of independent variables. In the first model, the independent variable is the total number of appliances owned, while in the second, it is the defined appliance tier. This comparison of models serves as an assessment meant to prove which explanatory variable can provide inference with more resolution. Afterwards, we applied another ANOVA test to determine whether the differences between the means of electricity expenditure among the different appliance ownership levels were statistically significant and performed the subsequent Tukey's HSD test. 2.5.2. Multiple linear regression MLR is employed to examine the factors influencing the attainment of the highest appliance tier. Fundamentally, MLR estimates the effect of a unit change in one variable on another. The MLR function shows the linear relationship between the dependent variable (y) and several independent variables or functions of independent variables (x), presented in Eq. (2): y=β0+β1x1+β2x2+…+βnxk+ ε (2) where y is the dependent variable, β 0 is the intercept (the value of y when all the variables are 0), β 1 …n are the regression coefficients, x 1 … k the independent variables, and ε represents the error term. MLR is a typical approach employed to estimate appliance ownership [17]. We assessed the significance of variables using a stepwise selection approach that involved building separate models, each with a distinct set of variables, and comparing adjusted R-squared values to gauge their relative impact. As a result, we scrutinized six distinct MLR models with the appliance tier as the dependent variable. Although non-linear models are typically used with discrete choice outcome variables, which is the case of the appliance tier, we decided to use the linear model to allow straightforward interpretation. At this stage, we are primarily interested in using MLR for understanding directional and relative relationships between time and the appliance tier, and less concerned with bias in our linear model using ordered outcome variables. Nonetheless, as we have coded each household as 1 to 5 depending on which tier of appliances the household has reached (see Fig. 3), the outcome variable mimics a continuous variable in the MLR models. We assessed multi-collinearity computing variance inflation factors (VIF) for each variable in our preferred models. All variables showed low VIF, with a maximum value below 2, indicating no multicollinearity concerns. We tested the assumption of homoscedasticity using the Breusch-Pagan/Cook-Weisberg test. If the p-value was >0.05, the null hypothesis of homoscedasticity was not rejected, indicating that the residuals had constant variance and no further action was necessary. For models where the p-value was <0.05, robust standard errors were applied to account for heteroscedasticity. 2.5.3. Ordered logistic regression To ensure the robustness of the results in determining factors influencing the highest tier reached, we performed ordered logistic regression (OLR) in addition to MLR. Logistic regressions are widely utilized in studies involving categorical dependent variables [29–31]. While Table 2 Variables used for the empirical analyses. Summary Statistics of the variables (N =192) Variable/statistic Data type Variable definition/description Mean St. Dev Min Max Gender Categorical Dummy: 1 =Male, 0 =Female 0.24 0.42 0 1 Education Categorical 0 =Not schooled, 1 =Primary, 2 =Secondary, 3 =Tertiary 1.3 0.83 0 3 Age Continuous Age of the household head in years 41 13.3 17 75 Household size Continuous Total number of rooms in the house 3.8 1.56 1 10 Children Continuous Total number of children in the house 2.5 1.16 1 4 Adults Continuous Total number of adults in the house 3.2 2.18 1 13 Income Continuous Household's income (KES) 12,620 10,474 500 50,000 Expenditure on other energy sources Continuous Household's monthly electricity expenditure on other energy sources (KES) 1425 1537 0 9200 Expenditure on electricity Continuous Household's monthly electricity expenditure on electricity (KES) 1633 1676 100 12,000 Years of MG connection Continuous Years since the household has electricity connection to the MG 4.78 1.48 0 6 Ownership of solar PV system Categorical Dummy: 1 =Yes, 0 =No 0.09 0.29 0 1 Home business Categorical Dummy: 1 =If the customer performs commercial activity from home, 0 = Otherwise 0.27 0.44 0 1 Use of microcredit Categorical Dummy: 1 =If the customer has used credit to purchase appliances, 0 = Otherwise 0.1 0.30 0 1 Appliance tier Categorical Appliance level reached by 2023 3.27 1.35 1 5 L.T. Fillol et al. Energy Research & Social Science 122 (2025) 104015 5 multinomial logistic regression is a popular choice for categorical variables, it discards information when the dependent variable is ordinal, as it ignores the ordered nature of the outcome. Different OLR models exist to address the ordinal aspect of outcomes. Given our research objective, the proportional odds model (POM), which compares cumulated higher categories with the remaining lower categories, aligns logically and comprehensibly with the appliances tier outcome. The POM assumes uniform odds ratios across all categories. If this proportionality condition is violated, the partially proportional odds model (PPOM) is recommended [32]. We assessed the proportional-odds assumption using the Stata gologit2 command with the autofit option. The results indicated no violations of the proportional-odds assumption, as all variables met the parallel-lines criteria with non-significant p-values (e.g. p < 0.05) for all tests. Furthermore, the Wald test for the final model confirmed the assumption, yielding a chi-squared statistic of 23.87 and a p-value of 0.344, indicating no evidence to reject the null hypothesis of parallel lines. Consequently, the POM was deemed suitable for this analysis. As mentioned earlier, the categorization of each household's tier (Y) observation falls into one of five groups. The vector of co-variates is represented by (x i ), which denotes a vector of dimensions p (i =1, 2, …, p), encompassing observations for all p independent variables. Consequently, we can articulate the relationship between the response variable Y and explanatory variables x i through the following expression: P(Y≥yi|x) = 1 1+exp(− α j−xʹ iβ),j=1,2,3,4 (3) Alternatively, the representation of the cumulative probability can be expressed in terms of the log-odds, as follows: log P(Y≥yj |x) 1−P(Y≥yj |x) , = α j+xʹ iβ(4) where P (Y ≥y j ) represents the cumulative probability of an event (Y ≥ y j ) being greater than or equal to y j , α j denotes the respective constant term or intercepts, and β is the vector of regression coefficients with dimension (p) corresponding to the co-variates xi. The index j varies from 1 to J −1, where J is the total number of categories. 2.5.4. Welch's t-test We applied Welch's t-test to compare the means of energy expenditure between households with home businesses and standard households. This test was selected because it does not assume equal variances between groups and is robust to moderate violations of the normality assumption. Preliminary analyses, including the Shapiro-Wilk test (p = 0.014), indicated mild non-normality and the difference in group sizes further justified the use of Welch's test to ensure reliable results. 3. Results This section presents findings from econometric analyses and descriptive statistics. We evaluate the tier framework across electricity expenditure levels, analyze determinants of appliance adoption, and provide survey-based descriptive statistics to contextualize results. 3.1. Appliance tier framework evaluation The ANOVA model comparison reveals a statistically significant effect of both appliance tier and total number of appliances owned on the outcome, electricity expenditure. However, the model that uses the total number of appliances as the independent variable shows a lower Rsquared value of 0.4216, indicating that it explains less of the variance in the electricity expenditure compared to the model with appliance tier, which has an R-squared value of 0.5344. In addition, both models show significance. However, the ANOVA results show that the variable is highly significant (Prob >F =0.0000) in the tier model, while in the total count of appliances model the F statistic is lower (Prob >F = 0.0025), indicating a weaker overall fit. Thus, we conclude that the appliance tier has a statistically more significant effect on energy consumption and is more suitable for estimating energy demand. Subsequently, we performed Tukey's post-hoc test, corresponding to the ANOVA model with appliance tier as the independent variable, as it was deemed more suitable for estimating energy demand. Table 3 presents results of the comparison of electricity expenditure between tiers. The pvalue tells if there is a significant difference between tier comparisons. These results confirm that each successive tier level exerts a stronger impact on energy consumption compared to the preceding tiers. Consequently, as the tier level increases, there is a corresponding statistically significant increase in energy consumption compared to the reference tier, which is tier 1. Nonetheless, considering a threshold of 0.05 significance level, and despite the increase in the average expenditure as the tier increases, the mean difference between a tier level and the immediate successive level is not statistically significant in most of the cases. This is especially true between tiers 4 and 5, which have lower t-value and higher p-value. Thus, indicating a less pronounced difference. Nevertheless, there is still a positive trend in electricity expenditure. Looking ahead, it is reasonable to anticipate that over time tier 5 expenditures may increase, potentially aligning with the broader societal trend of gradually acquiring more appliances. On the other hand, high variability in energy consumption inside each tier implies that energy consumption levels are intersected between consecutive tiers. Fig. 4 shows the gradual increase in the monthly electricity expenditure as the tier advances. We observe that in intermediate tiers, such as tiers 3 and 4, a portion of households remain at very low levels of expenditure. Possible explanations for this will be explored further in Section 3.4. 3.2. Drivers of appliance ownership This section explores factors that influence the current tier reached by different households. The impact of key variables on the dependent variable, appliance tier, was explored through two different methods: MLR and OLR. 3.2.1. Multiple linear regression In the first step, we assess the impact of various variables on the appliance tier using MLR. Various models with distinct sets of independent variables were employed. We adopted an additive approach through the different models to systematically evaluate the influence of alterations in the independent variables on the outcome variable. Results are represented in Table 4. Models 1–4 encompassed the entire user sample, whereas model 5 and 6 exclusively comprised standard households and households with business, respectively. This distinction was motivated by the findings from model 4, which indicated that having a Table 3 Variation of electricity expenditure across tiers. Tukey's post-hoc test Tier comparison Contrast Std. err. t P <|t| [95 % conf. interval] T2 vs T1 0.559 0.276 2.02 0.265 0.008–1.109 T3 vs T1 1.019*** 0.252 6.98 0.000 0.222–1.816 T4 vs T1 1.761*** 0.286 3.56 0.005 1.058–2.464 T5 vs T1 2.040*** 0.282 7.22 0.000 1.253–2.828 T3 vs T2 0.460 0.232 1.98 0.287 −0.189 - 1.019 T4 vs T2 1.202*** 0.189 6.33 0.000 0.672–1.731 T5 vs T2 1.481 0.228 6.48 0.000 0.844–2.11 T4 vs T3 0.741*** 0.203 3.65 0.004 0.175–1.308 T5 vs T3 1.021*** 0.239 4.26 0.000 0.352–1.689 T5 vs T4 0.279 0.198 1.41 0.624 1.478–2.602 *** p < 0.01. L.T. Fillol et al. Energy Research & Social Science 122 (2025) 104015 6 home business has a greater impact on achieving higher appliance tiers compared to all other variables considered. By isolating these household types from the overall sample, we aim to determine whether the effects of the independent variables on the outcome persist within these specific subgroups. Models 4 and 5, with higher R-squared values of 0.41 and 0.40, respectively, and a more comprehensive set of control variables, are considered to provide the best fit to our data. A comparison between models 4 and 5 reveals that the key factors influencing the outcome remain consistent; however, their effects are amplified when the analysis is restricted to standard households in model 5, and additional variables become significant in model 5 compared to model 4. Model 6, which includes only households with businesses, resulted in a nonsignificant model with an F -statistic of 0.632. Consequently, it will not be considered for further evaluation. This outcome may be partially attributed to the limited number of data points. In both models 4 and 5, the duration of electricity access, represented by the variable “years of MG connection,” is strongly associated with ownership of appliances at higher tiers. This positive relationship confirms that the duration of access to electricity plays a key role in explaining higher levels of appliance ownership. However, its influence is reinforced by other significant factors. Model 4 demonstrates that higher education levels, increased income, access to microcredit, the presence of home businesses, and prior access to electricity via SHS significantly influence the likelihood of households owning appliances in higher tiers. Contrarily, increased age is slightly negatively correlated with the dependent variable. In model 5, the number of children negatively correlates with appliance tier, while the number of rooms shows a positive correlation; neither is significant in model 4. This suggests that larger homes and fewer children may indicate higher appliance tiers in standard households, while their impact is less relevant in households with commercial activities. Additionally, the number of rooms might act as a proxy for socioeconomic status, overshadowing the direct effect of income, which is not statistically significant in model 5. 3.2.2. Ordered logistic regression In the second step, we performed OLR to supplement and refine the insights gained from the MLR analysis. OLR has the benefit of less bias in estimating non-linear relationships in the data. In the logistic model, we use the entire sample of observations, encompassing both standard households and households with home businesses and include the variables that were statistically significant in model 4. Although models 4 and 5 of the MLR demonstrates similar model fit in terms of R 2 , we opted to analyze the full dataset to ensure more robust results in the econometric models rather than partitioning the sample. For comparison, the model performed using only standard households, analogous to MLR Table 4 Determinants of higher tier reached. Multiple Linear Regression Dependent variable: Appliance tier Model (1) (2) (3) (4) (5) (6) Years 0.24*** 0.210*** 0.167*** 0.151*** 0.196*** −0.152 (0.05) (0.05) (0.05) (0.06) (0.07) (0.14) Education 0.748*** 0.741*** 0.509*** 0.672*** 0.278 (0.18) (0.19) (0.17) (0.24) (0.23) Rooms 0.156*** 0.128** 0.078 0.145*0.046 (0.05) (0.06) (0.06) (0.07) (0.08) Age −0.012*−0.012*−0.015*** −0.027*** 0.011 (0.007) (0.007) (0.007) (0.008) (0.008) Children −0.156** −0.128 −0.065 −0.080** −0.062 (0.07) (0.07) (0.07) (0.09) (0.08) Income 0.319*0.433*0.409 0.065 (0.17) (0.26) (0.31) (0.42) Microcredit 0.79*** 0.627*** 0.900*** 1.024 (0.29) (0.23) (0.28) (0.17) Gender 0.105 0.167 0.557*−0.403 (0.21) (0.24) (0.32) (0.24) Adults 0.051 0.019 −0.157 −0.009 (0.04) (0.04) (0.05) (0.04) Solar 0.444*** 0.511** 0.040*** (0.17) (0.22) (0.175) Business 1.00*** (0.17) R 2 0.09 0.23 0.28 0.41 0.40 0.18 N189 185 185 168 118 50 Constant 2.04 2.3 1.89 0.57 0.71 4.28 Prob >F0.0000 0.0000 0.0000 0.0000 0.0000 0.632 * p < 0.1. ** p < 0.05. *** p < 0.01. Table 5 OLR higher tier reached. Ordered Logistic Regression Log likelihood = − 235.30 N 177 LR chi2(10) 90.23 Prob >chi2 0.000 Pseudo R2 0.1673 Dependent variable: Appliance tier Independent Variables Odds Ratio Std. Err. z P >| t| [95 % Conf. Interval] Years of MG connection 1.26** 0.131 2.23 0.026 1.028–1.548 Education 2.561*** 0.833 2.89 0.004 1.354–4.846 Age 0.788 0.116 −1.61 0.108 0.590–1.052 Income 1.408*0.282 1.70 0.088 0.949–2.087 Home business 7.170*** 2.686 5.27 0.000 3.445–14.926 SHS 2.481*** 0.742 4.04 0.000 1.380–4.462 Microcredit 3.780*** 1.848 2.72 0.007 1.449–9.861 * p < 0.1. ** p < 0.05. *** p < 0.01. L.T. Fillol et al. Energy Research & Social Science 122 (2025) 104015 7 model 5, is provided in the supplementary material B. The analysis results are presented in Table 5. The results from the OLR are consistent with linear regression in terms of the direction and significance of the impact of the variables on the outcome. The overall consistency across both models lends robustness to our findings. We interpret the results from the OLR model as follows. Except for age, all the covariates included in the model—years of electricity access, education level, income, presence of a home business, presence of SHS and the use of microcredit—were found to be positively correlated and statistically significant. Age was negatively correlated but did not show a statistically significant effect (p =0.109). However, the p-value, being close to 0.1, suggests that older families are less likely to reach higher tiers during the first six years of gaining access to electricity. The results highlight that having a home business and access to microfinance are the strongest predictors of higher appliance tier ownership, followed by prior ownership of SHS. Households with a home business are 7 times more likely to achieve higher appliance tiers compared to those without (OR =7.17, p <0.001), underscoring the critical role of small-scale entrepreneurship in driving appliance adoption. Similarly, a household that has used microcredit is about 4 times more likely to reach a higher tier (OR =3.780, p =0.007), demonstrating the importance of financial inclusion in enabling households for households to invest in advanced appliances. Prior ownership of SHS makes households 2.5 times more likely to reach higher appliance tiers (OR =2.481, p =0.002). This result aligns with the study outcomes of [11], where it is demonstrated that basic access to electricity, for example through SHS, stimulates increasing load demands. While ownership of SHS may reflect higher economic status, it is also a practical and valuable indicator, particularly in under-served communities, as it reliably predicts the likelihood of households advancing more rapidly to higher appliance tiers. Education also emerged as an important determinant of appliance tier mobility, with individuals holding secondary or tertiary education exhibiting over two times higher odds of being in a higher appliance tier (OR =2.342, p =0.007) compared to those with lower education levels. Moreover, a higher income level is associated with 1.4 times greater odds of reaching a higher appliance tier, with a marginally significant effect (p =0.088). This indicates that households with higher incomes have a 40 % increase in the odds of advancing to higher tiers compared to lower-income households. While this effect is not exceptionally large, it underscores the role of economic capacity in enabling households to afford appliances. Each additional year of electricity access increases the likelihood of reaching higher appliance tiers by 26 % (OR =1.261, p =0.026). This result confirms that households adopt appliances incrementally, progressing from lower tiers in the initial years to higher tiers over time as they adapt to and benefit from sustained electrification. However, factors such as access to microfinance, education, business opportunities and presence of SHS play a critical role in driving faster progression. Households move to higher tiers over time, increasing their demand for electricity as they purchase more advanced appliances. The findings emphasize the importance of long-term planning in electrification programs to account for the gradual increase in electricity demand. Moreover, the positive coefficient of income, suggests that households often remain in lower tiers during the initial years of electrification, potentially due to affordability constraints. 3.3. Comparative analysis of home businesses and standard households 3.3.1. Marginal effects of electrification years To better understand the practical implications of the results of the OLR model, we calculated marginal effects that offer interpretable insights into the relationship between predictors and tier membership, using command margins in Stata. Marginal effects measure the change in the predicted probability of being in a specific tier (e.g., tier 1, tier 2, etc.) for a one unit increase in the independent variable while holding all other variables constant. In this analysis, we are particularly interested in understanding how appliance adoption might change over time. Thus, the independent variable of interest is years of electrification. Given that productive use of energy activities at home and access to microfinance are strong predictors of higher appliance tiers, we considered it important to examine their effects in detail. Marginal effects were calculated for each tier (1 to 5) across standard households and households with home businesses. These households were further segmented based on whether they had access to microfinance. Fig. 5 shows how tier membership probabilities evolve with years of electrification, highlighting the influence of home businesses and microfinance access on appliance adoption. The calculations were performed using the margins command in Stata® with the probabilities for each tier derived from the model's predicted values. The error bars in Fig. 5 represent the 95 % confidence intervals for these probabilities, providing a measure of uncertainty around the predictions. We observed that the probability of being in tier 1 or tier 2 is reduced in all the cases as the years of electrification increase. However, tier mobility is very slow in standard households that do not use any financing service. Thus, after 6 years, a standard household has a 40 % probability of being in tier 4, but the probability of being in tier 5 remains very low, about 8 %, indicating that even after several years have passed since gaining access to electricity, many households will remain at very low levels of appliance ownership. When standard households employ microfinancing services, the probability of being in higher tiers after six years increases remarkably to almost 25 % for tier 5 and 51 % for tier 4. As discussed earlier, the probability of being in a higher tier increases when income-generating activities are performed at home, reaching 41 % probability of being in tier 5 after six years, and further, when home businesses use microfinancing services, 73 % probability. It should be noted that the variable representing the years of electrification is considered from the household perspective, rather than the mini-grid perspective. This implies that we are assessing the duration it takes for households to attain certain appliance ownership levels after getting electricity at home. However, not all households are connected in the first year when MG is available, which is another factor that should be considered when estimating overall demand growth in future studies. 3.3.2. Variations in key appliance ownership trends We examined patterns of uptake of the most commonly owned appliances, differentiating between standard households and households with businesses. Fig. 6 provides evidence of a hierarchy of preferences among electric appliances. Most appliances present an increasing trend over the years, except for radios, PCs, and heating, suggesting that time is not a factor that influences their ownership. As anticipated from the findings in previous sections, households engaged in commercial activities possess a greater number of appliances with a notably higher count for kitchen appliances and fans. ICT devices, such as mobile phones and TVs, along with fans, exhibit the fastest adoption, as seen in prior studies conducted in Rwanda [16] and India [17]. These are followed by electric cooking devices, including blenders, pressure cookers, and clothes irons. A discernible upward trend is observed for all the aforementioned appliances, following an S-shaped pattern, as described in earlier studies [33], albeit with varying rates of diffusion. Consistent with the outcomes in [13,16,17], essential welfare appliances like refrigerators or washing machines are less prevalent in standard households during the initial years of electrification. Nevertheless, we found that the likelihood of owning a refrigerator increases significantly in households engaged in productive use of energy activities, even within the early stages of electrification. Microwaves and other energy-intensive appliances, like ovens and washing machines, typically appear around the fifth year of electrification or later. While some households expressed interest in acquiring L.T. Fillol et al. Energy Research & Social Science 122 (2025) 104015 8 them, none owned a washing machine, and only four owned ovens, three of which were home businesses. 3.3.3. Variation on electricity expenditure In order to assess potential differentiation in electricity demand, we compared electricity expenditure between standard households and those with home businesses. The analysis reveals that standard household average electricity expenditures are as low as 50 % of those in households engaged in entrepreneurial activities. To determine whether such difference is statistically significant, Welch's t-tests were Fig. 5. Appliance tier predictive margins with 95 % CIs. Fig. 6. Key appliances uptake. L.T. Fillol et al. Energy Research & Social Science 122 (2025) 104015 9