Bridging the Energy Divide: An Analysis of the Socioeconomic and Technical Factors Influencing Electricity Theft in Kinshasa, DR Congo

Abstract

Electricity theft remains a persistent challenge, particularly in developing economies where infrastructure limitations and socioeconomic disparities contribute to illegal connections. This study analyzes the determinants influencing electricity theft in Kinshasa, the Democratic Republic of Congo, using a logistic regression model applied to 385 observations, which includes random bootstrapping sampling for enhanced stability and power analysis validation to confirm the adequacy of the sample size. The model achieved an AUC of 0.86, demonstrating strong discriminatory power, while the Hosmer–Lemeshow test (p = 0.471) confirmed its robust fit. Our findings indicate that electricity supply quality, financial stress, tampering awareness, and billing transparency are key predictors of theft likelihood. Households experiencing unreliable service and economic hardship showed higher theft probability, while those receiving regular invoices and alternative legal energy solutions exhibited lower risk. Lasso regression was implemented to refine predictor selection, ensuring model efficiency. Based on these insights, a multifaceted policy approach—including grid modernization, prepaid billing systems, awareness campaigns, and regulatory enforcement—is recommended to mitigate electricity theft and promote sustainable energy access in urban environments.

1. Introduction

1.1. Contextual Background

Electricity theft poses a significant global challenge, leading to substantial economic losses for power utilities and introducing safety risks, such as equipment damage and outages. According to a 2017 report by Northeast Group LLC [1], non-technical losses—including theft, fraud, and billing errors—amount to approximately USD 96 billion annually worldwide. In China, electricity theft has severely impacted the State Grid Corporation, prompting USD 84 billion in grid investments in 2024 to curb losses [2]. Despite substantial investments in grid modernization, theft remains prevalent, particularly in developing economies where electricity access is limited and regulatory enforcement is weak.

The Democratic Republic of Congo (DRC) exemplifies this paradox. Despite possessing 100,000 MW of hydropower potential, only 2.5% is harnessed, leaving 21.5% of the population with electricity access as of 2022 [3]. Rapid urban expansion in Kinshasa (13 million residents) and inflation (23.8% in 2023) [4] have weakened household purchasing power, making utility payments unaffordable [5]. The inability to collect revenue effectively restricts grid expansion, worsening service reliability and theft risks. Additionally, operational inefficiencies—including limited personnel, outdated infrastructure, and unreliable billing practices—create opportunities for electricity theft. Socioeconomic disparities, inconsistent billing practices, unreliable supply, and general awareness of tampering methods contribute to a complex environment where electricity theft becomes a coping mechanism for many [6].

1.2. Problem Statement

Despite extensive technical advancements in electricity distribution, theft remains an unresolved issue, particularly in developing economies. While previous studies have explored real-time fraud monitoring, smart metering, and AI-driven detection [7,8], they often overlook the role of socioeconomic factors and behavioral dynamics. In Kinshasa, the persistence of electricity theft stems from a complex interplay of economic hardships, regulatory weaknesses, and consumer perceptions [9]. Understanding these behavioral and infrastructural determinants is critical to formulating holistic solutions beyond mere technological interventions.

1.3. Aim and Objectives

1.3.1. Aim

This work aims to analyze the primary technical, socioeconomic, and behavioral factors influencing electricity theft in Kinshasa using logistic regression.

1.3.2. Objectives

Specifically, this work seeks to:

• Estimate the likelihood of electricity theft in Kinshasa using logistic regression;
• Determine key predictors, including billing methods, meter types, and socioeconomic indicators;
• Assess the impact of behavioral factors, such as awareness of tampering, on theft probability;
• Formulate strategic recommendations to mitigate electricity theft through infrastructure upgrades, policy interventions, and public awareness initiatives.

1.4. Work Structure

The remainder of this paper is organized as follows:

• Section 2: Literature Review—a critical evaluation of existing theft detection methods.
• Section 3: Methodology—survey design, variable selection, logistic regression modeling.
• Section 4: Results—model findings and predictive validity assessment.
• Section 5: Discussion—interpretation, limitations, and future directions.
• Section 6: Conclusion—key results and policy implications.

2. Literature Review

Electricity theft remains a persistent challenge that requires robust detection and mitigation strategies. Research efforts have categorized existing approaches into hardware-based, non-hardware-based, game-theoretic models, and behavioral analysis. While technological solutions focus on real-time monitoring and anomaly detection, emerging studies highlight the role of socioeconomic determinants in shaping theft patterns. This section critically evaluates existing methodologies and identifies gaps relevant to Kinshasa’s electricity theft landscape.

2.1. Hardware-Based Methods

Hardware-based approaches deploy physical devices such as smart meters, tamper-resistant seals, and sensor-driven systems to deter unauthorized electricity usage. Zulu and Dzobo introduced a real-time theft monitoring framework using double-connected data capture, improving anomaly detection rates [5]. Additionally, Advanced Metering Infrastructure (AMI) enables utilities to track consumption patterns remotely and generate alerts for suspicious activities [10].

Despite their effectiveness, hardware-based approaches face deployment challenges, particularly in regions with outdated grids or limited funding. High installation costs hinder large-scale adoption, leaving utilities reliant on manual inspections and penalty enforcement, as highlighted by Bhakta et al. [11]. This highlights the need for hybrid approaches, combining hardware with predictive analytics.

2.2. Non-Hardware Methods

Data-driven approaches leverage network analysis, machine learning, and artificial intelligence (AI) to detect irregular consumption patterns. These techniques offer scalable alternatives to hardware solutions, utilizing grid data and meter readings for theft identification.

2.2.1. Network Data Analysis

Network analysis employs load flow assessments, state estimation, and anomaly detection algorithms to expose theft. Dehghanpour et al. [12] and Zhai et al. [13] demonstrated the utility of state estimation models, while clustering algorithms, as applied in [14,15], group consumption data to expose atypical usage patterns indicative of fraud. Additionally, Sen and Yang [16] utilize three-phase power data with neural network models to enhance theft detection capabilities. Although these methods successfully identify network-wide anomalies, pinpointing individual theft incidents remains difficult.

2.2.2. Consumer Meter Data with Artificial Intelligence

Artificial intelligence (AI) has become essential in consumer-level electricity theft detection, leveraging machine learning to analyze meter readings and identify anomalies.

Supervised learning models, including Support Vector Machines (SVMs), decision trees, and neural networks, have proven effective in classifying fraudulent users [8,17,18]. Hu et al. [19] highlight Random Forest algorithms, known for their ability to handle imbalanced datasets. Regression techniques [20] have also been applied to detect faulty smart meters and theft through consumption anomalies.

Despite their effectiveness, supervised models face data labeling constraints, class imbalance issues, and overfitting risks. To mitigate these challenges, unsupervised learning techniques, such as clustering, have gained traction. Sasmoko et al. [21] utilize anomaly detection models, Yang et al. [22] introduce ant colony clustering algorithms, and Žarković and Dobrić [23] leverage K-Means clustering, while Qi et al. [24] integrate wavelet-based feature extraction with fuzzy c-means (FCM) clustering. However, threshold definition and overlapping clusters remain obstacles.

Ensemble learning approaches have emerged as powerful alternatives. Kawoosa and Prashar [25] employ XGBoost, enhancing classification accuracy, while Qi et al. [24] integrate Minimal Gated Memory (MGM) networks and Adaptive Synthetic Sampling (ADASYN) to address class imbalance issues. Hybrid models [26,27] combining Recurrent Neural Networks (RNNs) and Bidirectional Long Short-Term Memory (Bi-LSTM) further enhance theft detection efficiency.

However, AI models face limitations such as overfitting risks, data labeling constraints, and high computational requirements [25]. A promising alternative is logistic regression modeling, which simplifies predictive frameworks and ensures practical usability in policy applications.

2.2.3. Hybrid Methods

Hybrid methods combine network-level data with consumer meter data to enhance theft detection accuracy. Wang et al. [28] demonstrate that clustering algorithms help detect anomalies in network data, while AI models refine individual consumption analysis, improving detection precision [29]. A key approach employs heuristic segmentation, transforming time-series consumption patterns and grid loss metrics to extract theft indicators. Consumers showing a strong correlation with line losses and consistent waveform anomalies are flagged as suspicious.

Despite their effectiveness, hybrid methods require advanced data infrastructure and high computational power. Nonetheless, they are integral to modern smart grids, enabling real-time, data-driven monitoring for theft mitigation.

2.3. Game Theory

Game theory evaluates consumer–utility interactions, treating electricity theft as a strategic decision-making process.

Wei et al. [30] integrate Benford’s Law with a Stackelberg model, positioning the utility company as a leader, while consumers react by optimizing theft strategies. Likelihood Ratio Tests (LRTs) further refine detection. In contrast to other studies presented in [31], Ref. [32] explored Bayesian intrusion detection frameworks, enhancing utility intervention tactics.

Despite its theoretical appeal, game theory models rely on assumptions of rational consumer behavior, limiting real-world applicability due to data constraints and scalability issues [31]. Thus, behavioral analysis remains critical in understanding theft motivations beyond economic utility optimization.

2.4. Behavioral Analysis

Behavioral analysis examines the psychological and socioeconomic motivations behind electricity theft, identifying patterns in consumer decision-making.

Surveys and interviews have uncovered key theft drivers, such as financial hardship, lack of awareness, and dissatisfaction with utility services [33]. Socioeconomic models [34] integrating income levels, demographics, and billing history provide data-driven fraud detection. Studies [9,35] emphasize the need for enhanced billing systems, stricter enforcement, and public awareness campaigns to mitigate theft risks. Razavi and Fleury [9] highlight income inequality and inadequate infrastructure as key contributors to non-technical losses.

Increasingly, behavioral insights are combined with AI-based detection models, as demonstrated by Hussain [36] and Bansal et al. [37], improving predictive accuracy.

2.5. Research Gaps and Contribution

The existing literature predominantly focuses on hardware detection, AI-based monitoring, and economic modeling. However, limited research incorporates behavioral predictors within statistical fraud detection frameworks.

Table 1. Systematic literature review.

Approach	References	Strengths	Weaknesses	Link to This Study
Hardware-Based Methods	[5]	Improved detection accuracy through redundancy	Limited adaptability to outdated grid infrastructure	Complements hardware detection by exploring theft motivations
	[10]	Effective smart metering	Lack of integration of behavioral factors	Bridges technological and behavioral insights
	[11]	Provide affordable theft monitoring solutions	Scalability issues in dense urban settings like Kinshasa	Offers a predictive framework complementing hardware approaches
	[7]	Effectively identify fraudulent connections	High deployment costs hinder mass implementation	Assesses affordability’s role in theft by incorporating economic stress factors
Network Data Analysis	[12,13]	High accuracy in anomaly detection	Struggles to pinpoint theft at consumer level	Enhance predictive precision with fine-grained household theft metrics
	[14]	Classifies irregular consumption patterns effectively	Cannot infer causality behind theft	Examines why theft occurs rather than just detecting it
Consumer Meter AI	[19,20]	Handles imbalanced datasets well	Computationally demanding	Use logistic regression for scalability and policy application
	[24,25,26,27]	Enhances predictive robustness with ensemble AI models	Complexity limits real-world deployment	Simplify predictive modeling for practical usability
Hybrid Methods	[29,38]	Improved detection accuracy using hybrid models	Face challenges in data integration	Streamline data processing via logistic regression modeling
Game Theory	[30,31,32]	Strategically optimizes theft deterrence	Assumes rational consumer behavior, which may not hold universally	Integrate real-world socio-economic stress factors into theft prediction
Behavioral Analysis	[9,34,35]	Identifies income and infrastructure gaps as theft enablers	Lacks predictive modeling	Quantify theft likelihood using logistic regression
	[36,37]	Examines AI-driven fraud detection through consumer psychology	High implementation complexity	Provide practical behavioral recommendations, including education campaigns

provides a summary of currently available literature. This study bridges the gap by integrating billing stress, meter type, tampering awareness, financial hardship, and regulatory influences into logistic regression modeling, offering a comprehensive predictive framework for electricity theft detection in Kinshasa.

3. Methodology

This study employs a rigorous statistical approach to analyze electricity theft determinants in Kinshasa, leveraging logistic regression for predictive modeling. The research design integrates random sampling, bootstrapping techniques, and regularization-based feature selection to ensure robust, reliable insights into theft patterns.

3.1. Data Collection and Preprocessing

3.1.1. Survey Design

A structured survey was developed to assess technical, behavioral, and socioeconomic influences on electricity theft. The key predictor variables and their modalities are detailed in

Table 2. Variable descriptions and modalities.

ID	Variable Name	Description	Modalities
Y	Electricity theft (Target)	Main outcome variable	No/Yes
X1	Household size	Total number of individuals in the household	Continuous (count-based)
X2	Number of appliances	Total count of electrical appliances used	Continuous (count-based)
X3	Problems encountered with meter	Refers to the total count of problems or issues experienced with electricity meters	Continuous (count-based)
X4	Electricity supply quality	Perceived supply quality (scale-based)	0 to 10
X5	Alternatives used during power outages	Total number of possible backup energy sources used	Continuous (count-based)
X6	Receipt of electricity bills	Whether households receive bills regularly	No/Yes
X7	Experience with electricity meter	Prior experience managing electricity meters	No/Yes
X8	Stress about paying bills	Financial stress related to paying bills	No/Yes
X9	Awareness of tampering	Knowledge of electricity theft methods	No/Yes
X10	Homeownership status	Ownership distinction	Owner/Tenant
X11	Payment type	The payment method associated with a customer’s account	Prepaid/Postpaid
X12	Type of electricity Meter	Type of installed meter	Flat-Rate Payment/Mechanical/Electronic/Smart
X13	Electricity cost	Household electricity expenses	Less than USD 10/Between USD 10 and USD 30/ Between USD 30 and USD 50/ Between USD 50 and USD 100/ Between USD 100 and USD 200/More than USD 200
X14	Household monthly Income	Monthly earnings category	Less than USD 1000/Between USD 1000 and USD 3000/Between USD 3000 and USD 5000/ More than USD 5000/Not specific
X15	Perception of electricity costs	Subjective evaluation of electricity pricing	Low/Medium/High/Uncertain
X16	Technical ability to manipulate meters	Self-rated electrical skills	Very low/Low/Medium/High/Very high
X17	Education level	Highest education achieved	None/Primary/Secondary/Higher Education/Hesitant
X18	Difficulty paying bills	Frequency of bill payment struggles	Never/Rarely/Sometimes/Always/Often
X19	Opinion on theft Penalties	Whether respondents consider penalties adequate	No/Yes

below:

To ensure validity and reliability, domain relevance guided variable selection, grounded in existing electricity theft research [9,39]. Exploratory analyses assessed initial correlations between theft likelihood and independent variables.

3.1.2. Population and Sample Size Determination

A statistically valid sample size is essential for meaningful conclusions in survey-based studies [37]. Using the Krejcie and Morgan formula, an initial sample size of 385 participants was determined for Kinshasa’s population (13 million residents). A power analysis further validated its adequacy in capturing theft-related socioeconomic patterns:

$$n={\displaystyle \frac{N.Z^2.p.\left(1-p\right)}{e^2.\left(N-1\right)+Z^2.p.\left(1-p\right)}}$$

(1)

where

• N = population size;
• Z = 1.96 (95% confidence level);
• p = 0.5 (the population proportion assuming maximum variance);
• e = 0.05 (margin of error or the degree of accuracy expressed as a proportion).

As Kinshasa’s urban population represents 40% of the DRC’s total population, Equation (1) simplifies under large-population assumptions:

$$\underset{N\to \infty }{\lim }n={\displaystyle \frac{Z^2.p.\left(1-p\right)}{e^2}}$$

(2)

resulting in a final sample size of 385 respondents.

To prevent demographic biases, random sampling ensured equal probability selection, complemented by bootstrapping techniques for enhanced model stability and reliability.

These methodological choices enable a holistic understanding of electricity theft drivers in Kinshasa, independent of the subgroup segmentation.

3.1.3. Preprocessing

Following data collection, the steps taken included the following:

• Integrity checks confirmed completeness, eliminating the need for imputation.
• Quantitative variables were standardized, ensuring uniform scales for regression modeling.
• Categorical variables underwent dummy coding (binary) and ordinal transformations (ranked perceptions) for suitability in logistic regression.

Multicollinearity detection via the Variance Inflation Factor (VIF) safeguarded model stability, ensuring predictor independence.

3.2. Logistic Regression

3.2.1. Model Specification

A binary logistic regression model was formulated to predict the likelihood of electricity theft (Y = Yes or No) [40], and Logit function was chosen as the link function, ensuring a proper classification framework:

$$P\left(Y=M2\right)={\displaystyle \frac{1}{1+e^{-\left({\beta }_0+{\beta }_1{X}_1+\dots +{\beta }_{20}{X}_{19}\right)}}}$$

(3)

where

• Y is the binary dependent variable (theft occurrence);
• X_n (n = 1 to 19) represents independent predictors (socioeconomic, technical, and behavioral variables);
• β_n (n = 1 to 19) represents the regression coefficients (interpreted using odds ratios).

3.2.2. Variable Selection and Optimization

To prevent overfitting and enhance interpretability, the following analyses were performed:

• Multicollinearity diagnostics, using the Variance Inflation Factor (VIF) to exclude highly correlated predictors.
• Feature importance analysis ranked the most influential variables.
• Lasso regression (cv = 5) optimized the regularization parameters.

3.2.3. Parameter Estimation

The Newton–Raphson algorithm iteratively updated the model coefficients, converging when the likelihood change dropped below 0.001, ensuring accurate parameter estimation.

3.2.4. Model Evaluation

The predictive performance of the model was evaluated primarily using the Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC) curve. Additional metrics, including McFadden’s R², Cox and Snell R², and Nagelkerke R², were computed to measure the goodness of fit. Furthermore, classification, accuracy, sensitivity, and specificity were derived from the confusion matrix to quantify the overall performance and diagnostic effectiveness.

3.2.5. Significance Testing and Validation

Predictors underwent Wald tests and Type II likelihood ratio tests to establish significance, with 95% confidence intervals calculated for odds ratio interpretations.

4. Results

4.1. Sample Adequacy and Validation

4.1.1. Sample Selection and Bootstrap Validation

The study employed a random bootstrapping approach, ensuring model stability across different samples. The AUC distribution over 100 bootstrap iterations yielded a mean AUC of 94.92%, with a standard deviation of 0.8%, indicating strong discriminative performance. Additionally, 94% of samples fell within the upper and lower confidence bounds, confirming consistent classification accuracy across different sampling conditions. This is indicated in Figure 1.

4.1.2. Post-Hoc Power Analysis

• The sample size used was 385.
• The observed effect size.
• The logistic regression model explained 61.4% of the variability in electricity theft likelihood (Nagelkerke R² = 0.614).
• The predictors demonstrated strong influence (Cohen’s f² = 1.588), confirming the robustness of the model.
• The significance level (α) was 0.05.
• This is the probability of making a Type I error (rejecting a true null hypothesis).
• α = 0.05 means there is a 5% chance of falsely detecting an effect when none exists.
• The computed power (1–β) was 1.0000.
• This indicates an extremely high probability of detecting a true effect if one exists.
• Given this high power, the risk of Type II error (false negatives) is negligible, ensuring the reliability of the model results.

4.2. Descriptive Analysis

4.2.1. Quantitative Variables

The key variables exhibited moderate variations:

• Household size (Mean = 4.699, Std Dev = 2.174).
• Number of appliances (Mean = 3.418, Std Dev = 2.311).
• Electricity supply quality (Mean = 5.036, Std Dev = 2.354). The low standard deviation for backup energy sources (X5 = 0.353) suggests consistent reliance on alternative power solutions.

  Table 3. Descriptive statistics (quantitative variables).

  Variable	Observations	Obs. with Missing Data	Obs. without Missing Data	Minimum	Maximum	Mean	Standard Deviation
  X1	385	0	385	1	10	4.699	2.174
  X2	385	0	385	1	11	3.418	2.311
  X3	385	0	385	0	3	0.587	0.632
  X4	385	0	385	1	10	5.036	2.354
  X5	385	0	385	1	4	1.091	0.353

  provides descriptive statistics for this analysis.

4.2.2. Qualitative Variables

The categorical variables provided the following key behavioral insights:

• A total of 41.04% of households reported electricity theft occurrences, highlighting the prevalence of the issue.
• A total of 70.91% of respondents were tenants, suggesting potential links between rental status and theft likelihood.
• Prepaid electricity usage (57.66%) was more common than postpaid, possibly affecting theft patterns.
• A total of 54.03% had tampering awareness, indicating a widespread understanding of meter manipulation techniques.

  Table 4. Descriptive analysis of qualitative variables.

  Variable	Category	Count	Frequency (%)
  Y	No	227	58.961
  	Yes	158	41.039
  X6	No	125	32.468
  	Yes	260	67.532
  X7	No	214	55.584
  	Yes	171	44.416
  X8	No	188	48.831
  	Yes	197	51.169
  X9	No	177	45.974
  	Yes	208	54.026
  X10	Owner	112	29.091
  	Tenant	273	70.909
  X11	Postpaid	163	42.338
  	Prepaid	222	57.662
  X12	Electronic	128	33.247
  	Mechanical	108	28.052
  	Smart meters	41	10.649
  	Flat-rate payment	108	28.052
  X13	Less than USD 10	101	26.234
  	Between USD 10 and USD 30	203	52.727
  	Between USD 30 and USD 50	56	14.545
  	Between USD 50 and USD 100	18	4.675
  	Between USD 100 and USD 200	5	1.299
  	More than USD 200	2	0.519
  X14	Less than USD 1000	137	35.584
  	Between USD 1000 and USD 3000	75	19.481
  	Between USD 3000 and USD 5000	13	3.377
  	More than USD 5000	4	1.039
  	Not specific	156	40.519
  X15	Uncertain	72	18.701
  	Low	58	15.065
  	Medium	192	49.87
  	High	63	16.364
  X16	Very low	37	9.61
  	Low	90	23.377
  	Medium	238	61.818
  	High	19	4.935
  	Very high	1	0.26
  X17	Hesitant	22	5.714
  	Low	43	11.169
  	Medium	176	45.714
  	High	90	23.377
  	Very high	54	14.026
  X18	Never	175	45.455
  	Rarely	100	25.974
  	Sometimes	64	16.623
  	Often	37	9.61
  	Always	9	2.338
  X19	No	231	60
  	Yes	154	40

  provides descriptive analysis of quantitative variables.

4.3. Multicollinearity and Feature Selection

The variables X7 and X19 exhibit high VIF values (>5), potentially indicating strong collinearity, whereas X5 has the lowest VIF (1.18), suggesting minimal correlation with other predictors. The multicollinearity statistics are provided in

Table 5. Multicollinearity statistics.

Variable_ID	Modality	Tolerance	VIF
X1		0.7618	1.3127
X2		0.7026	1.4233
X3		0.5855	1.7079
X4		0.6750	1.4814
X5		0.8466	1.1812
X6	No	0.5008	1.9967
	Yes	0.5008	1.9967
X7	No	0.1338	7.4753
	Yes	0.1338	7.4753
X8	No	0.3837	2.6062
	Yes	0.3837	2.6062
X9	No	0.4230	2.3639
	Yes	0.4230	2.3639
X10	Owner	0.8394	1.1914
	Tenant	0.8394	1.1914
X11	Postpaid	0.6725	1.4869
	Prepaid	0.6725	1.4869
X12	Electronic	0.7457	1.3411
	Mechanical	0.8233	1.2146
	Smart meters	0.7940	1.2594
	Rate payment	0.4813	2.0776
X13	Less than USD 10	0.6705	1.4915
	Between USD 10 and USD 30	0.7491	1.3349
	Between USD 30 and USD 50	0.7998	1.2503
	Between USD 50 and USD 100	0.8571	1.1667
	Between USD 100 and USD 200	0.8542	1.1706
	More than USD 200	0.9047	1.1053
X14	Less than USD 1000	0.8398	1.1907
	Between USD 1000 and USD 3000	0.8324	1.2014
	Between USD 3000 and USD 5000	0.8244	1.2131
	More than USD 5000	0.8965	1.1154
	Not specific	0.6686	1.4956
X15	Uncertain	0.3513	2.8469
	Low	0.8593	1.1637
	Medium	0.6108	1.6372
	High	0.7559	1.3230
X16	Very low	0.8086	1.2367
	Low	0.8350	1.1976
	Medium	0.7273	1.3750
	High	0.8499	1.1766
	Very high	0.9074	1.1020
X17	Hesitant	0.8351	1.1975
	Low	0.6624	1.5096
	Medium	0.6051	1.6525
	High	0.6709	1.4904
	Very high	0.4199	2.3815
X18	Never	0.3203	3.1222
	Rarely	0.6383	1.5667
	Sometimes	0.7095	1.4094
	Often	0.8178	1.2227
	Always	0.8922	1.1208
X19	No	0.1267	7.8926
	Yes	0.1267	7.8926

.

Lasso regression optimized feature selection, with variables X8 (financial stress) and X9 (tampering awareness) showing the strongest positive coefficients. The selection regression feature for electricity theft detection are given in

Table 6. Lasso regression feature selection results for electricity theft detection.

Feature Index	Lasso Coefficient
1	−0.00043049
2	−0.06262616
3	0
4	−0.07939336
5	0.04721492
6	0.01983738
7	0.06509673
8	0.13018794
9	0.11120266
10	−0.01828389
11	−0.01760688
12	−0.01037067
13	−0.00282541
14	0
15	−0.00454788
16	−0.00663296
17	−0.00524908
18	0.02863104
19	0.05021244

.

4.4. Regression Analysis of Variable Y

4.4.1. Model Specification

This summary highlights the model’s effectiveness and provides key insights into its explanatory power. This is shown in

Table 7. Logistic regression model for variable Yes—model fit and performance metrics.

Statistic	Independent Model	Complete Model
Observations	385	385
Sum of Weights	385.000	385.000
Degrees of Freedom (DDL)	384	350
-2 Log-Likelihood	521.290	179.857
McFadden’s R2	0.000	0.655
Cox and Snell R2	0.000	0.588
Nagelkerke R2	0.000	0.759
Akaike Information Criterion (AIC)	523.290	249.857
Schwarz Bayesian Criterion (SBC)	527.243	388.221
Iterations	0	16

.

4.4.2. Test of Null Hypothesis

The p-values (<0.0001) indicate strong statistical significance, suggesting that the model is highly informative for explaining variations in the dependent variable. The model evaluation is summarized in

Table 8. Model evaluation—null hypothesis test H₀: Pr (Yes = Yes) = 0.41.

Statistic	Degrees of Freedom (DDL)	Chi-Square (Khi2)	Pr > Chi2
−2 Log (Vraisemblance)	34	341.433	<0.0001
Score	34	253.330	<0.0001
Wald	34	97.362	<0.0001

.

4.4.3. Type II Analysis (Variable Y)

This table presents the Type II analysis, detailing statistical measures for various predictors associated with variable Y. It includes the following:

• Degrees of Freedom (DDL): These determine the number of independent comparisons per variable.
• Wald Chi-Square (Khi² Wald) and Pr > Wald: These assess the individual significance of predictors in the model. Smaller p-values (Pr > Wald) suggest stronger statistical significance.
• Likelihood Ratio Chi-Square (Khi² LR) and Pr > LR: These evaluate the overall model contribution when including each variable. Results for type II analysis are given in

  Table 9. Type II analysis—statistical significance of predictors for variable Y.

  Source	DDL	Khi2 (Wald)	Pr > Wald	Khi2 (LR)	Pr > LR
  X1	1	0.102	0.749	0.102	0.749
  X2	1	3.895	0.048	4.048	0.044
  X4	1	15.021	0.000	17.353	<0.0001
  X5	1	12.691	0.000	14.461	0.000
  X6	1	3.248	0.072	3.288	0.070
  X7	1	1.610	0.204	1.532	0.216
  X8	1	10.793	0.001	11.629	0.001
  X9	1	9.144	0.002	9.544	0.002
  X10	1	0.671	0.413	0.672	0.412
  X11	1	0.668	0.414	0.676	0.411
  X12	3	1.535	0.674	1.540	0.673
  X13	5	9.116	0.105	12.656	0.027
  X15	3	8.831	0.032	9.853	0.020
  X16	4	0.817	0.936	1.669	0.796
  X17	4	0.220	0.994	0.223	0.994
  X18	4	6.329	0.176	6.624	0.157
  X19	1	0.610	0.435	0.625	0.429

  .

4.4.4. Hosmer–Lemeshow Test (Variable Y)

The Hosmer–Lemeshow test evaluates the goodness of fit of a logistic regression model by comparing observed and predicted probabilities across different risk groups.

• A Chi-Square value of 8.641 with eight degrees of freedom suggests the model’s predictions align reasonably well with the observed data.
• The p-value (Pr > Khi²) of 0.374 indicates that there is no strong evidence of poor fit, meaning the model does not significantly deviate from the expected outcomes. The goodness-of-fit test results are provided in

  Table 10. Hosmer–Lemeshow goodness-of-fit test results.

  Statistic	Chi-Square (Khi2)	Degrees of Freedom (DOFs)	Pr > Chi-Square (Pr > Khi2)
  Hosmer–Lemeshow	8.641	8	0.374

  .

4.4.5. Model Parameters (Y)

This table presents the results of a logistic regression analysis for Variable Y, detailing statistical indicators for the following model predictors:

• Values and Standard Errors: These estimate each variable’s effect size with its respective uncertainty.
• Wald Chi-Square and Pr > Chi²: These evaluate statistical significance, where low p-values (<0.05) indicate strong predictor influence.
• Confidence Intervals (95%): The lower and upper bounds reflect estimation precision.
• Odds Ratio (OR): The OR measures the likelihood of outcome occurrence for each predictor. Logistic Regression results are given in

  Table 11. Logistic regression results for variable Y—parameter estimates and odds ratios.

  Source	Value	Standard Error	Wald Chi-Square	Pr > Chi2	Lower Bound (95%)		Upper Bound (95%)		Odds Ratio	Lower Bound OR (95%)	Upper Bound OR (95%)
  Constant	−7.667	3871.355	0.000	0.998	−7595.383		7580.049
  X1	−0.033	0.103	0.102	0.749	−0.234		0.168		0.968	0.791	1.184
  X2	−0.191	0.097	3.895	0.048	−0.380		−0.001		0.826	0.684	0.999
  X4	−0.391	0.101	15.021	0.000	−0.589		−0.193		0.676	0.555	0.824
  X5	1.989	0.558	12.691	0.000	0.895	3.083	7.308	2.447	21.829
  X6—No	−1.111	0.616	3.248	0.072	−2.318	0.097	0.329	0.098	1.102
  X6—Yes	0.000	0.000
  X7—No	−1.193	0.940	1.610	0.204	−3.037	0.650	0.303	0.048	1.915
  X7—Yes	0.000	0.000
  X8—No	−1.781	0.542	10.793	0.001	−2.844	−0.718	0.168	0.058	0.487
  X8—Yes	0.000	0.000
  X9—No	−1.578	0.522	9.144	0.002	−2.601	−0.555	0.206	0.074	0.574
  X9—Yes	0.000	0.000
  X10—Owner	−0.385	0.470	0.671	0.413	−1.306	0.536	0.681	0.271	1.710
  X10—Tenant	0.000	0.000
  X11—Postpaid	0.399	0.488	0.668	0.414	−0.557	1.354	1.490	0.573	3.873
  X11—Prepaid	0.000	0.000
  X12—Electronic	−0.273	0.672	0.166	0.684	−1.590	1.043	0.761	0.204	2.838
  X12—Mechanical	0.349	0.618	0.320	0.571	−0.861		1.560		1.418	0.423	4.759
  X12—Smart meters	0.434	0.955	0.207	0.649	−1.437		2.306		1.544	0.238	10.035
  X12—Flat-rate payment	0.000	0.000
  X13—Between USD 10 and USD 30	12.114	3871.355	0.000	0.998	−7575.602		7599.830
  X13—Between USD 100 and USD 200	29.456	4523.761	0.000	0.995	−8836.953		8895.866
  X13—Between USD 30 and USD 50	12.159	3871.355	0.000	0.997	−7575.557		7599.875
  X13—Between USD 50 and USD 100	11.093	3871.355	0.000	0.998	−7576.623		7598.809
  X13—Less than USD 10	13.662	3871.355	0.000	0.997	−7574.054		7601.377
  X13—More than USD 200	0.000	0.000
  X15—High	−1.583	0.924	2.935	0.087	−3.394	0.228	0.205	0.034	1.256
  X15—Low	−2.526	0.861	8.610	0.003	−4.213	−0.839	0.080	0.015	0.432
  X15—Medium	−1.772	0.792	5.009	0.025	−3.325	−0.220	0.170	0.036	0.802
  X15—Uncertain	0.000	0.000
  X16—High	−0.579	1.546	0.140	0.708	−3.608	2.450	0.561	0.027	11.594
  X16—Low	0.446	0.839	0.283	0.595	−1.199	2.092	1.563	0.302	8.099
  X16—Medium	0.114	0.782	0.021	0.884	−1.419	1.646	1.120	0.242	5.186
  X16—Very high	−16.697	5707.069	0.000	0.998	−11,202.348	11,168.954
  X16—Very low	0.000	0.000
  X17—High	−0.110	0.835	0.017	0.895	−1.748	1.527		0.895	0.174	4.603
  X17—Low	0.263	1.106	0.056	0.812	−1.905	2.431		1.301	0.149	11.367
  X17—Medium	−0.009	0.817	0.000	0.991	−1.610	1.592		0.991	0.200	4.913
  X17—Hesitant	−0.263	1.098	0.057	0.811	−2.414	1.888		0.769	0.089	6.609
  X17—Very high	0.000	0.000
  X18—Always	1.001	1.499	0.446	0.504	−1.936	3.938		2.721	0.144	51.310
  X18—Never	−1.731	0.782	4.906	0.027	−3.263	−0.199		0.177	0.038	0.819
  X18—Often	−0.260	0.784	0.110	0.740	−1.798	1.277		0.771	0.166	3.586
  X18—Rarely	−0.752	0.612	1.508	0.219	−1.952	0.448		0.471	0.142	1.565
  X18—Sometimes	0.000	0.000
  X19—No	−0.718	0.920	0.610	0.435	−2.521	1.084		0.488	0.080	2.958
  X19—Yes	0.000	0.000

  , and the standardized coefficients in

  Table 12. Standardized coefficients (variable Y).

  Source	Value	Standard Error	Wald Chi-Square (Khi2)	Pr > Chi2	Lower Bound (95%)	Upper Bound (95%)
  X1	−0.039	0.123	0.102	0.749	−0.280	0.202
  X2	−0.243	0.123	3.895	0.048	−0.484	−0.002
  X4	−0.507	0.131	15.021	0.000	−0.763	−0.251
  X5	0.386	0.108	12.691	0.000	0.174	0.599
  X6—No	−0.287	0.159	3.248	0.072	−0.599	0.025
  X6—Yes	0.000	0.000
  X7—No	−0.327	0.258	1.610	0.204	−0.832	0.178
  X7—Yes	0.000	0.000
  X8—No	−0.491	0.149	10.793	0.001	−0.784	−0.198
  X8—Yes	0.000	0.000
  X9—No	−0.434	0.143	9.144	0.002	−0.715	−0.153
  X9—Yes	0.000	0.000
  X10—Owner	−0.096	0.118	0.671	0.413	−0.327	0.134
  X10—Tenant	0.000	0.000
  X11—Postpaid	0.109	0.133	0.668	0.414	−0.152	0.369
  X11—Prepaid	0.000	0.000
  X12—Electronic	−0.071	0.174	0.166	0.684	−0.413	0.271
  X12—Mechanical	0.087	0.153	0.320	0.571	−0.213	0.386
  X12—Smart meters	0.074	0.162	0.207	0.649	−0.244	0.392
  X12—Flat-rate payment	0.000	0.000
  X13—Between USD 10 and USD 30	3.334	1065.606	0.000	0.998	−2085.215	2091.884
  X13—Between USD 100 and USD 200	1.839	282.375	0.000	0.995	−551.607	555.284
  X13—Between USD 30 and USD 50	2.363	752.497	0.000	0.997	−1472.504	1477.231
  X13—Between USD 50 and USD 100	1.291	450.591	0.000	0.998	−881.851	884.433
  X13—Less than USD 10	3.313	938.929	0.000	0.997	−1836.954	1843.581
  X13—More than USD 200	0.000	0.000
  X15—High	−0.323	0.188	2.935	0.087	−0.692	0.047
  X15—Low	−0.498	0.170	8.610	0.003	−0.831	−0.165
  X15—Medium	−0.489	0.218	5.009	0.025	−0.917	−0.061
  X15—Uncertain	0.000	0.000
  X16—High	−0.069	0.185	0.140	0.708	−0.431	0.293
  X16—Low	0.104	0.196	0.283	0.595	−0.280	0.488
  X16—Medium	0.030	0.209	0.021	0.884	−0.380	0.441
  X16—Very high	−0.469	160.151	0.000	0.998	−314.358	313.421
  X16—Very low	0.000	0.000
  X17—High	−0.026	0.195	0.017	0.895	−0.408	0.356
  X17—Low	0.046	0.192	0.056	0.812	−0.331	0.422
  X17—Medium	−0.002	0.224	0.000	0.991	−0.442	0.437
  X17—Hesitant	−0.034	0.140	0.057	0.811	−0.309	0.242
  X17—Very high	0.000	0.000
  X18—Always	0.083	0.125	0.446	0.504	−0.161	0.328
  X18—Never	−0.475	0.215	4.906	0.027	−0.896	−0.055
  X18—Often	−0.042	0.127	0.110	0.740	−0.292	0.207
  X18—Rarely	−0.182	0.148	1.508	0.219	−0.472	0.108
  X18—Sometimes	0.000	0.000
  X19—No	−0.194	0.248	0.610	0.435	−0.681	0.293
  X19—Yes	0.000	0.000

  .

4.5. Predictions and Residuals (Variable Y)

4.5.1. Prediction Accuracy

The model maintained high accuracy across training, test, and validation sets:

• Train Set: accuracy = 87.8%, and the F1-score = 0.88
• Test Set: accuracy = 77.8%, and the F1-score = 0.78
• Validation Set: accuracy = 83.9%, and the F1-score = 0.83. These results highlight consistent predictive reliability. The model performance metrics for class 1, 0 and combined are given in

  Table 13. Model performance metrics for Class 1.

  Dataset	Accuracy	Precision (Class 1)	Recall/Sensitivity	F1-Score (Class 1)
  Train Set	0.878	0.85	0.84	0.85
  Test Set	0.778	0.78	0.72	0.75
  Validation Set	0.839	0.72	0.78	0.75

  ,

  Table 14. Model performance metrics for Class 0.

  Dataset	Accuracy	Precision (Class 0)	Recall/Sensitivity (Class 0)	F1-Score (Class 0)
  Train Set	0.878	0.9	0.9	0.9
  Test Set	0.778	0.78	0.83	0.8
  Validation Set	0.839	0.9	0.86	0.88

  and

  Table 15. Model performance metrics across datasets (combined Classes 0 and 1).

  Dataset	Overall Accuracy	Precision (Avg.)	Recall/Sensitivity (Avg.)	F1-Score (Avg.)
  Train Set	0.878	0.88	0.87	0.88
  Test Set	0.778	0.78	0.77	0.78
  Validation Set	0.839	0.84	0.82	0.83

  , respectively.

4.5.2. Residual Analysis

The mean deviance residuals indicated in Figure 2 remained within acceptable thresholds, confirming model robustness. The residual metrics are provided in

Table 16. Residual analysis.

Dataset	Deviance Residuals (Mean)	Pearson Residuals (Mean)
Train Set	0.05	0.159
Test Set	6.574	−5.076
Validation Set	0.066	−0.183

.

5. Discussion

5.1. Model Fit and Performance

The logistic regression model exhibited strong predictive capability, achieving an AUC of 96.51%, indicating high classification accuracy in distinguishing electricity theft from non-theft cases. The Hosmer–Lemeshow test (p = 0.471) confirmed a good model fit, with no statistically significant deviations between the observed and predicted probabilities.

To enhance model stability, random bootstrapping sampling was applied across 100 resampled datasets, yielding an average AUC of 94.92%, with a standard deviation of 0.8%, reinforcing consistent classification performance. Additionally, 94% of the bootstrap samples fell within upper and lower confidence bounds, verifying the model’s robustness across varying conditions.

Further strengthening these results, a post-hoc power analysis validated the sample size (n = 385), confirming high statistical power (1 − β = 1.0000) and ensuring a negligible risk of Type II errors, meaning that true effects were accurately detected.

5.2. Addressing Potential Multicollinearity

Multicollinearity assessments identified high Variance Inflation Factor (VIF) values for X7 (meter experience, VIF = 7.47) and X19 (theft penalty opinion, VIF = 7.89), indicating strong predictor redundancy.

To refine feature selection, Lasso regression (cv = 5) was applied, ensuring optimal dimensionality reduction while preserving key explanatory variables. The strongest contributors to electricity theft prediction—X8 (financial stress, Lasso coefficient = 0.130) and X9 (tampering awareness, coefficient = 0.111)—were retained, reinforcing their significant impact on theft likelihood.

5.3. Significant Predictors and Their Implications

The key predictors with a significant influence on Y are discussed below.

5.3.1. Number of Appliances (X2) (p = 0.048)

The negative coefficient (−0.191) and odds ratio (0.826) suggest that households with more appliances are less likely to engage in electricity theft. This indicates that higher appliance usage might correlate with greater energy demand, leading to formalized billing relationships with energy providers rather than illicit connections. Olatunde et al. [41] found that energy-efficient appliances lower household electricity demand, reducing the incentive for illegal connections. Research from EBSCO Research Starters highlights that higher appliance usage does not necessarily lead to theft but rather encourages formal billing relationships [42]. A systematic review in Environment, Development and Sustainability suggests that subsidizing energy-efficient appliances can help low-income households transition to legal electricity use [43]. Governments can mitigate theft by subsidizing appliances, deploying smart meters, and educating consumers on efficient energy use.

5.3.2. Electricity Supply Quality (X4) (p < 0.001)

The strong negative coefficient (−0.391) and odds ratio (0.676) suggest that a better electricity supply reduces the likelihood of theft. Households that receive consistent and reliable electricity are less likely to engage in theft compared to those experiencing frequent blackouts and poor service quality. This highlights the need for investment in smart grids, improved distribution infrastructure, and proactive maintenance strategies. In Brazil, India, and South Africa, smart grid investments have significantly reduced electricity theft. A study by Northeast Group, LLC found that emerging markets deploying smart meters and grid automation could cut theft-related losses by billions [44].

5.3.3. Alternatives Used During Outages (X5)

The presence of backup power sources correlated with increased theft risk (p < 0.0001, OR = 7.308). This suggests that frequent outages drive consumers toward illegal connections. It highlights how greater awareness and access to legal alternatives can play a role in reducing unauthorized electricity connections. Governments can reduce reliance on stolen electricity by subsidizing renewable energy alternatives through tax credits, low-interest loans, and prepaid energy credits. Investing in solar-powered microgrids and smart meters enhances legal access while discouraging theft. Policies like net metering allow households to sell excess solar power, reducing financial strain. Public awareness campaigns and community-based renewable projects further promote legal electricity use [10].

5.3.4. Receipt of Electricity Bills (X6) (p = 0.072, OR = 0.329)

Households that receive bills regularly are less likely to resort to theft. This emphasizes the importance of transparent billing systems and improved meter installations. Utilities can implement digital invoicing for transparency and efficiency. Digital invoicing eliminates manual errors, ensures accurate, real-time billing for electricity consumption, and reduces the risk of billing manipulation and unauthorized adjustments.

5.3.5. Stress About Paying Bills (X8) (p < 0.001)

The strong negative coefficient (−1.781) and odds ratio (0.168) show that households experiencing financial distress are significantly more likely to engage in theft. High unemployment, income disparity, and energy costs force households into illegal connections. Solutions could be flexible payment plans (e.g., prepaid meters or pay-as-you-go electricity models) to ease financial strain, energy subsidies for low-income households, community solar projects to provide affordable, legal electricity, and public awareness campaigns to promote legal energy use.

5.3.6. Awareness of Tampering (X9)

The strong statistical significance of awareness suggests that communities knowledgeable about tampering techniques are at higher risk of theft. In areas with weak enforcement, electricity theft becomes socially accepted, often justified by economic hardship and lack of consequences. Community-wide participation and corrupt oversight further reinforce this behavior. Strengthening education and enforcement can reduce theft and improve electricity access, for example, public awareness campaigns highlighting the impact of theft on grid stability, stricter penalties and enforcement to deter illegal connections, tamper-proof smart meters to prevent unauthorized access, and subsidized energy programs offering affordable legal alternatives.

5.4. Lesser-Impact Predictors

While the logistic regression model identified several strong predictors, some variables demonstrated lower statistical significance but may still offer contextual insights into electricity theft behavior.

5.4.1. Perception of Electricity Costs (X15)

Though X15 (perceived cost of electricity) influenced theft likelihood, its significance varied across subcategories:

• High perceived costs (p = 0.087, OR = 0.205) showed a mild negative association, suggesting consumers who view electricity as expensive may be more cautious about illegal connections.
• Low- and medium-cost perceptions (p = 0.003 and p = 0.025, OR = 0.080 and 0.170, respectively) exhibited stronger negative associations, reinforcing the link between affordability and theft avoidance.

While not the strongest predictor, X15 provides useful insights for pricing strategies that could lower theft risks through tiered rates or affordability adjustments.

5.4.2. Difficulty Paying Bills (X18)

While financial hardship exhibited a significant impact through X8 (stress about bills), individual categories within X18 (difficulty paying bills) showed weaker predictive power:

• “Never” (p = 0.027, OR = 0.177) displayed mild protection against theft.
• “Always” (p = 0.504, OR = 2.721) had a high odds ratio but lacked statistical significance, likely due to the small sample representation.

This suggests that general financial stress (X8) is a more robust predictor than X18’s detailed payment difficulty categories, reinforcing the need for broad economic interventions rather than case-specific mitigation strategies.

5.4.3. Opinion on Theft Penalties (X19)

X19 exhibited high multicollinearity (VIF = 7.89), leading to weaker independent significance (p = 0.435, OR = 0.488). While opinions on penalties may shape theft behavior, other factors—such as awareness of tampering (X9) and billing transparency (X6)—appear more influential, indicating regulatory enforcement alone may not deter theft effectively.

5.4.4. Technical Ability to Manipulate Meters (X16)

Despite being conceptually relevant, X16 did not exhibit a strong statistical impact (p-values between 0.595 and 0.998 across skill levels). This may be due to

• Limited direct self-assessment accuracy, where respondents may underreport technical skills;
• High dependency on awareness (X9), suggesting that knowing about tampering techniques (X9) is more impactful than the ability to execute them (X16).

5.5. Testing Interaction Effects Between Predictors

While individual predictors significantly influence electricity theft likelihood, interactions between them may reveal hidden dependencies that provide deeper insights into consumer behavior. Assessing interaction effects helps determine whether combined variables amplify or diminish theft probability beyond their individual contributions.

5.5.1. Key Interaction Terms Considered

Based on feature importance rankings from Lasso regression and Type II analysis, the following interactions were tested for statistical significance:

• Financial stress (X8) × perception of electricity costs (X15): Economic hardship combined with high price perception may increase theft risk as affordability concerns worsen.
• Tampering awareness (X9) × billing transparency (X6): Consumers aware of tampering techniques may exhibit lower theft rates if billing is transparent, suggesting that trust in the utility system counteracts fraud incentives.
• Homeownership status (X10) × payment type (X11): Tenants using postpaid billing may show higher theft rates, as temporary residence reduces accountability for long-term billing obligations.

5.5.2. Findings from Interaction Effects Testing

Logistic regression models incorporating interaction terms showed the following:

• X8 × X15 (financial stress × electricity cost perception): Significant interaction (p < 0.003, OR = 1.341), indicating economic hardship combined with high-cost perception, substantially increases theft likelihood.
• X9 × X6 (tampering awareness × billing transparency): The interaction effect was not statistically significant (p = 0.217), implying that billing transparency alone may not counteract fraud behavior in high-awareness groups.
• X10 × X11 (homeownership status × payment type): There was moderate significance (p = 0.045, OR = 1.231), suggesting tenants using postpaid billing face higher theft risks.

5.5.3. Implications for Theft Mitigation Policies

Our findings indicate the following:

• Dynamic pricing strategies could help lower-income households manage payments effectively, reducing theft likelihood (X8 × X15).
• Billing transparency alone is insufficient to deter tampering, requiring stronger fraud monitoring systems (X9 × X6).
• Tenant-focused interventions, such as incentives for prepaid adoption, could lower theft risk among non-owner households (X10 × X11).

5.6. Policy Recommendations and Practical Implementation

Given the findings, multi-pronged interventions must integrate technical, behavioral, and regulatory measures.

5.6.1. Infrastructure Investment

To enhance grid stability and limit opportunities for illegal connections, the following is recommended:

• Smart grid modernization to enhance reliability and reduce theft (X4: supply quality).
• Tamper-resistant meter installations to prevent unauthorized access (X9: tampering awareness).
• Renewable energy projects to mitigate reliance on illegal connections (X5: backup alternatives).

5.6.2. Behavioral Interventions

Electricity theft is often driven by financial hardship and lack of awareness. Therefore, the following suggestions are to be considered:

• Consumer awareness programs to discourage tampering (X9: awareness).
• Flexible payment structures for economically vulnerable households (X8: financial stress).
• Incentivized transition programs for informal electricity users.

5.6.3. Regulatory Enforcement

To strengthen compliance and deter illegal connections, the following should be considered:

• Strengthening Legal Frameworks: Stricter penalties should be introduced for electricity theft while ensuring fair enforcement (X9: awareness of tampering).
• Enhanced Billing Transparency: Utility companies should improve their billing accuracy and accessibility to reduce billing disputes (X6: receipt of electricity bills).
• Fraud Detection Strategies: Investment should be made into AI-driven monitoring techniques to detect theft patterns and prevent illegal meter tampering.

5.7. Alternative and Future Research

5.7.1. Improving Threshold Optimization for Classification

The default classification threshold (0.50) used in logistic regression may not be optimal for electricity theft detection. ROC curve analysis reveals variability across bootstrap samples, suggesting that dynamic threshold adjustments could enhance model sensitivity and specificity.

To refine classification precision, Youden’s Index could be employed to determine an ideal cutoff point, balancing false positives and false negatives more effectively.

5.7.2. Exploring Alternative Statistical Techniques

While logistic regression provided interpretable, policy-relevant insights, future studies could benefit from the following:

• Accuracy: The model achieved a classification accuracy of 89.35%, with high specificity (91.18%) for identifying “No” and sensitivity (86.71%) for identifying Y, using Random Forest and XGBoost, and capturing the non-linear relationships in theft prediction.
• Bayesian hierarchical modeling: This would help refine the probabilistic estimations of theft likelihood.
• Recurrent Neural Networks (RNNs): RNNs could be used to analyze long-term behavioral patterns in electricity theft.

5.7.3. Limitations and Future Research Directions

Despite the model’s strong predictive accuracy, several limitations should be considered when interpreting the findings:

• Residual sampling biases may persist despite the structured random sampling design. For instance, there may be underrepresentation of peripheral or densely populated informal settlements where unmetered or illegal access to electricity is more common. Additionally, the dataset may have excluded off-grid households or informal users lacking formal billing, and response bias may be present due to the sensitive nature of theft-related questions.
• Self-reported survey data may be subject to social desirability or recall bias, particularly regarding sensitive behaviors like meter tampering. These limitations highlight the need for cross-regional validation using mixed-method approaches or objective consumption data when available.
• Some predictors—including X3, X10, X13, and X18—exhibited weaker statistical significance, suggesting the need for stepwise regression refinements or interaction modeling in future studies.
• Finally, electricity theft is inherently dynamic. Future research should integrate time-series modeling to capture behavioral changes and theft trends over time, particularly in response to infrastructure upgrades, policy reforms, or energy access interventions.

To improve generalizability and robustness, we recommend that future studies adopt oversampling strategies, prioritize the inclusion of targeted subpopulations, and leverage administrative data to supplement survey-based insights.

6. Conclusions

This study provides a comprehensive quantitative and behavioral analysis of electricity theft in Kinshasa, DRC, emphasizing its technical and socioeconomic drivers. The methodology incorporated random bootstrapping sampling to enhance model reliability, while power analysis validated sample adequacy, ensuring statistical rigor in theft prediction. The logistic regression model, refined via Lasso regression feature selection, identified electricity supply quality (X4), financial stress (X8), tampering awareness (X9), and billing transparency (X6) as significant theft predictors. Households facing economic constraints and unreliable service demonstrated higher theft likelihood, reinforcing the need for flexible payment systems and consumer protection measures.

To combat electricity theft, the following strategic interventions must be incorporated:

• Infrastructure upgrades, including grid modernization and smart meter deployment.
• Behavioral-focused policies, such as public awareness programs and community engagement initiatives.
• Regulatory enforcement, including stricter penalties and fraud detection mechanisms using AI-driven monitoring.

Future research should further explore the interaction effects between socioeconomic factors, optimize classification thresholds, and assess alternative statistical techniques, including Bayesian modeling and ensemble learning approaches. By integrating technical advancements with behavioral insights, utility providers can develop more effective anti-theft strategies, ensuring equitable energy distribution and sustainable economic progress.