The analysis uses Secondary data from the 2021 FinAccess Household Survey, conducted jointly by the Central Bank of Kenya (CBK), the Kenya National Bureau of Statistics (KNBS) and Financial Sector Deepening (FSD) Kenya, covering 22,025 households in all 47 counties [2]. The survey instrument collects information from one adult respondent per household, who answers on behalf of the household. Financial outcome variables reflect the household’s financial situation as reported by the responding adult. Household-level sampling weights (hhweights) provided in the dataset are applied throughout all descriptive and regression analyses to account for differential household selection probabilities. Strata and cluster identifiers were not incorporated into the variance estimation; accordingly, standard errors may be modestly underestimated relative to design-based estimates, which is an acknowledged limitation.

The digital loan access variable is derived from the FinAccess module on digital financial services, capturing use of any mobile-based loan product, including mobile banking credit (e.g., M-Shwari, KCB M-Pesa) and mobile application loans (e.g., Tala, Branch, Fuliza) within a 12-month recall period preceding the survey. Eligible respondents were adults aged 18 and above residing in Kenya with access to a mobile phone. Households were classified into three mutually exclusive categories: Currentlyuse (N = 367), Usedtouse (N = 387), and Neverused (N = 21,271). Together, current and former users constitute 754 identified digital loan households, representing 3.42% of the full sample.

The default risk analysis (Section 3.3) uses a broader analytical sample of 11,550 households with complete risk indicator data. Critically, this sample is not restricted to the 754 digital-loan households. The variable defaultloan is therefore interpreted as default on any loan product throughout; all headings and interpretations reflect this distinction explicitly.

All cleaning and analysis were conducted in Python. The following processing decisions are documented for reproducibility: 1) categorical variables were recoded into dummy variables, with string categories standardized to remove whitespace (affecting the “technical training” educational category); 2) monthly income was winsorized at the 1st and 99th percentiles to reduce outlier influence; 3) income brackets for the interaction model were defined as Low Income (≤KES 5000), Middle Income (KES 5001 - 20,000), and High Income (>KES 20,000), consistent with the sample quartile structure; 4) loan purpose categories were retained in original survey form for descriptive analysis; and 5) listwise deletion was applied in all regression models, yielding N = 2102 for the financial recovery model and N = 11,550 for the default risk analysis. The difference reflects distinct variable sets and item non-response patterns across survey modules.

Financial Recovery (Overcome_binary). The primary outcome is Overcome_binary, derived from the FinAccess household survey item: “Ihavebeenabletoovercomefinancialproblems.” Five response options were recoded as follows: Overcome_binary = 1 if the responding adult selected “Agree” (interpreted as the household having overcome financial problems); Overcome_binary = 0 for all other responses (“Disagree”, “Neither agree nor disagree”, “Don’t know”, “Refused to answer”). This conservative recoding treats uncertainty, disagreement, and non-response as a single reference category. Households with a missing value on this item were excluded via listwise deletion.

Default Status (defaultloan). Default status is a binary household-level indicator: 1 = the household reported missing at least one scheduled loan repayment on any loan product; 0 = did not default.

Credit History Variables. NegativeCRBlisting: binary (1 = household negatively listed with a Credit Reference Bureau; 0 = otherwise). Badcredithistory: self-reported binary (1 = poor repayment history on any prior loan; 0 = otherwise). Both variables may be outcomes of prior digital loan borrowing and are therefore treated as potentially endogenous (see Section 3.4).

The primary analytical model is logistic regression that estimates the probability of household recovery:

where Y i is Overcome_binary for household i ; X loan , i is a vector of the characteristics of digital loans; X demo , i is a vector of demographic characteristics of the head of the household; and X credit , i is a vector of credit history variables. To address the potential endogeneity of credit history variables, which may result from prior borrowing, two models are estimated: Model 1 (Baseline) excludes bad credit history and CRB listing; Model 2 (Adjusted) adds these variables and is interpreted as an adjusted association rather than a direct loan effect. Coefficients are reported as odds ratios.

The 22,025 surveyed households include 59% male and 41% female respondents. The age distribution peaks in the 25 - 35 years range. Monthly income is heavily right-skewed (modal income: KES 5000), reflecting Kenya’s predominantly informal economy: casual work (27.6%), family support (25.2%), and farming (21.1%) are the three largest income sources, while formal employment accounts for only 9.3%. The overwhelming majority of households (96.58%) had never used digital loans; 1.76% used to use them, and 1.67% currently use them.

Table 1 presents the percentage cross-tabulation of digital loan access against household financial recovery. A clear gradient is evident: households that previously used digital loans report the highest recovery rate (37.24%), followed by current users (34.25%), and non-users (22.79%). The chi-square test confirms a statistically significant association ( χ 8 2 = 108.05 , p < 0.001). However, this bivariate relationship does not control for confounding factors addressed in Section 3.4.

Table 1. Percentage cross-tabulation: digital loan access vs. overcoming financial problems.

Income differs significantly across access groups: current and past digital loan users report a median income of KES 10,000 and KES 9000, respectively, compared to KES 4000 for non-users (ANOVA:F = 156.55, p < 0.001), confirming that higher-income households are more likely to engage with digital credit.

Financial recovery rates increase monotonically with educational attainment: from 13.51% among those with no formal education to 44.99% among university graduates (Table 2). Similarly, households dependent on investment income (53.3% agree) and rental income (47.0% agree) report higher recovery rates than those relying on casual work (19.8%) or family support (18.3%).

Table 2. Financial recovery by educational level (percentage).

The default risk analysis is conducted on a broader analytical sample of 11,550 households with complete risk indicator data. The defaultloan variable captures default on any loan product and is not restricted to digital loans, since restricting to the 754 identified digital-loan households would yield too small a sample for stable multivariate estimation.

Monthly income is a significant predictor of default: households that defaulted have a median income of KES 4500 compared to KES 5000 for non-defaulters, and an independent samples t-test confirms a statistically significant difference in mean income ( t = − 9.66 , p < 0.001). Box and violin plots confirm that non-defaulters have a broader upper income distribution.

Default rates decline with educational attainment: from 71.76% among households with no formal education to 40.77% among university-educated households ( χ 6 2 = 293.73 , p < 0.001). Default rates also vary significantly by income sources ( χ 8 2 = 158.10 , p < 0.001): farming (60.62%), NGO or government income (59.14%), and casual work (58.50%) record the highest rates, while investment income (33.33%), pension (37.68%), and formal employment (41.60%) show the lowest. These patterns are consistent with the income stability hypothesis: predictable income streams enable consistent loan servicing [7].

The correlation heatmap of default risk indicators confirms that bad credit history and negative CRB listings are the most relevant linear predictors of default (r = 0.12 each), and are themselves moderately correlated (r = 0.21), reflecting the pathway through which a bad credit history leads to a negative CRB listing. Absence of a payslip (r = 0.00) and guarantor (r = 0.01) shows a negligible association with default (Table 3).

Table 3. Quartiles and median of monthly income by default status (KES).

Two logistic regression models are estimated.

Model 1 (Baseline) excludes bad credit history and CRB listing to avoid conditioning on potentially endogenous post-borrowing outcomes.

Model 2 (Adjusted), presented in Table 4, adds these credit history variables and should be interpreted as an adjusted association rather than a direct loan effect. The difference in digital loan access coefficients between the two models indicates the extent to which credit damage mediates the relationship between digital borrowing and financial recovery.

Table 4. Logistic regression: Predictors of household financial recovery (Sorted by odds ratio).

The adjusted logistic regression model (Table 4) was estimated on 2102 complete household observations with 24 predictors, converging in 6 iterations (Pseudo R 2 = 0.069 , LLR p-value = 1.269 × 10⁻²⁴).

Educational attainment is the dominant predictor of financial recovery. Even primary school completers are 4.3 times more likely to report overcoming financial problems relative to those with no formal education (OR = 4.318, p < 0.001), rising to 5.3 times for university graduates (OR = 5.319, p < 0.001). Formal employment (OR = 1.797, p < 0.001) and self-employment (OR = 1.517, p = 0.004) significantly improve recovery odds. Bad credit history reduces odds by 37.1% (OR = 0.629, p = 0.005) and negative CRB listing by 48.1% (OR = 0.519, p = 0.001).

Crucially, digital loan access is not a statistically significant predictor of financial recovery in the adjusted model (Never used: OR = 0.849, p = 0.440; Used to use: OR = 1.535, p = 0.129). The marginal effect of digital loan access is approximately −0.031 for never users (p = 0.440) and +0.080 for past users (p = 0.128), with confidence intervals crossing zero in both cases. This indicates that the bivariate recovery advantage among digital loan users is attributable to their more favourable socioeconomic profiles rather than to credit access itself, consistent with [1] and [5].

The interaction model (Table 5), estimated on N = 20,546 with household weights, reveals important heterogeneity. In isolation, current users (OR = 1.963, p < 0.001) and past users (OR = 2.057, p < 0.001) show significantly higher recovery odds. However, these benefits are significantly attenuated among middle-income households: the interaction of current use with middle income reduces recovery probability by 19.4 percentage points (dy/dx = −0.194, p < 0.001), and past use with middle income by 12.0 percentage points (dy/dx = −0.120, p = 0.021), suggesting that middle-income households face repayment pressures that offset the liquidity benefits of digital credit. No significant attenuation is observed for high-income households.

Table 5. Interaction model: Digital loan access × income bracket.