Écrire un commentaire

Summary -

This study examines digital financial service adoption among 663 active microfinance borrowers in Ranchi district, Jharkhand, using a TAM/UTAUT-based survey administered through interviewer-assisted CAPI in local languages. Its central empirical observation is a 72.8-percentage-point gap between digital loan repayment (84.9%) and fund-transfer use (12.1%), which the author interprets not as evidence of broad digital adoption but as a Compliance-Autonomy Gap the difference between completing institutionally required digital transactions and independently using digital finance for self-directed purposes. A second descriptive finding, termed Trust Asymmetry, is that borrowers report high trust in the microfinance institution (M = 4.37) alongside considerably lower trust in the technology itself (M = 3.38). An OLS model reports perceived usefulness, social influence, and facilitating conditions as positively associated with behavioural intention, and perceived risk negatively associated, though the author explicitly cautions that severe multicollinearity and heteroscedasticity limit coefficient-level interpretation. The paper closes with a conceptual staged model, the Trust-Enabled Phygital Adoption Framework (TEPAF), positioning human assistance as a bridge from institutional compliance toward financial autonomy.

The contribution is primarily conceptual and measurement-oriented rather than inferential. Its most useful move is reframing digital financial inclusion as a question of who performs the transaction and how independently, rather than transaction volume. A distinction with real practical consequences for how MFIs and regulators report inclusion metrics. Separating institutional trust from technological trust as distinct constructs, rather than assuming trust transfers from the former to the latter, is also a worthwhile refinement of the trust-adoption literature in a relationship-based lending context. The author is commendably candid about the statistical limitations throughout, which is rarer than it should be.

Major issues

1. The reported model fit statistics are mutually inconsistent, and the paper’s headline R² appears to be incorrect. The manuscript reports R² = .985, adjusted R² = .984, and F = 768.6 with n = 663 and 8 predictors. These cannot all be true. An R² of .985 with k = 8 and df = 654 implies F should be 5,368, not 768.6. F = (R²/k)/((1−R²)/(n−k−1)). So, F should be 5,368. Conversely, F = 768.6 implies R² ~ .904. Reconstructing the standard errors from the reported SDs and VIFs also yields values consistent with R² ~ .90 and not with .985 (e.g., for facilitating conditions, R² = .904 implies SE ~ .107 against the .112 reported; R² = .985 implies SE ~ .042). The reported R² should be recomputed and corrected, and the “very high in-sample fit” discussion revised accordingly.The implied R² ~.90 is anomalously high for a behavioural intention model and warrants investigation rather than interpretation.

2. VIFs between 35 and 89 are not merely “severe” - at this level the regression is uninterpretable as specified. The author acknowledges this, which is to their credit, but the manuscript still reports hypothesis “Supported / Not supported” decisions from those coefficients and carries them into the discussion and conclusion. With a VIF of 89 for perceived risk and 79 for trust in technology, the individual coefficients (including the counterintuitive negative TTEC estimate) carry essentially no information about unique effects. The recommendation would be to drop the hypothesis-decision table in its current form, or to re-estimate via PLS-SEM, penalised regression, or a reduced-dimension specification, and report hypothesis outcomes only from that model.

3. The central conceptual contribution is never actually tested. The Compliance-Autonomy Gap is the paper’s headline idea, but the regression models behavioural intention, not the gap. Nothing in the inferential analysis explains why some borrowers transact autonomously and others do not. The obvious analysis - modelling fund-transfer use (or a use-breadth count) as the outcome, with the same predictors - is available in these data and is not reported. As it stands, the paper’s strongest concept rests on a descriptive comparison of two single survey items.

4. The cluster sampling design is not reflected in the inference. Respondents were recruited at JLG centre meetings under a multi-stage cluster design, so observations within a group are almost certainly correlated (shared field officer, shared meeting, shared local infrastructure, peer effects - the very social-influence mechanism the paper theorises). Standard errors that ignore this clustering will be understated. Cluster-robust standard errors at the JLG or centre level are needed, and the number of clusters should be reported. This compounds the separately acknowledged heteroscedasticity issue.

5. Ethics documentation is absent for a human-subjects study involving a vulnerable population. The manuscript states that IRB approval status, protocol code, and the informed consent procedure “were not specified in the source manuscript” and must be finalised before submission. For face-to-face interviews with low-income borrowers whose participation is linked to an ongoing credit relationship, ethics approval and a consent statement are not administrative afterthoughts - they should be documented in the preprint itself, along with how voluntariness was protected given that recruitment occurred at institutional group meetings.

Minor issues

1. The digital repayment measure may not capture what the paper needs it to capture. In JLG microfinance, repayment is frequently collected at the group meeting with the field officer initiating or completing the transaction. If some of the 84.9% reflects transactions performed for the borrower rather than by them, this strengthens the paper’s argument but weakens the measure. The author flags this for future work; it is worth stating more directly as a limitation of the current figure.

2. Low fund-transfer use may reflect low need rather than low capability. A borrower with no remittance requirement and no counterparty has no occasion to transfer funds. Without an item on whether the borrower wanted to transfer and could not, the 12.1% figure conflates capability and demand.

3. No confidence intervals are reported for any coefficient, and no correction is applied across eight hypothesis tests. Two results sit at p = .043, which is fragile under any multiplicity adjustment.

4. Several demographic distributions are strikingly round (60/40 gender, 40/30/30 education, 50/40/10 occupation), with n = 265 recurring across three unrelated categories. This is likely coincidental, but given the other reporting inconsistencies, a verification pass against the raw dataset would be reassuring.

5. Three minor citation and consistency items: a systematic review of low-income populations is described in the introduction without an attached citation; Gefen et al. (2003), Hasan et al. (2024), Rizvi et al. (2024), Sant'Anna & Figueiredo (2024), Sreelakshmi & Prathap (2024), Tiwari et al. (2021), Ha et al. (2025), Hilale & Chakor (2024) and Amnas et al. (2024) appear in the reference list without corresponding in-text citations.

6. The abstract reports the OLS coefficients as substantive findings before the multicollinearity caveat appears. Given how severely those estimates are qualified in the body, the abstract would be more accurate leading with the descriptive gap and trust findings and characterising the regression results as provisional.

Competing interests

The author declares that they have no competing interests.

Use of Artificial Intelligence (AI)

The author declares that they did not use generative AI to come up with new ideas for their review.