This is the final part of a 3-part series on the recently released NCAER-MFIN report, “Assessing the Effectiveness of Regulated Small Borrowing in India” (March 2026). In the previous blog, we looked at claims of ancillary benefits of microfinance documented in the report – namely displacement of informal credit, awareness of insurance products, and digital financial inclusion. In this blog, we examine issues related to measurement and empirical strategy.
In the microfinance sector, quantitative survey data often serves as the primary benchmark for evaluating institutional performance and social impact. High repayment rates, increased loan uptake from formal sources, or self-reported improvements in household consumption levels are frequently interpreted by investors and policymakers as indicators of program success. However, errors in the data collection and analysis, can lead to situations where the true on-ground realities get obfuscated. We have already discussed the problems with this report that stem from the manner and timeline of data collection in Blogs 1 and 2. In this blog, we explore how the survey instrument design, indices constructed from the collected data, and the empirical methodology adopted in the report fare under a critical assessment.
Issues in Survey Instrument Design
The report attempts to capture the assets and liabilities of borrowers’ households through a self-reported survey. While the survey captures detailed information on household assets and loan characteristics, issues remain in the framing of certain survey questions, as well as regarding the methods used to aggregate this data into easily interpretable metrics, both of which are elaborated below.
- Limitations in Quesitonnaire Design
There exist some critical design flaws in some of the survey questions designed to capture patterns of loan usage and access to finance. These weaknesses likely carry over into the collected data, potentially leading to inaccurate findings and limiting the reliability of the conclusions drawn.
For example, Questions 3.34 and 3.35, which examine loan usage and repayment sources, explicitly allow multiple responses. Essentially, it acknowledges that microfinance borrowers often use the same loan for multiple purposes and rely on several repayment mechanisms. However, in presenting the findings, the insights are reduced to a single reported purpose or repayment source. This oversimplifies borrower behaviour and does not reflect the complex financial realities of households, where loans may be used for different purpose than stated or are repaid through multiple channels.
Similarly, Question 4.5, asks “Do you think that the interest rate and charges on the loans provided by the parent RE are competitive with market rates?” This is a fair line of inquiry. However, the responses allowed for the question are (1) “Does not matter, as interest rates from informal sources are exorbitant,” (2) “Yes,” (3) “No,” and (99) “Don’t know”. The options presented amalgamate different ideas together and is likely to confuse the respondent. To illustrate, Option (1) introduces a comparison with informal credit markets rather than directly measuring whether respondents perceive the lender’s rates as competitive. This can bias responses and make it difficult to draw clear conclusions about whether the rates were competitive. Essentially, the first option conflates categories. Though the question is about one entity’s interest rate, it contrasts “informal rates” to that of “formal rates”, creating a false parallel. Essentially the comparison is no longer between “an entity” and “the market”, but it becomes between “two parts of the credit market”. Further, the use of the phrase “exorbitant” will likely lead to priming of the respondents, as the question appears to be loaded. Additionally, the differences in options may also not be clear for the respondents. Choosing (1) “Does not matter…” can simply reflect no real comparison or lack of alternatives, which in practice is also close to (99) “Don’t know”. Incidentally, proportion of respondents selecting these options was also broadly similar (29.6% and 30.2% respectively). This makes it difficult to draw meaningful conclusions about perceptions of pricing or access to credit.
- Limitations of Collected Data in Constructing an Asset-Based Wealth Index
Another key limitation is around the measurement of household wealth. In the report, it is done using an asset-based index of ten indicators[1] which includes land, housing amenities, durable goods among other things, with households ranked into wealth quartiles. While this is a practical alternative when data is hard to come by or unreliable, it has important limitations that the report does not appear to appreciate or acknowledge.
First, the index assumes equal value across assets (and amenities, like water supply, separate kitchen, etc.) by coding ownership in binary terms (owned/not owned). This overlooks major differences in economic value and productive utility of an asset since a television is treated similarly to productive assets like livestock, which can distort any reliable quantification of the economic and productive value of assets at the household’s disposal. For example, a person may own land, and another may just own a bicycle, but according to this index, their score will be identical, which is not in tandem with either their economic value, or their productive capacity, or their monetary resale value.
Second, the wealth indicator may be endogenous to the borrowing process itself. Because many of the assets included in the index can be acquired through microfinance loans or other forms of credit, the measure may partially capture the effects of indebtedness rather than an independent underlying level of household wealth. This means that borrowing can mechanically increase a household’s asset score even where there has been no corresponding improvement in its long-term economic position or repayment capacity. This creates an important problem of interpretation.
Finally, asset-based measures conflate wealth with financial resilience by failing to distinguish illiquidity from insolvency. Households may hold assets and appear wealthy on paper while experiencing severe cash flow stress, making this an incomplete proxy for borrower well-being.
Issues of Methodological Oversimplification
The methodology outlined in Chapter 6, “Behavioural Analysis of Microfinance Borrowers,” relies on constructed variables and classifications such as the Product Literacy Index (PLI), Debt-to-Income (DTI) categories, and the classification of “productive use.” Since the validity of such constructs are not self-evident, the burden lies with the authors to establish their theoretical basis or provide rationale for the same. However, the chapter does not provide adequate justification for the construction or use of these measures, nor does it cite prior studies employing comparable constructs.
- Concerns in Variable Construction and Threshold Selection
The behavioural analytical framework, presented in the report, is presented without discussing the motivation or reasoning behind the choice of some of the constructed variables. In constructing Product Literacy Index (PLI)[2], equal weight is applied to all five components without explaining why each indicator should be treated as equally important. The implicit assumption is that knowledge of interest rates, loan tenure, credit bureau reporting, and other product features contribute identically to overall awareness of financial products. Such an assumption overlooks the possibility that some dimensions may be more salient than others in shaping borrower orientation towards financial products.
This problem is compounded by the subsequent reduction of this constructed index into three categorical groups for econometric analysis. The report assigns the category of “low” literacy if the index value is between 0 and 2, “medium” when the value is 3 and “high” for the values between 4 and 5. The report does not justify why these cutoffs are chosen. Further, given that the econometric specification in the report employs an ordered Logit Model, it is unclear why the original ordered 0–5 score of the index was not used directly instead of collapsing it into three arbitrary categories. Alternatively, if the score is treated as a summary measure of product literacy, simpler regression methods such as Ordinary Least Squares (OLS) could have provided similar interpretability. The absence of any explicit justification for these modelling choices is puzzling.
Similarly, the analysis of repayment behaviour across different levels of indebtedness relies on Debt-to-Income (DTI) categories of ≤15%, 15–25%, and >25%. However, the rationale for these thresholds used in categorisation is not clearly explained and appears to be randomly assigned. Without an explanation for why these specific cut-offs meaningfully distinguish different levels of indebtedness, the corresponding analysis risks imposing artificial groupings to explain the relationship between indebtedness and repayment behaviour.
Comparable issues arise in the treatment of loan utilisation patterns. The binary classification of borrowing into “productive” and “non-productive” rests on strong normative assumptions. For example, expenditures on healthcare, household consumption, or social obligations are classified as non-productive, despite the fact that such spending may be necessary for income preservation and managing economic shocks[3], both of which can influence future repayment outcomes. Hence, this binary categorisation inadvertently risks mischaracterising borrower behaviour and drawing pre-determined conclusions about responsible credit use. Had the classification been based on a higher theorical standard, the composition of clusters would have changed, consequently, perhaps the null hypothesis (that “there is no association between the productive use of loans and repayment”) would have been validated, or there would be a stronger case for rejecting the same. Either way, the resulting conclusions would rest on a more firmer footing.
- Need For a More Rigorous Empirical Strategy
The empirical strategy adopted in the report raises concerns regarding the conclusions drawn from specific hypotheses. While, every statistical exercise embeds a set of limitations, the problem with the report is that these limitations are seldom discussed and the statistical findings were subsequently presented as key findings, leading to a possibility of a biased narrative. We illustrate below some prominent hypotheses tested in the report that fall short, in terms of both the statistical methodology adopted as well as the robustness of findings.
For instance, the hypothesis, “Do less than 50% of borrowers consider interest rate as a factor in choosing a lender?” is examined using a simple one-sample proportion test, which substantially limits the analytical scope. Here, the choice of ‘50%’ seems to function more as a convenient reference point rather than one grounded in existing studies or having any policy relevance. More importantly, it overlooks the potential for a more robust multivariate analysis. A more informative approach would have been to model the probability that a borrower considers interest rates in lender choice as a function of observable characteristics such as income, education, indebtedness, geography, or prior borrowing experience using a Probit or Logit specification. Admittedly, given that we haven’t studied the data, it is possible that for some reasons the aforementioned approaches couldn’t have been adopted. If so, then the authors should have laid out the reasons they chose a specific test over other potential candidates.
Similarly, the examination of the hypothesis, “Are borrowers who value interest rates more likely to choose a particular type of lending institution?” relies on a Multinomial logit model, but the empirical specification appears to be oversimplified. The specification does not control for relevant borrower or lender characteristics such as loan size, tenure, income levels, regional differences, or debt burden, all of which could plausibly shape lender choice. Additionally, the interpretation of these models is limited to reporting coefficients without presenting marginal effects[4], which are generally more meaningful for understanding the magnitude of relationships.
To summarise, it remains unclear why the report does not make fuller use of available data to incorporate relevant control variables and improve the model specification. Without accounting for potentially confounding factors, the robustness of the behavioural conclusions drawn from these empirical exercises remains uncertain.
Conclusion
To conclude, while the report aims to make a valuable empirical contribution to our understanding the contemporary microfinance landscape, its methodological limitations raise concerns about the robustness of its findings. From the design of survey instruments to the construction of composite variables to the selection of appropriate statistical techniques, all could have benefited from a more transparent discussion focusing on the rationale and objectives of the report.
———————————-
Footnotes
[1] Wealth index is an asset/amenity-based measure built from 10 variables that show sufficient variation across households – agricultural land, separate kitchen, water supply, bicycle, scooter/motorcycle, sewing machine, cooler, refrigerator, TV, and livestock
[2] Product Literacy Index (PLI) is constructed from five indicators – knowledge of loan interest, loan tenure, processing fee, insurance benefits linked to loan and credit bureau role.
[3] Schicks, J. (2010). Microfinance Over-Indebtedness: Understanding its drivers and challenging the common myths. CEB Working Paper No. 10/048. Université Libre de Bruxelles
[4] The coefficients multinomial logit models are expressed in log-odds relative to a reference category and do not directly convey the magnitude of changes in outcome probabilities.
———————————-
Part 1 of the series is available here
Part 2 of the series is available here
