How to Tell What the Rule Changed
The Colombian study is not a simple before-and-after count of consumer loans. It asks whether the January 2023 maturity-specific provisioning rule changed the terms of newly originated covered loans once the ordinary differences between short and long loans are taken seriously. Understanding the data and comparison strategy lets readers judge which conclusions are supported, rather than treating a regression result as a black box.
What the data can see
The study uses Format 341, an administrative record maintained by Colombia’s Financial Superintendence. Its unit is a debtor–institution observation with detailed information about the loan and the lender. The records include the loan amount, interest rate, provisions, origination and maturity dates, collateral value and type, days past due, and the inputs used to estimate expected loss: probability of default, loss given default, and exposure at default. They also contain borrower characteristics and recent credit-history indicators, as well as institution-level measures such as size, portfolio quality, capital adequacy, return on assets, and structural liquidity.
That detail matters because the policy changes a lender’s incentive at origination. The study is not trying to infer a loan’s future performance from a few aggregate banking ratios. It observes the terms attached to the new loan itself and the characteristics of the institution that granted it. This makes it possible to ask whether a long-maturity loan became smaller, more expensive, or more heavily collateralized after the rule, while also asking whether the lender built a larger provision buffer.
The final cleaned dataset covers 2022 through 2024 and contains 8,120,390 newly originated consumer loans. “Newly originated” is a crucial limitation as well as a strength. The rule applied to new covered loans, so this sample is aligned with the intervention. But the data do not follow every loan through its full life. They cannot by themselves show whether a borrower eventually defaulted, whether the household’s repayment burden became sustainable, or whether a lender’s long-run profitability changed.
The study narrows the sample to products for which the maturity-specific rule is relevant. Credit cards and revolving credit are excluded. Payroll loans to pensioners are also excluded, and records classified as “Other” that cannot be reliably identified within the relevant consumer-loan categories are removed. The authors also address a technical feature of the administrative record: when a debtor has several obligations in the same segment, rating, collateral class, and institution, the system can consolidate them. The analysis retains unique debtor–institution observations in order to avoid treating a combined record as if its maturity and collateral described every underlying loan.
This cleaning choice protects the meaning of the outcomes. If one row combined several loans but displayed the maturity and collateral of only the largest one, treating that row as a single clean observation could attach the wrong characteristics to the reported amount or provision. The study therefore gives up some observations in exchange for a clearer link between the measured contract and the maturity tier assigned to it. The resulting sample is still very large, but its size does not erase the need to understand what each observation represents.
The policy creates two treatment groups. One contains loans with maturities above 72 months and up to 108 months. The other contains loans above 108 months. Both are compared with a common control group of loans with maturities of 72 months or less. This is not a comparison between identical products by construction. It is a way of organizing the maturity tiers that the rule treats differently.
The raw data show why organization alone is not enough. Longer loans tend to have higher average loan amounts, while shorter loans tend to have higher interest rates. The groups also differ in sub-segment: shorter loans are more concentrated in personal lending, while longer loans include larger shares of vehicle and payroll-deducted loans. Collateral patterns differ, as do probability of default, loss given default, loan-to-value ratios, economic sectors, and the institutions that originate the loans. The longest-maturity group is especially different from the short-loan control group.
Those patterns are economically plausible even without a regulatory change. A vehicle loan may have a different maturity and collateral structure from an unsecured personal loan. A lender that specializes in payroll-deducted credit may serve a different borrower population from one that specializes in personal loans. If the longer group has larger balances before the reform, a raw comparison after the reform could falsely attribute that difference to higher provisioning. If its interest rates are lower in the unmatched data, the comparison could also misread product composition as a pricing response.
Consider the two questions hidden inside a raw before-and-after comparison. First, did long loans already have different terms because they were different products or served different borrowers? Second, did those terms change more than they would have changed absent the new provision factor? The first question concerns selection into a maturity group; the second concerns the policy effect. A large difference in loan amount after January 2023 could answer neither question unless the pre-existing composition and the common movement of credit conditions are addressed.
The study therefore treats the data as a problem of comparability, not merely a large sample. Millions of observations help with precision, but they do not automatically remove selection bias. The relevant question is whether loans being compared are similar enough on the characteristics that influence both maturity choice and the contract terms under study.
From raw difference to bounded inference
The first safeguard is exact matching. Instead of comparing every treated loan with every short loan, the study retains treated and control observations that share the same specified observable characteristics. The matching includes the reporting quarter, the financial institution, the loan sub-segment, collateral type, credit rating, probability of default, loss given default, debtor type, and economic sector. Matching within the same institution is particularly important: it holds constant many lender-specific practices, such as risk appetite, credit-scoring systems, and business model, that would otherwise be mixed into the maturity comparison.
The study uses several matching specifications. A baseline specification matches on the main time, product, risk, borrower, and institution characteristics. More demanding versions also use percentile categories for continuous characteristics such as collateral value, loan amount, and interest rate. The most comprehensive specification adds recent borrower credit history, including prior relationships with the financial system, previous delinquency, and whether the borrower was recurring at the same institution. These stricter comparisons discard observations for which no sufficiently similar counterpart exists. That reduces the sample but improves balance on the selected variables.
There is an important discipline in this process: the study does not match on the outcome it is trying to estimate. If the outcome is the interest rate, the rate itself cannot be used to force treatment and control loans to look alike. If the outcome is loan amount or coverage, the matching variables must avoid mechanically building the outcome into the comparison. For LTV, which is constructed from loan amount and collateral, those components are handled carefully rather than used in a way that would make an effect impossible to identify. The purpose is to create a credible counterfactual, not to guarantee that the answer is zero.
After matching, the study uses weighted difference-in-differences. In plain language, it compares the change in the treated group with the change in the matched control group across the implementation boundary. The control loans show how outcomes moved over the same period for loans that were not subject to the additional maturity factor. The treatment loans show how outcomes moved for the maturity bands that were subject to it. The difference between those changes is interpreted as the policy effect, conditional on the comparison assumptions.
A simple illustration makes the subtraction intuitive. Suppose loan amounts fall in both the treated and control groups during a period of high interest rates. The fall in the control group represents a common credit-cycle movement that is not specific to the new maturity factor. The treated group’s additional movement, after the two changes are compared, is the part the design attributes to the reform. If both groups move together, the estimated policy effect can be close to zero even though the market as a whole is shrinking. If they diverge only after implementation, the difference becomes evidence of a maturity-specific response.
The weights matter because exact matching can produce different numbers of comparable control observations for different treated observations. Weighting preserves the intended balance instead of allowing whichever subgroup has the most records to dominate the comparison. It does not create information where no comparable loan exists; it makes the retained comparison represent the treatment and control relationship the design specifies.
The design uses the fourth quarter of 2022 as the period immediately before implementation. It includes time fixed effects to absorb shocks shared across loans in a quarter and financial-institution fixed effects to absorb stable differences between lenders. It also includes time-varying bank controls, such as size, nonperforming-loan ratios, portfolio quality, return on assets, capital adequacy, structural liquidity, and the countercyclical-provision phase. These controls do not make the study omniscient. They reduce the risk that a change in a lender’s condition is mistaken for a maturity-specific effect.
The event-study checks address a central requirement of difference-in-differences: before the policy, treatment and control groups should not already be moving on different trends for the outcome being studied. The paper reports that the coefficients for periods before implementation are statistically indistinguishable from zero across the observed outcomes. That supports the plausibility of parallel trends in the matched samples. It does not prove that every unobserved factor was balanced, nor does it guarantee that the same relationship would hold outside the study window.
The stricter the comparison becomes, the more carefully its scope must be described. Matching on more characteristics can make the treatment and control loans more alike, but it also removes observations that have no close counterpart. The estimate then speaks most directly about the subset of loans for which the study can construct that comparison. It may be more credible for that subset without representing every long-maturity loan in the market. This is a trade-off between comparability and coverage, not a flaw that can be solved by simply adding more observations.
The controls also have different jobs. Matching addresses observable differences at the loan and borrower level. Time effects absorb shocks shared across a quarter. Institution effects absorb stable lender differences. Time-varying bank controls account for changing institutional conditions. None of these controls substitutes for a clear treatment boundary or for the assumption that no unmeasured, policy-specific event moved only the treated maturity groups. Good evidence makes those assumptions visible instead of hiding them behind technical language.
This sequence explains why the result should be read as a disciplined quasi-experiment rather than as a controlled laboratory test. Exact matching addresses observed differences. Fixed effects and time controls address common or stable influences. The event-study checks make the pre-policy comparison more credible. None of these steps can guarantee that an unmeasured policy-specific factor did not change at exactly the same time and affect only one maturity group.
The distinction between an observed outcome and an unobserved decision is especially important for credit allocation. If a lender reduces the principal on loans it still originates, the study can see that change. If it rejects more applications, offers a shorter term, moves customers to an exempt product, or loses them to another lender, the same record may not reveal the decision. A lower average loan amount could reflect smaller approved loans; a stable average could coexist with fewer approvals if the remaining loans are unchanged. The study’s measures are therefore evidence about originated contracts, not a complete census of access to credit.
The same boundary applies to household and institutional consequences. The study does not directly estimate future default rates, lifetime repayment burdens, borrower welfare, lender profitability, or long-run market concentration. It also cannot fully separate every concurrent macroeconomic influence, even though it controls for time and lender conditions and checks robustness against monetary policy. The evidence can support a claim about how the rule affected selected origination terms during 2022–2024. It cannot support a claim that the reform made all households safer, made all lenders more profitable, or permanently changed the structure of Colombian consumer finance.
That is the standard the results must meet. The next question is not whether the study contains a large number of observations, but whether the measured outcomes move in the directions the mechanism predicts after the comparison has been made more credible. First comes the direct buffer outcome: provision coverage. Then come the average borrower-facing outcomes: interest rates, loan amounts, and LTV. Only after those averages are understood is it safe to ask whether different institutions carried the adjustment differently.