When Bank Policy Becomes Visible in the Mail
The central question in the U.S. credit-card market is not whether a survey bar and a market statistic point in the same direction in the abstract. It is whether the same lender that reports changing its standards also changes a behavior it controls: the volume of credit-card offers it sends. A Federal Reserve study makes that test at the bank level, allowing readers to judge both what the evidence establishes about tightening and why its result is not a one-for-one forecast of new card originations.
Matching banks' reports to their actions
The study joins two data sources that were built independently. The first is the Federal Reserve's Senior Loan Officer Opinion Survey, in which banks report whether their credit-card lending standards tightened, eased, or remained about unchanged over the previous three months. The second is Mintel Comperemedia's record of credit-card offers mailed to consumers. Mintel identifies the financial institution making an offer and records the offer material, including terms. The researchers use the institution's name to hand-match it to a contemporaneous bank identifier in the survey.
That match changes the question being asked. An aggregate comparison might show that total mailed offers and a broad survey measure moved together during a recession, but it cannot show whether the banks reporting tighter standards were the banks sending fewer offers. A lender-level panel can make that connection. It compares a bank's reported policy direction with the change in that same bank's observed mail volume, while also accounting for persistent differences among banks and common conditions across time.
The same-bank comparison also reduces a common aggregation trap. A large lender can dominate total mail volume without representing the behavior of a smaller lender, and the mix of lenders in the market can change over time. If the survey responses came mostly from one group while the offer data reflected another, a market-wide correlation would blur the relationship the study was trying to test. Matching the named lender keeps the policy observation and the action observation attached to the same decision-maker. It does not make the data complete, but it makes the comparison conceptually aligned.
The resulting panel contains 73 unique banks and observations from the first quarter of 2000 through the fourth quarter of 2019. On average, about 19 banks appear in each quarter. It is not a balanced census: bank acquisitions, coverage changes, and entries and exits from the credit-card market mean that not every lender is observed in every period. That imperfection is important, but it does not make the comparison meaningless. It means that the study's conclusion must be stated about the matched lenders and period that the data actually cover.
The distribution of responses also tells us what the comparison is—and is not. In the matched sample, 80.2 percent of bank-quarter observations reported unchanged general credit-card standards, 11.8 percent reported tightening, and 8.0 percent reported easing. Most observations therefore form the no-change comparison group. A reported tightening is a directional difference from that reference, not a measure of how many percentage points an individual bank changed its approval rate.
Nor does the general standard response identify a single contract change. The study compares it with separate survey responses about credit limits and interest-rate spreads. Among the bank-quarters reporting tighter general standards, 60.2 percent also reported tightening credit limits, 36.3 percent reported unchanged limits, and 3.5 percent reported easing limits. The spread responses were more mixed. This imperfect alignment is not a flaw to be hidden; it is evidence that a broad underwriting posture can be implemented through different combinations of limits, prices, and customer selection.
This is why the study does not ask whether every sign of restriction should appear at once. A bank might report a general tightening while leaving limits unchanged for the customers it still wants, or it might adjust limits without changing the amount of mail it sends. The broad response is the explanatory starting point; the offer data provide a way to see which margin moved. Treating the cross-tabulation as a promise that all banks use the same recipe would erase the very variation that makes the lender-level comparison useful.
The empirical comparison can be summarized as a chain from a bank's statement to an observed action and then to a bounded inference.
The design is stronger than treating mailed offers as a proxy by assumption. The survey response is not being called “supply” merely because researchers want a supply measure, and the mail record is not being called “demand” merely because consumers forward the material. The two are connected at the lender and quarter, then tested against one another. This does not eliminate all measurement error. It does give the reader a reason to believe that the relationship, if present, is about the lender-side conditions the study intends to examine rather than only about coincident movements in the overall economy.
What the estimate says—and what it does not
The baseline regression uses quarterly mail-volume growth as its outcome. It compares bank-quarters in which the lender reported tightening credit-card standards with bank-quarters in which it reported no change, while including bank fixed effects and time controls for year and quarter of year. The expected direction is negative: if approval standards become more restrictive, fewer households may qualify for a solicitation, so the lender may send fewer offers or reduce the growth of its mailing program.
The no-change group is the interpretation anchor. The estimate does not say that a tightening bank mails 18 percentage points fewer offers than every other bank, or that its total volume is 18 percentage points lower forever. It says that, conditional on the model's controls, the growth rate associated with a tightening response is lower than the growth rate in the comparison condition of reported no change. A growth-rate result can coexist with a large existing mail program, a small existing program, or subsequent recovery. Keeping the comparator and outcome visible prevents the headline number from becoming more general than the regression.
The estimate is economically meaningful. In the baseline specification, quarterly mail-volume growth is 18.367 percentage points lower when a bank reports tightening than when it reports no change. The table gives a standard error of 6.635 and marks the estimate as statistically significant at the 1 percent level. The paper describes the size as larger than one quarter of the sample standard deviation of mail-volume growth. For a reader, the essential unit is “percentage points of quarterly mail-volume growth,” not “18 percent fewer cards.” The study is measuring a change in a lender-controlled activity rate during a particular quarter.
The paper's public abstract describes the result as mail-volume growth being 18 percent lower, while the detailed regression table and results prose present a coefficient of -18.367 on a percentage-growth outcome. The most careful summary is therefore “about 18 percentage points lower in the regression.” That wording preserves the detailed unit and avoids turning an imprecise abstract shorthand into a claim about the number of accounts opened.
The demand controls address the most important competing explanation. Suppose consumers suddenly wanted fewer credit cards. A bank might then mail fewer offers because it anticipated weaker response, and a simple correlation would make the decline look like tighter supply. The study adds bank-reported changes in credit demand to the model. The tightening estimate is essentially unchanged, at -18.352, while the coefficients on stronger and weaker demand are numerically small and statistically insignificant. Within this design, the observed connection between reported tightening and slower mail-volume growth is not explained away by the measured demand responses.
That result should be understood as evidence about identification, not as evidence that demand never matters. The survey's demand question is itself a bank report, and the model cannot observe every reason consumers might apply or not apply. What the comparison supports is narrower: adding the available bank-level demand indicators does not materially alter the estimated relationship in this sample. The design makes a supply interpretation more credible; it does not create a complete decomposition of every change in credit-card borrowing.
The controls are especially useful because they test the study's main alternative explanation without pretending to solve all alternatives. They distinguish the reported direction of a bank's standards from the bank's own report of whether demand strengthened or weakened. They do not observe unmeasured changes in consumer preferences, the full set of acquisition channels, or all differences between people who receive mail and people who do not. A careful reader should therefore update toward a supply interpretation, not jump from “demand controls did not change the coefficient” to “the coefficient is purely causal.”
The researchers test whether the result depends on unstable or unusually small lenders. Restricting the sample to banks that appear for more than 20 quarters leaves the estimate robust and, if anything, stronger. A separate restriction to lenders with average quarterly mail volume above 500,000 offers produces a consistent result. These checks reduce concern that a few brief appearances or very volatile small-lender observations are driving the headline association. They do not turn the panel into a census of issuers that never appear in the data.
There is a useful distinction between robustness and representativeness here. Robustness asks whether the estimated relationship survives reasonable changes to the observed sample. Representativeness asks whether the observed sample stands for every issuer, every customer, and every period a reader cares about. The paper's checks speak mainly to the first question. The historical matched panel remains weighted toward the banks and mail activity that can be linked in the available records, so applying its exact magnitude elsewhere would require new evidence.
The study also examines whether tightening operates very differently across the credit-score distribution. Because score information is available only from 2007 onward, this is a shorter sample. The relationship between standards changes and mail-volume growth is broadly similar for prime, near-prime, and subprime offer groups; the interaction terms are not statistically significant. The paper notes suggestive evidence of a stronger subprime effect during easing episodes, but that observation is not enough to establish a general rule. Here again, the absence of a statistically clear difference is not proof that every borrower experiences the same effect. It means the study does not detect a systematic difference with the available data and specification.
History matters as well. Around the Credit Card Accountability Responsibility and Disclosure Act, the relationship between reported tightening and lower mail-volume growth was subdued before the Act and became sizeable and statistically significant afterward. The paper does not find a discernible relationship between easing and higher mail-volume growth in either period. This comparison suggests that institutional context can change how a reported policy stance appears in observed lender behavior. It is not a reason to assume that a post-2019 market will reproduce the same pattern without testing it.
The asymmetry is one of the study's most informative boundaries. Tightening has a clear negative association with mail-volume growth, but reported easing does not have a statistically significant positive association. A bank may cut outreach quickly when it becomes cautious, while an easing response may be implemented through limits, prices, approval decisions, or a gradual rebuilding of acquisition activity rather than an immediate increase in mail. The data do not identify which mechanism explains the missing mirror image. They do establish that the evidence is not symmetric enough to support the simple rule “tightening subtracts X, easing adds X.”
Several limits remain even after the checks. The panel ends in 2019, before the pandemic, and the FEDS paper is preliminary research circulated for discussion. The outcome is mailed-offer growth, not all acquisition channels, all approvals, total originations, or outstanding balances. Bank fixed effects account for stable differences among observed lenders, and time controls account for common calendar patterns, but they do not prove that a reported tightening caused the observed mail change. Banks may adjust their survey responses and mail strategies in response to another factor that the model does not measure.
The correct inference is consequently two-sided. The study validates the idea that SLOOS responses contain information about actual lender behavior: when banks report tighter card standards, their observed mail-volume growth is materially lower in the matched historical sample. At the same time, the result does not tell us that total card originations fell by 18 percentage points, that every issuer reduced every channel, or that every borrower lost the same opportunity. To understand household access, the analysis must move from the aggregate mail total to who receives an offer and what that offer contains.