Information and Financialization: Credit Markets as a New Source of Inequality

Abstract

Driven by financialization and rising demand for credit, household sector debt in OECD countries has risen sharply. We argue that this rise in private debt has become a significant driver of inequality because access to, and the terms of, credit vary by the risk of default, which is closely tied to income. The effect is magnified by a trove of new data that allow lenders to more accurately assess individual risks, thereby linking interest rates more closely to the underlying risk distribution. This inequalizing logic is conditioned by social transfers and by government regulation of financial markets. We test our model with data on mortgage interest rates and access to credit, using the government takeover of Fannie Mae and Freddie Mac (FM/FM) in the United States (resulting in regulatory change) and the Hartz-IV reform in Germany (resulting in changes to social transfers) as exogenous changes in important parameters of our model.

Keywords

Political economy social welfare programs Big Data credit lending financialization income inequality home ownership financial regulation welfare state Freddie Mac Fannie May

Introduction¹

At the turn of the last century, banking was personal. Banks made lending decisions based on personal knowledge of borrowers; a fact that made credit often haphazard and not infrequently biased toward friends and family, and against minorities. The small-town banker and horse trader David Harum, the main character in Edward Noyes Westcott’s 1898 novel by the same name, described his approach to lending in the 1932 movie adaptation of the novel (played by Will Rogers): “I go a long way on a man’s character. And then I go a longer way on his collateral. And if he’s got character and collateral both, I let him have about half what he asked for … anybody can get along on half of what they think they can.”

The use of information has come a long way since then, but the objective is the same: separate good risks from bad and lend to the former on the best possible terms (for the bank). The massive improvement in data, a large expansion of risk-sharing financial instruments, securitization, and a huge increase in demand have resulted in loans and credit to the household sector expanding exponentially (Figure 1). In less than 25 years, from 1995 to 2019, private debt in advanced democracies increased from an average of 90% to about 150% of disposable income (with some notable exceptions), extending a trend that started in the 1980s in the United States and the United Kingdom (the rise in consumer lending mostly occurred earlier in these countries). A growing portion of personal income now goes to servicing debt, and this has a sizable effect on discretionary income. With an average interest rate of 10%, it would amount to 15% of disposable income, but obviously with huge variation across countries, time, and individuals.

Moreover, access to credit has become an important determinant of individual welfare in a new economy where credit is used to smooth income across increasingly nonlinear life-cycles, with frequent changes in jobs, time off for retraining and additional schooling, and moves back and forth between work and family (Iversen & Soskice, 2019, Chapter 4; Wiedemann, 2021).

Owning a home has also become a marker of middle-class success in many countries, and access to mortgage finance is therefore increasingly seen as an important tool for aspirational voters. For all these reasons, access to credit and the cost of such access are emerging as important determinants of individual welfare. This paper explores the consequences of financialization of the household sector for economic inequality.

Specifically, we argue that credit has become a significant—and largely overlooked—driver of inequality. This is because terms of access to credit vary with individual risks of default, which is tied to socio-economic status. Risk assessments in turn depend on individual data on the likelihood of experiencing catastrophic life events—significant loss of income due to unemployment, illness, or involuntary job switches—and ability to financially weather such events. Such data have been greatly facilitated by the information revolution. “Big Data” combine information disclosed by borrowers with a trove of data on residence, demographic indicators, credit history, income, employment history, and so on, which are often kept in central credit registries. As Hurley and Adebayo (2017, p. 151) note, the credit-scoring industry takes an “all data is credit data” approach, “combining conventional credit information with thousands of data points mined from consumers’ offline and online activities.”

As the data available to lenders improve, they can make more differentiated risk-of-default assessments, which means that interest rates increasingly reflect the underlying risk distribution. As interest payments come out of disposable income, and insofar as disposable income is negatively correlated with default risk, the distribution of discretionary income (which is disposable income net of debt service) becomes more unequal. Those deemed too risky will not qualify for large loans (such as home mortgages) in the first place, and while “cheap money” is plentiful, access has been rationed in the wake of the financial crisis.

While nearly all research on inequality focuses on market or disposable income, and increasingly on wealth inequality, what ultimately matters to the welfare of most people is their ability to consume the goods and services that define a middle-class life-style (Fligstein & Goldstein, 2015). Financialization combined with the data revolution have enabled many to do so, but it has also increased discretionary income inequality (even when disposable income inequality is held constant), and those who are excluded from credit markets cannot enjoy the benefits of either income-smoothing or homeownership.²

The combined effect of information and financialization, however, is conditioned by national-level financial and social institutions. Income losses are cushioned by the social protection system, and financial regulations can absorb some of the default risk by subsidizing debt repayments or providing lender-of-last-resort guarantees. A notable example of the latter is when governments step in to buy and securitize mortgage debt, thereby absorbing risks that would otherwise fall on lenders. Lenders are in turn able to offer loans to more people, and on more equal terms. Quasi-public financial institutions such as Fannie Mae and Freddie Mac (FM/FM) in the United States are major cases in point.

The welfare state also matters. Specifically, when people become unemployed some of their lost income is replaced by unemployment benefits, and the higher the replacement rate the more likely the unemployed are to service their debt. Lenders know this and compete down rates to reflect the lower risk of default. The welfare state, by directly reducing the effect of adverse life events on disposable income thus has a similar effect on discretionary income as state loan guarantees or interest subsidies do by improving the terms of borrowing at the lower part of the distribution.

The underlying political coalitions that sustain these different institutions have distinct historical origins, creating sometimes surprising cross-national patterns that do not conform to standard typologies of welfare states or varieties of capitalism. Institutions and policies in Denmark and the United States produce surprisingly similar results, and while the differences in social protection institutions are familiar, it is Denmark and not the United States that has the most liberal, market-based mortgage system. These are not outliers. The organization of the mortgage lending market varies a great deal across advanced democracies, and it does not correlate with any widely used typologies of political economies (Blackwell & Kohl, 2018, 2019; Fernandez & Aalbers, 2016; Schwartz & Seabrooke, 2008; Wood, 2019).

Our paper contributes to a burgeoning literature on the politics of financialization. Ahlquist and Ansell (2017) argue that borrowing is used to compensate for high inequality, and that credit has been expanded more in inegalitarian countries as a result. In a similar vein, Wiedemann (2021) argues that increasing access to credit has been used to insure against new life-cycle risks, which in turn undercuts support for the traditional welfare state. Ansell (2014) shows that house ownership can serve as a form of long-term insurance that reduces demand for redistribution. Consistent with this idea, Hariri et al. (2020) find that when people are cut off from credit, short-term liquidity constraints drive up demand for social transfers.

Our paper builds on this work but shifts the focus away from the relationship between credit markets and social insurance toward the direct effects of financialization and information on inequality, and how these effects are conditioned by existing social institutions. Specifically, we make four contributions to the existing literature: (i) we develop a model that shows how financialization shapes discretionary income inequality; (ii) we show that this effect is magnified by better lender information about borrowers; (iii) we show how social protection and financial regulation can mitigate some of the inequalizing effects of financialization and data; and (iv) we present (quasi-experimental) evidence for large effects of information, financial regulation, and social protection on both access to credit and the terms of such access.

The Logic

We motivate our analysis with a simple formal model that is developed in more detail in Appendix A. Much lending is for purposes of “income smoothing,” where current consumption (say, of a car or daycare services) or investment (say, in an additional educational degree) are enabled by anticipation of higher income in the future (Hall, 1988). This is known as the permanent income hypothesis and presents no problems in terms of repayment of loans. People who stay on their anticipated income path would never default. Unanticipated drops in income because of long-term layoffs or illness, however, can lead to default, and this is what lenders worry about. That is true as well when borrowing to buy a home, which for most people is a long-term investment that greatly enhances welfare while also potentially generating wealth (Ansell, 2014). As in the case of other loans, discretionary income is reduced by the interest on the mortgage. We consider a world in which people’s welfare is enhanced by access to credit, but where the terms of credit vary because lenders assess a premium on higher-risk individuals. Our aim is to understand the distributive effects of how risk is assessed.

Discretionary Income and Welfare

Utility to the individual, i, is equal to discretionary income ( $D_{i}$ ), which is disposable income minus spending on necessities and interest payments, plus the utility of the consumption that borrowing in credit markets $u (L_{i})$ enables:

U_{i} = D_{i} + u (L_{i}) .

(1)

We focus on discretionary income, but since access to loans is determined by the same factors that shape the terms of borrowing, we generalize the logic to access.³ We assume that spending on necessities is constant in order to identify the effect of borrowing.⁴ Discretionary income for individual i,

D_{i}

, over the term of a loan is then equal to

D_{i} = Y_{i} - L_{i} \cdot r_{i},

(2)

where

Y_{i}

is disposable income excluding necessities, and

r_{i}

is the interest rate.

Assuming that people borrow at an optimal rate, as long as the elasticity of demand for credit is higher than −1, higher interest rates will lead to lower discretionary income. The standard assumption is that the elasticity of demand for credit is close to zero.⁵ Since utility is rising in both credit and discretionary income and since a rise in interest rates reduces discretionary income as well as borrowing, it also reduces utility.

Given this demand function, the loan amount and the total cost of borrowing are determined by the interest rate, and discretionary income will be a function of the risk of defaulting, $p_{i}$ . We show in Appendix A that

D_{i} = Y_{i} \cdot (1 - α \cdot \frac{\bar{r} + 2 p_{i}}{1 + \bar{r} + p_{i}}),

(3)

where

\bar{r}

is the competitive rate in a market with no default risk, and

α

is a weight that determines the demand for credit. We see that

d D_{i} / d p_{i} < 0

, so that discretionary spending is decreasing in the default risk. If the probability of default is declining in income—which is strongly supported by the data (more on this below, and in Appendix C)—then the greater the dispersion of the distribution of risk, the greater the dispersion of the distribution of discretionary income. This is our first result, and it shows a heretofore overlooked effect, via interest rates, of increasing inequality of risk, even though the latter is well-documented (Hacker, 2008; Hacker et al., 2013; Häusermann et al., 2015; Palier, 2010; Rueda, 2007).

Lenders are not always able to lend at the optimal rate because they are constrained by usury laws or other lending regulations or because it is too difficult to determine actual default risks above a certain level. In consequence, the lender may adopt a cutoff rule to limit exposure to bad loans. In the presence of such a rule, an increase in the dispersion of observed risk will also lead to more people being denied credit. At the same time, the dispersion of the distribution among those who can obtain loans will increase (under standard assumptions about the shape of the risk distribution).⁶ Consistent with this logic, in our data we observe an effect of risk distributions on both access to loans and interest rate dispersion.

It directly follows from our first result that countries with more unequal risk of default distributions have more unequal discretionary income distributions, after controlling for disposable income inequality. This is not captured by the effect of risk on (expected) future income (as in standard insurance models); it is a direct effect of credit markets on the distribution of current consumption.

The Effect of Information

We have assumed that borrowers and lenders are all fully informed about the risk of default. In practice, default risks are hard to observe for the lender and usually not possible to signal credibly by the borrower. This creates a classic adverse selection problem. If the lender has no information about risk type, it will have to set an average interest rate that is proportional in equilibrium to the amount of defaulted loans among all borrowers (which can be observed). This average rate is denoted $\bar{\bar{r}}$ (distinct from $\bar{r}$ , which is the competitive rate charged if all loans were repaid with interest).⁷

This common interest rate means that high-risk types will face lower interest rates than low-risk types compared to the case of full information. The consequence is a shift in lending toward high-risk types so that the total amount of defaulted debt increases, and the average interest rate rises. This is an efficiency cost, but at the same time, it reduces inequality in discretionary spending because those with higher income and lower risks now pay more for credit while those with lower income and higher risks pay less.

The inequalizing effect of information can be established more generally if we assume that lenders learn about individual risks by observing credit history. Such history is constructed by collecting information about the speed of debt accumulation, timeliness of repayments, past instances of default, etc., and we can conceive of such information as signals in a Bayesian updating game where “observed” risk is a weighted function of a prior and a signal. If $p_{l}^{o} = [p_{\min}^{o}, p_{\max}^{o}]$ is observed risk of individual i by lender l, we can write

p_{l}^{o} = ι \cdot p_{i}^{s} + (1 - ι) \cdot \bar{p}

(4)

where

p_{i}^{s}

is a noisy signal drawn from a distribution that is centered on the individual’s true risk,

p_{i}

, and

\bar{p}

is the mean among all borrowers, which is the prior. The parameter

ι

is a measure of the “precision” of the signal, which equals the information about i available to the lender. With no information (

ι = 0

) the lender only observes the population mean,

p_{i}^{o} = \bar{p}

, and the range is therefore zero. At the other extreme, with complete information,

p_{i}^{o} = p_{i}

, the range equals the difference between the individual with the lowest and the individual with the highest underlying risk.

If we use the range as a measure of dispersion, we therefore have that

[p_{\min}^{o}, p_{\max}^{o}] < [p_{\min}^{i}, p_{\max}^{i}],

(5)

and the difference in the range is falling in information:

(p_{\max}^{i} - p_{\min}^{i}) - (p_{m ax}^{o} - p_{\min}^{o}) = f (\bar{ι})

(6)

Alternatively, we could treat the difference in the variance of underlying and observed risks as a function of information. Keeping in mind that discretionary income is a function of default risk, the implication is that more information increases the inequality of discretionary income (i.e., increases the range or the variance in income). This is the second implication of the model.⁸ Closely related, more information also leads to more inequality in access since a more dispersed distribution increases the share who are above the lending threshold (i.e., considered too risky).

The Role of the Welfare State

So far, we have assumed that any “catastrophic” loss of income leads to default, but people have an incentive to try hard to avoid defaults, which will cut them off from future borrowing (or significantly raise the cost of such borrowing). In the case of defaulting on a mortgage, people will lose their home. We do not explicitly model the individual decision to default but instead assume that if private funds available in the bad state of the world, $k_{i}$ , are at or below some threshold, $T_{i}$ , the borrower will default; otherwise not:

If {\begin{cases} k_{i} \leq T_{i} then default \\ k_{i} > T_{i} then do not default \end{cases}} .

(7)

We can think of

k_{i}

as income from savings, selling assets or bringing forward long-term pension accumulations, etc. It is natural to think that

k_{i}

must be high enough to cover basic needs as well as essential fixed expenses (such as medicine) before debt servicing is possible. But there are clearly also subjective aspects to what individuals consider acceptable sacrifices, and the lender cannot observe these directly. Some people will try to repay their loans at great sacrifice; others will be more willing to let go.

In Appendix A, we derive the interest rate for the cases where (i) the lender cannot observe either risk of income loss, $p_{i}$ , or individual thresholds, $T_{i}$ , and (ii) the lender knows $p_{i}$ but not individual thresholds. In the former case there will be a common interest rate for all (see equation (A11) in Appendix A), but in the latter case, it will vary according to

r_{i} = \frac{\bar{r} + 2 \cdot p_{i} \cdot p_{(k_{i} < T)}}{1 - p_{i} \cdot p_{(k_{i} < T)}} .

(8)

Intuitively, the interest rate is rising in individual risks and the probability of default.⁹ Since the latter depends on personal assets,

k_{i}

, such assets are a source of discretionary income inequality, even in the good state, as long as they are rising in income.

Social protection mediates this relationship, however, by adding a transfer, $b_{i}$ , to personal funds in the bad state, which has the exact same effect as raising $k_{i}$ (and thus reducing $p_{(k_{i} < T)}$ ). Even if $b_{i}$ is a lump-sum benefit paid to everyone by a flat-rate tax (as in a Meltzer-Richard model), we show in Appendix B that the distribution of interest rates, and hence the distribution of discretionary income, becomes less dispersed as $b_{i}$ rises.¹⁰ This holds for a flat rate benefit; ipso facto it also holds for benefits that are targeted to those with low income (“means-tested”).

The conclusion is that the welfare state dampens the inequalizing effects of financialization and information, and that this effect is in addition to the direct effect of the welfare state on disposable income inequality. This is our third result. The existing literature only considers the direct effect of social spending on disposable income through redistribution; not the indirect effect through interest rates.

The Role of Financial Regulation

Social protection systems were not created to reduce default rates or to equalize discretionary spending through a lower dispersion of interest rates. They were created to alleviate poverty or to mitigate the risk of income loss, and it is only with financialization that the indirect effect of the welfare state on income has become important. For this reason, we treat social spending as an exogenous variable that is not caused by the credit regime.

Financial regulation, on the other hand, is specifically designed to shape the terms of lending, as well as the risks that lenders and borrowers assume. Regulations are complex, but what concerns us here is the extent to which they facilitate the transfer of default risk to the state. A common form is credit guarantee schemes (CGSs), where a state agency steps in to provide collateral and some repayment guarantees (typically less than 100%). State-guaranteed educational loans or government-backed loans to small businesses are examples. If these guarantees are credible, it reduces the risk of lending, and since risks are concentrated at the bottom of the income distribution it has the same pro-poor/pro high-risk effect as government transfers.

To illustrate the logic, we use the regulation of the American mortgage market as an example; it is perhaps the most important case of transferring default risks to the state. At the center of the system are two government-sponsored enterprises (GSEs)—FM/FM—which are required by law to purchase all mortgages that meet certain minimum requirements issued by commercial banks, savings and loan associations, and other originators, and to securitize them by issuing bonds in the secondary bond market. Before recent reforms, the quasi-public role of FM/FM had two effects on private lenders. First, they became less concerned about default risks because these were largely absorbed by FM/FM. Lenders were given considerable discretion and minimum requirements were often finessed by the banks since they knew loans were rarely returned. Second, less concerned about risk, lenders stopped acquiring detailed and costly information about individual borrowers and effectively treated all would-be homeowners equally (over and above the vague minimum requirements set by FM/FM). Once approved, “conforming loans” were offered at essentially the same terms to nearly everyone.¹¹

This equalizing effect masks significant subsidization of high-risk (usually lower income) borrowers.¹² The 1990 amendment of the FM/FM charter made it an explicit goal to “facilitate the financing of affordable housing for low- and moderate-income families,” a provision used aggressively under the Clinton administration to expand loans to low-income families (Acharya et al., 2011). It was thought, or at least hoped, that FM/FM’s strong market position and the large margins they had been able to sustain between borrowing costs in the securities market and mortgage interest rates were enough to cushion them from the risks of bad debt. Although private corporations since 1968, it was also widely believed that FM/FM loans were implicitly guaranteed by the government, which enabled the GSEs to borrow very cheaply. Apparently confirming this logic, China and other countries with saving surpluses poured large sums of money into the FM/FM-issued bonds, pushing average interest rates down (Eichengreen, 2008).

This cozy consensus was shattered with the crash of the sub-prime mortgage market, after which the stock prices of FM/FM collapsed. FM/FM were subsequently placed into conservatorship in September 2008. Before and after the government takeover, a series of reforms were implemented to reduce the risk-exposure of FM/FM and shift more of it to banks and other mortgage originators, as well as to a third government entity, Ginnie Mae, which securitizes mortgages directly guaranteed by the Federal Housing Administration. Below, we use these measures as a natural experiment whereby lenders are strongly incentivized to acquire more information and use it to screen out or raise interest rates on risky borrowers. The financial crisis is thus a window into both the effect of government regulation and the effect of information (one is a cause of the other).

Empirical Tests

Our theoretical model makes three empirical predictions:

• (H1) More information increases the spread of interest rates (and hence the inequality of discretionary income and/or access to loans).

• (H2) The government acting as a backstop in loan markets reduces the spread of interest rates (and hence the inequality of discretionary income and/or access).

• (H3) More generous public income support facilitates access to loans and decreases the spread of interest rates (and hence the inequality of discretionary income and/or access).

For the model and all three hypotheses, the underlying assumption is that risks are correlated with income. We do not think that this is a controversial assumption, but we show in Appendix C that it is strongly supported by the data.

A Note on the Relationship Between Information and Regulatory Incentives

It is difficult to test H1 and H2 separately because while information is increasing over time as a result of the data revolution, discontinuous exogenous shifts in information typically occur only as a result of regulatory changes that incentivize lenders to seek more information (or not). Conversely, changes in public subsidies for lending changes the risks that lenders face but at the same time also their incentives to acquire information. In this section, we briefly show that changes in incentives, under certain assumptions, can be treated as equivalent to changes in information. We use this equivalence to infer the effect of information from sharp regulatory changes.

Like Westcott’s small-town banker, lenders crave information because it allows them to separate good risks from bad and thus to (i) cut out potential borrowers who are likely to default and (ii) differentiate interest rates among borrowers to reflect individual risks. Yet, the benefits of information have to be weighted against the cost of acquiring information. Furthermore, when the state assumes some of the default risks, the incentive to acquire information falls. We can capture this logic using a very simply lender utility function

U_{L} = ι (δ) - c (ι, A)

where the benefit of information (

ι

measures information as before) is a negative function of

δ

, which we can think of as the probability that the regulator will assume responsibility for defaulted loans. The cost of information is a rising function of the level of information, moderated by an “information technology” multiplier, A. Big Data, faster processors, and better algorithms make the rise in cost “flatter.” A simple concave representation of this utility function is

U_{L} = ι \cdot (1 - δ) - c (A) \cdot (ι^{2})

which implies a maximum investment in information of:

ι^{*} = \frac{1 - δ}{2 \cdot c (A)} .

The expression shows that changes in the regulatory framework that affect the cost to the lender of defaults,

δ

, have the same effect on information as changes in the cost of information, c, due to new technology. The latter is mostly driven by secular changes in ICT technology that reduce the costs of compiling and analyzing data. The former is driven by regulatory changes that can be abrupt. We know that the cost of information is declining—as implied by Moore’s law—but the gradual nature of this decline makes it hard to identify its effect on interest rates. Sudden changes in the regulatory framework, by contrast, can be used to gauge the causal effect of information, even if it can only capture this effect indirectly through changes in the incentives to acquire information. In the next section, we provide a simultaneous test of H1 and H2 using this logic. We are also able to confirm that the gradual drop in the cost of information is correlated with a gradual increase in the dispersion of interest rates, although it is of course not possible to establish causality using this evidence.

Regulation, Information, and Inequality in Mortgage Interest Rates

To test H1 and H2, we use a dataset that contains all single-family loans that Freddie Mac purchased or guaranteed from the first quarter of 1999 to the last quarter of 2019—36,269,139 mortgages. As described above, Freddie Mac is one of the two main GSEs—along with Ginnie Mae, a government agency—that purchase “conforming mortgages” from lenders and sell them in the secondary bond market.

The main reason GSEs return mortgages is delinquency or default—even several years after closing—but it is at the discretion of the GSE. Mortgages closed in 2007 and 2008 saw a dramatic rise in put-back rates, which far exceeded the rise in defaults (Goodman et al., 2014, p. 60); the aggregate amount of repurchase requests increased tenfold. In terms of the notation used in the previous section, an increase in put-backs is equivalent to a decrease in $δ$ : the probability that the regulator will assume the default risk. Moreover, the GSEs tightened the underwriting guidelines for conforming mortgages that lenders had to adhere to¹³ and increased their quality controls in various ways.

These changes were rolled out starting in early 2008, and the beginning of that year therefore serves as a break after which lenders had strong incentives to use more and better information to accurately assess mortgage applications. From the perspective of our theoretical framework, the subprime mortgage crisis is a discontinuity, at which the effort lenders expend and the amount of information they use to assess mortgage quality sharply increased. Again, the trigger for lenders to acquire more information was regulatory reforms, put-backs in particular, that raised the costs of not accurately identifying default risks. For this reason, we expect the spread of interest rates to increase at the discontinuity.¹⁴

This increased scrutiny and intensified information collection clearly shows up in the data as a sharp rise in the number of days to close a loan. In an interesting account of the role of technological innovation in mortgage underwriting, Foote et al. (2019) show that mortgage processing times dramatically dropped between 1995 and 1998—from close to 50 (1994) to under 30 (1998) days—and continued to trend downward until 2005—to about 17 days.¹⁵ They attribute this decline in processing times to technology-augmented innovation; very consistent with the cost of information gradually dropping. More interestingly for our purposes is the sharp increase in processing times in 2008 and 2009, from about 18 (2007) to about 26 (2008) to almost 40 days (2009). It is worth citing their explanation in some detail:

“After the US housing boom ended, refinance timelines increase sharply as various lender and governmental policies changed. One of the most significant policy changes involved the repurchase policies of the GSEs. Fannie Mae and Freddie Mac occasionally require mortgage originators to repurchase loans that do not meet the agencies’ underwriting guidelines. After housing prices fell, both Fannie and Freddie increased their repurchase requests to originators that had incorrectly underwritten loans. This prompted originators to follow GSE policies more carefully, which likely lengthened origination timelines” (Foote et al., 2019, p. 14).

We consider these findings and conclusions by Foote et al. (2019) as evidence in favor of our assertion that the beginning of 2008 marks a discontinuity at which lenders were strongly incentivized to seek more and better information on mortgage applications. As argued, we expect an increase in the spread of interest rates at this discontinuity.

To assess the propositions that the interest rate spread increases over time in general, and at the discontinuity in particular, we start by calculating the Gini coefficient of interest rates for each year-month between 1999 and 2019, using Freddie Mac’s “Single Family Loan-Level Dataset.” On average, each cell (year-month) contains about 60,000 mortgages (the median cell size is 55,714, the minimum and maximum are 11,910 and 207,049, respectively). Figure 2 plots these Gini coefficients over time. While the figure shows an upward trend, there only seems to be a short-lived increase in the spread of interest rates at the discontinuity (January 2008).

However, balance tests reveal that the samples to the left and right of the discontinuity are very different. Most importantly, the composition changes in terms of the distribution of FICO scores (and FICO scores are highly correlated with interest rates), right at the discontinuity. This can be seen in Figure 3.

This rise in FICO scores is itself very consistent with the claim that early 2008 was a discontinuity at which lenders engaged in more careful screening since it implies that an increased number of potential borrowers with low FICO scores were denied loans.¹⁶ But in addition to such censoring, lenders began to differentiate more between borrowers with good FICO scores in the loan terms they were offered. The obvious interpretation is that lenders acquired additional information among borrowers with similar FICO scores. We focus our analysis on the change in the spread within FICO tranches to circumvent the potential problem of a changing composition of borrowers. Note that this gives us a conservative estimate of the effect of the discontinuity because (i) we do not capture the rise in rejected mortgage applications (which would increase dispersion if they stayed in the pool) and (ii) we do not capture the rise in the interest-rate spread across FICO groups.

Specifically, to balance the samples before and after the discontinuity—to compare apples with apples—we restrict the sample to mortgages that fulfill the following criteria, and we also shift the analysis from the year-month-level to the year-month-FICO-2d level:

- Credit (FICO) scores in the range of 620–819. We drop cases with scores below 620 because this is the minimum score required by Freddie Mac to qualify for a conforming mortgage, at least under normal conditions in most years (this drops 2.19% of the sample). We drop cases with FICO scores in the 820–850 range (the very top-end of the FICO-score distribution) because only few mortgages are in this category, and they are unevenly distributed over time—leading to unreliable and infrequent estimates of the spread of interest rates for FICO scores above 819 (this drops 0.03% of the original sample).

- 30-year mortgages (applies to 67.1% of the original sample).

- Fixed-rate mortgages (applies to 100% of the original sample).

- No mortgage insurance (applies to 82.9% of the original sample).

- Loan-to-value ratio of a maximum of 80% (i.e., minimum of 20% down payment) (applies to 80% of the original sample).

- Single-family units that are owner occupied (applies to 90.6% of the original sample).

- US states only (applies to 99.8% of the original sample).

This leaves us with a sample of about 15 million mortgages. Balance tests show that the composition of mortgages before and after January 2008 is very similar even across bins, and assuredly so within bins. We use this dataset to calculate measures of interest rate dispersion—such as the Gini coefficient, the Coefficient of Variation, and others—at the year-month-FICO-2d level.¹⁷ Figure 4 plots the Gini coefficient of interest rates within FICO-2d-bins over time. Therefore, a dot in the figure represents the Gini coefficient of a year-month-FICO-2d-bin.

At least three aspects are noteworthy about the patterns in Figure 4. First, the spread of interest rates—as measured by the Gini coefficient within 10-point FICO-bands—clearly increases over time, as hypothesized. The spread of interest rates roughly doubled within the 20-year period under consideration.

Second, there is a clear increase of the spread of interest rates beginning in January 2008 (the discontinuity). The figure shows local polynomial fit lines of orders 1 through 3 (fitted over the entire support of the pre- and post-treatment time periods, respectively). All indicate a visual break in the series. To test whether there is, indeed, a break in the spread of interest rates, we rely on regression discontinuity (in time) analysis. The outcome variable is the interest rate spread (measured by Gini coefficients within FICO-2d bins at the month-year level). The score/running variable is month-years, with January 2008 as the discontinuity. We employ a (data-driven) mean square error (MSE) optimal bandwidth selection procedure—imposing the same bandwidth on each side of the cutoff—for local polynomial estimation of and inference on treatment effects and report robust bias-corrected confidence intervals (Calonico et al., 2017), with standard errors clustered at the FICO-2d level. Observations are weighted via a triangular kernel function (i.e., observations closer to the cut-off are weighted more heavily). Our dataset has repeated observations in the running variable (20 FICO scores per year-month), for which the estimator controls. Table 1 reports the RD estimates based on local polynomials of orders 1 through 4.

Table 1.

Regression Discontinuity Estimates.

	(1)	(2)	(3)	(4)
Order of local polynomial	1	2	3	4
RD estimate	1.003^*** (0.0976)	1.300^*** (0.113)	1.032^*** (0.105)	1.063^*** (0.112)
Robust 95% CI	[.901; 1.283]	[1.14; 1.609]	[.75; 1.176]	[.782; 1.259]
BW type	mserd	mserd	mserd	mserd
Kernel	Triangular	Triangular	Triangular	Triangular
Order Loc. Poly. (p)	1	2	3	4
Order bias (q)	2	3	4	5
N	5025	5025	5025	5025
N (l)	2147	2147	2147	2147
N (r)	2878	2878	2878	2878
Eff. N (l)	340	300	320	500
Eff. N (r)	360	320	340	520
BW est. (l)	17.15	15.78	16.52	25.35
BW est. (r)	17.15	15.78	16.52	25.35
BW bias (l)	28.13	25.24	26.03	38.35
BW bias (r)	28.13	25.24	26.03	38.35

Note: Standard errors in parentheses (clustered at the FICO-2d level).

^* p < 0.05.

^** p < 0.01.

^*** p < 0.001.

Estimates adjusted for mass points in the running variable.

The outcome variable is the Gini coefficient of interest rates at the FICO-2d level.

The running variable is time (month-year) with the cutoff in January 2008.

These estimates are based on the user-written Stata commands—rdrobust—(version 8.4.0 from 2021-08-30) (Calonico et al., 2017).

Source: Freddie Mac’s “Single Family Loan-Level Dataset” (http://www.freddiemac.com/research/datasets/sf_loanlevel_dataset.page, last accessed June 3, 2021 [https://perma.cc/KU32-9TXH]), Q1-1999 to Q3-2020.

Table 1 shows that the RD estimate ranges from about 1 to about 1.3, depending on the order of the local polynomial. These estimates are statistically significant at p < 0.001. The literature usually recommends (Gelman & Imbens, 2019) and chooses (Pei et al., 2021) lower-order over higher-order polynomials. Therefore, we prefer the first model (which uses local linear regressions). The RD estimate of about 1 implies roughly a 40% increase at the threshold in the interest rate spread (from about 2.5 before the threshold on average).

In Appendix D, we perform a wide variety of additional tests and show that the finding of a statistically significant (and substantively meaningful) increase in the interest rate spread at the discontinuity is very robust. In particular, we perform the following additional/robustness checks:

- Different bandwidth selection procedures.

- Different kernel functions.

- Covariate adjusted estimates (controlling for month dummies; lagged dependent variable; average interest rate) (Hausman & Rapson, 2018).

- Different specification of the running variable (months, quarters, trimesters, half-years, and years), using the year-month-FICO-2d-level data.

- Different specification of the running variable (months, quarters, trimesters, half-years, and years), with the outcome variable recalculated at the respective unit level (months, quarters, trimesters, half-years, and years).

- Sensitivity to observations near the cutoff (donut hole approach).

- Placebo outcomes.

- Placebo cutoffs.

- Mass points adjustments.

The third noteworthy aspect in Figure 4 is the increasing range of the interest rate spread, with a clear jump at the discontinuity. For example, in 2007, the Gini coefficient of interest rates ranged from 2.3 to 3.5 while it ranged from 2.4 to 6 in 2008. This might suggest that the increase in the interest rate spread at the discontinuity was higher for some of the 20 FICO-2d groups. The obvious hypothesis is that lenders focused their increased screening efforts on applicants with lower FICO scores because it is well-documented that the spread of risks, measured by default rates, is greater for lower FICO tranches (VantageScore, 2020).¹⁸ If so, lower FICO scores essentially received a higher dosage of the treatment (scrutiny from lenders). To explore this supposition, Figure 5 reproduces Figure 4, but with separate panels for each of the 10-point FICO bands. Therefore, Figure 4 is a pooled version of Figure 5 and a dot within the panels of Figure 4 indicates a year-month. Within each panel, the mortgages are very similar. Most importantly, their FICO-scores are within 10 points of each other (by construction) and the samples before and after January 2008 are balanced well–the figure therefore offers something close to an apples-to-apples comparison.

Visual inspection of Figure 5 suggests that the increase in the spread of interest rates at the threshold was particularly pronounced at lower FICO-scores—roughly in the 620–679 range. In contrast, at higher FICO levels, the increases seem to be more minor.

Statistical discontinuity tests confirm this pattern. In particular, Figure 6 summarizes the regression discontinuity estimates for each of the 20 FICO-bands as coefficient plots. The three columns display estimates based on linear, quadratic, and cubic local fit lines (the estimates are equivalent to the models 1, 2, and 3 in Table 1—they employ the same bandwidth selection procedure, kernel function, and so on—but they are derived from each FICO-2d bin separately). The figure shows that while all estimates are positive and almost all estimates are statistically significant, they tend to be larger at the lower end of the FICO score distribution. This test therefore adds further evidence that is consistent with the hypotheses. It also ameliorates one of the weaknesses of regression discontinuity in time designs by adding cross-sectional evidence (Hausman & Rapson, 2018).

Summarizing, we interpret the general upward trend in interest rate inequality (Figure 4 and Figure 5), the increase in early 2008 (Figure 4), and the sharp increase in early 2008 among lower FICO scores (Figure 5 and Figure 6) as evidence consistent with our framework and hypotheses H1 and H2 above. Increasing information, whether gradually rising over time or induced by abrupt regulatory change, does indeed increase interest rate dispersion, as predicted.

Regression discontinuity-in-time designs—such as our approach above—face challenges (Hausman & Rapson, 2018). For example, January 2008 was during a tumultuous time and there might be other candidate explanations for the increase in interest rate inequality. But starting in early January, the evidence clearly suggests the importance of information: the time to close sharply increased; the average FICO scores of loans supported by Freddie Mac’s lower FICO score loans became much less common in Freddie Mac’s portfolio; and lenders began to make more fine-grained distinctions between borrowers. We can infer that they relied on information that went well beyond FICO scores since the spread in rates increased notably even within narrow (2-digit) FICO tranches. Lenders did this, we argue, because regulators provided them with powerful new incentives to separate good from bad risks.

But there are good reasons to believe that the financial industry has been continuously improving its information both before and after 2008. It is a frontrunner in adopting new ICT technologies for that purpose (Foote et al., 2019), and even with the increased role of Ginnie Mae (which reduces industry exposure to bad risks), there is a clear upward trend in the spread. This is consistent with the price of information falling over time, which is theoretically predicted to have the same effect on lender behavior as a rise in the cost of defaults.

The Welfare State and Homeownership

Our model’s third prediction is that more (less) generous public income support expands (contracts) access to lending and decreases (increases) the spread of interest rates. The unemployed and those at high risk of unemployment are a greater risk to lenders unless a generous unemployment benefit system enables people to keep servicing their debt. Because unemployment risks are higher for lower-skilled, lower-paid workers, low replacement rates will disproportionately raise borrowing costs and rejection rates at the bottom of the income distribution. Those at the higher end will instead benefit from lenders screening potential borrowers more carefully.

For an initial exploration of this hypothesis, we exploit the profound changes in the German unemployment benefit system resulting from the Hartz-IV reforms in 2005 (Arent & Nagl, 2013). Because the reform affected the ability of unemployed to service debt in the event of unemployment, we can compare changes in homeownership rates across groups (un)affected by the reform, from before to after the reform—a difference-in-differences approach.

Unlike the US system, government entities play no prominent role in the lending market in Germany,¹⁹ and banks offer mortgages, which are typically fixed-rate, on a competitive basis. The system has strong build-in prudential safeguards, including low loan-to-value ratios and limited equity release options, so any changes in the assessed creditworthiness of borrowers show up immediately in lending decisions, and because of civil usury law, lenders tend to cut off risky prospects rather than charge high interest rates.²⁰

The most consequential changes of the Hartz-IV reform occur after a year of unemployment benefit recipiency. Before the reform, the unemployed could qualify for “unemployment assistance,” a benefit proportional to previous wages—67% in the first year and ca. 57% in the second—that could be collected indefinitely, subject to annual renewal. To qualify, assets had to be below a certain threshold (∼520 Euro times age), though certain assets were protected (“Schonvermögen”), including owner-occupied housing of reasonable size. For banks, default was not a major concern if the income-to-loan ratio was high enough because unemployment assistance was proportional to previous wages at a fairly generous rate and paid indefinitely.

After the reform, the reasonably generous, proportional, perpetual unemployment assistance in the second year was replaced with a meager, flat rate, conditional benefit (Arbeitslosengeld II). Owner-occupied housing of reasonable size is treated as a protected asset as before, but the overall limit for other protected assets is significantly lower (∼150 Euro times age). This somewhat odd (Kaiser, 2018) differential treatment of assets means that a mortgage provides an opportunity to protect assets from the government by shifting them into owner-occupied housing. Consequently, for those at risk of unemployment, incentives for seeking a mortgage increased with the Hartz-IV reforms. For banks, however, there is now much more reason to worry about default among those with risk of long-term unemployment.

Overall, the reforms had two potential effects on financial markets: (i) they made it more difficult for some people to qualify for a loan and (ii) they increased the risk of default among some borrowers who did qualify. The logic is illustrated in Figure 7, where the solid line is the pre-reform distribution of default-risk, and the dashed line is the post-reform distribution. The share with observed risk above the threshold for approval increases, and the distribution of those below the threshold “flattens” (becomes more dispersed), reflecting a more right-skewed default-risk distribution. The implications are a reduction in the number of loans granted among higher risks and an increase in the interest rate spread of loans that are granted. We have data that can illuminate the former effect.

Unfortunately, we do not have data on declined mortgage applications, but we have data on homeownership rates that allow us to shed some light on the fortunes of the unemployed compared to the employed (or poor vs. rich) before and after the reform. We do so in three steps. First, we compare homeownership rates of the employed versus the unemployed before and after the reform. The data show a sharp drop in homeownership among the unemployed that coincides with the cuts of unemployment benefits.

Second, to gauge the effects of reforms on relinquishing ownership, we track homeownership rates by employment status for those that were homeowners before the reform. The data indicate that for this subsample, ownership rates did not differently drop among either the unemployed or employed. We infer from this that the unemployed rarely relinquished housing assets and that the drop in homeownership rates for the unemployed must be due to lower rates of home acquisition after the reform, presumably because of reduced access to credit. It may seem surprising that the reform is not associated with more widespread selloff among the unemployed, but as we noted, the Hartz-IV reforms gave people incentives to hold onto their housing wealth, which is exempt from the requirement to spend down personal savings before drawing benefits (“Schonvermögen”). Those who were already homeowners before the reform were undoubtedly also in a stronger financial situation than those who were not and therefore less likely to default. The point for our purposes is simply that the drop in homeownership among the unemployed must be because people are less likely to obtain mortgages after the reform.

Third, among the employed we find that homeownership rates of the poor and the rich diverged after the reform. That is especially true for those who were not already owners before the reform. Since the poor are at much higher risk of becoming unemployed, the obvious explanation is that mortgage lenders increasingly avoided bad risks after the reform.

The following four tables display the evidence just summarized. Each table shows the difference in homeownership between two groups, before and after the reform (“diff-in-diff”). The first is a comparison of homeownership rates by employment status. The most authoritative data for this information is the “Sample Survey of Income and Expenditure” (EVS), which is based on about 60,000 respondents and conducted every 5 years (we have data for 1993, 1998, 2003, 2008, 2013, and 2018). We do not have access to the micro-level data, but the Federal Statistical Office publishes—or provided us with—aggregate data on homeownership²¹ for all households and for the unemployed. This allows us to compare homeownership rates among the “Unemployed” (treated group) versus the “Employed”²² (control group), before and after 2005. Table 2 display the results of a difference-in-differences test of this comparison. The results show that homeownership rates among the employed did not significantly change while they sharply declined among the unemployed. The difference-in-differences estimate is about −12.7 percentage points for homeownership and statistically significant.²³

Table 2.

Homeownership (EVS).

Homeownership	Pre-2005	Post-2005	Difference
Employed	0.423	0.455	0.032 (0.023)
Unemployed	0.222	0.126	−0.096^** (0.023)
Difference	0.201^** (0.023)	0.329^** (0.023)	−0.127^** (0.033)

Note: Standard errors in parentheses. N = 12.²⁴

^* p < 0.05.

^** p < 0.01.

Source: Sample Survey of Income and Expenditure (EVS).

The German Socio-Economic Panel (GSOEP) (Liebig et al., 2020) is another source that allows us to track homeownership rates among the unemployed versus the employed.²⁵ The EVS and the GSOEP use somewhat different definitions of homeownership²⁶ and employment status,²⁷ and they cover different time-periods.²⁸ So, while estimates from the two sources are not directly comparable, we do expect that they reveal similar patterns. Table 3, which displays difference-in-differences estimates—comparing the unemployed with the employed before and after 2005—shows that this is, indeed, the case: homeownership rates dropped markedly among the unemployed, comparing 2000–2004 with 2005–2010, and the difference-in-differences estimate is about −7.5 percentage points.

Table 3.

Homeownership (GSOEP).

Homeownership	Pre-2005	Post-2005	Difference
Employed	0.402	0.384	−0.018^** (0.005)
Unemployed	0.256	0.162	−0.094^** (0.014)
Difference	0.146^** (0.013)	0.221^** (0.013)	−0.075^** (0.015)

Note: Standard errors (clustered at household-level) in parentheses. N = 100,667.

^* p < 0.05.

^** p < 0.01.

Source: German Socio-Economic Panel (GSOEP).

The second step in our three-pronged approach compares the development of homeownership for the subsample of respondents who already were homeowners before the reform (during 2000–2004), using the GSOEP data. We again use difference-in-differences estimates even though, by construction, there are no differences between the unemployed and employed before the reform since the sample is restricted to homeowners before the reform. Table 4 shows that there are no meaningful differences after the reform, either—the estimated difference-in-differences is essentially zero, and not statistically significant. This suggests that the unemployed in the GSOEP sample are not disproportionally relinquishing their homes after the reform and that the divergence in ownership rates between the unemployed and employed documented above is driven by the inability of the unemployed to secure mortgages credit after the reform.

Table 4.

Homeownership (GSOEP), Conditional on Being Homeowner Pre-Reform.

Homeownership	Pre-2005	Post-2005	Difference
Employed	1	0.963	−0.037** (0.003)
Unemployed	1	0.968	−0.032** (0.011)
Difference	0	0.005 (0.012)	0.005 (0.012)

Note: Standard errors (clustered at household-level) in parentheses. N=39,170.

^* p < 0.05.

^** p < 0.01.

Source: German Socio-Economic Panel (GSOEP).

It could still be the case, however, that the unemployed simply decide that they cannot afford a mortgage after the reform and do not apply. That is consistent with the model, but not speaking to the role of lenders. Therefore, in the third step, we compare changes in homeownership among the employed only, comparing rich and poor employed respondents. The Hartz reforms made lower income groups more likely to default—because they are at higher risk of unemployment²⁹—and we therefore expect homeownership rates among the employed poor and the employed rich to diverge after the reform. This is what the data show, with a statistically significant difference-in-differences estimate of about −3.2 percentage points (Table 5). As for the unemployed, the effect is mostly due to a relative drop in homeownership among those poor who were not already owners before the reform, suggesting that they faced tighter access to credit after the reform.

Table 5.

Homeownership (GSOEP), Rich versus Poor Employed.

Homeownership	Pre-2005	Post-2005	Difference
Rich	0.577	0.583	0.006 (0.009)
Poor	0.249	0.223	−0.026^** (0.008)
Difference	0.328^** (0.014)	0.360^** (0.017)	−0.032^** (0.012)

Note: Standard errors (clustered at household-level) in parentheses. N = 55,798.

^* p < 0.05, ^** p < 0.01.

Source: German Socio-Economic Panel (GSOEP).

Overall, the patterns in both datasets (the EVS and the GSOEP) are consistent with the hypothesis that access to credit worsened for the unemployed as well as those at higher risk of unemployment, after the Hartz-reforms lowered unemployment benefits. Different data sources and our three-pronged empirical approach support this conclusion, but since we do not have data on mortgage applications (and rejections), we cannot be certain that lending decisions drove the results. Future research will have to (dis)confirm that interpretation.³⁰

Conclusion

Financialization of advanced economies has made ability to access credit markets, and the terms of such access, increasingly important for individual welfare and economic inequality. Creditworthiness affects who can purchase a home and who can move between work and family and between work and further education, and it also affects the interest rate spread and therefore the dispersion of discretionary income. More plentiful information strengthens this relationship and empowers lenders to differentiate between high- and low-risk groups, raising interest rates for low-income groups or excluding them from credit markets altogether. The combination of financialization and Big Data is therefore a double whammy for the poor: like everyone else, they increasingly depend on borrowing to smooth income and acquire assets, but they are increasingly identified as bad risks and face worse access to, and terms of, borrowing.

Yet these inequalizing effects are strongly conditioned by the regulatory regime and by the welfare state. Where the state assumes some of the risks of lending—for example, by acting as a backstop in mortgage markets—or where the social protection system is generous, the effects of financialization and Big Data are muted. Our evidence from the housing market backs up these claims. While our evidence is based on isolated cases, they suggest a consistent (possibly causal) story across very diverse contexts, consistent with more descriptive evidence, that lends credence to our model. We believe the model captures an important consequence of the massive shifts toward financialization and individualized data, which has been mostly neglected in the CPE literature and points well beyond credit markets to social insurance and the welfare state.

An important question for future research is whether the improved capacity of markets to differentiate between risk groups will lead to a weakening of the regulatory regime and possibly also the welfare state. A major difficulty in keeping together progressive coalitions is that the underlying risk distribution is strongly right-skewed, which means that the median in the distribution—who is likely to be politically influential—is someone who would benefit from greater differentiation. Is it too pessimistic to suppose that the rising importance of credit combined with better information about the shape of the distribution will lead to intensified calls for the state to step back?

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

Philipp Rehm

Notes

References

Acharya

V. V.

Richardson

van Nieuwerburgh

White

L. J.

(2011). Guaranteed to Fail: Fannie Mae, Freddie Mac, and the Debacle of Mortgage Finance. Princeton University Press.

Ahlquist

J. S.

Ansell

B. W.

(2017). Taking credit: Redistribution and borrowing in an age of economic polarization. World Politics, 69(4), 640–675. DOI: 10.1017/S0043887117000089

Ansell

(2014). The political economy of ownership: Housing markets and the welfare state. American Political Science Review, 108(2), 383–402. DOI: 10.1017/S0003055414000045

Arent

Nagl

(2013). Unemployment compensation and wages: Evidence from the German Hartz reforms. Jahrbücher Für Nationalökonomie Und Statistik/Journal of Economics and Statistics, 233(4), 450–466. DOI: 10.1515/jbnst-2014-0504

Blackwell

Kohl

(2018). The origins of national housing finance systems: A comparative investigation into historical variations in mortgage finance regimes. Review of International Political Economy, 25(1), 49–74. DOI: 10.1080/09692290.2017.1403358

Blackwell

Kohl

(2019). Historicizing housing typologies: Beyond welfare state regimes and varieties of residential capitalism. Housing Studies, 34(2), 298–318. DOI: 10.1080/02673037.2018.1487037

(2017). rdrobust: Software for regression-discontinuity designs. The Stata Journal, 17(2), 372–404. DOI: 10.1177/1536867X1701700208

DeFusco

A. A.

Paciorek

(2017). The interest rate elasticity of mortgage demand: Evidence from bunching at the conforming loan limit. American Economic Journal: Economic Policy, 9(1), 210–240. DOI: 10.1257/pol.20140108

Eichengreen

(2008). Origins and responses to the current crisis. CESifo Forum, 9, 6–11.

(2016). Financialization and Housing: Between Globalization and varieties of capitalism. Competition & Change, 20(2), 71–88. DOI: 10.1177/1024529415623916

11.

Fligstein

Goldstein

(2015). The emergence of a finance culture in American households, 1989–2007. Socio-Economic Review, 13(3), 575–601. DOI: 10.1093/ser/mwu035

12.

Foote

Loewenstein

Willen

(2019). Technological innovation in mortgage underwriting and the growth in credit: 1985–2015 (Federal Reserve Bank of Boston Research Department Working Papers) [Federal Reserve Bank of Boston Research Department Working Papers]. Federal Reserve Bank of Boston. DOI: 10.29412/res.wp.2019.11

(2010). Escaping the tyranny of earned income? The failure of finance as social innovation. New Political Economy, 15(1), 147–164. DOI: 10.1080/13563460903553723

14.

Gelman

Imbens

(2019). Why high-order polynomials should not be used in regression discontinuity designs. Journal of Business & Economic Statistics, 37(3), 447–456. DOI: 10.1080/07350015.2017.1366909

15.

Godechot

(2020). Financialization and the increase in inequality. In Mader

Mertens

van der Zwan

(Eds.), The Routledge International Handbook of Financialization (pp. 413–424). Routledge. DOI: 10.4324/9781315142876-37

16.

Goering

Wienk

(2018). Mortgage lending, racial discrimination and federal policy. Routledge.

17.

Goodman