Negative interest rates and bank credit risk-taking

Abstract

We examine the impact of the negative interest rate policy (NIRP) on bank credit risk-taking. Employing a triple difference (TD) methodology and a dataset of 1958 banks from 29 member countries of the Organisation for Economic Cooperation and Development (OECD) over 2011–2017, we find that banks in countries adopting NIRP exhibit a contraction in loan loss provisioning. Moreover, this NIRP effect depends on country- and bank-specific characteristics such as inflation, bank size, and bank specialisation. We also employ other methods, such as the quadruple difference (QD) model and propensity score matching (PSM), to check the robustness of our findings from the TD model.

JEL Classification: E43, G21, G28

Keywords

Bank credit risk-taking loan loss provisions negative interest rates

1. Introduction

Since the Global Financial Crisis (GFC), policymakers have faced a profound challenge: how to manage the situation when the next cyclical slump comes. Conventional measures, such as lowering interest rates, have been widely used during economic downswings (Cukierman, 2013). With the aim of providing further economic stimulus, many economies have also adopted a range of unconventional monetary policies (Molyneux et al., 2019; Wu and Xia, 2020). The most debated among these is the implementation of the negative interest rate policy (NIRP).

Extant research vigorously examines the impact of NIRP. Bottero et al. (2022) show that NIRP has expansionary effects on credit supply via a portfolio rebalancing channel. However, some suggest the opposite. Banks may be reluctant to pass on negative interest rates to depositors, leading to a compression in net interest margin and bank profitability (Boungou, 2020; Brunnermeier and Koby, 2018; Heider et al., 2019; Molyneux et al., 2019). Regarding bank risk-taking behaviours, some studies report that NIRP is associated with increased bank risk (Heider et al., 2019; Hong and Kandrac, 2021). Conversely, other evidence indicates that banks subject to NIRP may reduce their risk-taking activities (Bongiovanni et al., 2021).

Overall, the evidence is mixed regarding whether NIRP has been beneficial. In addition, there is limited research examining NIRP’s specific impact on bank credit risk-taking. Despite the increasing number of papers on the NIRP’s impacts on bank risk-taking, most studies discuss the influence on undercapitalisation risk (Nucera et al., 2017) or risk-taking on an aggregate level, encompassing credit risk along with other relevant risk categories (Bongiovanni et al., 2021; Hong and Kandrac, 2021).

Given its critical role in predicting bank collapses and financial crises (Bedendo and Bruno, 2012; Chang and Chen, 2016; Jiménez et al., 2014), investigating bank credit risk-taking under a negative interest rate environment is an indispensable area of study. On the one hand, bank credit risk-taking is closely linked to credit supply, which is directly influenced by monetary policy decisions. Examining changes in banks’ credit risk appetite offers valuable insights into the effectiveness of monetary transmission mechanisms under NIRP regimes. On the other hand, lessons from past financial crises highlight the importance of bank credit risk-taking in maintaining financial market stability. NIRP may impair bank profitability, leading banks to grant riskier loans in search of higher returns (Tang et al., 2025). Such a search-for-yield motive can raise credit risk and weaken bank stability.

We investigate the effectiveness of NIRP by examining the following research questions. Does NIRP affect bank credit risk-taking behaviours? What is the role of the business cycle in affecting bank credit risk-taking? How do bank- and country-specific characteristics influence the impact of NIRP on credit risk-taking? We answer these questions by employing a triple difference (TD) methodology and datasets comprising 1958 banks from 29 member countries of the Organisation for Economic Cooperation and Development (OECD) over the 2011-2017 period. The datasets include country- and bank-level information, respectively collected from World Bank databases and Bank Focus. Our results show that NIRP has a negative effect on banks’ credit risk-taking by reducing, on average, 8.1% of loan loss provisions to total loans (LLPTL) and 13% of loan loss reserves to total loans (LLRTL). Notably, our subsample analyses suggest that small banks operating in low-inflation economies experience more contraction in LLPTL and LLRTL following the adoption of NIRP.

We contribute to the empirical literature on NIRP’s impact on bank credit risk-taking in four ways. First, while prior research has studied aggregate bank risk-taking and undercapitalisation risk, only a few studies have focused on bank credit risk-taking. Credit risk-taking measures the risks arising from the credit supply, while the primary objective of NIRP is to incentivise banks to drain down their excess reserve balances and stimulate the credit supply in the financial market (Molyneux et al., 2019). From a regulatory standpoint, scrutinising bank risk-taking is to prevent potential bank failures (Calem and Rob, 1999). It is thus crucial to examine the impact of NIRP on credit risk-taking to assess whether its implementation could pose potential threats to the banking sector. Our results suggest that the transmission of NIRP to banks and financial intermediaries may not be effective. While studies on conventional monetary policies indicate that lower positive policy rates can motivate banks to take on more credit risk by granting loans to riskier firms (Diamond and Rajan, 2006; Jiménez et al., 2014), we show that this effect does not hold under NIRP.

Second, many empirical studies evaluate the influences of NIRP by employing the Difference-in-Differences (DiD) model, which is fundamentally based on the parallel trends assumption. However, we find that the validity of the parallel trends assumption is questionable. Therefore, we adopt TD models that relax the parallel trends condition. Furthermore, by taking one more time-wise difference than TD, we utilise the quadruple difference (QD) model to verify the findings from the TD model. Our analysis suggests that the results of the TD model are credible.

Third, to control for the potential cyclical nature of banks’ provisioning (Laeven and Majnoni, 2003; Silva, 2021), we employ a measure of macroeconomic conditions to detrend the business cycle and isolate the exogenous shock. The results suggest that, within our sample period, the negative NIRP effect may be biased due to the cyclicality of LLPTL and LLRTL. We disentangle the impact of exogenous monetary shocks to estimate the true impact of NIRP on credit risk-taking and to provide a more robust evaluation of this unconventional monetary policy. To the best of our knowledge, our study may be the first one that disentangles the impacts of the business cycle and NIRP on credit risk-taking.

Fourth, we examine the heterogeneity of the NIRP effect on bank credit risk-taking. Our results suggest that smaller banks experience a more negative NIRP effect on LLPTL. In addition, banks with greater reliance on deposits as a funding source are likely to experience a greater contraction in LLPTL and LLRTL after the adoption of NIRP. Finally, we document that NIRP has a more significant impact on banks located in countries with lower inflation.

The rest of the paper is structured as follows. Section 2 reviews the literature. Section 3 presents the data and methodology. Section 4 analyses results with robustness discussions, and Section 5 concludes the paper.

2. Literature review

2.1. Conventional monetary policy and bank risk-taking

Since NIRP was implemented for a relatively short time, we rely on established theories in the context of conventional monetary policies. Some suggest that the interest rate is negatively correlated with bank credit risk. To investigate the connection between money, banks, and aggregate credit, Diamond and Rajan (2006) examine the ‘liquidity version of the lending channel’ and find that the mismatch of the demand and supply of liquidity exposes banks to monetary policy surprises. When banks utilise liquid deposits as a source of illiquid long-term loans, the aggregate liquidity condition with monetary intervention would impact banks’ risk appetite. Specifically, an expansionary monetary policy can improve the overall liquidity in financial markets, and banks with high liquidity are willing to grant more risky loans. Jiménez et al. (2014) empirically investigate the relationship between bank credit risk and expansionary monetary policies. The findings suggest that lower overnight rates motivate banks to ease lending standards and supply loans to borrowers with a bad or no credit history. Lower interest rates would increase a firm’s net worth so that a borrower with fewer pledgeable assets is viewed as less risky compared to the past. Besides borrowers’ financial condition, banks’ own net worth, which changed along with interest levels, can also explain this phenomenon.

However, an opposite view also prevails. For instance, Smith (2002) indicates that a higher interest rate level naturally increases the opportunity cost for banks to hoard cash, which motivates them to pursue riskier alternatives. In addition, there are divergent opinions on the effect of banks’ net worth. It is suggested that the reduced net worth, when interest rates are high, may make a ‘gambling and resurrection’ strategy more attractive to a bank (Hellmann et al., 2000; Kane, 1989). Therefore, banks may be less incentivised to pursue a riskier portfolio given a lower interest rate level. Overall, the inconclusive findings suggest a need for more empirical studies on monetary policies and bank credit risk-taking.

2.2. NIRP and bank risk-taking

Despite a vigorous discussion about the connection between the interest rate level and bank risk-taking, the adoption of NIRP has called for a new evaluation of the monetary transmission mechanisms in a negative interest environment. Figure 1 represents the policy rates of five NIRP-adopter economies. The experience with unconventional monetary policies in these economies reveals both benefits and potential side effects. By charging interest on reserves, central banks manage to lower and flatten the yield curve to cut borrowing costs for households and firms to stimulate aggregate demand. Nevertheless, this mechanism puts pressure on bank profitability (Dell’Ariccia et al., 2018; Molyneux et al., 2019).

Figure 1.

Official interest rates of NIRP adopters.

Many papers have pointed out the uniqueness of the transmission of NIRP. Negative interest rates may make banks reluctant to either pass through to retail depositors or lend with a ‘cost of lending’. If there are no special frictions unique to NIRP, we should observe the interest rate transmission through usual channels. For example, lower positive official interest rates set by central banks transmit to lower rates on both deposits and market-based debt. However, Heider et al. (2019) show that this is not the case anymore for deposit rates after the implementation of NIRP. Theoretically, negative return deposits would become inferior to cash, which has a zero nominal return, and depositors would withdraw.

Given the unique features of NIRP and the incomplete pass-through, over the past decade, there has been an increasing number of papers exploring the influences of NIRP on bank risk-taking. Should NIRP affect bank risk-taking, it is possible that banks’ credit risk-taking behaviours may have been similarly affected.

Existing evidence on NIRP’s impact on bank risk-taking remains inconclusive. Heider et al. (2019) document that negative policy rates have a significant impact on the nexus between the deposit-to-asset ratio and risk-taking. With a DiD model, they find that negative interest rates motivate banks with more deposits to grant loans to riskier firms. On the contrary, Nucera et al. (2017) suggest that there is a beneficial effect of NIRP on bank risk and NIRP has led to a lower risk of undercapitalisation. Bongiovanni et al. (2021) support the de-leverage hypothesis, which posits that banks are currently in recovery from the GFC and European sovereign debt crisis, and they are still focusing on repairing capital positions by investing in safer assets. Banks are thus reluctant to take excess risk under a negative interest rate environment, where they perceive greater potential threats.

A few studies have explored the impact of NIRP on bank credit risk-taking, with findings suggesting that NIRP may have an adverse effect (Boungou, 2020; Tang et al., 2025). Under NIRP, the main debate concerns whether the de-leverage motive or the search-for-yield motive dominates. The search-for-yield motive aligns with the bank risk-taking channel, suggesting that lower interest rates encourage banks to pursue riskier portfolios. In contrast, the de-leverage motive implies the opposite, as highlighted by Bongiovanni et al. (2021). Tang et al. (2025) identify the de-leverage motive as the main transmission channel, as they observe slower loan growth and no significant rise in loan rates, implying an absence of search-for-yield behaviour. Decreased loan growth may reflect changes on both the supply and demand sides. On the supply side, banks may deliberately reduce lending. On the demand side, borrowers may choose to repay their debts earlier due to lower interest rates. However, these results should be interpreted with caution, as they may be affected by the sample period or methodological choices. While a longer sample period provides richer data, it can also make it harder to isolate the effects of major events such as the GFC and COVID-19. In addition, the validity of DiD models relies on the parallel trends assumption, which requires careful examination to ensure the robustness of results.

Based on the above discussions, our first and second hypotheses are as follows:

H1: NIRP has a positive impact on bank credit risk-taking due to the search-for yield motive.

H2: NIRP has a negative impact on bank credit risk-taking due to the de-leverage motive.

2.3. Loan loss provisions (LLP) and bank credit risk-taking

LLP is one of the principal measures of bank credit risk-taking. Despite banks’ potential income-smoothing, LLP and loan loss reserves (LLR) constrained by regulations are arguably among the most important variables reflecting the expected credit losses and, thereby, bank credit risk-taking (Silva, 2021). Since the enforcement of the Basel I Capital Accord in 1988, it has been widely perceived that bank risk exposures should be reflected in the capital level if a bank pursues stable performance and avoids regulatory arbitrage (Laeven and Majnoni, 2003). In particular, there is a consensus that LLP and LLR, as part of the regulatory capital, are related to expected loan losses (Curcio and Hasan, 2015). In addition, unlike non-performing loans, LLP allows for estimating credit losses before they are actually realised. It is especially efficient in providing insights into assets that are often opaque (Beatty and Liao, 2014).

Given its effectiveness in capturing expected credit risk, many existing studies have employed LLP-related proxy variables for assessing bank credit risk-taking. Jiménez et al. (2017) investigate the impact of macroprudential policy on credit supply cycles and real effects by analysing the dynamic provisioning. Molyneux et al. (2019) use the ratio of LLP to total assets as a control variable to account for credit risk when examining the relationship between NIRP and bank profitability. Silva (2021) also utilises LLP to explore the impact of fiscal deficits on banks’ aggregate credit risk.

While LLP can be an effective proxy for bank credit risk-taking, some studies highlight the cyclical nature of banks’ provisioning behaviour. There have been debates about the manipulation of LLP as banks may smooth the earnings due to regulatory purposes or agency problems (Greenawalt and Sinkey, 1988). LLP, as a buffer, may thus follow the inherent cyclicality of loan market fluctuations. Banks tend to provision less during expansionary economic conditions and more when economic conditions deteriorate (Laeven and Majnoni, 2003). An opposite effect is also possible when banks provision more in expansions and less during downturns. We thereby include a new measure of macroeconomic shock to account for potential cyclicality embedded in LLP and detrend the business cycle. By detrending the business cycle, we expect LLP to reflect the bank credit risk-taking untainted by managers’ acts in response to existing macroeconomic conditions.

Taking into consideration banks’ potential income-smoothing, we have the third hypothesis as follows:

H3: The NIRP effect can be moderated due to the cyclicality of LLP.

2.4. Heterogeneous impact of NIRP

Prior studies suggest that banks with different characteristics in different countries exhibit rich heterogeneity in their responses to the NIRP (Bongiovanni et al., 2021; Molyneux et al., 2019, 2020). In terms of bank overall risk-taking in the context of NIRP, Nucera et al. (2017) show that fee-focused and large universal banks experienced reductions in the risk of a capital shortfall, while smaller banks with a more traditional business model had an increased risk of being undercapitalised. Bongiovanni et al. (2021) suggest that the effect of NIRP depends on bank capitalisation and market competition. In terms of the bank capitalisation, the capital channel of monetary policy suggests that banks’ responses to monetary shocks vary significantly with capitalisation (Dell’Ariccia et al., 2017; Van den Heuvel, 2002). Bongiovanni et al. (2021) find that banks with higher capitalisation levels and larger capital buffers see a greater increase in the volumes of risky assets after the adoption of NIRP. From a competition perspective, the loan market channel suggests that less competition across the banking sector may lead to less favourable terms on business loans, which could increase the default risk and result in greater bank instability.

Based on the above studies, our fourth hypothesis is as follows:

H4: The NIRP effects are heterogeneous among banks and countries.

3. Data and methodology

3.1. Data

Our sample covers 1958 financial institutions, including commercial banks, savings banks, cooperative banks, real estate and mortgage finance institutions, bank holding companies, finance companies, as well as fintech banks, from 29 OECD countries over 2011-2017.¹ The treated group consists of countries that introduced NIRP in 2014.

Many studies employing DiD models do not differentiate the variation in the timing of intervention implementation. However, varied treatment timing can lead to the use of an earlier treated unit as a control for a later treated unit, which may be deemed inappropriate in a DiD setting (Baker et al., 2022; Wooldridge, 2021). Therefore, to avoid the issue of staggered intervention, we do not include Denmark, Norway, Sweden, Switzerland, or Japan in the treated group. We intend to keep observations concentrated around 2014 when the treated countries in our sample adopted the NIRP regime, and at the same time, we try to enlarge the sample period and obtain more information than previous papers. On the one hand, if observations are too deviated from the intervention, other unobserved and omitted variables are likely to threaten the validity of the model (Bertrand et al., 2004; Roberts and Whited, 2013). On the other hand, this sample covers the longest period that we can include and enables us to alleviate the influence of the GFC and the Covid-19 pandemic.

We use public information on central banks’ official websites to identify the dates of adoption of NIRP in each country and to construct a dataset that differentiates pre- and post-adoption periods. The data on three macroeconomic indicators are collected from the World Bank Databases. The GDP growth and the inflation rate are in real terms, and the unemployment rate is the unemployment rate of the total labour force in a given year. Bank balance sheet data are from Bank Focus.²

Descriptive statistics for bank credit risk-taking variables, macroeconomic indicators, and other bank balance sheet variables are shown in Table 1. All variables are winsorised at 1% and 99% level to avoid the influence of outliers. Panels A and D of Table 1 describe bank credit risk-taking measured by proxy variables that are related to LLP and LLR. LLPTL is the ratio of bank loan loss provisions to total loans. Loan loss provisions to total assets (LLPTA) is the ratio of bank loan loss provisions to total assets. LLRTL is the ratio of bank loan loss reserves to total loans. Loan loss reserves to total assets (LLRTA) is the ratio of bank loan loss reserves to total assets. Non-performing loans to total loans (NPLTL) is the ratio of bank non-performing loans to total loans, which allows us to measure the realised credit risk. Given that NIRP may enhance borrowers’ debt repayment capacity, LLPTL and LLRTL naturally address this concern, as repaid loans proportionally reduce both LLP (or LLR) and total loans. Therefore, we use LLPTL and LLRTL as our main dependent variables, while LLPTA, LLRTA, and NPLTL serve as alternative measures for robustness checks.

Table 1.

Descriptive statistics of the treatment and control groups prior to and after the introduction of NIRP.

Variables	Treatment
	Pre-NIRP					Post-NIRP
	Obs.	Mean	Std. Dev.	Min	Max	Obs.	Mean	Std. Dev.	Min	Max
Panel A: Bank credit risk
LLPTL	2,877	0.427	1.269	−2.381	7.521	3,832	0.376	1.150	−2.381	7.521
LLPTA	2,889	0.243	0.692	−1.230	4.089	3,852	0.212	0.644	−1.230	4.089
NPLTL	1,345	7.168	8.337	0.000	43.778	2,414	6.294	8.881	0.000	43.778
LLRTL	1,113	4.296	5.709	0.006	27.170	1,484	4.698	6.274	0.006	27.170
LLRTA	1,113	2.091	2.269	0.000	10.759	1,484	2.245	2.553	0.000	10.759
Panel B: Bank balance sheet
Size	2,889	14.528	2.090	10.847	20.214	3,852	14.489	2.046	10.847	20.214
E/TA	2,889	9.505	7.177	1.800	82.176	3,852	10.206	6.872	1.800	82.176
ROE	2,889	6.226	8.472	−24.339	44.214	3,852	6.104	7.562	−24.339	44.214
Liquidity	2,889	21.562	16.775	0.388	83.145	3,848	21.343	17.342	0.388	83.145
Funding structure	2,885	89.017	17.704	0.000	99.653	3,846	90.540	16.958	0.003	99.653
Income structure	2,886	64.276	18.769	−2.523	111.524	3,846	61.859	18.987	−2.523	111.524
Loan growth	2,716	6.276	14.271	−29.165	102.663	3,833	3.158	19.355	−29.165	102.663
Total loans	2,877	13.920	2.117	9.851	19.461	3,832	13.909	2.080	9.851	19.461
Loan rate	2,458	0.149	0.596	0.014	4.686	3,271	0.123	0.535	0.014	4.686
Panel C: Macroeconomic indicators
GDP	2,889	1.006	1.668	−2.516	6.084	3,852	1.763	0.836	−0.487	6.084
Inflation	2,889	2.132	0.666	−0.094	5.652	3,852	0.870	0.602	−0.094	3.417
Unemployment	2,889	6.829	2.771	4.560	17.220	3,852	6.794	3.084	3.750	17.220
Variables	Control
	Pre-NIRP					Post-NIRP
	Obs.	Mean	Std. Dev.	Min	Max	Obs.	Mean	Std. Dev.	Min	Max
Panel D: Bank credit risk
LLPTL	2,898	0.655	1.120	−2.381	7.521	3,863	0.433	1.018	−2.381	7.521
LLPTA	2,985	0.388	0.632	−1.230	4.089	3,976	0.271	0.581	−1.230	4.089
NPLTL	2,303	2.471	3.600	0.000	43.778	3,098	1.566	3.591	0.000	43.778
LLRTL	3,041	1.962	2.428	0.006	27.170	4,056	1.462	2.150	0.006	27.170
LLRTA	3,135	1.143	1.326	0.000	10.759	4,180	0.884	1.106	0.000	10.759
Panel E: Bank balance sheet
Size	2,985	15.094	1.543	10.847	20.214	3,976	15.367	1.470	10.847	20.214
E/TA	2,981	13.359	13.444	1.800	82.176	3,952	13.244	12.703	1.800	82.176
ROE	2,981	12.081	8.695	−24.339	44.214	3,952	12.141	7.082	−24.339	44.214
Liquidity	2,978	25.554	15.469	0.388	83.145	3,967	21.698	14.447	0.388	83.145
Funding structure	2,964	84.488	22.415	0.000	99.653	3,946	85.270	22.600	0.000	99.653
Income structure	2,985	72.042	18.802	−2.523	111.524	3,980	72.939	17.869	−2.523	111.524
Loan growth	2,660	9.908	19.021	−29.165	102.663	3,866	11.941	17.473	−29.165	102.663
Total loans	2,898	14.506	1.605	9.851	19.461	3,863	14.848	1.546	9.851	19.461
Loan rate	2,478	0.131	0.511	0.014	4.686	3,306	0.113	0.480	0.014	4.686
Panel F: Macroeconomic indicators
GDP	2,973	2.104	0.896	−0.785	6.084	3,964	2.379	0.729	0.659	6.084
Inflation	2,973	2.424	1.078	0.938	7.493	3,964	1.543	1.306	−0.094	7.493
Unemployment	2,973	7.866	1.039	3.750	10.330	3,964	5.350	1.120	3.750	10.840

LLPTL is the ratio of bank loan loss provisions to total loans. LLPTA is the ratio of bank loan loss provisions to total assets. NPLTL is the ratio of non-performing loans to total loans. LLRTL is the ratio of bank loan loss reserves to total loans. LLRTA is the ratio of bank loan loss reserves to total assets. Size is the natural logarithm of the bank’s total assets. E/TA is the ratio of bank equity to total assets. ROE is the ratio of bank pre-tax profits to total equity. Liquidity is the ratio of bank liquid assets to total assets. Funding structure is the ratio of bank deposits and other short-term funding to total liabilities. Income structure is the ratio of bank net interest income to operating revenues. Loan growth is the yearly growth rate of total loans. Total loans is the natural logarithm of the bank’s total loans. Loan rate is the ratio of bank interest income to total loans. GDP is the yearly real growth rate of GDP. Inflation is the yearly inflation rate measured by the Consumer Price Index in real terms. Unemployment is the yearly unemployment rate of the total labour force. All variables are winsorised at the 1% and 99% level to avoid the influence of outliers.

Bank balance sheet data are presented in panels B and E of Table 1. Extensive literature suggests that bank size is associated with bank lending and provisioning (Abou El Sood, 2012; Berger and Black, 2011; Kishan and Opiela, 2000). Larger banks are capable of reaching more clients and granting a greater amount of loans, which means more LLP would be required if we presume banks are conservative and prudent. Therefore, we include the size variable measured by the natural logarithm of bank total assets to control for the size effect. E/TA is the ratio of bank equity to total assets. ROE is the ratio of bank pre-tax profits to total equity. Liquidity is the ratio of bank liquid assets to total assets. Loan growth is the yearly growth rate of total loans. These four variables reflect bank performance that affects the quality and riskiness of loans, thereby capturing credit-risk managers’ attitudes or appetite towards risk-taking (Greenawalt and Sinkey, 1988). We also include two variables that reflect the business model of a bank. Funding structure is the ratio of bank deposits and other short-term funding to total liabilities. Income structure is the ratio of bank net interest income to operating revenues. Olszak et al. (2017) investigate the role of bank specialisation in loan loss provisioning and find that LLP of commercial banks are more likely to be procyclical. Given the fact that the business model determines how a bank responds to macroeconomic conditions, we herein control for the funding and income structures to distinguish the influence of NIRP on banks by different operating characteristics. Finally, we also utilise total loans and the loan rate to examine possible transmission channels. Total loans is the natural logarithm of the bank's total loans. Loan rate is the ratio of bank interest income to total loans.

Macroeconomic data are shown in panels C and F of Table 1. Olszak et al. (2017) point out that banks’ provisioning is negatively associated with economic conditions measured by GDP and unemployment. Therefore, we include three indicators to reflect the macroeconomic conditions: GDP is the yearly real growth rate of GDP; inflation is the yearly inflation rate measured by the Consumer Price Index in real terms; unemployment is the yearly unemployment rate of the total labour force.

3.2. Methodology

The adoption of NIRP by the European Central Bank (ECB) presents an opportunity to study the impact of this monetary policy. Many existing papers on this topic employ a DiD methodology (Bongiovanni et al., 2021; Hong and Kandrac, 2021; Molyneux et al., 2019, 2020), which allows for dealing with the omitted variable bias and thus alleviating the endogeneity problem by differencing away common trends that affect both the treated and control group. For instance, changes in the regulatory environment, such as Basel III post-crisis regulatory reforms, may similarly influence banks’ behaviours in treated and untreated countries.

However, previous studies assume the parallel trends without a robust check. It should be noted that the DiD model requires the untreated response changes be equivalent across the treated and control groups if control units serve as a counterfactual scenario for treated units (Goodman-Bacon, 2021; Lee, 2016; Wooldridge, 2021). In the absence of sufficient evidence demonstrating that the pre-treatment gap between the treated group and the control group is constant, the practice of taking a difference between the post-treatment gap and pre-treatment gap (i.e. the DiD model) will lead to biased estimates of the NIRP effect. Therefore, we follow Khandelwal et al. (2013), Lee (2016), and Demir et al. (2017), and adopt a TD model that takes one more time-wise difference than the DiD model. While the DiD model requires equivalence in the changes between the treated and control groups, should treatment not occur, the TD model relaxes this requirement, but instead assumes common changes in changes between the two groups.³ It is worth emphasising that if the parallel trends condition holds, it implies that the common changes in changes between the two groups. In this sense, TD models nest DiD models.

We illustrate an example of the presence of a non-zero but constant time effect in Figure 2 and suppose that a treatment is applied at the beginning of t = 3. We allow for a constant, but not zero, time effect for the variable $Y$ as

Y_{1}^{T} - Y_{0}^{T} \neq Y_{1}^{C} - Y_{0}^{C}

and (Y_{2}^{T} - Y_{1}^{T}) - (Y_{1}^{T} - Y_{0}^{T}) = (Y_{2}^{C} - Y_{2}^{C}) - (Y_{1}^{C} - Y_{0}^{C})

or equivalently, Δ Y_{2}^{T} - Δ Y_{1}^{T} = Δ Y_{2}^{C} - Δ Y_{1}^{C}

where $Y_{t}^{T}$ denotes $Y$ of the treated unit at the time $t$ , and $Y_{t}^{C}$ denotes $Y$ of the control unit at the time $t$ .

Figure 2.

Constant non-zero time effects in the TD model.

Then, in the post-treatment period, we can derive the unobserved counterfactual outcome $Y_{3}^{T^{'}}$ by the following equation,

Δ Y_{3}^{T^{'}} - Δ Y_{2}^{T} = Δ Y_{3}^{C} - Δ Y_{2}^{C}

and calculate the treatment effect as

β_{T D} = Y_{3}^{T} - Y_{3}^{T^{'}}

In a DiD regression model, we estimate the treatment effect $β_{D i D}$ with the following regression equation (Bongiovanni et al., 2021; Molyneux et al., 2019, 2020):

Y_{i j t} = α + β_{D i D} D_{i j t} + β_{1} X_{i j t} + β_{2} G_{i j t} + γ_{j} + δ_{t} + e_{i j t}

where $Y_{i j t}$ is the credit risk-taking variable of the bank $i$ in country $j$ at time $t$ , $D_{i j t}$ is a dummy variable that equals 1 if a bank has been affected by the NIRP, $X_{i j t}$ is included as a vector of country and bank characteristics, $G_{i j t}$ is the $D_{i j t}$ -interacting elements of $X_{i j t}$ , and $γ_{j}$ and $δ_{t}$ respectively represent country and time fixed effects. The treatment effects are represented by $β_{D i D}$ and $β_{2}$ . While $β_{D i D}$ represents the average treatment effect over all countries, $β_{2}$ represents changes in the marginal effect of $X_{i j t}$ on $Y_{i j t}$ due to the treatment.

With one more time-wise differencing than DiD, the TD model can be constructed as follows:

Δ Y_{i j t} = β_{T D} D_{i j t} + β_{1} Δ X_{i j t} + β_{2} Δ G_{i j t} + Δ δ_{t} + Δ e_{i j t}

Formally, we propose our TD model as follows.

Δ C R_{i j t} = β_{T D} (T r e a t m e n t_{i j} * P o s t_{j t}) + β_{1} Δ X_{i j t} + β_{2} Δ G_{i j t} + δ_{t} + e_{i j t}

(1)

$Δ C R_{i j t}$ is credit risk-taking proxied by changes in LLPTL or LLRTL of the bank $i$ in country $j$ at time $t . T r e a t m e n t_{i j}$ is a dummy variable that equals 1 if NIRP affects the bank $i$ in country $j$ and 0 instead. $P o s t_{j t}$ is a dummy variable that equals 1 if country $j$ at time $t$ is after the period that they decided to adopt a negative interest rate policy (2014) and 0 prior to the period.⁴ $β_{TD}$ is the treatment effect denoting the average difference in changes in LLPTL or LLRTL between non-NIRP-adopters and NIRP-adopters. To capture cross-country and cross-bank heterogeneity that also has effects on credit risk, $Δ X_{i j t}$ is included as a vector of country and bank characteristics. Specifically, country-specific variables are macroeconomic performance indicators, while bank-specific variables include the balance sheet and key financial measures. $Δ G_{i j t}$ is the interaction term of $T r e a t m e n t_{i j} * P o s t_{j t}$ and $Δ X_{i j t}$ . As a consequence of the additional time-wise difference, this TD model obviates country-specific dummies. We keep $δ_{t}$ that represents changes in fixed effects that control for time-variant shocks over the sample period on bank credit risk-taking.

3.3. Plausibility of parallel trends

We examine the two assumptions for the traditional DiD model and the TD model. First, the parallel trends assumption states that, although we allow for a pre-treatment gap between treatment and control groups, the trends of bank credit risk-taking variables in the two groups should be identical in the absence of NIRP. For the DiD model, as illustrated in Figures 3 and 4, the pre-intervention trends observed in LLPTL and LLRTL across the two groups do not appear to exhibit alignment.

Figure 3.

LLPTL before and after the NIRP.

Figure 4.

LLRTL before and after the NIRP.

To further validate the common changes of trends condition, we follow Lee (2016) and conduct identification tests for the DiD and the TD models using only the pre-intervention periods. Table 2 presents the results of identification tests. In the DiD setting (columns 1 and 3), the coefficient on Treatment for LLPTL is statistically significant, while that for LLRTL is not. This result suggests that the NIRP explains the changes in LLPTL before the intervention. In other words, the pre-intervention responses of LLPTL change differently across the treated and control groups. This provides further evidence on the violation of the parallel trends assumption for LLPTL in the DiD model. In the TD setting (columns 2 and 4), the coefficients of Treatment are not statistically significant. Therefore, the NIRP does not account for the changes in LLPTL and LLRTL before 2014, implying a possible non-zero but constant time effect. Overall, these findings support the use of the TD model over the DiD model.⁵

Table 2.

DiD and TD identification tests.

	(1) $Δ L L P T L_{2012, 2013}$	(2) $Δ^{2} L L P T L_{2013}$	(3) $Δ L L R T L_{2012, 2013}$	(4) $Δ^{2} L L R T L_{2013}$
Treatment	0.432*** (0.105)	0.202 (0.228)	0.086 (0.135)	0.105 (0.328)
$Δ$ Size	0.369 (0.231)	0.166 (0.611)	−0.668*** (0.254)	−1.946*** (0.523)
$Δ$ E/TA	−0.017 (0.020)	−0.016 (0.037)	−0.035 (0.028)	−0.042 (0.045)
$Δ$ ROE	−0.044*** (0.007)	−0.031*** (0.009)	−0.010*** (0.004)	−0.007 (0.006)
$Δ$ Liquidity	−0.015** (0.006)	0.001 (0.010)	0.004 (0.005)	0.022** (0.009)
$Δ$ Funding structure	−0.002 (0.011)	−0.031* (0.017)	0.004 (0.010)	−0.010 (0.016)
$Δ$ Income structure	−0.009 (0.005)	0.006 (0.008)	−0.005 (0.004)	−0.002 (0.006)
$Δ$ Loan growth	−0.004*** (0.002)	−0.002 (0.003)	−0.003 (0.002)	0.001 (0.002)
$Δ$ GDP	−0.032 (0.028)	−0.046 (0.056)	−0.015 (0.029)	−0.059 (0.038)
$Δ$ Inflation	0.104* (0.058)	0.039 (0.095)	0.043 (0.059)	−0.003 (0.094)
$Δ$ Unemployment	0.051 (0.061)	−0.351 (0.244)	0.120* (0.068)	0.086 (0.190)
Treatment $\times Δ$ Size	−1.052** (0.456)	−0.042 (1.125)	−0.373 (0.617)	0.015 (1.200)
Treatment $\times Δ$ E/TA	−0.175*** (0.042)	−0.107 (0.073)	0.028 (0.051)	−0.027 (0.064)
Treatment $\times Δ$ ROE	0.023** (0.010)	0.010 (0.015)	0.001 (0.008)	−0.003 (0.011)
Treatment $\times Δ$ Liquidity	0.021** (0.010)	−0.017 (0.018)	0.035*** (0.012)	−0.003 (0.024)
Treatment $\times Δ$ Funding structure	0.018 (0.015)	0.052** (0.026)	0.006 (0.015)	0.028 (0.023)
Treatment $\times Δ$ Income structure	0.007 (0.008)	−0.013 (0.013)	0.001 (0.006)	0.009 (0.009)
Treatment $\times Δ$ Loan growth	−0.002 (0.003)	−0.009 (0.006)	0.005 (0.004)	−0.002 (0.005)
Treatment $\times Δ$ GDP	−0.016 (0.044)	−0.134 (0.111)	0.105** (0.051)	0.152* (0.090)
Treatment $\times Δ$ Inflation	0.101 (0.096)	−0.028 (0.198)	−0.094 (0.127)	0.158 (0.223)
Treatment $\times Δ$ Unemployment	−0.073 (0.073)	0.153 (0.346)	0.218** (0.088)	−0.264 (0.306)
Observations	3,422	1,512	2,321	981
Adj. $R^{2}$	0.122	0.070	0.144	0.069

Treatment is a dummy variable that equals 1 if the NIRP is implemented in a country and 0 otherwise. LLPTL is the ratio of bank loan loss provisions to total loans. LLRTL is the ratio of bank loan loss reserves to total loans. Size is the natural logarithm of the bank's total assets. E/TA is the ratio of bank equity to total assets. ROE is the ratio of bank pre-tax profits to total equity. Liquidity is the ratio of bank liquid assets to total assets. Funding structure is the ratio of bank deposits and other short-term funding to total liabilities. Income structure is the ratio of bank net interest income to operating revenues. Loan growth is the yearly growth rate of total loans. GDP is the yearly real growth rate of GDP. Inflation is the yearly inflation rate measured by the Consumer Price Index in real terms. Unemployment is the yearly unemployment rate of the total labour force. $Δ$ denotes the change in variables, whereas $Δ^{2}$ denotes the change in the change in variables. For TD identification tests (columns 2 and 4), covariates have $Δ^{2}$ , not $Δ$ . ***, **, and * indicate statistical significance at 1%, 5%, and 10%, respectively.

3.4. Validity of counterfactual controls

The second requirement for the DiD and the TD model is the counterfactual assumption that requires us to examine whether the control group constitutes a valid counterfactual scenario. Ideally, if we can observe the same bank in a counterfactual world and compare the two banks’ credit risk-taking, we are able to estimate what would have happened in the counterfactual scenario, which in turn shows the causal effect of NIRP. As it is impossible for one bank to be in and out of treatment at once, existing studies typically try to build two groups that are ‘equivalent’ to each other (Schwerdt and Woessmann, 2020). Based on this logic, we estimate Pearson correlation coefficients for inflation rate, unemployment rate, and GDP growth, which are three macroeconomic variables in the treated and control groups. As shown in Table 3, the results of GDP and inflation are statistically significant, suggesting that the macroeconomic environment for the two groups is similar and that the control group does serve as a valid counterfactual group. In addition, we provide a further robustness check for this counterfactual assumption. By using propensity score matching (PSM), each bank in the treatment group is matched to a control unit with similar characteristics. Our PSM-TD model is provided in Section 4.2.1.

Table 3.

Pearson correlation tests for macroeconomic variables in control and treatment groups.

	Pearson corr.
GDP	0.7395***
Inflation	0.7573***
Unemployment	0.2863***

This table shows three macroeconomic indicators and the Pearson correlation test for the control and treatment groups during the period 2007-2020. We arbitrarily chose a longer time period (in comparison with the sample period) to highlight that these macroeconomic indicators move together for several years after the GFC. ***, **, and * indicate statistical significance at 1%, 5%, and 10%, respectively.

4. Empirical results

4.1. Baseline results

We estimate NIRP’s effects on bank credit risk-taking using the TD model specified in equation (1). Table 4 displays the baseline results (columns 3 and 6). Average differences in credit risk-taking variables between NIRP adopters and non-adopters are measured by the ‘NIRP effect’, which is an interaction of the Treatment dummy variable and the Post dummy variable. Banks that have been affected by NIRP reduce their LLPTL and LLRTL by 8.1% and 13%, respectively. Overall, the results are consistent with our second hypothesis, which suggests that banks in NIRP-adopter countries are less willing to take credit risk.

Table 4.

The NIRP effect on LLPTL and LLRTL.

	(1) $Δ$ LLPTL	(2) $Δ$ LLPTL	(3) $Δ$ LLPTL	(4) $Δ$ LLRTL	(5) $Δ$ LLRTL	(6) $Δ$ LLRTL
NIRP effect	−0.065*** (0.011)	−0.042** (0.018)	−0.081*** (0.022)	0.057 (0.038)	−0.071** (0.036)	−0.130*** (0.045)
$Δ$ Size		0.090(0.106)	0.116(0.114)		−0.470*** (0.114)	−0.335*** (0.104)
$Δ$ E/TA		−0.074*** (0.013)	−0.062*** (0.012)		−0.007(0.013)	−0.009(0.012)
$Δ$ ROE		−0.037*** (0.004)	−0.031*** (0.005)		−0.010*** (0.002)	−0.009*** (0.003)
$Δ$ Liquidity		−0.001(0.003)	−0.003(0.003)		0.011*** (0.003)	0.008*** (0.003)
$Δ$ Funding structure		0.006(0.004)	0.006(0.005)		0.009*** (0.004)	0.008** (0.004)
$Δ$ Income structure		−0.005** (0.002)	−0.002(0.003)		−0.004** (0.002)	−0.004** (0.002)
$Δ$ Loan growth		−0.003*** (0.001)	−0.003*** (0.001)		−0.002*** (0.001)	−0.003*** (0.001)
$Δ$ GDP		−0.076*** (0.011)	−0.091*** (0.013)		0.033*** (0.010)	0.027** (0.011)
$Δ$ Inflation		−0.006(0.016)	0.007(0.016)		−0.036** (0.018)	−0.040** (0.019)
$Δ$ Unemployment		0.066*** (0.015)	0.065*** (0.018)		0.223*** (0.022)	0.246*** (0.025)
NIRP effect $\times Δ$ Size			−0.196(0.235)			−0.596(0.495)
NIRP effect $\times Δ$ E/TA			−0.039(0.031)			0.000(0.043)
NIRP effect $\times Δ$ ROE			−0.014** (0.007)			−0.003(0.005)
NIRP effect $\times Δ$ Liquidity			0.009(0.005)			0.012(0.009)
NIRP effect $\times Δ$ Funding structure			0.001(0.008)			0.003(0.009)
NIRP effect $\times Δ$ Income structure			−0.007* (0.004)			0.000(0.005)
NIRP effect $\times Δ$ Loan growth			0.001(0.002)			0.001(0.002)
NIRP effect $\times Δ$ GDP			0.105*** (0.025)			0.017(0.031)
NIRP effect $\times Δ$ Inflation			0.017(0.034)			−0.040(0.080)
NIRP effect $\times Δ$ Unemployment			0.014(0.038)			−0.168* (0.089)
Observations	11,542	11,039	11,039	8,307	7,660	7,660
Adj. $R^{2}$	0.002	0.095	0.101	0.008	0.075	0.077

NIRP effect is the interaction between the Treatment dummy variable and the Post dummy variable. It takes the value of 1 if the bank i in country j has been affected by the NIRP after the monetary shock and 0 otherwise. LLPTL is the ratio of bank loan loss provisions to total loans. LLRTL is the ratio of bank loan loss reserves to total loans. Size is the natural logarithm of the bank's total assets. E/TA is the ratio of bank equity to total assets. ROE is the ratio of bank pre-tax profits to total equity. Liquidity is the ratio of bank liquid assets to total assets. Funding structure is the ratio of bank deposits and other short-term funding to total liabilities. Income structure is the ratio of bank net interest income to operating revenues. Loan growth is the yearly growth rate of total loans. GDP is the yearly real growth rate of GDP. Inflation is the yearly inflation rate measured by the Consumer Price Index in real terms. Unemployment is the yearly unemployment rate of the total labour force. ∆ denotes the change in variables. NIRP Effect × Covariates represent interaction terms between the NIRP dummy and covariates. All regressions include fixed time effects. Robust standard errors clustered by banks in parentheses. ***, **, and * indicate statistical significance at 1%, 5%, and 10%, respectively.

In terms of bank balance sheet covariates, E/TA, ROE, and loan growth are significantly associated with LLPTL. All bank-related covariates except E/TA have significant relationships with LLRTL. Bank size is negatively associated with LLRTL, suggesting that smaller banks set aside higher reserves. This may indicate that smaller banks engage in greater credit risk-taking in pursuit of higher returns. The E/TA variable is negative and significant for LLPTL, implying that leveraged banks with lower capital ratios tend to have more motivations to grant riskier loans (Bongiovanni et al., 2021). ROE, as a profitability-related variable, exhibits a negative association with LLPTL and LLRTL, suggesting that banks tend to have more provisions and reserves when their revenues are impaired. This also indicates that banks may not manipulate their provisioning to smooth the revenue. Therefore, our finding does not support the ‘income-smoothing hypothesis’ which indicates banks’ tendency to save a proportion of their earnings as a provision for future loan losses–the notion of saving for the rainy days (Greenawalt and Sinkey, 1988; Laeven and Majnoni, 2003). The liquidity variable is positive and significant for LLRTL, indicating that banks with higher liquidity ratios tend to take more credit risk. This may suggest less prudent credit risk management when sufficient funds are available. The funding structure is positively correlated with the LLRTL, revealing that banks with a higher proportion of deposits and other short-term funding in their liabilities will have more reserves accumulated. The income structure variable is negative and significant for the LLRTL, suggesting that banks relying on interest income have fewer reserves. In other words, this implies that banks with diversified business models exhibit a greater appetite for credit risk-taking. Loan growth is negatively associated with both LLPTL and LLRTL, suggesting that low-growth banks may engage in a potential ‘gambling for resurrection’ strategy (Hellmann et al., 2000; Kane, 1989).

Among the macroeconomic indicators, unemployment has a significant and positive association with both provisions and reserves. Such a relationship supports the procyclicality hypothesis, suggesting that LLPTL and LLRTL increase during economic downturns and decrease during economic upswings (Olszak et al., 2017).

To further identify whether the search-for-yield or de-leverage transmission channel dominates, we test the impact of NIRP on bank loan rate and total loans, respectively. The dependent variables in Equation (1) are replaced with the loan rate and total loans. Table 5 reports the results. Following the NIRP, banks experienced a reduction in loan rates of approximately 31.4%, indicating a shift towards safer lending. This finding does not support the search-for-yield hypothesis, which suggests that banks pursue riskier loans for higher profits. In addition, total loans decreased by around 1.4%. Despite the modest change, this result aligns with the de-leverage hypothesis, which posits that banks may prioritise repairing their capital base and cleansing balance sheets after the NIRP.

Table 5.

The NIRP effect on the loan rate and total loans.

	(1) $Δ$ Loan rate	(2) $Δ$ Total loans
NIRP effect	−0.314*** (0.106)	−0.014*** (0.003)
Observations	9,452	11,039
Adj. $R^{2}$	0.140	0.835

NIRP effect is the interaction between the Treatment dummy variable and the Post dummy variable. It takes the value of 1 if the bank $i$ in country $j$ has been affected by the NIRP after the monetary shock and 0 otherwise. Loan rate is the ratio of bank interest income to total loans. Total loans is the natural logarithm of the bank's total loans. $Δ$ denotes the change in variables. Both regressions include fixed time effects. Robust standard errors clustered by banks in parentheses. ***, **, and * indicate statistical significance at 1%, 5%, and 10%, respectively.

4.2. Robustness tests

4.2.1 PSM-TD model

We provide PSM as a further robustness check for the counterfactual assumption. The predicted probability of a country adopting negative policy rates is estimated by a probit model. This paper utilises GDP growth, inflation rate, and unemployment rate to pair banks in NIRP adopter countries with those in non-NIRP adopter countries. To successfully control for bank-specific differences between treatment and control groups with the propensity score, this PSM model also includes all the bank-specific control variables to estimate the propensity score. The PSM model is as follows:

p_{i} = \Pr (D_{i} = 1 | X_{i j t}) = F ({X^{'}}_{i j t} β)

(2)

$D_{i}$ is a dummy variable that equals 1 if NIRP has influenced the bank and 0 otherwise. $X_{i j t}$ is a vector of bank- and country-specific control variables. $F (\cdot)$ is a standard normal cumulative distribution function.⁶

The probit model results for PSM, from which we get the propensity scores of banks affected by the NIRP, are shown in Table 6. Except for bank size and funding structure, all bank-specific and macroeconomic covariates are statistically significant for both LLPTL and LLRTL. Overall, banks characterised by lower capitalisation, profitability, and liquidity, along with lower reliance on interest income and slower loan growth, are prone to being affected by NIRP. Since a bank’s performance is unlikely to influence policymakers’ decision-making process, a proper way of interpreting the results would be that countries with banks with these characteristics are likely to adopt the NIRP. Among the macroeconomic indicators, inflation is negative, suggesting that countries with a lower level of inflation have a greater probability of being affected by the NIRP. GDP and unemployment variables are positive, signifying that countries with higher GDP growth and unemployment rates are more likely to implement the NIRP. Overall, except for the GDP variable, the results are consistent with the consensus that suitable NIRP adopters are countries where banks are faced with weaker economic prospects.

Table 6.

PSM-probit model test results.

	(1)LLPTL	(2)LLRTL
Size	−0.213*** (0.009)	−0.020(0.014)
E/TA	−0.034*** (0.003)	−0.024*** (0.003)
ROE	−0.051*** (0.002)	−0.024*** (0.002)
Liquidity	−0.011*** (0.001)	−0.009*** (0.001)
Funding structure	0.004** (0.001)	−0.002(0.001)
Income structure	−0.022*** (0.001)	−0.011*** (0.001)
Loan growth	−0.008*** (0.001)	−0.006*** (0.001)
GDP	0.117*** (0.015)	0.226*** (0.026)
Inflation	−1.060*** (0.021)	−1.041*** (0.029)
Unemployment	0.036*** (0.007)	0.221*** (0.011)
Observations	12,965	9,026
Pseudo R-squared	0.3790	0.427
Log Likelihood	−4885.1007	−2353.0006
LR test (chi-square)	5963.90	3500.95

LLPTL is the ratio of bank loan loss provisions to total loans. LLRTL is the ratio of bank loan loss reserves to total loans. Size is the natural logarithm of the bank's total assets. E/TA is the ratio of bank equity to total assets. ROE is the ratio of bank pre-tax profits to total equity. Liquidity is the ratio of bank liquid assets to total assets. Funding structure is the ratio of bank deposits and other short-term funding to total liabilities. Income structure is the ratio of bank net interest income to operating revenues. Loan growth is the yearly growth rate of total loans. GDP is the yearly real growth rate of GDP. Inflation is the yearly inflation rate measured by the Consumer Price Index in real terms. Unemployment is the yearly unemployment rate of the total labour force. Robust standard errors clustered by banks in parentheses. ***, **, and * indicate statistical significance at 1%, 5%, and 10%, respectively.

Utilising propensity scores and weights derived from the probit model, we re-estimate the NIRP effects with a weighted regression. This approach ensures observations are weighted to achieve comparability based on observed bank and macroeconomic characteristics, which is similar to the inverse probability of treatment weighting (Stuart et al., 2014). The results of the PSM-TD model are presented in Table 7. The impact of NIRP on matched banks’ loan loss provisioning is negative and significant, while the result for LLRTL is not statistically significant. NIRP reduces LLPTL, which is in line with our results in Table 4. More importantly, after the matching, the NIRP effect is more sizable than our baseline results. In that sense, banks with similar characteristics in a similar macroeconomic environment witness a more negative impact of NIRP.

Table 7.

PSM-TD model test results.

	(1) $Δ$ LLPTL	(2) $Δ$ LLRTL
NIRP effect	−0.213*** (0.066)	−0.132(0.133)
Observations	11,039	7,660
Adj. $R^{2}$	0.128	0.205

NIRP effect is the interaction between the Treatment dummy variable and the Post dummy variable. It takes the value of 1 if the bank $i$ in country $j$ has been affected by the NIRP after the monetary shock and 0 otherwise. LLPTL is the ratio of bank loan loss provisions to total loans. LLRTL is the ratio of bank loan loss reserves to total loans. $Δ$ denotes the change in variables. Both regressions include fixed time effects. Robust standard errors clustered by banks in parentheses. ***, **, and * indicate statistical significance at 1%, 5%, and 10%, respectively.

4.2.2 Detrending the business cycle

We are faced with the endogeneity issue that the NIRP effect on bank credit risk-taking is biased due to the cyclicality of LLPTL and LLRTL. In other words, the change in provisions and reserves might simply be a result of the business cycle rather than being influenced by NIRP. To address this concern, we follow Romer and Romer (2004) and include a Taylor-like regression specified in equation (3) below.

P R_{j t} = α + β_{1} G D P_{j t} + β_{2} I n f l a t i o n_{j t} + β_{3} U n e m p l o y m e n t_{j t} + ε_{j t}

(3)

$P R_{j t}$ denotes the policy rate in the country $j$ at time $t$ presented as a function of GDP growth, inflation, and unemployment. Since most of the central banks only adjust the official interest rates irregularly, we employ the policy rate data at the end of each year. The policy rate at the time $t$ is thus conceived as policymakers’ opinion on macroeconomic conditions at the same period.⁷ By taking the residuals of this regression, the resulting series are able to capture the unsystematic changes in central banks’ intentions for policy rates, which is an exogenous component in policymakers’ decision-making process. Controlling for it in our baseline regressions, we are able to detrend the business cycle.

The results are displayed in Table 8. We find that the NIRP effect with business cycle detrended is still negative and statistically significant for LLPTL and LLRTL at 1% level. Moreover, the NIRP effects are more pronounced in this model, reflecting the response of banks’ provisioning to business cycle trends. The finding supports our third hypothesis. Figure 5 illustrates the actual NIRP effect without the impact of the business cycle. As demonstrated, the estimated true effect on LLPTL and LLRTL surpasses the baseline effect. It is postulated that NIRP has an expansionary effect on the economy, partly attributed to the functioning of a bank bond financing channel (Onofri et al., 2023). In that sense, the greater contraction in LLPTL and LLRTL with business cycle detrended may indicate a countercyclical tendency of banks’ provisioning behaviours. This is in line with Fonseca and Gonzalez (2008), suggesting that banks should recognise potential credit risk and accumulate provisions and reserves during economic booms to be drawn during downturns. Overall, our analysis suggests that our baseline results are robust and reliable.

Table 8.

The NIRP effect after detrending the business cycle test results.

	(1) $Δ$ LLPTL	(2) $Δ$ LLRTL
NIRP effect	−0.102*** −0.03	−0.182*** −0.046
Observations	11,039	7,660
Adj. $R^{2}$	0.101	0.092

NIRP effect is the interaction between the Treatment dummy variable and the Post dummy variable. It takes the value of 1 if the bank $i$ in country $j$ has been affected by the NIRP after the monetary shock and 0 otherwise. LLPTL is the ratio of bank loan loss provisions to total loans. LLRTL is the ratio of bank loan loss reserves to total loans. $Δ$ denotes the change in variables. Both regressions include fixed time effects. Robust standard errors clustered by banks in parentheses. ***, **, and * indicate statistical significance at 1%, 5%, and 10%, respectively.

Figure 5.

Components of negative effects.

4.2.3 QD model

Using QD models by taking one more time-wise difference than TD models, we are able to weaken the identification condition by allowing for the time effect to change over time. Similar to the association between DiD models and TD models, a nested relationship also exists between TD and QD models. Accordingly, we employ the following QD model to further examine the robustness of the treatment effects.

Δ^{2} C R_{i j t} = β_{1} (T r e a t e d_{i j} * P o s t_{j t}) + β_{2} Δ^{2} X_{i j t} + β_{3} Δ^{2} G_{i j t} + φ_{t} + ε_{i j t}

(4)

$Δ^{2}$ denotes changes in variables. The rest of the notation and identification of variables remain the same as our baseline model. With an additional differencing on the time dimension, we only keep from 2013 to 2017 in our sample. The results of the QD model are presented in Table 9. The estimated NIRP effects are negative and significant, which are largely consistent with the earlier estimates derived from the TD model. This provides additional support for the robustness of our baseline results.

Table 9.

QD model results.

	(1) $Δ^{2}$ LLPTL	(2) $Δ^{2}$ LLRTL
NIRP effect	−0.120*** −0.024	−0.137*** −0.051
Observations	9,120	6,308
Adj. $R^{2}$	0.101	0.043

NIRP effect is the interaction between the Treatment dummy variable and the Post dummy variable. It takes the value of 1 if the bank $i$ in country $j$ has been affected by the NIRP after the monetary shock and 0 otherwise. LLPTL is the ratio of bank loan loss provisions to total loans. LLRTL is the ratio of bank loan loss reserves to total loans. $Δ^{2}$ denotes the change in variables. Covariates also have $Δ^{2}$ , not $Δ$ . Both regressions include fixed time effects. Robust standard errors clustered by banks in parentheses. ***, **, and * indicate statistical significance at 1%, 5%, and 10%, respectively.

4.2.4 Alternative measures

We utilise LLPTA, LLRTA, and NPLTL as alternative measurements of bank credit risk-taking. These variables have been commonly used in existing literature to measure bank loan loss provisioning in relation to the level of total loans and total assets (Angklomkliew et al., 2009; Bushman and Williams, 2012; Curcio and Hasan, 2015; Molyneux et al., 2019). Among them, NPLTL reflects the realised level of credit risk.

We re-run the baseline and robustness tests, and Table 10 presents the results. Panel A reports the results from the baseline model. The NIRP effects on all the alternative measures are negative and significant. The realised credit risk measured by NPLTL has the most significant contraction following NIRP. This suggests that banks’ credit risk appetite is not imprudent but is based on a rational assessment of expected loan losses. Panels B-D display results of the PSM-TD model, the TD model after detrending the business cycle, and the QD model, respectively. Results from robustness tests are in line with the results in Panel A, which further validate the utilisation of the TD model. Overall, the findings from the alternative measures are consistent with our baseline results in Table 4.

Table 10.

Alternative measures.

Panel A: TD model
	(1) $Δ$ LLPTA	(2) $Δ$ LLRTA	(3) $Δ$ NPLTL
Treatment	−0.055*** −0.011	−0.073*** −0.024	−0.432*** −0.071
Observations	11,045	7,664	7,397
Adj.	0.137	0.092	0.094
Panel B: PSM-TD model
	(1) $Δ$ LLPTA	(2) $Δ$ LLRTA	(3) $Δ$ NPLTL
NIRP effect	−0.079*** −0.025	−0.077** −0.036	−0.732*** −0.188
Observations	11,045	7,664	7,397
Adj.	0.2	0.21	0.305
Panel C: NIRP effect after detrending the business cycle
	(1) $Δ$ LLPTA	(2) $Δ$ LLRTA	(3) $Δ$ NPLTL
NIRP effect	−0.052*** −0.014	−0.104*** −0.025	−0.482*** −0.084
Observations	11,045	7,664	7,397
Adj.	0.137	0.119	0.105
Panel D: QD model
	(1) $Δ^{2}$ LLPTA	(2) $Δ^{2}$ LLRTA	(3) $Δ^{2}$ NPLTL
NIRP effect	−0.083*** −0.011	−0.104*** −0.03	−0.329*** −0.102
Observations	9,125	6,312	5,959
Adj.	0.125	0.084	0.027

Treatment is a dummy variable that equals 1 if the NIRP is implemented in a country and 0 otherwise. NIRP is the interaction between the Treatment dummy variable and the Post dummy variable. It takes the value of 1 if the bank $i$ in country $j$ has been affected by the NIRP after the monetary shock and 0 otherwise. LLPTA is the ratio of bank loan loss provisions to total assets. LLRTA is the ratio of bank loan loss reserves to total assets. NPLTL is the ratio of non-performing loans to total loans. $Δ$ denotes the change in variables, whereas $Δ^{2}$ denotes the change in variables. For the QD model, covariates have $Δ^{2}$ , not $Δ$ . All regressions include fixed time effects. Robust standard errors clustered by banks in parentheses. ***, **, and * indicate statistical significance at 1%, 5%, and 10%, respectively.

4.3. Subsample analyses

Following Molyneux et al. (2019, 2020), we re-estimate our baseline model using subsamples based on percentiles of bank- and country-specific covariates to examine the heterogeneity among banks and countries. We utilised data in 2013 to determine medians and quartiles of covariates. In this manner, we focus on the pre-treatment characteristics and alleviate the endogeneity issue that may arise from the potential impact of NIRP on covariates such as bank size and inflation.

We first investigate the difference in the NIRP effect among subsamples based on bank size. Panel A of Table 11 presents the results. Overall, banks with total assets below the 50th percentile experience a more negative NIRP effect on provisioning, particularly for LLPTL. For LLRTL, however, the difference in NIRP effects between small and large banks is not pronounced. The effect on LLPTL for smaller banks is negative and statistically significant, while it is insignificant for larger banks. This may be because larger banks are less sensitive to macroeconomic changes and tend to maintain stable balance sheet performance. In addition, their more rigorous and prudent risk assessment processes may prevent a substantial downward revision of credit risk expectations, even under NIRP.

Table 11.

Subsample analyses results.

Panel A: The NIRP effect on bank credit risk-taking by percentile regressions based on bank size
	Bank size < 50^th percentile								Bank size > 50^th percentile
	(1)∆LLPTL				(2)∆LLRTL				(3)∆LLPTL				(4)∆LLRTL
NIRP effect	−0.165*** (0.029)				−0.127* (0.072)				−0.022(0.035)				−0.130** (0.056)
Obs.	5,581				4,685				5,458				2,975
Adj. $R^{2}$	0.147				0.085				0.096				0.078
Panel B: The NIRP effect on bank credit risk-taking in subsamples based on bank capitalisation, funding structure, and income structure.
	Less capitalised			More capitalized		Less reliant on deposits			More reliant on deposits			Less interest-focused			More interest-focused
	(1)∆LLPTL	(2)∆LLRTL		(3)∆LLPTL	(4)∆LLRTL	(5)∆LLPTL		(6)∆LLRTL	(7)∆LLPTL	(8)∆LLRTL		(9)∆LLPTL	(10)∆LLRTL		(11)∆LLPTL		(12)∆LLRTL
NIRP effect	−0.045(0.033)	−0.222*** (0.066)		−0.118*** (0.031)	−0.022(0.068)	−0.053(0.036)		0.028(0.068)	−0.099*** (0.031)	−0.318*** (0.063)		−0.113*** (0.035)	−0.097(0.067)		−0.040(0.031)		−0.158** (0.069)
Obs.	3,995	2,151		7,032	5,497	6,056		3,354	4,978	4,306		4,350	2,418		6,689		5,242
Adj. $R^{2}$	0.082	0.049		0.134	0.113	0.103		0.071	0.107	0.103		0.088	0.116		0.122		0.068
Panel C: The NIRP effect on bank credit risk-taking in subsamples based on bank specialization
	Commercial bank				Savings bank				Cooperative bank					Bank holding company
	(1)∆LLPTL		(2)∆LLRTL		(3)∆LLPTL		(4)∆LLRTL		(5)∆LLPTL		(6)∆LLRTL			(9)∆LLPTL		10)∆LLRTL
NIRP effect	−0.189*** (0.044)		−0.090(0.072)		0.131(0.110)		−0.263*** (0.089)		−0.177*** (0.036)		−0.241*** (0.064)			−0.032(0.210)		0.040(0.108)
Obs.	3,959		3,764		1,449		448		3,605		1,607			1,123		1,095
Adj $R^{2}$	0.137		0.092		0.255		0.295		0.221		0.223			0.105		0.111
Panel D: The NIRP effect on bank credit risk-taking by percentile regressions based on inflation
	Inflation < 50^th percentile								Inflation > 50^th percentile
	(1)∆LLPTL				(2)∆LLRTL				(3)∆LLPTL				(4)∆LLRTL
NIRP effect	−0.144*** (0.041)				−0.140** (0.063)				0.045(0.037)				−0.135* (0.073)
Obs.	6,328				6,231				4,711				1,429
Adj. $R^{2}$	0.129				0.105				0.110				0.069

NIRP effect is the interaction between the Treatment dummy variable and the Post dummy variable. It takes the value of 1 if bank i in country j has been affected by the NIRP after the monetary shock and 0 otherwise. LLPTL is the ratio of bank loan loss provisions to total loans. LLRTL is the ratio of bank loan loss reserves to total loans. ∆ denotes the change in variables. Panel A presents subsample regressions based on 50^th percentiles of bank size using our baseline model. Panel B displays subsample regressions based on 50^th percentiles of bank capitalisation, funding structure, and income structure. Less and more capitalised banks respectively refer to banks with equity to total assets (E/TA) lower and higher than the 50^th percentile. Banks that are less and more reliant on deposits, respectively, refer to those with a deposit and other short-term funding to total liabilities ratio (funding structure) lower and higher than the 50^th percentile. Less and more interest-focused banks refer to those with a net interest income to operating revenue ratio (income structure) lower and higher than the 50^th percentile. Panel C shows subsample regressions based on bank specialisation. We do not include the real estate & mortgage finance institutions, finance companies, and fintech company subsamples due to the small number of observations. Panel D presents subsample regressions based on the inflation rate. All regressions include fixed time effects. Robust standard errors clustered by banks in parentheses. ***, **, and * indicate statistical significance at 1%, 5%, and 10%, respectively.

We then examine the NIRP effect in subsamples based on bank capitalisation, funding structure, and income structure. Panel B of Table 11 reports the results. The findings for capitalisation and income structure are mixed across LLPTL and LLRTL, with significance observed in different groups for each measure. However, it is worth noting that banks that rely heavily on deposits as a funding source experience a greater contraction in both LLPTL and LLRTL following NIRP. This suggests that, although the transmission of NIRP to depositors is incomplete, lower interest rates may still discourage deposit activity, thereby constraining lending for banks more dependent on deposit funding.

To further explore the role of business models in banks’ loan loss provisioning, we also split the sample by bank specialisation. Panel C of Table 11 shows the results. The NIRP effects on LLPTL and LLRTL are negative and significant for cooperative banks. The results for commercial banks are mixed and not consistent across the two measures, which is inconsistent with the findings of Olszak et al. (2017), who suggest that the provisioning of commercial banks is more prone to business-cycle fluctuations and macroeconomic changes. One possible explanation could be that, unlike commercial banks, cooperative banks have a relatively smaller area of operation and less variety of services. This constrains the capacity of cooperative banks to generate profits through diverse channels. In cases when NIRP impairs the bank's profitability (Molyneux et al., 2019), cooperative banks may therefore have stronger incentives to engage in riskier lending to recover revenue.

Finally, we examine the NIRP effect in subsamples divided by inflation rate. Panel D of Table 11 displays the results. The NIRP effect is negative and significant for banks in countries with inflation below the 50th percentile. This suggests that in low-inflation environments, which are often accompanied by weaker economic growth, banks tend to take less credit risk under NIRP. This may reflect more cautious lending behaviours in uncertain macroeconomic conditions.

Overall, the results of subsample analyses show that banks’ responses to the NIRP depend on bank- and country-specific characteristics, which is consistent with our fourth hypothesis. Specifically, small banks with more reliance on deposits see more contraction in LLPTL and LLRTL after the adoption of NIRP. Banks in a low-inflation environment are also more sensitive to the NIRP.

5. Concluding remarks

Despite the large literature on the bank credit risk model, few studies have examined the effects of NIRP using a cross-country analysis. In addition, utilising a TD model, we present results that are less susceptible to bias arising from the parallel trends assumption. Our findings could provide practical implications for banks’ credit risk-taking in a negative interest rate environment. Banks in NIRP-adopter countries are found to take less credit risk than those in non-NIRP-adopter countries. Negative interest rates may thus bring more pressure to the monetary policy pass-through channel as banks in NIRP adopter countries become more risk-averse than non-adopter countries. In other words, banks may not perceive enough stimulus or promising prospects of the economy from the adoption of NIRP, raising questions about the effectiveness of NIRP.

Policymakers should also consider the heterogeneity across countries and banks when evaluating the effectiveness of NIRP. Within our sample, banks’ provisioning behaviour appears countercyclical. Provisions and reserves tend to decline during expansionary periods, suggesting that banks act more cautiously when faced with the uncertain effects of NIRP. Our findings further indicate that the impact of NIRP on credit risk-taking is not uniform across banks. Small cooperative banks operating in low-inflation economies are more sensitive to NIRP, as evidenced by their larger contraction in provisioning. Such heterogeneity in bank responses may weaken the overall transmission of monetary policy to credit risk-taking and, consequently, to financial stability.

Key practical and research implications

Theoretical and practical research contributions

This study contributes to the literature on unconventional monetary policy by providing new evidence on the impact of the negative interest rate policy (NIRP) on bank credit risk-taking. Theoretically, it challenges the conventional risk-taking channel of monetary policy by showing that, unlike positive rate cuts, NIRP leads to a contraction in banks’ credit risk-taking, as reflected in lower loan loss provisions and reserves. By focusing specifically on credit risk rather than aggregate bank risk, this paper refines existing theories on monetary transmission under negative rates. Moreover, by disentangling business cycle effects from exogenous monetary shocks, the study enhances understanding of the procyclicality of bank provisioning. From a practical perspective, the findings suggest that NIRP may be less effective in stimulating bank risk-taking and credit supply, particularly for smaller, deposit-dependent banks and those operating in low-inflation environments, offering important insights for policymakers and regulators.

Footnotes

Appendix 1

Table 12.

Number of banks by country.

Control		Treatment
Australia	17	Austria	266
Calvados (France)	1	Belgium	13
Canada	35	Estonia	2
Chile	15	Finland	5
Czech Republic	11	France	136
Iceland	3	Germany	377
Israel	4	Greece	8
Loire-Atlantique (France)	2	Hungary	2
Maine-et-Loire (France)	1	Ireland	5
Mexico	30	Italy	94
New Zealand	12	Luxembourg	22
Poland	11	Netherlands	6
Republic of Korea	10	Portugal	6
Turkey	23	Slovakia	4
United Kingdom	31	Slovenia	4
United States of America	789	Spain	13
Total	995	Total	963

Acknowledgements

The authors would express sincere thanks to the Deputy Editor, Chelsea Liu, the Associate Editor, Elizabeth Zhu, and the anonymous reviewers for their valuable comments and suggestions, which have greatly enhanced the quality of our manuscript in the review process.

Final transcript accepted on 13 February 2026 by Elizabeth Zhu (AE – Finance).

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

ORCID iDs

Zixuan Dai

Lei Xu

Chandra Krishnamurti

Zenghua Lu

Notes

References

Abou El Sood

(2012) Loan loss provisioning and income smoothing in US banks pre and post the financial crisis. International Review of Financial Analysis 25: 64–72.

Angklomkliew

George

Packer

(2009) Issues and developments in loan loss provisioning: The case of Asia. BIS Quarterly Review, December.

Baker

Larcker

Wang

(2022) How much should we trust staggered difference-in-differences estimates? Journal of Financial Economics 144: 370–395.

Beatty

Liao

(2014) Financial accounting in the banking industry: A review of the empirical literature. Journal of Accounting and Economics 58: 339–383.

Bedendo

Bruno

(2012) Credit risk transfer in US commercial banks: What changed during the 2007–2009 crisis? Journal of Banking & Finance 36: 3260–3273.

Berger

Black

(2011) Bank size, lending technologies, and small business finance. Journal of Banking & Finance 35: 724–735.

Bertrand

Duflo

Mullainathan

(2004) How much should we trust differences-in-differences estimates? The Quarterly Journal of Economics 119: 249–275.

Bongiovanni

Reghezza

Santamaria

, et al. (2021) Do negative interest rates affect bank risk-taking? Journal of Empirical Finance 63: 350–364.

Bottero

Minoiu

Peydró

, et al. (2022) Expansionary yet different: Credit supply and real effects of negative interest rate policy. Journal of Financial Economics 146: 754–778.

10.

Boungou

(2020) Negative interest rates policy and banks’ risk-taking: Empirical evidence. Economics Letters 186: 108760.

11.

Brunnermeier

Koby

(2018) The ‘‘reversal interest rate’’: An effective lower bound on monetary policy, Working Paper, https://www.nber.org/system/files/working_papers/w25406/w25406.pdf

12.

Bushman

Williams

(2012) Accounting discretion, loan loss provisioning, and discipline of banks’ risk-taking. Journal of Accounting and Economics 54: 1–18.

13.

Calem

Rob

(1999) The impact of capital-based regulation on bank risk-taking. Journal of Financial Intermediation 8: 317–352.

14.

Chang

Chen

(2016) Government capital injection, credit risk transfer, and bank performance during a financial crisis. Economic Modelling 53: 477–486.

15.

Cukierman

(2013) Monetary policy and institutions before, during, and after the global financial crisis. Journal of Financial Stability 9: 373–384.

16.

Curcio

Hasan

(2015) Earnings and capital management and signaling: The use of loan-loss provisions by European banks. The European Journal of Finance 21: 26–50.

17.

Dell’Ariccia

Laeven

Suarez

(2017) Bank leverage and monetary policy’s risk-taking channel: Evidence from the United States. Journal of Finance 72: 613–654.

18.

Dell’Ariccia

Rabanal

Sandri

(2018) Unconventional monetary policies in the Euro area, Japan, and the United Kingdom. Journal of Economic Perspectives 32: 147–172.

19.

Demir

Michalski

Ors

(2017) Risk-based capital requirements for banks and international trade. The Review of Financial Studies 30: 3970–4002.

20.

Diamond

Rajan

(2006) Money in a theory of banking. American Economic Review 96: 30–53.

21.

Fonseca

Gonzalez

(2008) Cross-country determinants of bank income smoothing by managing loan-loss provisions. Journal of Banking & Finance 32: 217–228.

22.

Goodman-Bacon

(2021) Difference-in-differences with variation in treatment timing. Journal of Econometrics 225: 254–277.

23.

Greenawalt

Sinkey

(1988) Bank loan-loss provisions and the income-smoothing hypothesis: An empirical analysis, 1976–1984. Journal of Financial Services Research 1: 301–318.

24.

Heider

Saidi

Schepens

(2019) Life below zero: Bank lending under negative policy rates. The Review of Financial Studies 32: 3728–3761.

25.

Hellmann

Murdock

Stiglitz

(2000) Liberalization, moral hazard in banking, and prudential regulation: Are capital requirements enough? American Economic Review 90: 147–165.

26.

Hong

Kandrac

(2021) Pushed past the limit? How Japanese banks reacted to negative rates. Journal of Money, Credit and Banking 54: 1027–1063.

27.

Jiménez

Ongena

Peydró

, et al. (2014) Hazardous times for monetary policy: What do twenty-three million bank loans say about the effects of monetary policy on credit risk-taking? Econometrica 82(2), 463–505.

28.

Jiménez

Ongena

Peydró

, et al. (2017) Macroprudential policy, countercyclical bank capital buffers, and credit supply: Evidence from the Spanish dynamic provisioning experiments. Journal of Political Economy 125: 2126–2177.

29.

Kane

(1989) The S&L Insurance Mess: How Did It Happen? Washington, DC: The Urban Institute Press.

30.

Khandelwal

Schott

Wei

(2013) Trade liberalization and embedded institutional reform: Evidence from Chinese exporters. American Economic Review 103: 2169–2195.

31.

Kishan

Opiela

(2000) Bank size, bank capital, and the bank lending channel. Journal of Money, Credit and Banking 32: 121–141.

32.

Laeven

Majnoni

(2003) Loan loss provisioning and economic slowdowns: Too much, too late? Journal of Financial Intermediation 12: 178–197.

33.

Lee

(2016) Generalized difference in differences with panel data and least squares estimator. Sociological Methods & Research 45: 134–157.

34.

Molyneux

Reghezza

Xie

(2019) Bank margins and profits in a world of negative rates. Journal of Banking & Finance 107: 105613.

35.

Molyneux

Reghezza

Thornton

, et al. (2020) Did negative interest rates improve bank lending? Journal of Financial Services Research 57: 51–68.

36.

Nucera

Lucas

Schaumburg

, et al. (2017) Do negative interest rates make banks less safe? Economics Letters 159: 112–115.

37.

Olszak

Pipień

Kowalska

, et al. (2017) What drives heterogeneity of cyclicality of loan-loss provisions in the EU? Journal of Financial Services Research 51: 55–96.

38.

Onofri

Peersman

Smets

(2023) The effectiveness of a negative interest rate policy. Journal of Monetary Economics 140: 16–33.

39.

Roberts

Whited

(2013) Endogeneity in empirical corporate finance. Handbook of the Economics of Finance 2: 493–572.

40.

Romer

(2004) A new measure of monetary shocks: Derivation and implications. American Economic Review 94: 1055–1084.

41.

Schwerdt

Woessmann

(2020) Empirical methods in the economics of education. In: Bradley

Green

(eds) The Economics of Education. 2nd ed. Amsterdam: Elsevier Science, pp. 3–20.

42.

Silva

(2021) Fiscal deficits, bank credit risk, and loan-loss provisions. Journal of Financial and Quantitative Analysis 56: 1537–1589.

43.

Smith

(2002) Monetary policy, banking crises, and the Friedman rule. American Economic Review 92: 128–134.

44.

Stuart

Huskamp

Duckworth

, et al. (2014) Using propensity scores in difference-in-differences models to estimate the effects of a policy change. Health Services and Outcomes Research Methodology 14: 166–182.

45.

Tang

Chen

, et al. (2025) Negative interest rate policy and bank risk-taking: Search for yield or de-leverage? International Journal of Finance & Economics 30: 2450–2469.

46.

Van den Heuvel

(2002) Does bank capital matter for monetary transmission? Economic Policy Review 8: 259–265.

47.

Wooldridge

(2021) Two-way fixed effects, the two-way mundlak regression, and difference-in-differences estimators. Empirical Economics 69: 2545–2587.

48.

Xia

(2020) Negative interest rate policy and the yield curve. Journal of Applied Economics 35: 653–672.