Journal of Risk

Monitoring transmission of systemic risk: application of partial least squares structural equation modeling in financial stress testing

Necmi K Avkiran, Christian M. Ringle and Rand Low

  • PLS-SEM explains transmission of systemic risk from shadow banking to regulated banking
  • Complex phenomena not directly observable are examined as microlevel and macrolevel linkages
  • Around 75% of the variation in systemic risk in regulated banking is explained by shadow banking
  • PLS-SEM can guide microprudential vs macroprudential regulatory decisions as markets change

Regulators need a method that is versatile, is easy to use and can handle complex path models with latent (not directly observable) variables. In a first application of partial least squares structural equation modeling (PLS-SEM) in financial stress testing, we demonstrate how PLS-SEM can be used to explain the transmission of systemic risk. We model this transmission of systemic risk from shadow banking to the regulated banking sector (RBS) using a set of indicators (directly observable variables) that are sources of systemic risk in shadow banking and consequences of systemic risk measured in the RBS. Procedures for predictive model assessment using PLS-SEM are outlined in clear steps. Statistically significant results based on predictive modeling indicate that around 75% of the variation in systemic risk in the RBS can be explained by microlevel and macrolevel linkages that can be traced to shadow banking (we use partially simulated data). The finding that microlevel linkages have a greater impact on the contagion of systemic risk highlights the type of significant insight that can be generated through PLS-SEM. Regulators can use PLS-SEM to monitor the transmission of systemic risk, and the demonstrated skills can be transferred to any topic with latent constructs.

1 Introduction

According to Calluzzo and Dong (2015), it is difficult to quantify systemic risk in integrated markets, and it changes dynamically. Further, research on how risk is transmitted is still in its early stages due to inadequate data and complex linkages (Liang 2013). We examine the exposure of US bank holding companies (BHCs) to systemic risk sourced from shadow banking (SB), where SB is comprised of less regulated transactions, also known as market-based financing, through nonbank channels such as real estate investment trusts, leasing companies, credit guarantee outlets and money market funds.

Given the intricate and often changing connections between SB and the regulated banking sector (RBS), we refrain from defining yet another network topology designed to explain the transmission of systemic risk (examples of network topology can be found in Boss et al (2004), Hu et al (2012), Oet et al (2013), Caccioli et al (2014), Hautsch et al (2014) and Levy-Carciente et al (2015)). Instead, we work with a set of key indicators (directly measurable variables) identified as capturing the sources of systemic risk in SB and the consequences of systemic risk in the RBS.

From a regulator perspective, as connections change in a complex cause–effect environment, it is easier to add or remove indicators from a predictive contagion model, rather than redefine another network topology. As Acharya et al (2013, p. 76) point out, “The analysis of network effects in a stress test is extremely complex, even if all of the data on positions [is] available”. The statistical method in this paper is more versatile and easier to use, compared with network-based analyses. It better accommodates data characteristics often found in the real world, such as multivariate nonnormality.

This paper illustrates how the transmission of systemic risk from SB to the RBS can be modeled using partial least squares structural equation modeling (PLS-SEM; Sarstedt et al (2017)) in an effort to help regulators better monitor and manage contagion. PLS-SEM is a nonparametric approach based on ordinary least squares (OLS) regression, designed to maximize the explained variance in latent constructs, eg, systemic risk that cannot be directly observed or measured but can be observed indirectly through related indicators.

In addition to being robust with skewed data, PLS-SEM is considered an appropriate technique when working with composite models (Henseler et al 2014; Sarstedt et al 2016). A variance-based SEM technique, such as PLS-SEM, has particular advantages when it comes to composite modeling over its better-known cousin, covariance-based structural equation modeling (CB-SEM) (Henseler et al 2009; Hair et al 2014; Hair et al 2017a; Sarstedt et al 2014).11CB-SEM can be used to investigate relationships or linkages among latent constructs indicated by multiple variables or measures, but it expects multivariate normal distribution and large samples. CB-SEM follows a confirmatory approach to multivariate analysis, where the researcher theorizes about causal relations among the variables of interest. For a highly readable introduction to CB-SEM, see Lei and Wu (2007). The prediction-oriented character and the capability to deal with complex models highlights PLS-SEM as the method of choice in a wide range of disciplines (Wold 1982; Lohmöller 1989; Cepeda Carrión et al 2016; Richter et al 2016a).

The results of various review and overview studies across different business research disciplines, including accounting (Lee et al 2011; Nitzl 2016), family business (Sarstedt et al 2014), management information systems (Hair et al 2016; Ringle et al 2012), marketing (Hair et al 2012b; Henseler et al 2009; Richter et al 2016b), operations management (Peng and Lai 2012), supply chain management (Kaufmann and Gaeckler 2015), strategic management (Hair et al 2012a) and tourism (do Valle and Assaker 2016), support the rising popularity of PLS-SEM. Beside its wide application in business research, the use of PLS-SEM as published in journal articles reveals that it has recently expanded into fields such as biology, engineering, environmental and political science, medicine and psychology (see, for example, Willaby et al 2015).

Gart (1994, p. 134) defines systemic risk as the clear hazard that difficulties with the operations of financial institutions can be quickly transferred to others, including markets, and cause economic damage. In the period leading up to the global financial crisis (GFC) of 2007–9, a large portion of the financing of securitized assets was handled by the SB sector (Gennaioli et al 2013). The collapse of SB during these years therefore played an important role in weakening the RBS. According to the Financial Stability Board’s (FSB’s) report, SB makes a significant contribution to financing the real economy; for example, in 2013, SB assets represented 25% of total financial system assets (Financial Stability Board 2014).

Because of the interconnectedness between SB and the RBS (Adrian and Ashcraft 2012), SB can become a source of systemic risk: a major concern to all regulators. A main motivation for mitigating systemic risk is minimizing a negative impact on the real economy. As systemic risk rises, distressed banks reduce lending to clients, who in turn invest less, which reduces employment. As part of the interaction between SB and the RBS, there is a concern that banks might be evading increased regulation by shifting activities to SB (Gennaioli et al 2013). As the Basel III Accord moves toward full implementation by 2019, with a focus on better preparing financial institutions for the next crisis, and the Dodd–Frank Wall Street Reform and Consumer Protection Act of 2010 (DFA) unfolds in the United States, the contribution of SB to systemic risk in the RBS needs to be closely monitored.

Gennaioli et al (2013) maintain that according to the regulatory arbitrage view, banks pursue securitization using special or structured investment vehicles (SIVs) to circumvent capital requirements. In the period leading up to the GFC, traditional banks’ entry into SB through SIVs and special purpose vehicles (SPVs) created strong interdependencies and enabled the RBS to engage in almost unrestricted leverage. Banks were able to maintain higher leverage and still comply with risk-weighted capital requirements by transforming assets into highly rated securities. Such a strategy makes banks more vulnerable to shocks. The FSB (2011, p. 5) reports that while Basel III addresses a number of failings, regulatory arbitrage is likely to rise as bank regulation becomes tighter. The main motivation behind this study is to examine to what extent the transmission of systemic risk from SB to the RBS can be monitored. To the best of our knowledge, this is the first use of PLS-SEM in the field of financial stress testing.

Despite extensive empirical literature on systemic risk and the accompanying transmission mechanisms, Weiß et al (2014) state that the evidence is inconclusive (Bisias et al (2012) provide an extensive survey of systemic risk analytics). Yet, tracking systemic risk is a core activity in enabling macroprudential regulation (Jin and De Simone 2014). Our indicator-based approach to modeling systemic risk is favored by international regulators such as the Basel Committee on Banking Supervision (BCBS) and reflects microprudential as well as macroprudential perspectives (microlevel and macrolevel linkages). Similar to Glasserman and Young (2015), we avoid starting the investigation with a predefined network structure or topology, because we consider financial networks to be dynamic.

There is a wealth of information on the interconnectedness of the financial system and regulation in finance and law journals. However, both of these disciplines appear to ignore the body of knowledge generated by the other when we examine the references in such articles. Motivated by this observation, we attempt to strike a balance by tapping into both disciplines as we explore the feasibility of monitoring the transmission of systemic risk. The rest of this paper unfolds with a conceptual framework that develops assumptions to be tested. This is followed by an outline of the PLS-SEM method and a description of data. After reporting the results, we offer concluding remarks.

2 Conceptual framework

There are two banking systems in the United States, and each is governed by a different legal regime. Financial institutions that carry a banking charter belong to the traditional depository banking system often evaluated as three tiers, namely city banks, regional banks and community banks. These are referred to as the RBS; most US banks are owned by BHCs supervised by the Federal Reserve (Fed). Those who do not have a charter belong to the SB system, such as investment banks, money market mutual funds (MMMFs), hedge funds and insurance firms. One of the key differences between regulated banks and shadow banks is that the former are allowed to fund their lending activities through insured deposits (capped at US$250 000 per account), whereas federal law prohibits the latter from using deposits. Shadow banks therefore depend on deposit substitutes in a mostly unregulated and uninsured environment.

Over the last thirty years or so, SB has become increasingly dependent on various forms of short-term funding that substitute for functionality of deposits, such as over-the-counter (OTC) derivatives (traded outside regulated exchanges), short-term repurchase agreements (repos are regarded as fully secured short-term loans), commercial papers, MMMF shares, prime brokerage accounts and securitized assets. During a financial crisis, the reliance on deposit substitutes can have a contagious effect in the wider economy. For example, multinational corporations use MMMFs to fund their day-to-day cash needs. During the GFC, MMMFs were the primary buyers of commercial paper used by financial institutions as well as nonfinancial corporations such as General Electric and Ford (Jackson 2013). When MMMFs failed, large corporations were unable to sell their commercial paper to raise cash for their operations. Chernenko and Sunderam (2014) argue that instabilities associated with MMMFs were central to the GFC, and Bengtsson (2013) provides similar evidence from Europe.

Given externalities or moral hazards such as implicit expectations on the part of institutions to be bailed out in crises, it is unlikely that either banking sector will implement optimal protection or fully hedge their risks. There is a strong argument in favor of regulating how the SB sector relies on deposit substitutes and the systemic risk channeled to the RBS. In the finance literature, scholars such as Beltratti and Stulz (2012) have shown evidence of fragility in banks financed with the short-term funding that is often the domain of SB.

This study models the transmission of systemic risk using PLS-SEM in an effort to help regulators to better predict what is likely to happen in the RBS we heavily depend on for a well-functioning society. Thus, the first assumption is as follows.

  1. (A1)

    Systemic risk in SB makes a significant contribution to systemic risk in the RBS.

The well-known prudential regulation’s main focus is on identifying and mitigating exposure to endogenous crises at individual financial institutions, regulating leverage through internal risk management policies overseen by boards of directors. Ellul and Yerramilli (2013) report that BHCs with stronger and more independent risk management functions before the GFC had lower tail risk, less impaired loans, better operating performance and higher annual returns in 2007–8. Importantly, prudential regulation addressed by the Basel Accords has recently been supplemented by the European Systemic Risk Board (ESRB) and the Office of Financial Research (OFR) from the United States working on macroprudential regulation designed to identify and mitigate systemic risks.

Macroprudential regulation – an emerging framework – is designed to investigate the interconnectedness between SB and the RBS by accounting for counterparty relationships; common models and metrics; correlated exposure to assets; and shared reliance on market utilities (Johnson 2013). Macroprudential policies designed by regulators such as the Fed recognize systemic risk as a negative externality, where firms lack private incentives to minimize it (Liang 2013). Macroprudential regulation complements prudential regulation by simultaneously focusing attention on institution-specific endogenous factors and network-related exogenous factors that give rise to systemic risk.

We continue by expanding on key linkages between SB and the RBS, with a view to laying the groundwork for a predictive systemic risk framework that could enable monitoring contagion. A good starting point is an article by Anabtawi and Schwarcz (2011) that discusses regulating systemic risk. The authors premise their extensive arguments on the need for regulatory intervention but highlight the absence of an analytical framework that could help the regulators, particularly regarding how systemic risk is transmitted. Anabtawi and Schwarcz (2011) express a strong concern about the market participants being unreliable in terms of interrupting and limiting the transmission of systemic risk.

First, Anabtawi and Schwarcz (2011) posit an intrafirm correlation between a firm’s exposure to the risk of low-probability adverse events that can cause economic shocks and harm a firm’s financial integrity. Second, the authors put forward the concept of an interfirm correlation among financial firms and markets, where interaction with the intrafirm correlation can facilitate the transmission of otherwise localized economic shocks. An example of intrafirm correlation from the GFC is the fall in home prices (a low-probability risk) leading to defaulting of asset-backed securities and erosion of the integrity of institutions that are heavily invested in such securities. An example of interfirm correlation is the failure to fully appreciate the interconnectedness among traditional financial institutions and institutions such as Bear Stearns (failed in 2008), Lehman Brothers (failed in 2008), AIG and other SB institutions.

According to Anabtawi and Schwarcz (2011, p. 1356), intrafirm and interfirm correlations give rise to a transmission mechanism that can take a local adverse economic shock and convert it to strong systemic concerns. Effective regulation that weakens the abovementioned correlations can reduce the cost associated with financial crises. Following our first assumption, our second and third assumptions are defined below.

  1. (A1A)

    Systemic risk sourced from intrafirm correlations or microlevel linkages emanating from SB makes a significant contribution to systemic risk in the RBS.

  2. (A1B)

    Systemic risk sourced from interfirm correlations or macrolevel linkages emanating from SB makes a significant contribution to systemic risk in the RBS.

Another publication that attempts to make sense of interconnectedness and systemic risk is by Judge (2012), which focuses on financial innovation and the resulting complexity that can lead to systemic risk. Judge (2012, p. 661) identifies four sources of complexity: “(1) fragmentation, (2) the creation of contingent and dynamic economic interests in the underlying assets, (3) a latent competitive tendency among different classes of investors, and (4) the lengthening of the chain separating an investor from the assets ultimately underlying its investment”. It is then argued that complexity contributes to information loss and stickiness (the latter refers to arrangements in markets that are difficult to modify), both of which are sources of systemic risk. In short, the longer the chain separating an investor from an investment, the more difficult it becomes for investors to exercise due diligence in assessing risk and value.

Rixen (2013) argues that SB is primarily incorporated in lightly regulated offshore financial centers (OFCs). SPVs and SIVs benefit from regulatory and tax advantages offered by OFCs. Rixen (2013, pp. 438–439) maintains that OFCs can increase financial risk in at least five ways by (1) making it easier to register SPVs and SIVs, (2) enabling onshore financial institutions to hide risks, (3) raising the incentives for risky behavior, (4) helping avoid quality checks on credit that it is to be securitized, and (5) nurturing the debt bias in investments.

In summary, regulators’ main tasks in mitigating systemic risk should be to encourage less fragmentation and shorter chains between investors and investments; monitor existing linkages while looking out for new linkages; and disrupt transmission mechanisms. Starting from the above discussion of linkages, Table 1 outlines the sources of systemic risk in SB and the consequences of systemic risk in the RBS in an effort to draft a list of potential indicators (manifest variables) that can be used for predictive modeling.

Table 1: Potential indicators of systemic risk. [SB: shadow banking; RBS: regulated banking sector; BHC: bank holding company; SR: systemic risk; MACRO: macroprudential perspective; MICRO: microprudential perspective; indicators used are in bold. *This variable is currently not available. It was expected to be implemented in the United States as of 2016 (Gibson Dunn Lawyers 2013). **Brämer and Gischer (2013) illustrate a practical adaptation of the indicator-based method proposed by the BCBS (2011b). ***See the core profitability model (CPM) in Avkiran and Cai (2014).]
Sources of systemic risk in SB and    
the corresponding potential    
formative indicators Consequences of systemic risk in RBS are potential reflective indicators

Excessive securitization through special purpose investment vehicles (Gennaioli et al 2013)

Number of associations with structured credit vehicles for a given BHC (higher values lead to higher SR) {MACRO}

Countercyclical capital buffer percentages, eg, 0-2.5% of risk weighted assets (BCBS 2011a) {MACRO}*


Excessive dependence on short-term funding (Baker 2012; Barr 2012)

Level of OTC derivatives associated with a BHC (US$) (Higher values lead to higher SR) {MICRO} Repos (US$) (higher values lead to higher SR) {MICRO} MMMFs (US$) (Higher values lead to higher SR) {MICRO} Total regulatory capital ratio (Berger and Bouwman 2013) {MICRO} Noninterest income scaled by interest income (Brunnermeier et al 2012) {MICRO} Nonperforming loans (NPLs) scaled by total loans (Laeven and Valencia 2008; Idier et al 2014) {MICRO}

Complexity of derivatives (Blyth 2003; Bryan and Rafferty 2006)

Level of specific complex derivatives such as collateralized debt obligations (CDOs) or loan obligations (CLOs) with tranches associated with a BHC (US$) (higher values lead to higher SR) {MICRO} Average length of the intermediation chain from investors to assets, measured by the number of counterparties (higher values lead to higher SR) {MACRO} Spread of credit default swaps (CDSs) reflecting default risk (Allen et al 2012; Eichengreen et al 2012; Huang et al 2009, 2012) {MICRO} Bank z-score measured as the ratio of ROA plus the capital-asset ratio, divided by the SD of ROA over six years, distance to default (Laeven and Levine 2009; Erel et al 2014) {MICRO}

Nonrobust (mispriced) credit/liquidity put options (Adrian and Ashcraft 2012)

Relationship of a BHC with financial performance of its insurer(s) providing put options measured by return on assets (lower ROA is a proxy for nonrobust puts and higher SR) {MACRO}

Spread of bank bonds defined as the difference between ten-year A-rated bond yield and ten-year US Treasury yield (Oet et al 2011, 2013; Bianco et al 2012) {MACRO}


Incorporation of SB facilities in OFCs (FSB 2011, 2013; Rixen 2013)

Number of SB facilities incorporated in OFCs associated with a BHC (higher values lead to higher SR) {MACRO}

Financial beta defined as volatility of bank share price relative to the overall stock market (Oet et al 2011, 2013; Bianco et al 2012) {MACRO}


Types of executive compensation (Anabtawi and Schwarcz 2011; Tung 2011; Kaal 2012; Johnson 2013)

For SB institutions associated with a BHC: Average duration of executive stock options in years (shorter duration leads to higher SR) {MICRO}; Number of compensation packages linked to a risk-weighted portfolio of firm’s securities (lower values lead to higher SR) {MICRO}; Contingent convertible executive bonds (US$) (lower values lead to higher SR) {MICRO} Results of annual stress testing required by Dodd–Frank Act (Barr 2012) {MACRO} Stressed value-at-risk (VaR) required by Basel Accords (BCBS 2011a) {MACRO} Modified BCBS’ score approximating domestic systemic importance of banks {MACRO}** Relative efficiency scores based on CPM {MICRO}***

Homogeneity of financial assets in SB (Elsinger et al 2006)

Extent of financial assets of a given SB institution associated with a BHC are correlated with similar SB institutions (higher values lead to higher SR) {MACRO}


3 Method and data

3.1 Partial least squares structural equation modeling

For the first time in the field of financial stress studies, we use the iterative OLS regression-based PLS-SEM (Lohmöller 1989; Wold 1982). PLS-SEM has become a key multivariate analysis method to estimate complex models with relationships between latent variables in various disciplines. For example, popular PLS-SEM applications focus on explaining customer satisfaction and loyalty, or technology acceptance and use (Hair et al (2014, Table 1) provides a breakdown of business disciplines that use PLS-SEM). The goal of the nonparametric PLS-SEM method is to maximize the explained variance of endogenous latent constructs (a latent construct explained by other latent constructs in the PLS path model), whereby the assumption of multivariate normality is relaxed. For instance, Hair et al (2011, 2012b, 2014, 2017a, 2018) and Sarstedt et al (2017) introduce users to PLS-SEM, while, for example, Lohmöller (1989) and Monecke and Leisch (2012) provide a step-by-step explanation of the mathematics behind its algorithm.

Given the extent of dynamic interconnectedness in the US financial system, we treat the transmission of systemic risk as a set of latent constructs representing phenomenons that cannot be directly observed or measured. Figure 1 represents a predictive model. This study’s main objective remains one of predictive modeling and understanding the transmission of systemic risk from SB to the RBS through a foundational illustration of PLS-SEM in this field.

We start with known sources of systemic risk in SB captured by formative indicators, and estimate the extent to which we can predict consequences of systemic risk in the RBS captured by reflective indicators (see Table 1 and Figure 1). According to Jöreskog and Wold (1982, p. 270), “PLS is primarily intended for causal-predictive analysis in situations of high complexity but low theoretical information”. In summary, using the PLS-SEM approach is recommended when (1) the objective is explaining and predicting target constructs and/or detecting important driver constructs, (2) the structural model has formatively measured constructs, (3) the model is complex (with many constructs and indicators), (4) the researcher is working with a small sample size (due to a small population size) and/or data is nonnormal, and (4) the researcher intends to use latent variable scores in follow-up studies (Hair et al 2017a; Rigdon 2016). The latter case has been demonstrated by importance-performance map analyses (Ringle and Sarstedt 2016) or the combination of PLS-SEM results with agent-based simulation (Schubring et al 2016).

Some other advantages of PLS-SEM over CB-SEM are its focus on predicting dependent latent variables (Evermann and Tate 2016; Shmueli et al 2016), which is often a key objective in empirical studies, and its ability to accommodate indicators with different scales. In this context, the distinction between formative and reflective indicators is particularly important (Hair et al 2011, 2012b; Sarstedt et al 2016).

  • Formative indicators form the associated exogenous latent constructs. We try to minimize the overlap among them because they are treated as complementary (Table 1’s left-hand column contains potential formative indicators likely to lead to systemic risk in SB). The exogenous latent constructs in Figure 1 are formed by the associated indicators, and the outer weights result from a multiple regression, with the construct as a dependent variable and its associated formative indicators as independent variables.

  • Reflective indicators are consequences or manifestations of the underlying target latent construct, meaning causality is from the construct to the indicator. Because of substantial overlap among the reflective indicators, they are treated as interchangeable, meaning they are expected to be highly correlated. Potential reflective indicators likely to capture the systemic risk in the RBS are indicated in the right-hand column of Table 1. The endogenous latent construct in Figure 1 becomes the independent variable in single regression runs to determine the outer loadings, where the reflective indicators individually become the dependent variable in each run.

Illustrative representation of a predictive model for SB's contribution to systemic risk in the RBS
Figure 1: Illustrative representation of a predictive model for SB’s contribution to systemic risk in the RBS. This is an illustrative depiction of PLS-SEM modeling; the actual diagrammatic model representing the results reported is shown in Figure 3. Circles represent the latent variables or constructs that comprise the structural model; left-hand rectangles (X1X5) house the formative indicators theorized as forming the two exogenous latent constructs (measurement model for systemic risk in SB); right-hand rectangles (X6X10) house the reflective indicators theorized as the consequences of the endogenous or target latent construct (measurement model for systemic risk in the RBS). W1W10 are the outer weights, and P1 and P2 are the proxies or path coefficients for Y1 and Y2 (exogenous latent constructs) explaining Y3 (endogenous latent construct). The number of indicators represented in Figure 1 is illustrative only and does not represent the actual indicator numbers used (the actual model is reported in Figure 3).

PLS-SEM models consist of two main components, namely the structural or inner model and the measurement or outer model, visible in Figure 1. A group of manifest variables (indicators) associated with a latent construct is known as a block, and a manifest variable can only be associated with one construct. According to Monecke and Leisch (2012, p. 2): “latent variable scores are estimated as exact linear combinations of their associated manifest variables and treat them as error free substitutes for the manifest variables … PLS path modeling is a soft-modeling technique with less rigid distributional assumptions on the data.” PLS-SEM requires the use of recursive models, where there are no circular relationships (Hair et al 2017a); nonrecursive models with circular relationships may use the latent variable scores and, in a second stage, estimate the circular relationships by using, for example, the two-stage least squares method (see, for example, Bollen 2001).

Figure 2 provides a diagrammatic representation of the PLS-SEM algorithm as described in Monecke and Leisch (2012). At the beginning of the algorithm, all the manifest variables in the data matrix are scaled to have a zero mean and unit variance. The algorithm estimates factor scores for the latent constructs by an iterative procedure, where the first step is to construct each latent construct by the weighted sum of its manifest variables. The inner approximation procedure (step 2) reconstructs each latent construct by its associated latent construct(s), as a weighted sum of neighboring latent constructs.

The outer approximation procedure (step 3) then attempts to locate the best linear combination to express each latent construct by its manifest variables, in the process generating coefficients known as outer weights. While the weights were set to one during initialization, in step 3 weights are recalculated based on latent construct values emerging from the inner approximation in step 2.

In step 4, latent constructs are put together again as the weighted sum or linear combination of their corresponding manifest variables to arrive at factor scores. The algorithm terminates when the relative change for the outer weights is less than a prespecified tolerance (following each step, latent constructs are scaled to have zero mean and unit variance).

Diagram depicting the PLS-SEM algorithm (adapted from Monecke and Leisch (2012, Figure 5))
Figure 2: Diagram depicting the PLS-SEM algorithm (adapted from Monecke and Leisch (2012, Figure 5)).

The PLS-SEM algorithm provides latent variable scores, reflective loadings and formative weights in the measurement models, estimations of path coefficients in the structural model, and R2 values of endogenous latent variables. These results allow computing many additional results and quality criteria, such as Cronbach’s alpha, the composite reliability, f2 effect sizes, Q2 values of predictive relevance (see, for example, Chin 1998; Tenenhaus et al 2005; Chin 2010; Hair et al 2017a) and the new HTMT criterion (heterotrait–monotrait ratio of correlations) to assess discriminant validity (Henseler et al 2015).

Nevertheless, PLS-SEM has been criticized for giving biased parameter estimates because it does not explicitly model measurement error, despite employing bootstrapping to estimate standard errors for parameter estimates (Gefen et al 2011). This potential shortcoming can be restated as PLS-SEM parameter estimates based on limited information not being as efficient as those based on full information estimates (Sohn et al 2007). Alternatively, CB-SEM is able to model measurement error structures via a factor analytic approach; however, the downside is covariance among the observed variables that need to conform to overlapping proportionality constraints, meaning measurement errors are assumed to be uncorrelated (Jöreskog 1979).

Further, CB-SEM assumes homogeneity in the observed population (Wu et al 2012). Unless latent constructs are based on highly developed theory and the measurement instrument is refined through multiple stages, such constraints are unlikely to hold. Therefore, secondary data – often found in business databases – is unlikely to satisfy expected constraints. In such a situation, CB-SEM that relies on common factors would be the inappropriate choice, and PLS-SEM that relies on weighted composites would be more appropriate because of its less restrictive assumptions. Further, using formative indicators is problematic in CB-SEM because it gives rise to identification problems and reduces the ability of CB-SEM to reliably capture measurement error (Petter et al 2007). Those interested in further critique/rebuttal of PLS-SEM are invited to read Henseler et al (2014), Marcoulides et al (2012), Rigdon (2016) and Sarstedt et al (2016).

Recapping, in addition to being robust with skewed data because it transforms nonnormal data according to the central limit theorem, PLS-SEM is also considered an appropriate technique when working with small samples (Henseler et al 2009; Hair et al 2017a). However, this argument is relevant when the sample size is small due to a small population size. Otherwise, using large data sets and normally distributed data are advantageous when using PLS-SEM. The literature review in Hair et al (2014, Table 1) lists the top three reasons for PLS-SEM usage as (1) nonnormal data, (2) small sample size and (3) presence of formative indicators (all of these conditions exist in this study’s data set).

Against this background of stated reasons, it is important to consider the arguments for and against the use of PLS-SEM that Rigdon (2016) puts forward. In summary, good reasons for using PLS-SEM are

  1. (1)

    to explain and predict the key target construct of the model,

  2. (2)

    to estimate complex models,

  3. (3)

    the inclusion of formatively measured constructs,

  4. (4)

    small populations and relatively small sample sizes, and/or

  5. (5)

    the use of secondary data.

Finally, PLS-SEM provides determinate latent variable scores, which can be employed in complementary methods (see, for example, Ringle and Sarstedt 2016; Schubring et al 2016).

3.2 Data

We dub the list of indicators found in Table 1 the “researchers’ theoretical wishlist”, because most of the data on formative indicators and some of the data on the reflective indicators cannot be accessed for various reasons. For example, in addition to commercial databases, we perused individual BHC submissions of FORM 10-K (annual report) required by the US Securities and Exchange Commission. We found inconsistent reporting formats and scant useful data for the project at hand.

We focus on BHCs because most banks in the United States, and particularly those at mature stages of operation, are owned by BHCs (Partnership for Progress 2011). The structures of BHCs allow them to diversify their portfolios and banking activities (Strafford 2011). Our working sample of sixty-three BHCs after removing those with missing values are for the year 2013, and those in the sample represent 82.35% of the cumulative total assets for all the BHCs in that year (sourced from BankScope).

For the purposes of illustrating predictive modeling, we start with seven reflective indicators and ten formative indicators from the potential list first summarized in Table 1 (see also Table A1 in the online appendix). The set of formative indicators is comprised of five indicators of microprudential focus capturing intrafirm relationships defining one of the two exogenous constructs, and five indicators of macroprudential focus capturing interfirm relationships defining the other exogenous construct (see the left-hand column of Table A1 in the online appendix, where the first five formative indicators are microprudential and the next five are macroprudential). In the run-up to the GFC of 2007–9, Acharya et al (2010) argue that the SB system was used to organize the manufacture of systemic tail risk (based on securitization) with inadequate capital in place; it is challenging for regulators to supervise this type of risk-taking by financial institutions.

After the initial run of PLS-SEM, we are left with four reflective indicators for the endogenous construct (two indicators of microprudential and two indicators of macroprudential perspective), and the same set of ten formative indicators for the two exogenous constructs.22The reflective indicators “relative efficiency scores based on CPM”, “total regulatory capital ratio”, and “noninterest income ratio” are sequentially removed because of their low outer loadings and the observed improvement in statistical criteria once they are omitted. As the maximum number of arrows pointing at a latent variable (in the measurement models or in the structural model) is five, we would need at least 5×10=50 observations to technically estimate the model (according to the ten-times rule of Barclay et al (1995)). Following the more rigorous recommendations from a power analysis (Hair et al 2016, p. 26, Exhibit 1.7), at least forty-five observations are needed to detect a minimum R2 value of 25% at a significance level of 5% and a statistical power level of 80%. Therefore, the sample size of sixty-three BHCs passes both technical minimum sample size requirements for estimating the underlying PLS path model. Summary statistics on the variables reported in Table 2 indicate nonnormal data, as evidenced by substantial skewness and kurtosis across about half the variables (observed as well as simulated).

Table 2: Summary statistics and correlations on variables used in all PLS-SEM tests (N=63 BHCs). [Refer to Table A1 in the online appendix for more details on the variables. aCoefficient of variation (standard deviation (SD)/mean). bRemoved following the initial PLS-SEM analysis.]
(a) Summary statistics
  Mean Median SD CV? Min. Max. Skew. Kurtosis
Observed reflective indicators
1. Total regulatory capital ratio (RCR)b 14.3766 14.3300 3.5822 0.2491 0.0000 25.6200 -1.3991 8.2999
2. Noninterest income (NII)b 0.7669 0.4445 1.0677 1.3920 0.0194 6.0739 3.1101 10.8240
3. Nonperforming loans (NPLs) 2.0053 1.3900 1.7451 0.8702 0.0700 9.4600 1.8585 4.7642
4. Bank z-score (BZS) 0.0005 0.0003 0.0005 0.9375 0.0000 0.0027 2.3059 6.3931
5. Relative efficiency scores (RESs)b 1.5797 1.3734 0.7217 0.4568 0.7036 3.4820 0.8405 -0.0460
6. Financial beta (FB) 0.9488 0.9650 0.2667 0.2810 0.4424 1.7139 0.5159 0.5501
7. Modified BCBS score (CBS) 1.5873 0.0820 4.3401 2.7343 0.0228 24.9039 3.5877 14.3240
Simulated formative indicators
1. Level of specific complex derivatives (CD) 1.1863 0.9047 0.8666 0.7305 0.0430 2.9411 0.5643 -0.9885
2. Repurchase agreements (RAs) 1.0297 1.0107 0.4005 0.3889 0.2094 1.8557 -0.0745 -0.5724
3. Average duration of executive stock (DES) 0.1381 0.1209 0.0449 0.3254 0.1013 0.2819 1.8351 2.5231
4. # of compensation packages (#CP) 0.0138 0.0120 0.0039 0.2889 0.0101 0.0312 2.2837 5.8411
5. Contingent convertible executive bonds (CEB) 0.1385 0.1295 0.0339 0.2451 0.1014 0.2600 2.1280 4.5771
6. # of counterparties (#CP) 5.5298 5.7266 1.6230 0.2935 2.2560 7.9147 -0.3859 -0.9434
7. # of SB facilities incorporated in OFCs (#OFC) 10.1111 10.0000 2.7126 0.2682 4.0000 15.0000 -0.1662 -0.4193
8. Extent that financial assets are correlated (FAC) 0.5822 0.5856 0.1816 0.3119 0.2577 0.9632 0.2088 -0.7347
9. # of associations with credit vehicles (#ACV) 8.4761 7.0000 5.2451 0.6188 2.0000 22.0000 1.1121 0.3193
10. Insurer’s return on assets (ROA) 0.2049 0.1032 0.6174 3.0131 -0.6998 4.4647 5.7099 38.0971
(b) Correlations
RCR 1.00                                
NII 0.14 1.00                              
NPL -0.03 -0.17 1.00                            
BZS -0.09 -0.01 0.51 1.00                          
RES -0.20 -0.22 0.30 0.18 1.00                        
FB 0.14 0.40 0.37 0.41 -0.05 1.00                      
CBS 0.15 0.35 0.06 -0.05 -0.31 0.48 1.00                    
CD 0.17 0.31 0.42 0.26 -0.03 0.58 0.53 1.00                  
RA -0.02 0.22 0.52 0.37 0.27 0.44 0.29 0.35 1.00                
DES 0.14 0.44 0.45 0.28 0.09 0.58 0.34 0.50 0.35 1.00              
#ComP -0.09 0.20 0.68 0.49 0.41 0.47 0.08 0.43 0.42 0.65 1.00            
CEB -0.06 0.07 0.70 0.49 0.43 0.50 0.13 0.41 0.43 0.58 0.92 1.00          
#CP 0.11 0.07 0.28 0.33 0.14 0.20 0.05 0.21 0.34 0.21 0.27 0.26 1.00        
#OFC 0.07 0.15 0.26 0.21 0.07 0.23 0.14 0.14 0.30 0.26 0.26 0.25 0.04 1.00      
FAC -0.09 0.36 0.21 0.17 -0.02 0.57 0.26 0.37 0.31 0.55 0.32 0.37 0.09 0.07 1.00    
#ACV 0.17 0.33 0.47 0.30 -0.05 0.62 0.65 0.64 0.35 0.60 0.56 0.60 0.26 0.21 0.38 1.00  
ROA -0.03 -0.15 0.35 0.21 0.13 0.07 -0.16 0.16 0.04 0.13 0.38 0.32 0.19 0.06 -0.13 0.06 1.00

In the absence of data on the formative indicators in the public domain, we simulate data such as that detailed in the online appendix by ensuring we use the systemic risk levels indicated by the observed data on reflective indicators, adjusting for firm size where relevant. Our simulation process for formative indicators starts by dividing each observed potential reflective indicator of a BHC into three quantiles (see the online appendix, Table A1, second column). These quantiles are defined as the upper, middle and lower ranges. Depending on the number of reflective indicators that each BHC exhibits in these quantiles, a BHC is assigned to one of eleven systemic risk categories (see Table A2 in the online appendix). The list of BHCs assigned to each systemic risk category is given in Table A3 (available online).

A random normal distribution for each formative indicator is simulated and bounded by the tiered range, as given by a set of rules based on the systemic risk category of each BHC (Table A3). The tiered ranges for the formative indicators are defined by the range of maximum and minimum values of formative indicators based on assumptions in the systemic risk literature for BHCs. Further, certain formative indicators require an additional simulation step to account for firm size captured by total assets. These are formative indicators 4, 5, 7 and 9. In this scenario, each of the original upper, middle and lower ranges for the formative indicator now has three quantiles each, creating nine quantiles. For example, if a BHC is considered to have a formative indicator that is in the middle range from step 1, and it is noted to be in the upper range in terms of firm size, the random simulation will occur in the sixth quantile. For more details on the simulation process, please see Section A2 of the online appendix.

4 Model estimation and results

Initially, we model ten formative indicators (five for each of the two exogenous constructs) and seven reflective indicators for the endogenous construct (see Table 1 and Table A1 in the online appendix). We then execute an additional run of PLS-SEM by removing three low-loading reflective indicators before reporting the final results. The reduced model in Figure 3 provides a diagram of the PLS-SEM final results. We used the software SmartPLS 3 (Ringle et al 2015) to conduct all the PLS-SEM analyses in this study.

Depiction of the reduced model in SmartPLS software (expanded variable names are in Table 3 and Table A1)
Figure 3: Depiction of the reduced model in SmartPLS software (expanded variable names are in Table 3 and Table A1).

4.1 Procedure followed for PLS-SEM model assessment

SmartPLS was set to 300 maximum iterations, with a stop criterion of 10-7 and analysis converged in thirty-six iterations. Hair et al (2017a, Exhibit 4.1) contains an outline of the procedure used below.

Reflective measurement model

Indicator reliability.

Hair et al (2012b) state that in exploratory research, loadings as low as 0.4 are acceptable. Outer loadings fall in the range 0.067–0.875. The three reflective indicators with the outer loadings of 0.067, 0.141 and 0.403 (“relative efficiency scores based on CPM”, “total regulatory capital ratio” and “noninterest income ratio”) are removed, as their indicator reliability is at relatively low levels. As a result of using the reduced model, outer loadings rise with a narrower range of 0.529–0.849. The rest of the testing is based on the reduced model.

Internal consistency.

According to Hair et al (2017a), we use Cronbach’s alpha as our lower boundary and composite reliability as our upper boundary to determine internal consistency; the formulas for Cronbach’s alpha and composite reliability are shown in Hair et al (2017a, pp. 111–112). Cronbach’s alpha is 0.627 and composite reliability is 0.784; 0.6 is acceptable in exploratory research (Hair et al 2012b). Similarly, values above 0.95 are undesirable (Hair et al 2017a). Overall, we establish internal consistency at a satisfactory level.

Convergent validity.

Average variance extracted (AVE) greater than 0.5 is preferred. When examining reflective indicator loadings, it is desirable to see higher loadings in a narrow range, indicating that all items explain the underlying latent construct, meaning convergent validity (Chin 2010). AVE is 0.482, suggesting that the endogenous construct accounts for 48.2% of the reflective indicators’ variance. Even though the AVE does not exceed the critical value of 0.5, we consider the result of 0.482 to be close enough to assume that convergent validity has been established. AVE could be increased above 0.5 by removing reflective indicators, but such an action is not recommended when the starting point is four indicators, because of its theoretical impact on the reflective measurement model.

Table 3: Abbreviated variable names in Figure 3. [Here, we provide names corresponding to the abbreviated variable names (also used in Table 2). RI-MICRO stands for “reflective indicator with microprudential perspective” and RI-MACRO stands for “reflective indicator with macroprudential perspective”; similarly, FI-MICRO stands for “formative indicator with microprudential perspective” and FI-MACRO stands for “formative indicator with macroprudential perspective”. *An alternative approach would be to use systemic risk scores by the Fed. See Benoit et al (2017) for systemic risk scoring used by the BCBS.]
(a) Four observed reflective indicators
(RI-MICRO)Non-perLoansRatio Nonperforming loans (NPL)
(RI-MICRO)BankZscore_Recip Bank z-score (BZS)
(RI-MACRO)FinBeta Financial beta (FB)
(RI-MACRO)BCBS Modified BCBS score (CBS)*
(b) Ten simulated formative indicators
(FI-MICRO)CDO_CLO Level of specific complex derivatives (CD)
(FI-MICRO)Repos Repurchase agreements (RA)
(FI-MICRO)DurExeStockOpt_Recip Average duration of executive stock (DES)
(FI-MICRO)#CompPkg_TA_Recip # of compensation packages (#CP)
(FI-MICRO)ContBonds_TA_Recip Contingent convertible executive bonds (CEB)
(FI-MACRO)#Counterparties # of counterparties (#CP)
(FI-MACRO)#SBfacilitiesOFC_TA # of SB facilities incorporated in OFCs (#OFC)
(FI-MACRO)extFinAssSBcorr Extent that financial assets are correlated (FAC)
(FI-MACRO)#CreditVehicles_TA # of associations with credit vehicles (#ACV)
(FI-MACRO)insurerROA_Recip Insurer’s return on assets (ROA)
Discriminant validity.

For reflective constructs, we aim to establish discriminant validity. Since the PLS path model only has one reflectively measured latent variable, we do not address this issue, for example, by applying the HTMT criterion.

Formative measurement model

Convergent validity.

Convergent validity is the degree to which a measure correlates positively with other measures (eg, reflective) of the same construct, using different indicators. When evaluating the convergent validity of formative measurement models in PLS-SEM, we use redundancy analysis (Chin 1998) to test whether the formatively measured construct is highly correlated with a reflective measure of the same construct (Hair et al 2017a). Since we do not have such reflective items of the formatively measured constructs in this study or a single-item measure of the same construct, we cannot conduct the redundancy analysis. We only find that the sign of the relationship between the formatively measured exogenous constructs and the reflectively measured endogenous construct is high and positive as expected, with path coefficients of 0.567 for MICRO and 0.342 for MACRO. Also as expected, the correlation between the formatively measured constructs is positive. We can therefore, at least to some extent, substantiate convergent validity.

Multicollinearity among indicators.

When collinearity exists, standard errors and variances are inflated. A variance inflation factor (VIF) of one means there is no correlation among the predictor variable examined and the rest of the predictors; therefore, the variance is not inflated. If the VIF is higher than five, the researcher should consider removing the corresponding indicator, or combining the collinear indicators into a new composite indicator. In this case, the VIF is 3.025. Since this number is less than five, multicollinearity is not an issue.

Significance and relevance of outer weights.

At a 5% probability of error level, the bootstrap confidence intervals indicate that the outer weights’ (an indicator’s relative contribution) significance of five out of ten formative indicators cannot be established. Checking outer loadings (an indicator’s absolute contribution) for these formative indicators, only two indicators are candidates for potential removal, namely “insurer’s return on assets” and “number of counterparties”. We do not remove these formative indicators, as they are important components of the theorized exogenous construct on macrolevel linkages.

Structural model

Establishing substantial measurement model(s) is a prerequisite for assessing the structural model. The latter provides confidence in the structural (inner) model. Analysis of the structural model is an attempt to find evidence supporting the theoretical/conceptual model, meaning the theorized/conceptualized relationships between the latent variables.

Size and significance of path coefficients.

The 95% bootstrap confidence intervals indicate that the path coefficient of 0.567 between the MICRO exogenous construct and the endogenous construct is significant; the MACRO path coefficient of 0.342 is also significant.

Predictive accuracy, coefficient of determination.

The R2 value is high, at 0.756 (adjusted to 0.748). This number indicates that the two exogenous constructs substantially explain the variation in the endogenous construct. According to Hair et al (2011, 2017a), as a rough rule of thumb, 0.25 is weak, 0.50 is moderate and 0.75 is substantial.

Assessing the “effect sizes”.

f2 measures the importance of the exogenous constructs in explaining the endogenous construct, and it recalculates R2 by omitting one exogenous construct at a time. A 0.435 (MICRO) is pleasingly high, indicating a large change in R2 if the exogenous construct on microlevel linkages were to be omitted; a 0.159 (MACRO) is lower but still substantial, implying that while the MACRO exogenous construct contributes relatively less to explaining the endogenous construct, both exogenous constructs are important. Hair et al (2017a) provide a rule of thumb, whereby an effect size of 0.02 is considered small, 0.15 is medium and 0.35 is large. The formula for effect size can be found in Hair et al (2017a, p. 201).

Predictive relevance, Q2.

This is obtained by the sample reuse technique called “blindfolding” in SmartPLS, where omission distance is set to eight (Hair et al (2012b) recommend a distance between five and ten, where the number of observations divided by the omission distance is not an integer). For example, setting the omission distance to eight, every eighth data point is omitted and parameters are estimated with the remaining data points. Omitted data points are considered missing values and replaced by mean values (Hair et al 2017a). In turn, estimated parameters help predict the omitted data points, and the difference between the actual omitted data points and predicted data points becomes the input in our calculation of Q2. In this case, Q2 emerges as 0.316. Since this number is larger than zero, it is indicative of the path model’s predictive relevance in the context of the endogenous construct and the corresponding reflective indicators.


the relative impact of predictive relevance (q2). Following from the above analysis of predictive relevance, q2 effect size can be calculated by excluding the exogenous constructs one at a time (see Hair et al (2017a, p. 207) for the formula). According to Hair et al (2013, 2017a), an effect size of 0.02 is considered small, 0.15 is moderate and 0.35 is large. The effect sizes following the respective exclusion of exogenous constructs are MACRO (0.0190) and MICRO (0.0833). The numbers indicate the dominance of MICRO in predicting systemic risk in the RBS.

In summary, our systematic evaluation of PLS-SEM results supports the establishment of substantial constructs via their measurement models, on which we build the analysis of the structural model. For the reflectively measured construct (systemic risk in the RBS), we can say we have construct validity (the extent we measure systemic risk as theorized) if both convergent and discriminant validity have been established. Convergent validity is the extent an indicator is positively correlated with alternative indicators measuring the same construct. For example, in the reflective measurement model, indicators are considered as reflecting the same endogenous construct. They are expected to share a high proportion of variance, where ideally the outer loadings exceed 0.7 (Hair et al 2011), although loadings as low as 0.4 are acceptable in exploratory research, our current study included (Hair et al 2012b). For the formatively measured constructs, we also need to examine the convergent validity by means of the redundancy analysis. Due to the lack of additional indicators we need for conducting the redundancy analysis, we could not carry out such an assessment in this study. However, we find that collinearity between indicators is not a critical issue and establish the significance and relevance of outer weights.

Based on the findings, we assess the PLS-SEM results of the structural model. Starting with the strongest finding reported under the structural model, R2 and adjusted R2 for our parsimonious model are substantial at 0.756 and 0.748, respectively, suggesting that the two exogenous constructs theorized significantly explain the variation in the endogenous construct. This means sources of systemic risk emanating from SB explain the consequences of systemic risk observed in the RBS, supporting A1.

Continuing with the properties of the structural model, predictive relevance is also satisfactory as measured by a Q2 of 0.316, meaning a value larger than zero shows that data points for reflective indicators are accurately predicted by the endogenous construct. Equally pleasing is the finding that the two path coefficients of 0.567 and 0.342 between the exogenous latent constructs and the endogenous latent construct are statistically highly significant, supporting A1A and A1B (we note that microlevel linkages play a larger role compared with macrolevel linkages). The f2 effect sizes of MICRO (0.435) and MACRO (0.159) suggest that microlevel linkages explain more of the variation in systemic risk in the RBS. A similar finding, but at a lower level, holds for the q2 effect sizes of predictive relevance.

As a result, we establish reliable and valid PLS-SEM results that allow us to substantiate our assumptions regarding the structural model. We find that the exogenous latent variables MICRO and MACRO explain 75.6% of the target construct (systemic risk in RBS), whereby MICRO is the somewhat more important explanatory construct. Also, we establish predictive relevance of the PLS path model.

4.2 Robustness testing

Hwang and Takane (2004, 2014) introduced generalized structured component analysis (GSCA) as an alternative to PLS-SEM (see Hair et al 2017b). We apply GSCA as a robustness test because it belongs to the same family of methods. PLS-SEM and GSCA are both variance-based methods, appropriate for predictive modeling, and they substitute components for factors. GSCA uses a global optimization function in parameter estimation with least squares (Hwang et al 2010). We restate that CB-SEM is not a meaningful alternative to PLS-SEM under the conditions of the current study, where the sample size is small, formative indicators are present and the study is exploratory rather than confirmatory.

GSCA maximizes the average or the sum of explained variances of linear composites, where latent variables are determined as weighted components or composites of observed variables. GSCA follows a global least squares optimization criterion, which is minimized to generate the model parameter estimates. GSCA is not scale-invariant and it standardizes data. GSCA retains the advantages of PLS-SEM, such as fewer restrictions on distributional assumptions, unique component score estimates and avoidance of improper solutions with small samples (Hwang and Takane 2004; Hwang et al 2010).

We use the web-based GSCA software GeSCA ( for robustness testing of the reduced model with ten formative indicators across two exogenous constructs and four reflective indicators attached to the reflective measurement model (Hair et al 2017b). As can be seen in Table 4, the PLS-SEM results are confirmed by GSCA. For example, AVE is identical; outer loadings are of a similar magnitude across the four reflective indicators; both path coefficients are also of a similar magnitude in the structural model; and the coefficients of determination are close to each other, with GSCA giving a slightly larger R2.

Table 4: Robustness testing of PLS-SEM with GSCA.
Measurement model
Average variance extracted (AVE) 0.482 0.482
Outer loadings of reflective indicators
Nonperforming loans 0.720 0.743
Bank z-score 0.641 0.646
Financial beta 0.849 0.835
Modified BCBS score 0.529 0.513
Structural model (path coefficients)
MICRO 0.567 0.561
MACRO 0.342 0.368
Coefficient of determination (R2) 0.756 0.765

5 Concluding remarks

We embarked on this project to illustrate how the transmission of systemic risk from SB to the RBS can be modeled using PLS-SEM, to help regulators monitor contagion without resorting to complex network topologies. To the best of our knowledge, using PLS-SEM in financial stress testing is the first such application. We have taken care to detail the procedure to be followed and how to interpret the results correctly. Behavioral finance is bound to provide a wealth of opportunities to apply PLS-SEM.

Following a literature review of finance and law disciplines, we identified various microlevel and macrolevel linkages between SB and the RBS. To address an extensive amount of missing data on sources of systemic risk in SB, we opted to simulate formative indicator data by establishing mathematical linkages to the observed reflective indicator data. The structural model to emerge consisted of two latent exogenous constructs of microlevel and macrolevel linkages embedded in SB, explaining the latent endogenous construct on systemic risk in the RBS. Based on partially simulated data, statistically significant results from PLS-SEM predictive modeling indicate that around 75% of the variation in systemic risk in the RBS can be explained by microlevel and macrolevel linkages that can be traced to SB.

While based on partially simulated data, the finding that microlevel linkages have a greater impact on the contagion of systemic risk (compared with macrolevel linkages) highlights the type of significant insight that can be generated through PLS-SEM. It suggests that internal risk management in BHCs has a greater role in reducing the likelihood of systemic risk events. Although central banks and other regulators can impose macroprudential frameworks on the markets, these appear to have a lower impact on reducing the likelihood of spread of systemic risk in the RBS. This finding is in line with the Dodd–Frank Act, which calls for stricter prudential regulation of systemically important financial institutions.

Regulators can use the approach in this paper to monitor the transmission of systemic risk. As Majerbi and Rachdi (2014) aptly point out in their study of the probability of systemic banking crises across a sample of fifty-three countries, stricter banking regulation, supervision and bureaucratic efficiency generally result in the reduced probability of crises. However, Hirtle et al (2016) draw a distinction between regulation and supervision, defining the latter as out-of-sight monitoring to identify unsound banking practices, which serves to complement regulation, ie, rules governing banks. Further, continued focus on transmission of systemic risk is warranted by the empirical evidence reported in Fink and Schüler (2015), where emerging market economies are shown to be negatively affected by systemic financial stress emanating from the United States.

Benoit et al (2016) conduct an extensive survey on systemic risk. The authors define systemic risk as a concept along the lines of “hard-to-define-but-you-know-it-when-you-see-it”. They continue to highlight that regulators need systemic risk measures that capture properly identified economic interactions in a timely manner, and that can be used in regulation. The authors highlight the fact that policy makers need reliable tools to monitor the escalation of systemic risks. They end their article with the comment that the search for a global risk measure that incorporates different sources of systemic risk and generates a single metric is still not over.

An extension of the study can include testing the stability of parameters over time. Other potential extensions may focus on smaller financial crises such as the eurozone sovereign debt crisis (2011–12) as well as the US debt ceiling crisis in 2011 (and to a lesser extent 2013). For example, the majority of redemptions resulted from flight-to-liquidity during the US debt ceiling crisis in 2011 (Gallagher and Collins 2016). A new model may be designed to understand the contribution of such actions to systemic risk. As the Basel III Accord is rolled out, it is feasible to collect data (post-2019) on additional variables, such as the thirty-day liquidity coverage ratio (LCR) and the net stable funding ratio (NSFR), and run PLS-SEM.

PLS-SEM is appropriate where (a) the nature of the underlying theory is predictive and exploratory rather than confirmative, (b) the types of latent constructs modeled include formative and reflective models and (c) the sample size is small due to a relatively small population, and the data exhibits nonnormal data characteristics. Against this background, we would like to reiterate the versatile, easy-to-use nature of PLS-SEM compared with network topologies and encourage others to use PLS-SEM in prediction-oriented and exploratory research. Recent advances such as PLSpredict (Shmueli et al 2016) further improve the prediction-oriented results evaluation and extend the PLS-SEM analysis.

Declaration of interest

The authors report no conflicts of interest. The authors alone are responsible for the content and writing of the paper. This paper uses the statistical software SmartPLS 3 ( Ringle acknowledges a financial interest in SmartPLS.


  • Acharya, V. V., Cooley, T., Richardson, M., and Walter, I. (2010). Manufacturing tail risk: a perspective on the financial crisis of 2007–2009. Foundations and Trends in Finance 4, 247–325 (
  • Acharya, V. V., Brownlees, C., Engle, R., Farazmand, F., and Richardson, M. (2013). Measuring systemic risk. In Managing and Measuring Risk: Emerging Global Standards and Regulations After the Financial Crisis, Roggi, O., and Altman, E. I. (eds), Volume 5, pp. 65–98. World Scientific Series in Finance. World Scientific.
  • Allen, L., Bali, T. G., and Tang, Y. (2012). Does systemic risk in the financial sector predict future economic downturns? Review of Financial Studies 25(10), 3000–3036 (
  • Anabtawi, I., and Schwarcz, S. L. (2011). Regulating systemic risk: towards an analytical framework. Notre Dame Law Review 86(4), 1349–1412.
  • Avkiran, N. K., and Cai, L. (2014). Identifying distress among banks prior to a major crisis using non-oriented super-SBM. Annals of Operations Research 217, 31–53 (
  • Baker, C. (2012). The Federal Reserve as last resort. University of Michigan Journal of Law Reform 46(1), 69–133.
  • Barclay, D. W., Higgins, C. A., and Thompson, R. (1995). The partial least squares approach to causal modeling: personal computer adoption and use as illustration. Technology Studies 2(2), 285–309.
  • Barr, M. S. (2012). The financial crisis and the path of reform. Yale Journal on Regulation 29(1), 91–119.
  • Basel Committee on Banking Supervision (BCBS) (2011a). Basel III: a global regulatory framework for more resilient banks and banking systems. Report, June, Bank for International Settlements.
  • Basel Committee on Banking Supervision (BCBS) (2011b). Global systemically important banks: assessment methodology and the additional loss absorbency requirement: rules text. Report, November, Bank for International Settlements.
  • Bengtsson, E. (2013). Shadow banking and financial stability: European money market funds in the global financial crisis. Journal of International Money and Finance 32, 579–594 (
  • Benoit, S., Colliard, J. E., Hurlin, C., and Perignon, C. (2016). Where the risks lie: a survey on systemic risk. Review of Finance 21(1), 109–152 (
  • Benoit, S., Hurlin, C., and Perignon, C. (2017). Pitfalls in systemic-risk scoring. Working Paper, Social Science Research Network.
  • Bianco, T., Oet, M. V., and Ong, S. J. (2012). The Cleveland financial stress index: a tool for monitoring financial stability. Working Paper 2012-04, March 21, Federal Reserve Bank of Cleveland.
  • Bisias, D., Flood, M., Lo, A. W., and Valavanis, S. (2012). A survey of systemic risk analytics. Unpublished Working Paper 0001, January 5, Office of Financial Research, US Department of Treasury.
  • Blyth, M. (2003). The political power of financial ideas. In Monetary Orders, Kirshner, J. (ed), pp. 239–259. Cornell University Press, Ithaca, NY.
  • Bollen, K. A. (2001). Two-stage least squares and latent variable models: simultaneous estimation and robustness to misspecifications. In Structural Equation Modeling: Present and Future: A Festschrift in Honor of Karl Jöreskog, Cudeck, R., du Toit, S., and Sorbom, D. (eds), pp. 199–138. Scientific Software International, Lincolnwood, IL.
  • Brunnermeier, M. K., Dong, G., and Palia, D. (2012). Banks’ non-interest income and systemic risk. Unpublished Working Paper, Princeton University.
  • Bryan, D., and Rafferty, M. (2006). Capitalism with Derivatives. Palgrave Macmillan.
  • Caccioli, F., Shrestha, M., Moore, C., and Farmer, J. D. (2014). Stability analysis of financial contagion due to overlapping portfolios. Journal of Banking and Finance 46, 233–245 (
  • Calluzzo, P., and Dong, G. N. (2015). Has the financial system become safer after the crisis? The changing nature of financial institution risk. Journal of Banking and Finance 53, 233–248 (
  • Cepeda Carrión, G., Henseler, J., Ringle, C. M., and Roldán, J. L. (2016). Prediction-oriented modeling in business research by means of PLS path modeling. Journal of Business Research 69(10), 4545–4551.
  • Chernenko, S., and Sunderam, A. (2014). Frictions in shadow banking: evidence from the lending behavior of money market mutual funds. Review of Financial Studies 27(6), 1717–1750 (
  • Chin, W. W. (1998). The partial least squares approach to structural equation modeling. In Modern Methods For Business Research, Marcoulides, G. A. (ed), pp. 295–358. Lawrence Erlbaum, Mahwah, NJ.
  • Chin, W. W. (2010). How to write up and report PLS analyses. In Handbook of Partial Least Squares: Concepts, Methods and Applications, Esposito Vinzi, V., Chin, W. W., Henseler, J., and Wang, H. (eds), pp. 655–690. Springer Handbooks of Computational Statistics Series, Volume 2. Springer.
  • do Valle, P. O., and Assaker, G. (2016). Using partial least squares structural equation modeling in tourism research: a review of past research and recommendations for future applications. Journal of Travel Research 55(6), 695–708 (
  • Eichengreen, B., Mody, A., Nedeljkovic, M., and Sarno, L. (2012). How the subprime crisis went global: evidence from bank credit default swap spreads. Journal of International Money and Finance 31, 1299–1318 (
  • Ellul, A., and Yerramilli, V. (2013). Stronger risk controls, lower risk: evidence from US bank holding companies. Journal of Finance 68(5), 1757–1803 (
  • Erel, I., Nadauld, T., and Stultz, R. M. (2014). Why did holdings of highly rated securitization tranches differ so much across banks? Review of Financial Studies 27(2), 404–453 (
  • Financial Stability Board (2011). Shadow banking: scoping the issues: a background note of the Financial Stability Board. Report, April 12, FSB.
  • Financial Stability Board (2013). Global shadow banking monitoring report 2013. Report, November 14, FSB.
  • Financial Stability Board (2014). Global shadow banking monitoring report 2014. Report, October 30, FSB.
  • Fink, F., and Schüler, Y. S. (2015). The transmission of US systemic financial stress: evidence for emerging market economies. Journal of International Money and Finance 55, 6–26 (
  • Gallagher, E., and Collins, S. (2016). Money market funds and the prospect of a US Treasury default. Quarterly Journal of Finance 6(1), 1640001-1–1640001-44 (
  • Gart, A. (1994). Regulation, Deregulation, Reregulation: The Future of the Banking, Insurance, and Securities Industries. Wiley.
  • Gefen, D., Rigdon, E. E., and Straub, D. (2011). An update and extension of SEM guidelines for administrative and social science research. MIS Quarterly 35(2), iii–xiv.
  • Gibson Dunn Lawyers (2013). Final Basel III capital rule issued by US bank regulators: some relief for community banks; for SIFIs, just the end of the beginning. Report, July 9, Gibson Dunn & Crutcher LLP.
  • Hair, J. F., Sarstedt, M., Pieper, T. M., and Ringle, C. M. (2012a). The use of partial least squares structural equation modeling in strategic management research: a review of past practices and recommendations for future applications. Long Range Planning 45, 320–340 (
  • Hair, J. F., Sarstedt, M., Ringle, C. M., and Mena, J. A. (2012b). An assessment of the use of partial least squares structural equation modeling in marketing research. Journal of the Academy of Marketing Science 40, 414–433 (
  • Hair, J. F., Ringle, C. M., and Sarstedt, M. (2013). Partial least squares structural equation modeling: rigorous applications, better results and higher acceptance. Long Range Planning 46, 1–12 (
  • Hair, J. F., Sarstedt, M., Hopkins, L., and Kuppelwieser, V. G. (2014). Partial least squares structural equation modeling (PLS-SEM): an emerging tool in business research. European Business Review 26(2), 106–121 (
  • Hair, J. F., Hollingsworth, C. L., Randolph, A. B., and Chong, A. Y. L. (2016). An updated and expanded assessment of PLS-SEM in information systems research. Industrial Management and Data Systems 117(3), 442–458 (
  • Hair, J. F., Hult, G. T. M., Ringle, C. M., and Sarstedt, M. (2017a). A Primer on Partial Least Squares Structural Equation Modeling (PLS-SEM), 2nd edn. Sage, Thousand Oaks, CA.
  • Hair, J. F., Hult, G. T. M., Ringle, C. M., Sarstedt, M., and Thiele, K. O. (2017b). Mirror, mirror on the wall: a comparative evaluation of composite-based structural equation modeling methods. Journal of the Academy of Marketing Science 45(5), 616–632.
  • Hair, J. F., Sarstedt, M., Ringle, C. M., and Gudergan, S. P. (2018). Advanced Issues in Partial Least Squares Structural Equation Modeling (PLS-SEM). Sage, Thousand Oaks, CA.
  • Henseler, J., Ringle, C. M., and Sinkovics, R. R. (2009). The use of partial least squares path modeling in international marketing. In New Challenges to International Marketing, pp. 277–319. Advances in International Marketing, Volume 20. Emerald Group Publishing (
  • Henseler, J., Dijkstra, T. K., Sarstedt, M., Ringle, C. M., Diamantopoulos, A., Straub, D. W., Ketchen, D. J., Hair, J. F., Hult, G. T. M., and Calantone, R. J. (2014). Common beliefs and reality about partial least squares: comments on Rönkkö & Evermann (2013). Organizational Research Methods 17, 182–209 (
  • Henseler, J., Ringle, C. M., and Sarstedt, M. (2015). A new criterion for assessing discriminant validity in variance-based structural equation modeling. Journal of the Academy of Marketing Science 43, 115–135 (
  • Hirtle, B., Kovner, A., and Plosser, M. (2016). The impact of supervision on bank performance. Staff Report, Federal Reserve Bank of New York.
  • Hu, D., Zhao, J. L., Hua, Z., and Wong, M. C. S. (2012). Network-based modeling and analysis of systemic risk in banking systems. MIS Quarterly 36(4), 1269–1291.
  • Huang, X., Zhou, H., and Zhu, H. (2012). Assessing the systemic risk of a heterogeneous portfolio of banks during the recent financial crisis. Journal of Financial Stability 8(3), 193–205 (
  • Hwang, H., and Takane, Y. (2014). Generalized Structured Component Analysis: A Component-Based Approach to Structural Equation Modeling. Chapman & Hall.
  • Idier, J., Lamé, G., and Mésonnier, J. S. (2014). How useful is the marginal expected shortfall for the measurement of systemic exposure? A practical assessment. Journal of Banking and Finance 47, 134–146 (
  • Jackson, B. F. (2013). Danger lurking in the shadows: why regulators lack the authority to effectively fight contagion in the shadow banking system. Harvard Law Review 127(2), 729–750.
  • Johnson, K. N. (2013). Macroprudential regulation: a sustainable approach to regulating financial markets. University of Illinois Law Review 2013(3), 881–918.
  • Jöreskog, K. G. (1979). Basic ideas of factor and component analysis. In Advances in Factor Analysis and Structural Equation Models, Jöreskog, K. G., and Sörbom, D. (eds), pp. 5–20. University Press of America, New York.
  • Jöreskog, K. G., and Wold, H. (1982). The ML and PLS techniques for modeling with latent variables: historical and comparative aspects. In Systems under Indirect Observation: Causality, Structure, Prediction, Part I, Jöreskog, K. G., and Wold, H. (eds), pp. 263–270. North-Holland, Amsterdam.
  • Judge, K. (2012). Fragmentation nodes: a study in financial innovation, complexity, and systemic risk. Stanford Law Review 64, 657–725.
  • Kaal, W. A. (2012). Contingent capital in executive compensation. Washington and Lee Law Review 69, 1821–1889.
  • Kaufmann, L., and Gaeckler, J. (2015). A structured review of partial least squares in supply chain management research. Journal of Purchasing and Supply Management 21, 259–272 (
  • Laeven, L., and Levine, R. (2009). Bank governance, regulation and risk taking. Journal of Financial Economics 93, 259–275.
  • Lee, L., Petter, S., Fayard, D., and Robinson, S. (2011). On the use of partial least squares path modeling in accounting research. International Journal of Accounting Information Systems 12, 305–328 (
  • Levy-Carciente, S., Kenett, D. Y., Avakian, A., Stanley, H. E., and Havlin, S. (2015). Dynamical macroprudential stress testing using network theory. Journal of Banking and Finance 59, 164–181 (
  • Lohmöller, J. B. (1989). Latent Variable Path Modeling with Partial Least Squares. Physica-Verlag, Heidelberg.
  • Majerbi, B., and Rachdi, H. (2014). Systemic banking crises, financial liberalization and governance. Multinational Finance Journal 18(3/4), 281–336.
  • Marcoulides, G. A., Chin, W. W., and Saunders, C. (2012). When imprecise statistical statements become problematic: a response to Goodhue, Lewis, and Thompson. MIS Quarterly 36(3), 717–728.
  • Monecke, A., and Leisch, F. (2012). semPLS: structural equation modeling using partial least squares. Journal of Statistical Software 48(3), 1–32 (
  • Nitzl, C. (2016). The use of partial least squares structural equation modelling (PLS-SEM) in management accounting research: directions for future theory development. Journal of Accounting Literature 37, 19–35 (
  • Oet, M. V., Eiben, R., Bianco, T., Gramlich, D., and Ong, S. J. (2011). Financial stress index: identification of systemic risk conditions. Unpublished Working Paper 11-30, Federal Reserve Bank of Cleveland.
  • Oet, M. V., Bianco, T., Gramlich, D., and Ong, S. J. (2013). SAFE: an early warning system for systemic banking risk. Journal of Banking and Finance 37, 4510–4533 (
  • Peng, D. X., and Lai, F. (2012). Using partial least squares in operations management research: a practical guideline and summary of past research. Journal of Operations Management 30, 467–480 (
  • Petter, S., Straub, D. W., and Rai, A. (2007). Specifying formative constructs in information systems research. MIS Quarterly 31(4), 623–656.
  • Richter, N. F., Cepeda Carrión, G., Roldán, J. L., and Ringle, C. M. (2016a). European management research using partial least squares structural equation modeling (PLS-SEM): editorial. European Management Journal 34(6), 589–597.
  • Richter, N. F., Sinkovics, R. R., Ringle, C. M., and Schlägel, C. (2016b). A critical look at the use of SEM in international business research. International Marketing Review 33, 376–404 (
  • Rigdon, E. E. (2016). Choosing PLS path modeling as analytical method in European management research: a realist perspective. European Management Journal 34(6), 598–605 (
  • Ringle, C. M., and Sarstedt, M. (2016). Gain more insight from your PLS-SEM results: the importance-performance map analysis. Industrial Management & Data Systems 116(9), 1865–1886 (
  • Ringle, C. M., Sarstedt, M., and Straub, D. W. (2012). A critical look at the use of PLS-SEM in MIS Quarterly. MIS Quarterly 36, iii–xiv.
  • Ringle, C. M., Wende, S., and Becker, J. M. (2015). SmartPLS 3. SmartPLS GmbH, Bönningstedt.
  • Rixen, T. (2013). Why deregulation after the crisis is feeble: shadow banking, offshore financial centers, and jurisdictional competition. Regulation & Governance 7, 435–459 (
  • Sarstedt, M., Ringle, C. M., Smith, D., Reams, R., and Hair, J. F. (2014). Partial least squares structural equation modeling (PLS-SEM): a useful tool for family business researchers. Journal of Family Business Strategy 5, 105–115 (
  • Sarstedt, M., Hair, J. F., Ringle, C. M., Thiele, K. O., and Gudergan, S. P. (2016). Estimation issues with PLS and CBSEM: where the bias lies! Journal of Business Research 69(10), 3998–4010 (
  • Sarstedt, M., Ringle, C. M., and Hair, J. F. (2017). Partial least squares structural equation modeling. In Handbook of Market Research, Homburg, C., Klarmann, M., and Vomberg, A. (eds). Springer (
  • Schubring, S., Lorscheid, I., Meyer, M., and Ringle, C. M. (2016). The PLS agent: predictive modeling with PLS-SEM and agent-based simulation. Journal of Business Research 69(10), 4604–4612 (
  • Shmueli, G., Ray, S., Velasquez Estrada, J. M., and Chatla, S. B. (2016). The elephant in the room: evaluating the predictive performance of PLS models. Journal of Business Research 69(10), 4552–4564 (
  • Sohn, S. Y., Han, H. K., and Jeon, H. J. (2007). Development of an air force warehouse logistics index to continuously improve logistics capabilities. European Journal of Operational Research 183, 148–161 (
  • Strafford (2011). Bank holding and financial holding companies: evaluating the legal and business implications. Webinar Transcript, Stafford. URL:
  • Tenenhaus, M., Esposito Vinzi, V., Chatelin, Y. M., and Lauro, C. (2005). PLS path modeling. Computational Statistics & Data Analysis 48(1), 159–205.
  • Tung, F. (2011). Pay for banker performance: structuring executive compensation for risk regulation. Northwestern University Law Review 105(3), 1205–1251.
  • Willaby, H. W., Costa, D. S. J., Burns, B. D., MacCann, C., and Roberts, R. D. (2015). Testing complex models with small sample sizes: a historical overview and empirical demonstration of what partial least squares (PLS) can offer differential psychology. Personality and Individual Differences 84, 73–78 (
  • Wold, H. O. A. (1982). Soft modeling: the basic design and some extensions. In Systems Under Indirect Observations: Part II, Jöreskog, K. G., and Wold, H. O. A. (eds), pp. 1–54. North-Holland, Amsterdam.
  • Wu, W. W., Lan, L. W., and Lee, Y. T. (2012). Exploring the critical pillars and causal relations within the NRI: an innovative approach. European Journal of Operational Research 218, 230–238 (

Only users who have a paid subscription or are part of a corporate subscription are able to print or copy content.

To access these options, along with all other subscription benefits, please contact [email protected] or view our subscription options here:

You are currently unable to copy this content. Please contact [email protected] to find out more.

You need to sign in to use this feature. If you don’t have a account, please register for a trial.

Sign in
You are currently on corporate access.

To use this feature you will need an individual account. If you have one already please sign in.

Sign in.

Alternatively you can request an individual account here: