A bond trustee is the party legally charged with protecting bondholders: it monitors the issuer's promises and declares default when they are broken. In China's domestic corporate bond market the trustee is often the same investment bank that underwrote the bond and earns fees from the issuer. A legal case study of real estate bonds (Difficulties and Optimization Paths of the Application of "Cross-default" Clause in Chinese Real Estate Corporate Bonds) found that, for Country Garden, the cross-default trigger on offshore USD bonds with independent trustees was only 12.2% of the domestic trigger. That study also describes issuers delaying revenue recognition to stay inside covenant limits. No one has measured whether this conflict changes how firms report earnings. Our proposal is to measure it in a way that tells a lax monitor apart from a badly governed firm.
The test rests on two predictions. First, managers facing a strict creditor monitor avoid manipulating accounting estimates and instead cut real activity, such as R&D or discounting, because accruals are what monitors check. Debt Covenant Violation and Real Earnings Management: Evidence From the U.S. shows exactly this in a regression discontinuity around covenant thresholds: accruals did not change and real manipulation rose about 0.05% per quarter. A conflicted trustee should flip that mix back toward cheap accounting manipulation. Second, this flip should appear only for managers whose pay rewards upside. Managers whose wealth behaves like the firm's debt, through deferred pay, pensions, or fixed salary (the "pension-induced risk aversion" of How Compensation and Liability Structures Distort Managerial Risk), already share bondholders' interests. They should not exploit a weak trustee, and they may push accruals down to build reserves instead.
Core Claim
The first claim is about the mix of manipulation. Affiliation between trustee and underwriter raises discretionary accruals (the part of reported accounting adjustments that firm fundamentals do not explain) relative to real earnings management. The second claim is about who responds. The shift is concentrated among issuers whose executives hold convex, equity-like payoffs. Among executives with debt-like payoffs the shift is absent or reversed, showing up as negative accruals in the lower quantiles.
This pattern is what separates our proposal from the obvious one, which would regress accruals on a trustee-duality dummy. A firm that is simply badly run should show more of both kinds of manipulation, whatever its executives' pay shape. A firm whose monitor is lax on accounting should show substitution toward accruals, and only where managers want to inflate earnings. The pay-shape dimension therefore acts as a built-in placebo for the trustee effect. In turn, the trustee variation gives the pension-aversion literature its first test of the open question flagged in the file: does debt-like executive wealth change the sign of accrual management?
Research Program
Data. We will build an issuer-year panel of Chinese A-share firms with outstanding bonds, using three linked sources. CSMAR supplies financials and executive pay. WIND supplies bond prospectuses, including the identities of the lead underwriter and the trustee, and the cross-default thresholds. CNRDS supplies governance variables. These are the same sources used by Deng & Tu, Chen, and Peng et al.
The treatment is the share of an issuer's outstanding bond principal whose trustee is the lead underwriter or an affiliate. We read this directly from prospectuses rather than inferring it from investor types, as Deng & Tu do.
For executive payoff shape, we compute the ratio of debt-like claims (deferred compensation, enterprise annuity or pension disclosures, and fixed salary) to equity-linked wealth (shares and options). This ratio comes from annual-report compensation tables. We acknowledge that Chinese disclosure of inside debt is thinner than U.S. disclosure, so this is a proxy. If the debt-like component turns out to be too sparsely disclosed, we will fall back on the post-reform state-owned enterprise pay structure as a coarser measure, and say so.
Accrual measurement. We first estimate the modified Jones model, which predicts normal accruals from revenue growth and plant and equipment within industry-year peer groups. We then apply the Kothari performance adjustment, which controls for return on assets, as used in Earnings Management in Small Listed Firms in Malaysia Using Quantile Regression.
One methodological change matters here. Corporate Fraud Detection in Rich-yet-Noisy Financial Graph reports that only 30% of Chinese accounting frauds are detected in the year they occur, with a gap of about two years on the Main Board. A standard peer regression therefore treats undetected manipulators, including conflicted-trustee firms, as "normal." That pulls the benchmark toward them and shrinks any measured effect. We will instead estimate the normal-accrual model on a clean peer set. These peers have only independent trustees and no regulatory enforcement in the following three years. We will then score all firms against this clean benchmark.
Real earnings management will be measured with the three standard abnormal-operations measures used by Yang & Zhang: abnormal cash flow, abnormal production costs, and abnormal discretionary expenses.
Identification. A firm's choice of trustee is not random, so the pooled comparison is only descriptive. Our main design uses an issuer's first offshore USD bond with an independent trustee, or a new domestic issue that replaces an affiliated trustee, as a staggered event. That event lowers the firm's conflicted-debt share. We will estimate it with a heterogeneity-robust staggered difference-in-differences estimator.
The central specification is a triple difference:
$$ Y_{it} = \beta_1 C_{it} + \beta_2 C_{it} D_{it} + \beta_3 D_{it} + \gamma' X_{it} + \alpha_i + \tau_t + \varepsilon_{it} $$
In this equation, $Y_{it}$ is either signed discretionary accruals or the share of total manipulation done through accruals rather than real activity, for firm $i$ in year $t$. $C_{it}$ is the conflicted-trustee share of bond principal. $D_{it}$ is the executives' debt-like pay ratio. $X_{it}$ contains governance and distress controls. $\alpha_i$ and $\tau_t$ are firm and year fixed effects. The prediction is $\beta_1 > 0$ and $\beta_2 < 0$, with $\beta_1 + \beta_2 \le 0$ at high $D$.
We will run the same specification in quantile regressions at the 10th, 25th, 50th, 75th and 90th percentiles. The reason is that Abdullah shows governance effects that are invisible at the mean but appear in the 70th–90th quantiles. The predicted downward provisioning by debt-like managers is a lower-tail phenomenon.
Baselines
The naive reference point is the obvious study. It regresses absolute modified-Jones discretionary accruals on a trustee-duality dummy using pooled OLS with standard size, leverage, and growth controls. It uses the same sample and the conventional all-peer benchmark. We report it to show how much of its estimate survives the fixed effects, the event design, and the clean benchmark. If the naive estimate is large and the event-study estimate is near zero, selection explains the naive result.
The strongest existing method is the monitor-independence design of Institutional Investor Heterogeneity, Accounting Information Quality and Enterprise Performance. That study classifies institutional holders as independent (mutual funds, social security funds, foreign institutions) or commercially tied (brokers, insurers, trusts). It relates these holdings to modified-Jones accruals and performance with fixed effects. It reports a coefficient of 0.03 for independent ownership on return on assets and an interaction of −0.511 between accruals and institutional ownership. It is the closest competitor because it already proxies "conflicted monitor" with commercial ties. If its holder classification captures the same variation as trustee affiliation, our measure adds nothing.
We will rebuild its independent and tied ownership variables on our panel and run both specifications with identical controls, fixed effects, and sample. Two results would count as beating it. First, trustee affiliation must keep a significant accrual-substitution effect after its ownership variables are included. Second, our model must explain more within-firm variance in the accrual share than the ownership-only model does.
The governance horse race adds the internal monitors that skeptics would point to. These are internal control quality and management power from Internal Control, Management Power and Accounting Information Quality: Based on Chimeric Perspective, and board independence and audit-committee expertise from Abdullah. Chen reports only that the interaction of internal control and power is significantly negative. No reported magnitude was found on Lacuna for that interaction, so the comparison is about whether trustee effects survive its inclusion, not about matching a number.
Two ablations isolate our key ingredients. The first re-runs everything with the conventional all-peer accrual benchmark instead of the clean benchmark. We expect attenuated trustee coefficients; if they are not attenuated, the detection-lag concern is empirically minor here. The second drops the pay-shape interaction. This shows whether the effect is general monitor laxity or specifically exploited by convex-pay managers.
For the secondary predictive check, the strongest reported comparison is KeGCN_R. On the Main Board it reaches an area under the ROC curve of 0.782, against 0.706 for XGBoost, in the fraud-graph paper. We will run XGBoost with and without our trustee-conflict and clean-benchmark accrual features on a matched train/test split of later enforcement cases. The goal is an incremental gain over XGBoost on our own sample, not a claim against 0.782, which was measured on a different task setup.
Why not just…?
Why not just regress accruals on the trustee-duality dummy? That is exactly the naive baseline, and it cannot separate a lax monitor from a bad firm. Issuers that accept their underwriter as trustee may be smaller, weaker, or more connected. Any of these would raise accruals without the trustee doing anything. Our proposal relies on the substitution prediction from Yang & Zhang and on the pay-shape placebo. A bad firm should raise both accrual and real manipulation regardless of pay, and that pattern is what would reject our mechanism.
Why not just use realized outcomes such as defaults, downgrades, or regulatory penalties instead of noisy accrual residuals? This is a real threat, because accrual residuals are estimated with error. Kim & Yoon's Corporate Bankruptcy Prediction with Domain-Adapted BERT shows that ratio-based models are lagging: their hazard-model R² is 16.23% and rises to 26.38% only when disclosure text is added. Can & Wang describe accounting discretion used precisely to keep ratios and ratings clean until collapse. Realized outcomes therefore arrive late and are themselves shaped by the trustee's enforcement choices. Still, we will use later regulatory penalties as a validation outcome. If clean-benchmark accruals at conflicted-trustee firms do not predict later enforcement, we will report that the accrual measure is not capturing manipulation.
Why not just rely on internal governance such as board independence or CEO-chair separation? Abdullah finds board independence has no significant association with accruals at any quantile. Can & Wang note that only the trustee can call bondholder meetings and trigger acceleration, so no board structure substitutes for it. Even so, Chen shows internal control does constrain accruals. We therefore enter these variables directly in the horse race rather than assume them away.
Why not just measure real earnings management alone, since accrual models miss it? Real manipulation alone cannot reveal substitution. The core signature is the share of manipulation done through accruals, which needs both measures.
Decisive Checks
Four results would decide the proposal. First, in the event study, a firm's first independently trusteed offshore issue should be followed by falling accruals and flat or rising real manipulation. Pre-event trends must be flat, and the effect should appear only among convex-pay executives. Second, in the quantile regressions, debt-like-pay executives at high conflicted-trustee shares should sit lower in the signed-accrual distribution, which is the cookie-jar prediction. Third, the clean benchmark should enlarge the trustee coefficient relative to the all-peer benchmark. Fourth, the trustee-conflict effect should survive the Deng & Tu ownership controls.
Two results would falsify the claim. One is equal effects across pay shapes. The other is parallel rises in accrual and real manipulation, which would point to firm quality rather than monitor laxity. Either outcome would itself be informative about the limits of the cross-default conflict that Can & Wang documented qualitatively.
References
- Difficulties and Optimization Paths of the Application of "Cross-default" Clause in Chinese Real Estate Corporate Bonds
- Debt Covenant Violation and Real Earnings Management: Evidence From the U.S.
- Controlling shareholders’ equity pledge and corporate innovation investment—empirical analysis based on pharmaceutical manufacturing
- Earnings Management in Small Listed Firms in Malaysia Using Quantile Regression
- Corporate Fraud Detection in Rich-yet-Noisy Financial Graph
- Institutional Investor Heterogeneity, Accounting Information Quality and Enterprise Performance
- Internal Control, Management Power and Accounting Information Quality: Based on Chimeric Perspective
- Corporate Bankruptcy Prediction with Domain-Adapted BERT