Does a Captured Bond Trustee Change How Firms Manage Earnings?

Firms that want to look healthier than they are can massage their reported profits in two ways. They can bend accounting estimates, for example by recognizing revenue early or under-provisioning for bad debts. Researchers call this accrual-based earnings management. Or they can bend real operations, for example by cutting research budgets, overproducing inventory so fixed costs are spread over more units, or discounting to pull sales forward. This is real earnings management. Accounting tricks are cheap but visible to a vigilant auditor or creditor. Operational tricks are hard to detect but destroy value. Evidence from U.S. loan covenant violations shows that when lenders tighten monitoring, managers stop moving accruals and switch to operational manipulation (Debt Covenant Violation and Real Earnings Management). That result assumes the monitor is actually vigilant.

In Chinese domestic corporate bonds, the bond trustee is often the same bank that underwrote the deal. The trustee is the party legally charged with watching the issuer for bondholders. The underwriter depends on the issuer for repeat fee business. This proposal tests a sharp prediction that follows from that conflict. Under a conflicted trustee-underwriter, firms facing debt pressure should keep using cheap accrual manipulation and should not pay for the costly switch to operational manipulation. The quantity of interest is therefore the mix of manipulation, not only its level. The design matches each conflicted-trustee issuer to an issuer that uses the same lead underwriter but appoints an independent trustee. This separates the effect of the dual role from the effect of which bank is involved.

Core Claim

The substitution from accruals to real operations documented under arm's-length creditor monitoring is conditional on monitor independence. Among bond issuers in high-pressure years (refinancing due within twelve months, or earnings just above zero), those whose trustee is also their lead underwriter will show higher discretionary accruals and a lower share of real manipulation in total manipulation, relative to matched issuers with independent trustees. This reconciles two findings that currently sit in tension. The U.S. covenant study finds that distress pushes firms toward real manipulation. The Vietnamese study (Financial distress and earnings management, HOSE) finds that distressed firms adjust accruals downward under creditor and auditor scrutiny. The proposal attributes both patterns to how strong the monitor is, rather than to distress itself.

The closest existing work is the legal case analysis of cross-default clauses in Chinese real estate bonds (Cross-default clause in Chinese real estate bonds). That paper documents the dual-role conflict and its visible contractual consequence: Country Garden's USD-bond cross-default trigger was 12.2% of its domestic threshold. It runs no estimation and covers only distressed developers. The U.S. covenant paper measures the substitution but never asks who the monitor is. Nothing in the file links trustee identity to firm-level accrual or operational manipulation.

Research Program

Sample. The sample is built from Chinese A-share issuers of domestic corporate bonds. Bond prospectuses compiled through Wind supply, for each issue, the lead underwriter, the trustee and the cross-default terms. CSMAR supplies annual and quarterly financials, governance, and institutional ownership. The unit is the issuer-year.

Treatment. An issuer-year is treated if any outstanding bond has a trustee identical to its lead underwriter.

Accrual outcome. Discretionary accruals come from the performance-matched Modified Jones model. This model regresses total accruals on revenue change net of receivables and on fixed assets within industry-year groups and treats the residual as the discretionary part.

Operational outcome. Real manipulation is measured the standard way. Abnormal operating cash flow, abnormal production cost and abnormal discretionary expense are the residuals from industry-year regressions of each item on sales levels and sales changes, and they are combined into one index.

Mix outcome. The headline outcome is the real-manipulation share: the real index divided by the sum of the absolute real index and absolute discretionary accruals, both standardized.

Detector that avoids the Jones model. The Jones residual misreads growth and distress as discretion. As a third detector, the program uses the fourth-quarter anomaly test from Testing earnings management. That paper fits periodic autoregressions and found a weakened fourth-to-next-first-quarter correlation in 29% of Compustat firms and inflated fourth-quarter variance in 38%. The program fits the same per-firm periodic model and compares the frequency of these anomalies across treated and control arms.

Matching. Matching is the methodological core. Propensity score matching collapses many firm characteristics into one estimated probability of treatment, $e(X)=P(T=1\mid X)$, and pairs treated and control units with similar scores. The firm-level covariates in $X$ are size, leverage, ROA, operating cash flow, Altman Z, growth, ownership type (state-owned or not), the institutional-investor independence split from Institutional Investor Heterogeneity, the internal control index from Internal Control, Management Power, and industry.

Plain matching on these variables treats firms as independent units. Firms sharing an underwriter are not independent: the underwriter's book, reputation and client pool confound both the trustee arrangement and reporting quality. Propensity Score Matching for Causal Inference with Relational Data shows that ignoring this kind of network dependence inflates false discoveries on a Foursquare benchmark. It also shows that adding aggregated neighbor covariates and blocking on shared relations removes the bias. That method is reused here. Matches are required to share the lead underwriter and year (relational blocking). The score is augmented with aggregates of the underwriter's other clients, such as the mean and dispersion of their accruals and their default history, and with underwriter degree (number of clients).

Inference and robustness. Treatment effects on the treated are estimated within and outside high-pressure years. Standard errors are clustered by underwriter. Sensitivity to an unobserved confounder is reported as the confounder strength needed to erase the effect.

Baselines

Naive reference point. The first reference point is the raw difference in each outcome between treated and untreated issuer-years. It shows how much any adjustment moves the answer. The file reports that on LinkedIn data a naive comparison overstated a lift by 76%, a 10% naive effect against 2.4% after matching (The Importance of Being Causal). That is a domain-external warning, not a target figure.

Strongest competitor: fixed-effects regression. This is an OLS regression of each outcome on the treatment dummy, the full covariate set, and issuer, year and underwriter fixed effects. It is the strongest competitor because underwriter fixed effects absorb much of the confounding that relational matching targets, at a fraction of the complexity. It runs on the identical issuer-year panel and covariates and uses the same clustering. The matching method beats it only if covariate-balance diagnostics show the regression is extrapolating outside common support and the matched estimate differs from it materially. If the two agree, the substantive finding stands and matching is reported as a robustness layer, not a contribution.

Matching ablation. Standard i.i.d. nearest-neighbor matching with a caliper, on the same firm covariates without underwriter blocking or neighbor aggregates, isolates whether the relational ingredient matters.

Prior quantitative result. The closest published estimate is the U.S. regression-discontinuity finding that real manipulation rises by about 0.05% per quarter after a covenant violation with no significant change in discretionary accruals (Debt Covenant Violation and Real Earnings Management). The independent-trustee arm should reproduce that sign pattern: real manipulation up, accruals flat in pressure years. For the effect of trustee-underwriter duality on accruals or operational manipulation, no reported result was found on Lacuna.

Why not just…?

Why not just run the fixed-effects regression and stop? This is a real threat, and the proposal does not argue around it. The fixed-effects regression is run as the primary competitor. If treated and untreated firms overlap well on covariates, it may give the same answer, and that answer is the paper's finding. Matching earns its place only where balance diagnostics show the treated firms are unusual, for example concentrated among state-owned or real estate issuers, so that a linear specification extrapolates. The comparison of the two estimates, together with balance tables, settles which applies.

Why not just measure discretionary accruals, the field standard? The hypothesis concerns substitution, so accruals alone cannot test it. A conflicted trustee could show "no accrual effect" either because it does not matter or because it shifts manipulation across channels. The U.S. evidence found exactly a null on accruals alongside a real-manipulation jump, and the Vietnamese study names its omission of real manipulation as its main limitation. The mix outcome, and the fourth-quarter detector that does not depend on the Jones model, are what make the hypothesis falsifiable.

Why not just use issuers that switch trustee arrangements across their own bond issues? This within-firm event study is cleaner on unobserved, time-fixed issuer traits, and the program runs it. Its weakness is that switching is itself a choice that may coincide with deteriorating credit. It is therefore reported alongside the matched cross-section, and agreement in sign between the two is required before any claim is made. If a regulation separating the trustee and underwriter roles exists in the sample period, it would supersede both designs. The file documents none, so this is checked rather than assumed.

Proposed Experiments and Decisive Checks

Main test. The decisive test is the treatment-by-pressure interaction on the real-manipulation share. The hypothesis predicts this interaction is negative, and that discretionary accruals rise in treated pressure years. The mirrored pattern in the independent arm (real manipulation up, accruals flat) replicates the U.S. result. If both arms substitute equally, the capture hypothesis fails.

Relational falsification. On a placebo treatment, randomly reassigned within underwriter-year blocks, i.i.d. matching should produce more spurious "effects" than relational matching. This mirrors the Type I error inflation in the relational matching paper, now measured on financial data.

Contract-channel check. Within treated issuers, looser domestic default triggers, measured as the domestic threshold relative to the same parent's offshore threshold where one exists, should predict a lower real-manipulation share. This is the mechanism the cross-default case study implies.

Consequence check. Treated firms with high accruals should show more later rating downgrades or defaults. Matched independent-trustee firms with high real manipulation should instead show declines in later R&D and gross margin. This reflects the claim that real manipulation leaves lasting operational damage while accrual manipulation tends to reverse.