Is the Governance U-Curve a Mixture? Latent Classes of Concentrated Power in Chinese Listed Firms

A widely cited result in Chinese corporate governance says that the relationship between ownership concentration and earnings manipulation is U-shaped. When ownership is diffuse, managers go unwatched. At moderate concentration, large shareholders monitor them. Past roughly one-third ownership, controlling owners start manipulating the books to hide the resources they extract. Another line of work, on the same market, fits a straight line and finds that concentration only makes things worse. This note proposes a different explanation that can be tested. The aggregate U may not describe how any single firm behaves. It may instead arise from mixing two unobserved kinds of firms. In one kind, concentrated owners monitor, so manipulation falls as their stake grows. In the other, concentrated owners entrench themselves, so manipulation rises. Which kind a firm belongs to is correlated with how concentrated it is. Pool such firms into one regression and it can draw a U with a spurious universal turning point.

To test this we borrow a tool from transportation safety and survival statistics: latent class regression. This is a model that assumes the population is secretly split into a few groups, each with its own coefficients, and learns the groups from the data. We apply it to firm panels, with each firm's group held fixed across years. The payoff is both substantive and methodological. If the mixture account holds, the "33% rule" should be retired, and the question for investors becomes which kind of firm is this, not how concentrated is it. The two accounts also make opposite predictions about how a given firm's reporting changes when its ownership stake changes. That lets us adjudicate between them with within-firm events rather than with goodness of fit alone.

Core Claim

The claim has three parts. First, discretionary accruals (the part of reported earnings not explained by normal business activity, a standard proxy for manipulation) respond to ownership concentration and management power with slopes that differ in sign across unobserved classes of firms. Second, within each class the relationship is close to monotone. Third, the pooled quadratic curve reported by Liu's study of private holding enterprises can be reproduced by a two-class mixture whose membership shifts with concentration. The observable consequence that separates the accounts is this. Under a true U, a firm's response to a rise in its largest shareholder's stake depends on whether it sits above or below the turning point. Under the mixture, the response depends on the firm's class, whatever its starting level.

The idea builds on three pieces of prior work. The motivating evidence comes from the governance papers. Liu fits a deterministic quadratic with a turning point near 33.3% on 6,045 firm-years from 2008 to 2014. Wei fits a linear model of inefficient investment on 5,192 firm-years from 2017 to 2020 and reports a concentration coefficient of 0.0121 and an interaction with internal control of −0.0015. Chen Chen shows that management power inflates discretionary accruals among 3,673 manufacturing firms, that internal control quality dampens this, and that the dampening differs between state-owned and private firms. All three treat the governance slope as one number, possibly within an observed subsample.

The method comes from the latent class accelerated hazard study of ring-road incidents. It found two latent classes in which the same factor flips sign: mobile-phone reporting cut clearance time by 17.6% in one class and raised it by 79.9% in the other. We reuse the finite-mixture idea but move it from single events to repeated firm-years, holding class membership fixed within a firm.

The constraint comes from the crash-frequency random parameters work, which reports that ignoring slope heterogeneity can make marginal effects more than 100% inaccurate. That warning cuts both ways. A mixture is only one form of slope heterogeneity, so we must show that it beats the continuous alternative rather than assume it does.

Research Program

For firm $i$ in year $t$, with latent class $c \in \{1,\dots,C\}$ fixed per firm, the within-class model is

$$DA_{it} = \alpha_c + \beta_c\,TOP_{it} + \gamma_c\,POWER_{it} + \delta^\top X_{it} + \lambda_t + \varepsilon_{it}, \qquad \varepsilon_{it} \sim N(0,\sigma_c^2).$$

The symbols are as follows. $DA_{it}$ is absolute discretionary accruals from the modified Jones model. $TOP_{it}$ is the largest shareholder's stake. $POWER_{it}$ is the management-power index used by Chen Chen, built from CEO duality, tenure and board control. $X_{it}$ holds standard controls: size, leverage, growth and auditor. $\lambda_t$ is a year effect. The probability that firm $i$ belongs to class $c$ follows a multinomial logit in slow-moving firm traits $Z_i$: state versus private control, industry, province, and the initial internal-control index from the DIB database. We estimate by expectation–maximization. The number of classes is chosen by held-out likelihood on firms withheld entirely from fitting, not by in-sample BIC alone.

Before touching real data, we run a simulation study. Data are generated from two monotone classes whose membership probability rises with concentration. We then check two things: whether a pooled quadratic finds a significant U with a stable turning point, and whether our estimator recovers the true classes across panel lengths typical of CSMAR, which is five to fifteen years per firm. This establishes that the confound we propose is possible at realistic sample sizes and that the method can detect it.

The empirical panel is CSMAR non-financial A-share firms from 2008 to the latest available year. This window covers Liu's period, so we can replicate his result, and extends past it. We link DIB internal-control scores. Following the temporal stability analysis of lighting conditions in crashes, we use likelihood ratio tests to check whether class-specific slopes stay stable across the 2008–2014 and post-2014 windows. We do this because the file documents that such stability is routinely rejected in pooled multi-year data. We keep this test narrow: it asks whether classes are an era artifact, not whether governance changed.

Baselines

The naive reference point is pooled OLS of accruals on concentration, power and controls with year and industry effects. This is the linear specification of Wei's type, and any heterogeneity model must at least beat its held-out likelihood.

The strongest existing competitor is Liu's pooled quadratic in concentration. The only figure reported for it on Lacuna is the turning point of approximately 33.3% ownership. No fit statistic or held-out error was found on Lacuna for it, so we re-estimate it ourselves on the identical panel with the identical controls. It is the most serious rival because it already explains the non-monotonicity with fewer parameters.

We add a flexible pooled curve: a spline or threshold regression in concentration, with the threshold chosen by grid search. This baseline gets the same tuning budget as our choice of class count. We also add a Gaussian random-slope mixed model with firm-level normal slopes on concentration and power and heterogeneous means driven by $Z_i$, in the spirit of the crash-frequency random parameters formulation. This is the continuous-heterogeneity alternative.

All models use the same firm-grouped five-fold splits, the same controls, and the same held-out firms. Beating the baselines requires two things. On held-out firm log-likelihood, the mixture must beat the quadratic, the spline and the Gaussian random-slope model. More importantly, it must win the event-study test described below.

The key ablations isolate two ingredients. The first lets a firm's class change every year; this checks whether persistence of class carries the result. The second allows quadratic terms inside each class. If within-class curvature remains large and significant, the mixture explanation fails and we will say so.

Why not just…?

Why not just add firm fixed effects? Fixed effects absorb differences in the level of manipulation across firms, but they hold the slope on concentration constant. They also identify that slope only from within-firm movement in $TOP$, which is slow and sparse for most Chinese listed firms. So fixed effects would largely discard the cross-sectional variation in which the U-shape was found. We report the fixed-effects estimate anyway. Its imprecision is informative, but it cannot tell a U apart from a mixture.

Why not just split by state ownership or internal-control quality? Chen Chen and the ESG-restatement study already do this. The latter reports ESG coefficients of −0.408 versus −0.259 across concentration halves. This is a real threat. If the latent classes coincide with state versus private control plus high versus low internal control, the latent model adds nothing beyond a four-cell split. We settle it directly by comparing held-out likelihood and class assignments between the mixture and the fully interacted observed-split model. If the observed split does as well, the finding shifts from "latent" to "the U is a state/private composition effect." That is still a refutation of the universal turning point, and we would report it as such.

Why not just use a flexible nonlinear curve? A spline will fit the U at least as well as the mixture in the pooled data, since both reproduce the same average relationship. Fit alone cannot decide the question. The decisive evidence has to come from firm-level dynamics, which is why the event study is the core experiment.

Why not just use normally distributed random slopes? A normal distribution of slopes centered near zero can mimic some sign heterogeneity. It cannot, however, produce a bimodal split between monitoring and entrenching firms tied to firm traits. We compare the two directly. If the Gaussian model matches the mixture out of sample and in the event study, we adopt the simpler continuous story.

Proposed Experiments and Decisive Checks

The first experiment is replication. We re-estimate Liu's quadratic on 2008–2014 and confirm a turning point near 33.3%, then check whether that turning point moves in later years or across provinces. Instability of the turning point is weak evidence for composition effects.

The second is the simulation-recovery study described above. It reports how often a pooled quadratic declares a significant U when the truth is two monotone classes, and how accurately our estimator recovers class slopes.

The third, and decisive, experiment is a within-firm event study. We use discrete jumps in the largest shareholder's stake: block transfers, private placements and share-pledge liquidations recorded in CSMAR. Firms are assigned to classes using only pre-event years. The pooled-U account predicts that the accrual response to a stake increase is negative below about 33% and positive above it, whatever the class. The mixture account predicts that the response is negative for monitoring-class firms and positive for entrenchment-class firms at any starting level. We estimate responses in a two-by-two grid of pre-event level by predicted class, using difference-in-differences against matched non-event firms. If the sign follows class rather than level, the mixture wins. If it follows level, the U is real at the firm level.

This event study does not solve endogeneity in general. Stake changes are chosen by the owners making them. It does, however, give the two accounts opposite predictions, which fit statistics cannot do.

The fourth check is external. We test whether classes learned on accruals predict investment inefficiency (Richardson residuals, as in Wei) and future restatements, neither of which was used to fit the model. The goal is to show that the classes capture a governance type rather than accrual measurement noise.