Measuring Earnings Management When the Regulation Moves the Fundamentals

Researchers who want to know whether firms "cook the books" under regulatory pressure usually measure discretionary accruals. Accruals are the accounting entries that do not correspond to cash in the same period, such as depreciation or receivables. The researcher regresses total accruals on business fundamentals, mainly revenue changes and fixed assets, and treats the leftover residual as the part managers chose. That residual is then used as the outcome in a policy evaluation, typically propensity score matching followed by difference-in-differences (PSM-DID). Propensity score matching pairs each regulated firm with an unregulated firm that looked similar beforehand. Difference-in-differences compares the two groups' changes before and after the policy.

The problem this note targets is simple to state. A compliance shock such as China's carbon emissions trading pilots forces firms to buy abatement equipment and changes their revenue. Those are exactly the fundamentals the accrual model uses to define "normal" accruals. The residual is therefore computed by conditioning on variables the treatment itself caused. In causal terms, the outcome has a mediator built into its definition. The common performance-matching step makes things worse: it adjusts for current profitability, and current profitability contains the accruals being measured, so it sits downstream of the outcome.

We propose to treat "discretionary accruals under a regulatory shock" as a direct-effect estimation problem, and to estimate it in a way that respects timing. We will build a semi-synthetic simulator where the true manipulation effect is known and use it to show how far standard pipelines miss. We then re-answer an open empirical question on the same CSMAR data: do carbon trading pilots push firms to manage accruals, beyond the tax avoidance already documented?

Core Claim

The standard discretionary-accrual pipeline has a bias whose size and sign are predictable. When a policy shifts the fundamentals that define expected accruals, a PSM-DID on residual accruals mixes three things: the effect of manipulation, the effect of mis-specified normal accruals on treatment-altered fundamentals, and contamination of the cross-sectional accrual coefficients by treated firms. Performance matching on contemporaneous return on assets adds collider bias on top.

The correction fixes the "normal accrual" mapping and the matching covariates to information that predates treatment. Concretely, three choices go together:

  1. The accrual expectation model is fitted only on never-treated firms and pre-policy years.
  2. The quantity of interest is defined as the natural direct effect of treatment on total accruals, holding fundamentals at the values they would have taken without treatment.
  3. Matching and outcome covariates are screened for post-treatment status before use.

Write total accruals for firm $i$ in year $t$ as $TA_{it} = g(M_{it}) + D_{it}$. Here $M_{it}$ is the vector of fundamentals (revenue change net of receivables, fixed assets), $g$ is the normal-accrual mapping, and $D_{it}$ is managerial discretion. The target is

$$\theta = E\left[TA_{it}(1, M_{it}(0)) - TA_{it}(0, M_{it}(0)) \mid T_i = 1, t \ge t_0\right].$$

In this expression, $T_i$ marks pilot-region firms, $t_0$ is the policy year, and $TA(a, m)$ is the potential total accrual under treatment status $a$ with fundamentals set to $m$. Standard practice instead estimates a DID on $TA - \hat g(M(1))$, with $\hat g$ fitted on pooled industry-year cross-sections. That equals $\theta$ only if $g$ is linear and correctly specified, and only if its coefficients are unaffected by treated firms. Both conditions fail precisely when compliance spending is large.

Relation to Prior Work

The closest empirical work is Does the Carbon Emissions Trading Policy Increase Corporate Tax Avoidance? Evidence from China (also indexed as Does the Carbon Emissions Trading Policy Increase Corporate Tax Avoidance? Evidence from China Does the Carbon Emissions Trading Policy Increase Corporate Tax Avoidance? Evidence from China). It applies PSM-DID to CSMAR industrial firms over 2008–2016, finds that the pilots increased tax avoidance through an operating-cash-flow squeeze, and strips discretionary accruals out of its book-tax measure as a nuisance. That paper motivates ours and supplies the design. We differ in two ways: we make accrual discretion the outcome, and we question whether the accrual filter itself is valid once fundamentals respond to the policy.

Three papers document the compliance squeeze that makes the bias large, and so serve as motivation:

  • Li Li (2020) on profit declines under China's New Environmental Protection Law.
  • Yi, Zhang and Feng (2024) on massive capital expenditures by coal firms under the dual-carbon targets.
  • Feng, Zeng and Qian (2018) on end-of-pipe spending crowding out green innovation.

Massive treatment-induced capital spending is exactly a shock to the fixed-asset term of the accrual model.

On the accounting side, Nguyen et al. (2024) model accruals on revenue growth and fixed assets and use System GMM for persistence. Abdullah (2022) motivates performance matching and shows that accrual behaviour concentrates in the tails. That evidence of heavy tails constrains our estimator: it must be checked at quantiles, not only at the mean.

The method comes from causal ML. Causal Foundation Models Perform Better without Post-treatment Variables decomposes the bias from conditioning on post-treatment variables into an indirect effect, an interaction term and selection bias. It also introduces TC-LD, a two-phase screen that keeps treatment-associated variables and prunes treatment descendants. We reuse that decomposition and screen, but on a different object. That paper prunes post-treatment columns from a covariate table; here the post-treatment variable is hidden inside the construction of the outcome, a residual from a first-stage regression. Simply deleting a column cannot fix that. The sequential adjustment criterion in snu-causality-lab/minSAC supplies a check on which time-indexed covariates form a valid adjustment set. CausalProfiler and SpaCE show the value of semi-synthetic benchmarks with known truth and warn about the gap between simulation and real data. Neither contains accounting identities, which is what our simulator adds.

Research Program

The first component is a semi-synthetic firm-panel simulator anchored to real data. We take pre-2013 CSMAR industrial-firm balance sheets as the starting state. Accounting structure is then enforced explicitly:

  • Accruals equal earnings minus operating cash flow.
  • Return on assets includes accruals.
  • Receivables link reported revenue to cash revenue.
  • Fixed assets accumulate capital expenditure.

A placebo "pilot" assignment is drawn from a selection model calibrated on pre-period firm size, age, leverage and region, mimicking the observed pilot selection. Treatment then acts through three tunable channels. The first is a real channel, where compliance capital expenditure raises fixed assets and depreciation accruals. The second is a cash channel, where quota purchases lower operating cash flow. The third is a manipulation channel, where managers shift a known amount of accruals, with heterogeneity by ownership and financial constraint as documented by Feng et al. (2022). Because the manipulation effect $\theta$ is set by us, we can measure bias, error and confidence-interval coverage for any pipeline. Normal accruals will be generated both linearly and nonlinearly in fundamentals, with heavy-tailed shocks, so that misspecification and tail behaviour are tested and not assumed away.

The second component is the estimator. It has three pieces:

  1. Matching on pre-period covariates. Propensity scores are estimated only on covariates measured before $t_0$. This covariate set is audited with the TC-LD screen and the sequential criterion, which flag any covariate that responds to the pilot.
  2. A frozen normal-accrual model. The normal-accrual mapping $g$ is fitted only on never-treated firms and pre-policy years. We allow a flexible form (gradient-boosted trees or splines) alongside the linear form, so that nonlinearity in the fixed-asset response is not read as discretion.
  3. Fundamentals at their untreated values. Each treated firm's post-period fundamentals are replaced by an estimate of $M(0)$: the matched controls' fundamentals, shifted by the treated firm's own pre-period level. Expected accruals are computed at this estimate.

The DID is then run on $TA - \hat g(\hat M(0))$. The share of the total-accrual effect that flows through fundamentals is reported separately as the non-discretionary part. Performance adjustment uses lagged, pre-policy return on assets, never the contemporaneous value.

The third component is the empirical re-analysis. The data and window match Feng et al. (2022): CSMAR industrial firms, 2008–2016, with the seven 2013–2014 pilot regions. The analysis will report effects separately for state-owned versus private firms and for financially constrained versus unconstrained firms. It will also add a test of the gap the file identifies: whether accrual management is larger where subsidies are at stake.

Baselines

The naive reference point is a two-way fixed-effects regression of the standard residual accruals on the pilot indicator, with contemporaneous firm controls. This is what a reader would do with no causal concern. It shows how much any careful method changes the answer.

The strongest existing empirical competitor is the Feng et al. (2022) pipeline with accruals substituted as the outcome. It uses 1:1 nearest-neighbour caliper propensity matching on firm size, age and leverage, followed by DID with firm controls. Accruals come from the standard revenue-and-fixed-assets residual model fitted per industry-year on the full sample. We add the performance-matched variant motivated by Abdullah (2022), because that is what a careful accounting referee would request. For this pipeline applied to accruals there is no reported result found on Lacuna. Feng et al. report only significance and parallel-trend patterns for their tax-avoidance outcome, not coefficients.

The strongest existing confounder-control competitor is TC-LD from Causal Foundation Models Perform Better without Post-treatment Variables, applied as its authors intend: a pre-filter on the covariate table, leaving the standard accrual residual untouched. On IHDP and ACIC-2016 with injected mediators, that paper reports recovering 85.5% to 92.6% of the oracle headroom and a 100% true-positive rate for injected mediators on IHDP. Those are the bar for what covariate pruning alone achieves when the post-treatment variable is a visible column. We expect it to fall well short on our simulator, because the contaminating variables are inside the outcome's construction.

The oracle is the true $\theta$ from the simulator. For fairness, every pipeline sees the same simulated panels and the same CSMAR sample. Every flexible accrual model gets the same hyperparameter search, cross-validated within the pre-period on control firms only.

Two ablations isolate the key ingredient:

  • Freeze the normal-accrual model to pre-period, never-treated firms but keep post-treatment fundamentals as its inputs.
  • Use $\hat M(0)$ but fit $g$ on the pooled sample.

The proposal counts as beating the baselines if, on the simulator, it cuts absolute bias in $\theta$ by at least half relative to the Feng-style pipeline across the high-capital-expenditure settings. It must also keep 95% intervals near nominal coverage, while doing no worse when the real channel is switched off.

Why not just…?

Why not just match on lagged covariates, as careful applied work already does? Lagging the propensity covariates fixes only the matching stage. The outcome is still a residual computed from post-treatment fundamentals with coefficients estimated partly on treated firms. Our second ablation, pre-period matching with the standard accrual residual, is exactly this alternative. If it matches the full estimator on the simulator, the contribution shrinks to a reporting recommendation, and we will say so.

Why not just use total accruals or operating cash flow as the outcome and skip the discretionary split? This avoids the first-stage problem entirely, and we report it. The cost is that it cannot separate manipulation from the mechanical depreciation and working-capital consequences of compliance spending. The open question in the file is precisely whether firms disguise the squeeze through discretion. The decomposition into a fundamentals-mediated part and a direct part is what this alternative cannot supply.

Why not just apply TC-LD, or a classical discovery algorithm, to all variables including the accrual inputs? Pruning revenue change and fixed assets from the accrual model would remove the very adjustment that defines "normal" accruals. Every capital-expenditure-driven depreciation change would then be counted as discretion, which biases the estimate in the opposite direction. The simulator shows this directly: run TC-LD over all variables and compare its error with the frozen-mapping estimator as the real channel grows.

Why not just use System GMM, as in Nguyen et al. (2024)? System GMM addresses persistence and reverse causality through lagged instruments. It still operates on whatever residual it is given, so it inherits the contaminated outcome, and its instruments face a structural break at the pilot date. We include a GMM specification on both residual definitions so that any difference in the answer is attributable to the outcome construction.

Proposed Experiments and Decisive Checks

On the simulator, we will sweep the strength of treatment-induced capital expenditure from zero to the scale implied by Yi, Zhang and Feng (2024). We will cross that sweep with linear versus nonlinear normal accruals and Gaussian versus heavy-tailed shocks. Metrics are bias, root mean squared error and interval coverage for $\theta$, plus the rate at which each pipeline gets the sign of $\theta$ wrong. Because the empirical literature on accruals is tail-driven, the same metrics will be computed at the 10th and 90th quantiles.

The decisive prediction is quantitative. The gap between the standard pipeline and the oracle should grow with the treatment-induced change in fixed assets. It should vanish when the real channel is switched off. If the gap does not track capital expenditure, the mechanism claimed here is wrong.

On CSMAR, there is no ground truth, so the checks are falsification tests:

  • Event-study pre-trends must be flat for every outcome definition.
  • Placebo pilot years (shifting $t_0$ to 2010–2011) must produce null effects.
  • Low-emission firms inside pilot regions, which face little compliance spending, should show nearly identical estimates under the standard and corrected residuals. High-emission firms should show the largest divergence. That within-region contrast is the real-data analogue of the simulator's capital-expenditure sweep.

The substantive deliverable is the answer to whether, once measured correctly, the carbon pilots raised accrual discretion. We will also report whether any such effect concentrates in private and financially constrained firms, as the tax-avoidance findings would suggest, or whether the apparent effect in the standard pipeline was compliance spending mislabelled as manipulation.