Trigger Timing
- The main regression results are already robust, but mechanism analysis is lacking
- Heterogeneity analysis only splits "state-owned / non-state-owned" or "eastern/central/western"—the granularity is not enough
- Reviewers ask "Why does this effect exist?"
Mechanism Analysis (Mechanism)
Three Mainstream Paths
- Mediation effect / moderation effect regression
- The classic Baron & Kenny approach has been weakened → recommend Bootstrap + direct estimation + causal mediation (imai-keele-yamamoto)
- Do not run only the Sobel test
- Replace the dependent variable with the mechanism variable
- Main regression: policy → Y
- Mechanism: policy → M (M is the intermediate mechanism)
- Then discuss the plausibility of M → Y
- Cross-subsample comparison
- Estimate separately in subsamples where "the mechanism holds" and "the mechanism does not hold"
- If the coefficients differ significantly, this constitutes mechanism evidence
Three-part mechanism writing
本文进一步从[机制名]角度提供证据。
首先，[机制变量 M] 与[处理变量 D]的关系：……
其次，在[M 强 / 弱] 子样本中，主效应分别为：……
这一结果支持……机制。
Heterogeneity Analysis (Heterogeneity)
Priority of splitting dimensions
Ordered by the expectations of Management World readers:
- Policy intensity / institutional environment — marketization indices in different regions, regulatory intensity in different industries
- Enterprise ownership / corporate governance — state-owned / private / foreign-funded, board independence
- Enterprise size / industry attributes
- Time window — early vs. late policy period
- Individual characteristics(if the data are at the individual level)
Split at least 3 dimensions, and compare at least 2 subsamples for each dimension.
Principles for dimension selection
- The splitting dimensions musthave theoretical guidance—one cannot "split for the sake of splitting"
- The split must include asignificance test of coefficient differences(Chow test or interaction term)
- After splitting, the samples must not be too small (each subsample N ≥ 100 is the bottom line)
Must-check list
- At least 1 mechanism + at least 3 heterogeneity dimensions
- The mechanism variable comes from theory / existing literature (not fabricated)
- The heterogeneity dimensions have theoretical guidance (not merely descriptive statistical differences)
- Report the significance test of coefficient differences
- Subsample size is sufficient
Anti-patterns
- Attaching an interaction term to a control variable and calling it "heterogeneity"
- Running only the Sobel test for the mediation effect
- Using future information when constructing the mechanism variable and the treatment variable
- Still interpreting a group with N=30 after splitting
This journal's review expectations table for mechanism and heterogeneity
Management World is sponsored by the Development Research Center of the State Council and upholds the empirical tradition of the trinity of "theoretical contribution + Chinese practice + policy implications." Mechanisms and heterogeneity are not decorations, but the key to moving from "the policy is effective" to "why it is effective and for whom it is effective," and they directly bear on the precision of policy implications. The table below anchors this journal's review expectations and common rejection patterns:
| Dimension | This journal's expectations | High-risk signals for rejection |
|---|---|---|
| Mechanism | An intermediate variable M that is theoretically guided and falsifiable | Only running Sobel, with the mechanism imagined out of thin air |
| Heterogeneity | Splits that connect to the institutional environment/property rights/strategic issues | Padding by attaching interaction terms to control variables |
| Rigorous testing | Testing coefficient differences with Chow/interaction terms | Only looking at the size of the two groups' coefficients without conducting a test |
| Policy linkage | Heterogeneity conclusions can derive differentiated policies | Heterogeneity is disconnected from policy implications |
Common rejection patterns (mechanism and heterogeneity dimensions)
- Mediation effects are tested only with Sobel, without Bootstrap or causal mediation, and are judged by reviewers as methodologically outdated
- Heterogeneity is split only into "SOEs/non-SOEs" and "east/central/west," with granularity insufficient to support differentiated policy recommendations
- After splitting, a subsample with N=30 is still being interpreted, and statistical power is insufficient
- The construction of mechanism variables uses future information, causing look-ahead bias
Mini walkthrough example: digital transformation and firms' total factor productivity
Following up on the digital transformation manuscript in mw-identification, walk through the mechanism and heterogeneity (the figures are illustrative):
- Mechanism (replacing the explained variable): Policy → degree of digital transformation M (digitalization word frequency in annual reports), coefficient 0.21 (illustrative); then argue the theoretical chain from M → TFP (resource allocation efficiency, innovation output). In the strong/weak M subsamples, the main effects are 0.048 and 0.012 respectively (illustrative), and the difference is significant (interaction term p<0.05, illustrative), constituting mechanism evidence.
- Heterogeneity dimension 1 (institutional environment): Grouping by high versus low marketization index, the effect is 0.056 in high-marketization regions and 0.020 in low-marketization regions (illustrative), and the Chow test is significant → corresponding to the policy anchor of "business environment reform."
- Heterogeneity dimension 2 (firm ownership): SOEs 0.025, private firms 0.051 (illustrative), suggesting private firms are more sensitive to digitalization subsidies.
- Heterogeneity dimension 3 (factor endowment): Labor-intensive vs. technology-intensive, with a stronger effect in technology-intensive firms (illustrative).
- Judgment: 1 mechanism + 3 theoretically guided heterogeneity dimensions + complete coefficient difference tests → can proceed to mw-tables-figures; the heterogeneity conclusions can already derive differentiated policies (see mw-policy-implication).
Reviewer follow-up patterns + revision approaches in this journal's context
- Follow-up: "Are the mechanism variable and the treatment variable from the same source, creating reverse causality?" → Revision approach: set the mechanism test as the reduced form of "policy→M," and make clear that M is measured later than the treatment and earlier than the outcome, avoiding look-ahead.
- Follow-up: "Is the heterogeneity difference statistical noise or a real structure?" → Revision approach: report the interaction term coefficient and the Chow test, and set a bottom-line check on subsample size (N≥100).
- Follow-up: "Why choose these heterogeneity dimensions?" → In this journal's context, the editorial board requires each dimension to have a theoretical or institutional basis and to be able to address "which type of region/enterprise policy is more effective," rather than exhaustive partitioning.
Calibration anchors
Empirical papers already published in this journal generally include "1 clear mechanism + 3 or more heterogeneity dimensions guided by theory + significance tests for coefficient differences," and the heterogeneity conclusions correspond one-to-one with the stratified policy recommendations at the end of the paper. For mechanism methods, causal mediation and subsample comparisons are the mainstream, and the standalone use of Sobel is rare. Specific expectations are subject to the editorial office's latest manuscript guidelines and the form of papers published in the same period.
Output format
【机制数】X 个
【异质性维度】X 个
【系数差异检验】是 / 否
【机制理论引用】[文献]
【异质性→政策衔接】强 / 中 / 弱
【下一步】mw-tables-figures