Mechanism analysis is an almost indispensable part of empirical articles in Economic Research. Butits methodology has undergone a fundamental shift in recent years— this skill is written according to the post-shift standard.
Trigger timing
- The main regression results are already robust, but mechanism analysis is missing
- The mechanism only has a verbal explanation of "we think it may be because..."
- The reviewer asks "why would this effect occur?"
- You are planning to (or have already) run the three-step mediation effect method / Sobel test —— please read the next section first
⚠ Paradigm shift: stop using the stepwise mediation effect method
江艇 (2022) "Mediation and moderation effects in empirical research on causal inference," China Industrial Economics, No. 5, pp. 100–120——has become the quasi-standard for mechanism analysis in the four major journals (Economic Research, Management World, China Industrial Economics, China Economic Quarterly). Core conclusion:
- It is recommended to stop using the stepwise mediation effect method (the Baron-Kenny three-step method) and the Sobel test. Under the causal framework, the treatment D can be regarded as exogenous through an identification strategy, butthe mediator variable M is almost never random。
- In the regression of Y on (D, M), the coefficient of M (that is, the "M→Y step") is contaminated by omitted variable bias and reverse causality between M and Y,and has no causal meaning; the "mediation effect" and "proportion of indirect effect" decomposed on this basis are also untrustworthy.
- The three-step method also implicitly assumes "no unobserved post-treatment confounding," which is anextremely strong and untestableassumption.
In one sentence: do not use an endogenous regression to "prove" a mechanism.
The correct approach: D→M + theoretical argument for M→Y
Mechanism testing = two things:
- Empirically, estimate only D→M: treat the mechanism variable M as the dependent variable,and use exactly the same identification strategy as the main regression D→Y(the same DID / IV / RDD, the same fixed effects and clustering). This link is a credibly identified causal relationship — "the treatment indeed changed the mechanism variable."
- Theoretically, argue for M→Y: use economic theory, institutional background, and existing literature to explain "why M affects Y." This linkdoes not run a regression, and relies on argumentation.
Supporting code:resources/code/stata/07_mechanism.do。
Three-paragraph structure for mechanism writing
本文进一步从[机制名]角度提供证据。
第一，理论上，[机制变量 M] 应通过[路径]影响[结果 Y]——[引理论/制度/文献，著者-出版年制，如（江艇，2022）]。
第二，实证上，本文以 [M] 为被解释变量，沿用基准识别策略估计 D→M，结果见表 X：
处理使 [M] [上升/下降] [量级]，与上述理论预期一致。
第三，结合 [理论/文献]，[M] 的变化进而通过[路径]作用于 [Y]，构成本文主效应的机制。
Note that the third step only performslogical closure, and does not report numbers from mediation decomposition such as "indirect effect = 0.0X, proportion Y%."
Three principles of mechanism design
- Theoretical guidance: the mechanism variable must be attachable to some branch of theory, not imagined out of thin air.
- Temporal consistency: the measurement time point of M is after treatment D and before outcome Y; constructing M must not use future information。
- Falsifiability: one can write "if the mechanism holds → D changing M should be observed; if it does not hold → it should not be observed."
Use heterogeneity as circumstantial evidence for the mechanism (a recommended alternative to the "mediation" idea)
More credible mechanistic evidence than mediation decomposition istheory-driven heterogeneity: if the mechanism is M, then in subsamples where the "M channel is more open/more constrained," the main effect should be stronger/weaker.
Example: if the effect comes from "alleviation of financing constraints," then infirms with tighter financing constraintsthe effect should be larger. Verifying this in that subsample is more persuasive than running an endogenous mediation regression of M.
This connects the mechanism (why there is an effect) with heterogeneity (under what conditions it is stronger)—see er-heterogeneity。
Benchmarking against model papers (confirmed to be in Economic Research Journal)
- Zhang Xun, Wan Guanghua, Zhang Jiajia, He Zongyue, "Digital Economy, Inclusive Finance and Inclusive Growth," Economic Research Journal, 2019, Issue 8—a model for mechanism analysis: digital finance -> income, layer-by-layer decomposition of the "entrepreneurship + financial accessibility" channel together with group heterogeneity.
Must-check list
- At least 1 mechanism (ideally 2–3 complementary mechanisms)
- Mechanism variables come from theory / literature, not fabricated
- Mechanism testonly runs D -> M, using the same identification strategy as the main regression
- M -> Y is argued by theory / institutions / literature, whilenotendogenous regression
- The timing logic of mechanism variable construction is correct (no use of future information)
- The mechanism section has at least 1 independent regression table
- (if applicable) uses theory-driven heterogeneity to provide side evidence for the mechanism
- Mechanism results are consistent with the main regression direction (if inconsistent, explain)
Anti-patterns
- Using a sentence like "we believe it may be because..." to replace mechanism analysis
- Running the three-step mediation method / Sobel / Bootstrap indirect effect proportion—has been widely abandoned by the top four journals
- Treating an endogenous regression of Y on M as "evidence that M affects Y"
- Using future information when constructing mechanism variables
- The mechanism result direction is opposite to the main effect but is not explained
- The mechanism variable and the treatment variable are essentially different measures of the same thing (tautology, not counted as a mechanism)
Output format
【机制数】X 个
【机制变量】[M1, M2, ...]
【D→M 识别策略】与主回归一致 / 不一致（需修）
【M→Y 论证方式】理论 / 制度 / 文献（不跑内生回归）
【中介三步法 / Sobel 命中】无 / 有（必须移除）
【时间一致性】通过 / 待修
【异质性侧证】有 / 无
【下一步】er-heterogeneity