Chapter 19: Mechanisms—How and Why

Opening Question

If we know that X causes Y, why isn't that enough? What do we gain by understanding how and why the effect operates?

---

Chapter Overview

The previous chapters have focused on whether X causes Y—identification, estimation, and inference for causal effects. But knowing that a policy "works" is often insufficient. Policymakers want to know why it works, so they can adapt it to new contexts. Scientists want to understand the mechanism, so they can build better theories. And practitioners want to know which components matter, so they can implement efficiently.

This chapter explores causal mechanisms—the pathways through which causes produce effects. Mechanism analysis goes beyond "does it work?" to ask "how does it work?" and "why does it work?" This pursuit is both scientifically valuable and practically important, but it introduces new identification challenges.

What you will learn:

• Why mechanism matters for policy, science, and generalization

• Traditional mediation analysis and its limitations

• Causal mediation analysis and the assumptions it requires

• Structural approaches that model mechanisms explicitly

• How qualitative process tracing complements quantitative mechanism analysis

• When mechanism analysis is feasible and when it's too ambitious

Prerequisites: Chapter 9 (Causal Framework), Chapter 11 (Selection on Observables)

---

19.1 Why Mechanisms Matter

Beyond the Average Treatment Effect

Suppose an RCT shows that a job training program increases employment by 10 percentage points. This is valuable, but incomplete:

• Policy adaptation: Will the program work in a different labor market? That depends on why it works. If it provides skills that employers need, transferability depends on skill demand. If it signals worker quality to employers, transferability depends on whether the signal is recognized.

• Program improvement: Which components drive the effect? Is it technical training, job search assistance, or credential certification? Knowing the mechanism helps allocate resources.

• Scientific understanding: What does this tell us about labor markets? Understanding mechanism connects the finding to theory.

Three Reasons to Study Mechanisms

1. Policy Relevance

Effects may be mediated by modifiable factors. If a health intervention works by improving nutrition, then nutrition programs may achieve similar benefits. If it works by providing social support, a different intervention is needed.

Example: Microfinance impacts on poverty

Early evidence suggested microfinance reduced poverty. But how?

• Business investment? (Suggests targeting entrepreneurs)

• Consumption smoothing? (Suggests broader targeting)

• Women's empowerment? (Suggests gender-focused design)

The optimal policy depends on the mechanism.

2. Scientific Understanding

Mechanism connects empirical findings to theory. An effect without a mechanism is a "black box"—we know the input-output relationship but not what happens inside.

Mechanism also enables prediction: if we understand how $X$ causes $Y$, we can predict effects of interventions that modify the mechanism.

3. External Validity

Will findings generalize to new contexts? This depends on whether the mechanism operates similarly elsewhere.

Example: Minimum wage effects

If minimum wages don't reduce employment because:

• Firms have monopsony power → Effect generalizes to similar labor markets

• Workers were already paid above minimum → Effect doesn't generalize to binding minimums

• Firms substitute technology for labor slowly → Effect generalizes only in short run

Mechanism determines generalization.

---

19.2 The Mediation Framework

Direct and Indirect Effects

Mediation analysis decomposes a total effect into pathways. Let:

• $D$ = treatment

• $M$ = mediator (intermediate variable affected by treatment)

• $Y$ = outcome

The total effect of $D$ on $Y$ may operate:

• Directly: $D \to Y$ (not through $M$)

• Indirectly: $D \to M \to Y$ (through the mediator)

Definition 19.1 (Mediation): Variable $M$ mediates the effect of $D$ on $Y$ if $D$ affects $M$, $M$ affects $Y$, and part of $D$'s effect on $Y$ operates through the $D \to M \to Y$ pathway.

The DAG Perspective

Figure 19.1: Basic Mediation Structure

Figure 19.1: Basic mediation structure. The total effect of treatment D on outcome Y can be decomposed into a direct effect (D → Y) and an indirect effect through mediator M (D → M → Y). For example, job training (D) might increase employment (Y) directly or indirectly through improved skills (M).

The total effect of $D$ on $Y$ combines:

• Path $D \to Y$ (direct effect)

• Path $D \to M \to Y$ (indirect effect through $M$)

Traditional (Baron-Kenny) Approach

Baron and Kenny (1986) proposed:

1. Show $D$ affects $Y$ (total effect)

2. Show $D$ affects $M$ (first stage)

3. Show $M$ affects $Y$ controlling for $D$ (mediator effect)

4. Indirect effect = (effect of $D$ on $M$) × (effect of $M$ on $Y$|$D$)

Regression implementation: $M_i = \alpha_1 + a D_i + \varepsilon_{1i}$ $Y_i = \alpha_2 + c' D_i + b M_i + \varepsilon_{2i}$

• $a$ = effect of $D$ on $M$

• $b$ = effect of $M$ on $Y$ (controlling for $D$)

• $c'$ = direct effect of $D$ on $Y$ (controlling for $M$)

• Indirect effect = $a \times b$

• Total effect = $c' + ab$

The SEM Tradition and Mediation

Baron-Kenny mediation emerged from the structural equation modeling (SEM) tradition in psychology—path analysis, latent variables, and simultaneous estimation of complex causal structures. For a fuller discussion of SEM's history and its relationship to modern causal inference, see Section 9.7.

For mediation analysis specifically, the key tension is this: SEM software estimates path coefficients, but causal interpretation of "indirect effects" requires assumptions that SEM cannot verify:

• Interventionist interpretation: The "indirect effect" through M is only meaningful if intervening on M is well-defined. For psychological constructs like "motivation" or "self-efficacy," what would it mean to experimentally manipulate them?

• Post-treatment confounding: SEM estimates associations conditional on model structure, but it doesn't flag when conditioning on a mediator opens backdoor paths through unobserved confounders.

The path coefficients are causal only if the model structure is causally correct—which requires the same identification arguments as any other method.

Problems with the Traditional Approach

1. Confounding of $M \to Y$

Even if $D$ is randomized, $M$ is not. The relationship between $M$ and $Y$ (controlling for $D$) is observational and subject to confounding.

Example: Training program ($D$) affects motivation ($M$) and employment ($Y$). But motivation is also affected by unobserved factors (personality, prior experience) that independently affect employment. Controlling for $D$ doesn't remove $M$-$Y$ confounding.

2. Post-treatment confounding

If $D$ affects a variable $L$ that confounds $M \to Y$, controlling for $L$ blocks the causal pathway while not controlling creates confounding. There's no regression specification that solves this.

3. Interaction effects

If $D$ and $M$ interact in affecting $Y$, the decomposition into "direct" and "indirect" is not unique.

---

19.3 Causal Mediation Analysis

Potential Outcomes Framework

Modern causal mediation analysis (Pearl 2001; Robins and Greenland 1992; Imai et al. 2010) uses potential outcomes to define causal mediation effects.

Nested potential outcomes:

• $M_i(d)$ = mediator value if unit $i$ receives treatment $d$

• $Y_i(d, m)$ = outcome if unit $i$ receives treatment $d$ and mediator $m$

• $Y_i(d) = Y_i(d, M_i(d))$ = outcome under natural mediator value

Definition 19.2 (Natural Direct Effect, NDE): The effect of treatment on outcome when the mediator is held at its untreated value: $NDE = E[Y_i(1, M_i(0)) - Y_i(0, M_i(0))]$

Definition 19.3 (Natural Indirect Effect, NIE): The effect of changing the mediator from its untreated to its treated value, holding treatment at the treated level: $NIE = E[Y_i(1, M_i(1)) - Y_i(1, M_i(0))]$

Decomposition: $ATE = NDE + NIE$

Intuition: NDE asks "what if we gave treatment but held the mediator fixed?" NIE asks "what if we changed the mediator without changing treatment directly?" Together they decompose the total effect.

Identification Assumptions

Identifying NDE and NIE requires strong assumptions beyond randomized treatment:

Assumption 19.1 (Sequential Ignorability):

1. Treatment is independent of potential outcomes and potential mediators: $(Y(d,m), M(d)) \perp D | X$

2. Mediator is independent of potential outcomes given treatment and covariates: $Y(d,m) \perp M | D, X$

Figure 19.2: Sequential Ignorability Assumptions

Figure 19.2: The two parts of sequential ignorability. Condition 1 (left) requires that treatment is independent of potential outcomes given covariates X—satisfied by randomization. Condition 2 (right) requires no unobserved M-Y confounders—the hardest condition, because treatment may affect variables that confound the M-Y relationship.

Translation:

1. Treatment assignment is as-if random (often satisfied by design)

2. Given treatment, the mediator is as-if random conditional on covariates

The second condition is demanding: it requires that all confounders of $M \to Y$ are measured, even though $M$ is post-treatment.

Box: Why Sequential Ignorability Almost Never Holds

Sequential ignorability is routinely invoked but rarely credible. The fundamental problem: the mediator is post-treatment, so anything affected by treatment that also affects outcomes will confound the $M \to Y$ relationship.

The logical trap:

• Randomizing treatment gives you Part 1 of sequential ignorability

• But Part 2 requires the mediator to be "as good as randomized" conditional on treatment

• If treatment affects any variable that also affects outcomes, and that variable correlates with the mediator, Part 2 fails

Example: Job training → Skills → Earnings A job training RCT randomizes training. The mediator is "skills" (test scores after training). But training also affects:

• Motivation (unobserved)

• Self-confidence (partially observed)

• Job search networks (unobserved)

• Time allocation (partially observed)

All of these affect earnings. All are correlated with measured skills. Sequential ignorability fails even though treatment was randomized.

The uncomfortable truth: In most substantive applications, we can list post-treatment confounders of $M \to Y$ that we cannot fully measure. This isn't a minor problem requiring sensitivity analysis—it's a fundamental threat to the entire enterprise.

Practical implications:

• Be deeply skeptical of point estimates from causal mediation analysis

• Sensitivity analysis should be the centerpiece, not an afterthought

• Consider whether your research question can be reformulated to avoid mediation (e.g., "Does treatment work through M?" might be better answered by studying treatment-by-moderator interactions)

• The experimental manipulation of mediators (Chapter 19.X) may be the only credible approach in many settings

Bottom line: Sequential ignorability is not an assumption to "assume and proceed." It is a strong claim that requires either explicit defense (rare) or frank acknowledgment that results are bounds-like ("if sequential ignorability held, which it probably doesn't...").

When Is Sequential Ignorability Credible?

More plausible:

• Mediator is measured accurately and soon after treatment

• Strong theory identifies the relevant confounders

• Treatment doesn't affect confounders of $M \to Y$

Less plausible:

• Long time between treatment and mediator measurement

• Mediator is a complex construct (e.g., "motivation")

• Treatment affects variables that confound $M \to Y$

Sensitivity Analysis

Given the strength of sequential ignorability, sensitivity analysis is essential. Imai et al. (2010) develop methods to bound mediation effects under violations.

Key parameter: $\rho$, the correlation between errors in the mediator and outcome equations (unobserved confounding).

• $\rho = 0$: sequential ignorability holds

• $\rho \neq 0$: there is unobserved confounding

Report mediation effects for a range of $\rho$ values; assess at what $\rho$ conclusions change.

Instrumental Variables for Mediation

When sequential ignorability fails because the mediator is endogenous, instrumental variables can sometimes identify mediation effects.

Box: IV Approaches to Mediation Analysis

The setup: Treatment $D$ affects mediator $M$ and outcome $Y$. We want to decompose the total effect into:

• Direct effect: $D \to Y$ (not through $M$)

• Indirect effect: $D \to M \to Y$

The problem: Unobserved $U$ confounds the $M \to Y$ relationship. Simple regression of $Y$ on $M$ is biased.

IV solution: Find an instrument $Z_M$ that:

1. Affects the mediator: $Z_M \to M$ (relevance)

2. Only affects $Y$ through $M$: $Z_M \not\to Y$ except via $M$ (exclusion)

3. Is unrelated to confounders: $Z_M \perp U$ (independence)

Two approaches:

Approach 1: Randomized encouragement for mediator

If you can randomize an encouragement that shifts the mediator without affecting outcomes directly, this provides an instrument.

Example: Studying whether education ($M$) mediates the effect of family income ($D$) on child outcomes ($Y$). A college scholarship lottery ($Z_M$) affects education but plausibly only affects outcomes through education.

Approach 2: Principal stratification (when treatment is randomized)

When $D$ is randomized, define strata by how units' mediators respond to treatment:

• Always-takers: $M(0) = M(1) = 1$

• Never-takers: $M(0) = M(1) = 0$

• Compliers: $M(0) = 0, M(1) = 1$

Within strata where $M$ doesn't change, any effect of $D$ on $Y$ must be direct.

Challenges:

• Finding valid instruments for mediators is difficult

• Exclusion restrictions are often implausible (anything affecting $M$ may affect $Y$ through other paths)

• Interpretation requires assuming effect homogeneity or settles for LATE-style local effects

Key references: Imai et al. (2013) for principal stratification; Dippel et al. (2020) for IV-based mediation.

Box: The Front-Door Criterion—Using Mediators to Identify Total Effects

Pearl's (2009) front-door criterion offers a striking result: sometimes a mediator can identify the total causal effect even when you can't control for confounders—a different use of mediation than decomposing direct and indirect effects.

Setup: Suppose you want the effect of $D$ on $Y$, but there's an unmeasured confounder $U$:

       U
      / \
     /   \
    D → M → Y

You can't adjust for $U$, so the backdoor path $D \leftarrow U \to Y$ is open. Normally, identification fails.

The front-door solution: If a mediator $M$ satisfies three conditions:

1. $M$ fully mediates $D \to Y$ (no direct path from $D$ to $Y$ except through $M$)

2. $D \to M$ is unconfounded (no backdoor paths to $M$)

3. Conditioning on $D$ blocks all backdoor paths from $M$ to $Y$

Then the total effect of $D$ on $Y$ is identified as: $P(Y | do(D=d)) = \sum_m P(M=m | D=d) \sum_{d'} P(Y | M=m, D=d') P(D=d')$

Intuition: We can identify $D \to M$ (unconfounded by assumption). We can identify $M \to Y$ by averaging over $D$ (this blocks the backdoor through $U$). Chaining these together gives $D \to Y$.

Why this works: The confounder $U$ affects both $D$ and $Y$ directly, but it doesn't affect $M$ except through $D$. So we can use $M$ as a "stepping stone" to trace the causal pathway.

Classic example: Does smoking ($D$) cause cancer ($Y$)? The confounder might be genetic predisposition ($U$). But tar deposits in lungs ($M$) satisfy the front-door conditions: smoking causes tar, tar causes cancer, and conditioning on smoking blocks any genetic path to tar.

In practice: The front-door criterion is rarely applicable because its conditions are demanding—especially the requirement that $M$ fully mediates the effect. But it illustrates a deep point: mediating structure can enable identification that no amount of covariate adjustment could achieve. This is the logic that underlies more flexible graphical identification methods.

Figure 19.3: The Front-Door Criterion

Figure 19.3: The front-door criterion. Even with unobserved confounding (U) between D and Y, the total causal effect can be identified if mediator M intercepts all causal paths from D to Y, D→M is unconfounded, and conditioning on D blocks backdoor paths from M to Y.

---

19.4 Estimation Methods

Regression-Based Estimation

Under sequential ignorability, mediation effects can be estimated by regression:

$M_i = \alpha_1 + a D_i + \gamma_1' X_i + \varepsilon_{1i}$ $Y_i = \alpha_2 + c' D_i + b M_i + \gamma_2' X_i + \varepsilon_{2i}$

• NIE = $\hat{a} \times \hat{b}$ (product of coefficients)

• NDE = $\hat{c}'$

Standard errors via bootstrap or delta method.

Weighting Methods

Analogous to propensity score methods for selection on observables:

1. Estimate propensity for mediator: $P(M|D,X)$

2. Reweight observations to balance $M$ between treatment groups

3. Compute weighted outcome means

This is more flexible than regression when functional forms are uncertain.

Randomization-Based Approaches

When possible, randomize both treatment and mediator:

• Crossover designs: Randomize treatment; within treated, randomize mediator levels

• Mechanism experiments: Directly manipulate hypothesized mediators

These provide strong identification but are often infeasible (many mediators can't be randomized).

---

19.5 Structural Approaches to Mechanism

Beyond Reduced-Form Mediation

Structural models specify the full mechanism explicitly:

1. How treatment affects mediators

2. How mediators affect outcomes

3. How other factors interact with the mechanism

This goes beyond decomposing effects into shares—it models the underlying economic or behavioral process.

Example: Returns to Education

Why does education increase earnings?

Human capital mechanism: Education builds skills → Skills increase productivity → Productivity increases wages

Signaling mechanism: Education signals pre-existing ability → Employers pay for the signal → Wages reflect signaling equilibrium

These mechanisms have different policy implications:

• Human capital → Investment in education quality matters

• Signaling → Only credentials matter; education is socially wasteful

A structural model can distinguish them by specifying the production function for skills and the employer learning process.

Equilibrium Effects

In markets, individual-level mechanisms may differ from aggregate effects due to equilibrium adjustments.

Example: Job training increases individual employment. But if all workers receive training:

• Human capital channel: All workers more productive → aggregate employment may rise

• Signaling channel: Signal becomes uninformative → no aggregate effect

Structural models can capture these equilibrium effects; reduced-form mediation cannot.

When Structural Models Help

• Complex mechanisms with multiple pathways

• Equilibrium effects matter

• Policy requires understanding counterfactuals not observed in data

• Strong theory guides model specification

When Structural Models Are Problematic

• Mechanisms are poorly understood

• Model misspecification leads to biased conclusions

• Computational demands are high

• Results depend heavily on functional form assumptions

---

19.6 Qualitative Bridge: Process Tracing

What Is Process Tracing?

Process tracing is a qualitative method for inferring causal mechanisms within cases. Rather than estimating mediation parameters across units, it examines the sequence of events and decisions within a single case.

Definition 19.4 (Process Tracing): A within-case method that examines the causal chain connecting a cause to an outcome by identifying observable manifestations of the hypothesized mechanism.

The Logic

If mechanism $M$ links $D$ to $Y$, then observable implications should be present:

• Temporal sequence: $D$ precedes $M$ precedes $Y$

• Actors' reasoning: Decision-makers should mention $M$ in explaining their choices

• Documentary evidence: Records should show $M$ operating

Example: Did the Marshall Plan aid European recovery through capital investment?

• Were funds actually invested in capital?

• Did investment rates rise?

• Did productivity increase following investment?

• Did contemporary observers attribute growth to investment?

Answering "yes" to these questions provides evidence for the investment mechanism.

Types of Evidence

Beach and Pedersen (2013) distinguish:

Straw-in-the-wind: Consistent with mechanism but not definitive

• Example: Politicians mention investment in speeches

Hoop test: Must be present if mechanism is true; absence eliminates mechanism

• Example: Investment data must show increase

Smoking gun: Confirms mechanism if present

• Example: Internal documents showing investment decisions driven by Marshall Plan funds

Doubly decisive: Both confirms the hypothesized mechanism and disconfirms alternatives

• Example: Rare in practice

Complementarity with Quantitative Methods

Quantitative strengths: Estimates average effects, assesses effect sizes, generalizes across cases

Qualitative strengths: Identifies mechanisms, explains how effects operate, handles complex causal chains

Combination strategies:

1. Use quantitative analysis to establish that $D \to Y$

2. Use process tracing to investigate how in specific cases

3. Use qualitative findings to refine quantitative models

Example: Microfinance Mechanisms

Quantitative RCTs establish that microfinance has modest average effects on poverty. But why?

Process tracing in specific communities can examine:

• How do borrowers actually use loans? (Business investment vs. consumption)

• What constrains business expansion? (Capital vs. skills vs. market access)

• How do repayment pressures affect household decisions?

• What social dynamics accompany microcredit?

Findings from qualitative mechanism research:

• Many loans finance consumption, not business

• Successful businesses often constrained by demand, not capital

• Group lending creates social pressure that may harm some borrowers

This explains the modest average effects and suggests targeting improvements.

---

19.7 Practical Considerations

When Is Mechanism Analysis Feasible?

More feasible when:

• Mediator is well-defined and measurable

• Treatment assignment is clean (randomized or strongly designed)

• Theory identifies specific mechanisms to test

• Data include mediator measurements

• Confounders of $M \to Y$ relationship are measured

Less feasible when:

• Multiple potential mediators that are hard to separate

• Mediator is vague or multi-dimensional ("attitudes," "culture")

• Treatment affects confounders of $M \to Y$

• Mechanisms operate at different levels (individual vs. market)

What Can Go Wrong

1. Post-treatment bias: Controlling for the mediator introduces bias if treatment affects confounders

2. Measurement error in mediator: Attenuates mediation estimates

3. Multiple mechanisms: Hard to separate when mechanisms are correlated

4. Equilibrium effects: Individual-level mechanisms may not aggregate

Recommendations

Be modest: Mechanism analysis is harder than effect estimation. Acknowledge uncertainty.

Use multiple approaches: Combine quantitative mediation with qualitative process tracing.

Focus on salient mechanisms: Don't try to decompose into every possible pathway.

Sensitivity analysis: Report how conclusions change under violations of sequential ignorability.

Consider experiments: When feasible, directly manipulate hypothesized mechanisms.

---

Practical Guidance

When to Pursue Mechanism Analysis

Situation	Recommendation
Policy requires understanding "why"	Pursue mechanism analysis
Clear, measurable mediator	Quantitative mediation feasible
Complex mechanism with multiple paths	Structural modeling or qualitative
Mediator is vague or post-treatment confounded	Proceed with extreme caution
Stakes are high, assumptions questionable	Report bounds under alternative assumptions

Common Pitfalls

Pitfall 1: Treating Baron-Kenny as causal Traditional mediation analysis estimates associations, not causal mediation effects. The "indirect effect" may reflect confounding, not mechanism.

How to avoid: Use causal mediation framework; acknowledge identification assumptions; conduct sensitivity analysis.

Pitfall 2: Ignoring post-treatment confounding If treatment affects variables that confound $M \to Y$, no regression specification identifies the mediation effect.

How to avoid: Draw the DAG; identify potential post-treatment confounders; acknowledge if the problem is present.

Pitfall 3: Over-interpreting small mediation shares "30% of the effect operates through M" sounds precise but may be driven by measurement error or specification choices.

How to avoid: Report sensitivity analysis; bound the mediation share; acknowledge uncertainty.

Pitfall 4: Conflating mechanisms and effect heterogeneity "The effect is larger for women" describes heterogeneity, not mechanism. The mechanism question is why it's larger for women.

How to avoid: Distinguish "for whom" (heterogeneity, Chapter 20) from "through what" (mechanism).

Implementation Checklist

• Clearly state the hypothesized mechanism

• Draw the DAG including potential confounders

• Assess whether sequential ignorability is plausible

• Estimate mediation effects using appropriate methods

• Report sensitivity analysis for unobserved confounding

• Consider qualitative process tracing for specific cases

• Discuss what would be needed to strengthen mechanism evidence

---

Integration Note

Connections to Other Methods

Method	Relationship	See Chapter
Selection on Observables	Mediation has similar confounding concerns	Ch. 11
Instrumental Variables	Can sometimes identify specific mechanisms	Ch. 12
Heterogeneity	Mechanisms may explain effect variation	Ch. 20
Triangulation	Combining methods strengthens mechanism evidence	Ch. 23

Triangulation for Mechanism

Mechanism claims are stronger when supported by:

1. Quantitative mediation analysis: Estimated indirect effects

2. Process tracing: Qualitative evidence from specific cases

3. Theory: Mechanism is consistent with established models

4. Experimental manipulation: Direct intervention on hypothesized mediator

5. Cross-context variation: Mechanism explains when effects are larger/smaller

---

Running Example: Microfinance Mechanisms

The Question

Microfinance RCTs show modest average effects on poverty indicators. But why these effects—or lack thereof? Understanding mechanisms is essential for program design.

Hypothesized Mechanisms

1. Business investment

• Loans fund business creation/expansion

• Business generates income

• Income reduces poverty

2. Consumption smoothing

• Credit allows smoothing consumption over income shocks

• Reduced volatility improves welfare

• Not through income growth, but through stability

3. Women's empowerment

• Loans to women increase their economic role

• Increased bargaining power changes household decisions

• May affect education, health investments

4. Social capital

• Group lending builds social networks

• Networks provide insurance, information, opportunities

• May have spillover effects

Evidence

Quantitative mediation:

• Banerjee et al. (2015): Most effects operate through business investment, but business effects are modest

• Crépon et al. (2015): Some evidence for consumption smoothing in Morocco

Process tracing:

• Ethnographic work shows many loans used for consumption, ceremonies, repaying other debts

• Successful borrowers often had businesses before; microcredit provides working capital, not startup funds

• Group meeting attendance generates information sharing but also social pressure

Structural analysis:

• Returns to capital are high for some businesses but heterogeneous

• Many potential entrepreneurs face demand constraints, not capital constraints

• Interest rates may be too high relative to business returns for poorest borrowers

Implications

Mechanism analysis suggests:

• Target lending to existing microenterprises with growth potential

• Combine credit with business training and market access

• Consider alternative products for consumption smoothing (savings, insurance)

• Recognize limits of microcredit for the very poor

This goes beyond "microfinance works" to "microfinance works for some people, through business channels, when constraints are capital rather than demand."

---

Summary

Key takeaways:

1. Mechanism analysis asks how and why effects operate, going beyond the average treatment effect.

2. Traditional mediation (Baron-Kenny) decomposes effects into direct and indirect, but is subject to confounding and doesn't identify causal mediation.

3. Causal mediation analysis defines natural direct and indirect effects using potential outcomes, but requires sequential ignorability—a strong assumption.

4. Structural approaches model mechanisms explicitly, enabling analysis of equilibrium effects and policy counterfactuals, at the cost of model dependence.

5. Process tracing complements quantitative mediation by examining mechanisms within specific cases, providing evidence for causal chains.

6. Mechanism analysis is hard: Post-treatment confounding, multiple mechanisms, and measurement challenges make causal mechanism claims more uncertain than treatment effect claims.

Returning to the opening question: Knowing that X causes Y is valuable but incomplete. Understanding mechanism enables policy adaptation, program improvement, scientific understanding, and assessment of external validity. But mechanism analysis faces additional identification challenges beyond effect estimation. Combining quantitative mediation with qualitative process tracing, while acknowledging uncertainty, provides the strongest basis for mechanism claims.

---

Further Reading

Essential

• Imai, Keele, and Tingley (2010), "A General Approach to Causal Mediation Analysis" - Modern causal mediation framework

• VanderWeele (2015), Explanation in Causal Inference - Comprehensive treatment

For Deeper Understanding

• Pearl (2001), "Direct and Indirect Effects" - Potential outcomes framework for mediation

• Robins and Greenland (1992), "Identifiability and Exchangeability for Direct and Indirect Effects" - Foundational paper

• Imai et al. (2011), "Unpacking the Black Box of Causality" - Sequential ignorability and sensitivity

Advanced/Specialized

• Acharya, Blackwell, and Sen (2016), "Explaining Causal Findings Without Bias" - Intermediate confounding

• Huber (2014), "Identifying Causal Mechanisms (Primarily) Based on Inverse Probability Weighting" - Weighting approaches

• Beach and Pedersen (2013), Process-Tracing Methods - Qualitative mechanism analysis

Applications

• Banerjee et al. (2015), "The Miracle of Microfinance?" - Mechanism analysis in development

• Ludwig et al. (2011), "Mechanism Experiments and Policy Evaluations" - Experimental approaches

• Deaton and Cartwright (2018), "Understanding and Misunderstanding Randomized Controlled Trials" - Mechanism and external validity

---

Exercises

Conceptual

1. Explain the difference between the natural direct effect (NDE) and the controlled direct effect (CDE). When do they differ? Which is more relevant for policy?

2. Why is sequential ignorability a strong assumption? Construct an example where treatment is randomized but sequential ignorability fails due to post-treatment confounding.

3. How does process tracing complement quantitative mediation analysis? What can each approach do that the other cannot?

Applied

1. Design a mechanism analysis for the following setting: An RCT shows that a financial literacy program increases savings. What mechanisms might explain this? How would you investigate them quantitatively and qualitatively?

2. Using data from a job training evaluation that includes mediator measurements (e.g., skill assessments, job search effort):

   • Estimate total, direct, and indirect effects using regression

   • Conduct sensitivity analysis for violations of sequential ignorability

   • Discuss the credibility of your mediation estimates

Discussion

1. Some researchers argue that mechanism analysis is too hard—identification assumptions are too strong, and we should focus on well-identified total effects. Others argue that effects without mechanisms are policy-irrelevant "black boxes." Where do you stand? Under what conditions is mechanism analysis worth pursuing?