Difference-in-Differences: Natural Experiments & Causal Estimation

Published: March 19, 2026

Article by Ryan O'Connell, CFA, FRM

When a major regulation like the Sarbanes-Oxley Act takes effect, audit costs rise across the board — but how much of that increase is caused by the regulation itself, and how much reflects broader trends affecting all public companies? Simply comparing costs before and after the law confuses the policy’s impact with every other change that occurred simultaneously. Difference-in-differences (DiD) solves this problem by comparing two groups over two time periods, subtracting out common trends to isolate the causal effect of a policy change. DiD has become the standard method for evaluating regulatory impacts in finance — from SOX compliance costs to banking deregulation and IFRS adoption. This article covers the DiD estimator, the parallel trends assumption, the regression framework, and a worked Sarbanes-Oxley example, drawing on the natural experiment framework from Wooldridge’s Introductory Econometrics (Chapter 13). For background on why controlling for confounders matters, see our guide on omitted variable bias.

What Is Difference-in-Differences?

A natural experiment occurs when an exogenous event — typically a change in government policy or regulation — alters the environment for some economic units but not others. Unlike a controlled laboratory experiment, the researcher does not assign treatment; instead, the policy itself creates a treatment group (units affected by the change) and a control group (units not affected). The difference-in-differences estimator exploits this structure to measure the causal effect of the treatment.

The logic involves two differences. The first difference — comparing the treatment group before and after the policy — captures the treatment effect plus any common time trends. The second difference — subtracting the control group’s before-after change — removes those common time trends. What remains is the estimated treatment effect: the change attributable to the policy itself, net of changes that would have occurred anyway.

DiD is one of the most widely used tools in the causal inference toolkit. In finance, it has been applied to study the effects of securities regulation, tax policy changes, exchange listing requirements, and central bank interventions — any setting where a clearly defined policy shock affects an identifiable subset of firms, investors, or markets.

Consider a concrete example: when the SEC adopted new disclosure requirements for executive compensation in 2006, one might compare the change in CEO pay transparency for firms subject to the rule (treatment) versus firms below the reporting threshold (control). If both groups were trending similarly before the rule, the differential change in transparency after the rule can be attributed to the regulation rather than to broader corporate governance trends.

The Parallel Trends Assumption in Difference-in-Differences

The credibility of every DiD estimate rests on a single critical assumption: parallel trends. This assumption states that, in the absence of treatment, the treatment and control groups would have experienced the same change in the outcome variable over time.

Why does this matter? If the treatment group was already trending differently from the control group before the policy, DiD will misattribute the pre-existing divergence to the treatment. For example, if larger firms were already experiencing faster audit cost growth before SOX, comparing large firms (treatment) to small firms (control) would overstate the regulation’s effect.

Researchers assess parallel trends using pre-treatment data. With multiple pre-treatment periods, you can plot the outcome variable for both groups over time and visually inspect whether the trends track each other before the policy takes effect. Formal pre-trend tests regress the outcome on leads of the treatment indicator — if leads are jointly significant, the parallel trends assumption is suspect. Event-study plots (also called leads-and-lags plots) display the estimated treatment effect at each time period relative to the policy date, providing a visual diagnostic for both pre-trends and dynamic post-treatment effects.

The 2×2 DiD Estimator

The simplest DiD design uses two groups and two time periods. The estimator can be computed directly from group means arranged in a 2×2 table:

Where:

• δ̂_DiD — the estimated DiD treatment effect
• Ȳ_T,after — sample mean of the outcome for the treatment group in the post-treatment period
• Ȳ_T,before — sample mean of the outcome for the treatment group in the pre-treatment period
• Ȳ_C,after — sample mean of the outcome for the control group in the post-treatment period
• Ȳ_C,before — sample mean of the outcome for the control group in the pre-treatment period

To make this concrete, consider an illustrative example with hypothetical compliance cost data (in arbitrary units):

Both groups saw increases, but the treatment group increased by 30 while the control group increased by only 15. Under parallel trends, the control group’s change of 15 represents the common time trend — the increase that both groups would have experienced absent the policy. The treatment group’s additional increase of 15 is the estimated causal effect of the treatment.

Graphically, imagine plotting the outcome over time for both groups. Before the policy, both groups trend upward at the same rate. After the policy, the treatment group jumps above its projected path while the control group continues along its original trajectory. The counterfactual for the treatment group is constructed by taking the treatment group’s pre-period level and adding the control group’s observed change — this is where the treatment group would have ended up under parallel trends if the policy had not occurred. A dashed line from the treatment group’s pre-period point to this counterfactual level shows the projected path. The vertical gap between the treatment group’s actual post-treatment outcome and this counterfactual level is the DiD estimate.

The estimator can equivalently be computed by first taking the cross-group difference in each period and then differencing across periods: (80 − 60) − (50 − 45) = 20 − 5 = 15. Both orderings produce the same result — hence the name “difference-in-differences.”

This equivalence is reflected in the algebraic structure of the 2×2 table. Following Wooldridge (Chapter 13, Table 13.3), the expected values are:

The bottom-right cell shows that β₃ — the DiD estimate — is obtained regardless of whether you difference first across time and then across groups, or first across groups and then across time.

Difference-in-Differences in Regression Form

While the 2×2 table is intuitive, researchers almost always implement DiD as a regression. The standard specification is:

Where:

• Y_it — outcome variable for unit i in period t (e.g., audit fees for firm i in year t)
• Treat_i — dummy variable equal to 1 if unit i is in the treatment group, 0 if in the control group
• After_t — dummy variable equal to 1 for the post-treatment period, 0 for the pre-treatment period
• Treat_i × After_t — interaction term that equals 1 only for treated units in the post-treatment period
• u_it — error term capturing all other factors affecting the outcome

Each coefficient has a clear interpretation:

• β₀ — mean outcome for the control group in the pre-treatment period
• β₁ — average difference between treatment and control groups in the pre-treatment period (baseline group difference)
• β₂ — change in the outcome from before to after for the control group (common time effect)
• β₃ — the DiD estimate: the additional change experienced by the treatment group beyond the common time effect

An important practical note: DiD can be estimated using either panel data (the same units observed before and after) or repeated cross-sections (different samples drawn from the same populations in each period). When using panel data with many time periods, researchers often add entity and time fixed effects — see Panel Data Analysis for the full fixed effects framework.

Standard errors in DiD require careful attention. When treatment is assigned at the group level (e.g., all firms above a regulatory threshold), standard errors should be clustered at the level at which treatment varies. With a limited number of clusters, conventional cluster-robust standard errors can be unreliable; in such cases, the wild cluster bootstrap provides more accurate inference. Serial correlation within units across time is the primary reason that unclustered standard errors tend to be too small in DiD settings with many time periods.

DiD Example: Sarbanes-Oxley and Audit Costs

Using illustrative data that reflects the magnitude of effects documented in the empirical literature:

Under the parallel trends assumption, the control group’s increase of $270,000 reflects common factors affecting all public companies during this period — market-wide increases in demand for auditing services, general improvements in corporate governance post-Enron, and inflation in professional fees. The DiD estimate of $860,000 isolates the incremental cost specifically attributable to SOX Section 404(b) compliance.

In regression form, the specification is:

Adding control variables does not change the identification strategy — DiD still relies on parallel trends — but controls reduce residual variance, producing smaller standard errors and more precise estimates. The numbers above are illustrative; empirical studies in this literature typically model log(audit fees) to account for the skewed distribution of fee data and to express the treatment effect as a percentage change rather than a dollar amount. For more on how regulatory frameworks shape financial institutions, see our guide on financial regulation and deposit insurance.

Difference-in-Differences vs. Other Causal Methods

DiD is one of several quasi-experimental methods available for causal inference. The right choice depends on the research setting — specifically, whether you have a control group, an instrument, a sharp cutoff, or panel data with sufficient pre-treatment periods.

In many empirical finance applications, these methods complement each other. A researcher studying SOX might use DiD to estimate the overall treatment effect, panel fixed effects to control for firm-level confounders, and instrumental variables to address remaining endogeneity concerns. The choice of method should be driven by the available data and the plausibility of each method’s identifying assumptions. For a broader overview of causal inference methods including regression discontinuity and propensity score matching, see Causal Inference in Econometrics.

Extensions: Staggered Treatment and Triple Differences

Staggered Treatment Timing

Many policy changes do not take effect at a single point in time for all treated units. When different groups receive treatment at different dates — for example, states adopting banking deregulation in different years — the design is called staggered DiD. Researchers have traditionally estimated staggered designs using two-way fixed effects (TWFE) regressions with unit and time fixed effects.

However, recent research has revealed a serious problem: when treatment effects vary across cohorts or over time, the standard TWFE estimator can produce biased estimates because it implicitly uses already-treated units as controls and assigns negative weights to some group-time treatment effects. Three prominent solutions have been developed:

• Callaway and Sant’Anna (2021) — estimates group-time average treatment effects and aggregates them with proper weighting
• Sun and Abraham (2021) — uses an interaction-weighted estimator that avoids contamination from heterogeneous effects
• de Chaisemartin and d’Haultfoeuille (2020) — provides a robust estimator that correctly handles the negative weighting problem

Event-study plots (leads-and-lags specifications) are the standard diagnostic tool in staggered DiD settings. These plots display the estimated treatment effect at each time period relative to the treatment date, allowing researchers to visually assess both pre-trends (the lead coefficients should be near zero and statistically insignificant) and the dynamic path of treatment effects after the policy takes effect.

Triple Differences (DDD)

When the parallel trends assumption is suspect — perhaps the treatment and control groups had different pre-trends for reasons unrelated to the policy — adding a third differencing dimension can strengthen the design. Triple differences (DDD) adds data from an additional comparison that helps absorb one source of differential trends.

For example, to study SOX’s effect on audit fees, a DDD design might compare the change in audit fees for accelerated vs. non-accelerated filers, in the U.S. vs. a country that did not adopt SOX-like requirements, before vs. after the regulation. The third difference nets out any differential trend between large and small firms that is common across countries and unrelated to SOX.

DDD is not a free lunch: it absorbs one additional source of bias but replaces the original parallel trends assumption with a new parallel-trends-style assumption across the third dimension. The identifying condition is still that remaining differential trends are zero — only now the condition is imposed on a higher-order difference.

Common Mistakes

1. Assuming parallel trends without testing. The most common error in applied DiD is proceeding without examining whether pre-treatment trends are parallel. Always plot outcome trends for both groups before the treatment date and conduct formal pre-trend tests using event-study specifications with lead indicators.

2. Ignoring statistically significant pre-trend evidence. When pre-treatment lead coefficients are statistically significant or show a clear pattern, the parallel trends assumption is violated and the standard DiD estimate is unreliable. Researchers sometimes dismiss pre-trend evidence by arguing it is “close enough to zero” — but a visible pre-trend calls the entire identification strategy into question.

3. Using too few clusters without adjusting inference. When treatment varies at the group level (e.g., state-level regulation) and there are few clusters, conventional cluster-robust standard errors become unreliable and typically reject the null hypothesis too often. The wild cluster bootstrap or similar small-sample corrections should be used when the number of clusters is limited.

4. Applying DiD when treatment timing is fuzzy. If the treatment does not switch cleanly from “off” to “on” at a known date — for example, a regulation that is announced, debated, partially implemented, and then fully enforced over multiple years — the simple before/after distinction breaks down. Anticipatory behavior by firms that adjust before the official implementation date can attenuate or reverse the estimated treatment effect.

5. Confusing DiD with a simple before-after comparison. Comparing the treatment group before and after the policy, without a control group, conflates the treatment effect with all other time-varying factors. The control group is what makes DiD a credible causal design — without it, you have a pre-post comparison that cannot separate the policy’s effect from concurrent changes in the economic environment.

6. Including bad controls. Adding post-treatment covariates that are themselves affected by the policy introduces bias rather than reducing it. For example, controlling for a firm’s internal control quality when studying SOX’s effect on audit fees is problematic because internal control quality is an outcome of the regulation, not a pre-determined confounder. Valid controls must be determined before or independently of the treatment.

Limitations of Difference-in-Differences

1. The counterfactual is unobservable. DiD estimates what would have happened to the treatment group absent treatment, but this counterfactual path is never observed. No amount of pre-treatment data can definitively establish that trends would have remained parallel in the post-treatment period. Sensitivity analyses and bounding exercises can help, but the assumption ultimately requires a qualitative argument.

2. SUTVA: no spillovers between treated and control units. The Stable Unit Treatment Value Assumption requires that the treatment of one unit does not affect the outcomes of other units. If SOX compliance requirements cause non-accelerated filers to voluntarily upgrade their internal controls or auditing standards (a spillover effect), the DiD estimate understates the true treatment effect because the control group’s outcomes are also partially affected by the policy.

3. External validity is limited. DiD estimates the effect of a specific policy, in a specific institutional context, at a specific point in time. The SOX audit cost estimate reflects the U.S. regulatory environment of the early 2000s and may not generalize to the cost of future regulations in different industries, countries, or economic conditions.

4. Composition changes can bias estimates. If the treatment causes units to enter or exit the sample — for example, firms delisting, merging, or going private in response to SOX compliance costs — the post-treatment sample represents a different population than the pre-treatment sample. This selection effect can bias the DiD estimate upward or downward depending on which firms leave.