This workflow helps business PhD and DBA researchers connect a causal question to an explicit identification strategy. Four independent synthetic studies illustrate digital adoption, a common-date rollout, encouragement to adopt, and a grant cutoff. They are not alternative estimators applied to one dataset, and their effect estimates should not be compared as a horse race.
You will define estimands, audit assignment and overlap, estimate effects with design-appropriate uncertainty, examine falsification and sensitivity results, and produce reporting language. Prerequisites are linear regression, basic R, and a substantive understanding of treatment timing and confounding. This is a teaching workflow; a significant coefficient does not establish identification.
For real data, set use_synthetic to FALSE and supply all
four input files with the schemas in the data dictionary. These are four
separate worked designs; keep only the relevant sections in a
study-specific copy. Recenter a real RD score so that the cutoff is
zero. A missing-data strategy must be chosen before running: this
workflow deliberately stops on incomplete required inputs. It does not
impute or silently delete cases.
| design | rows | firms | missing |
|---|---|---|---|
| observational | 900 | 900 | 0 |
| did | 1800 | 180 | 0 |
| iv | 1000 | 1000 | 0 |
| rd | 1400 | 1400 | 0 |
| Design | Target in this example | Identification requirement |
|---|---|---|
| Weighting | ATE in the sampled firm population | No unmeasured confounding given pretreatment size, age, and sector; overlap; consistency |
| Difference-in-differences | Average post-adoption ATT | Parallel untreated trends; no anticipation; no spillovers; common adoption date |
| Instrumental variables | LATE among encouragement compliers | Relevant randomized encouragement; exclusion; monotonicity; no interference |
| Sharp RD | Local effect at the grant cutoff | Continuity of potential outcomes; no precise manipulation; deterministic assignment |
Before adjustment, sketch a causal diagram. In the weighting example, size, age, and sector precede both adoption and productivity. Do not adjust for a mediator such as post-adoption training or a collider such as selection into a performance award. Sampling weights, propensity weights, and precision weights answer different questions.
| variable | before | after |
|---|---|---|
| size_z | 0.597 | -0.004 |
| age | -0.147 | 0.008 |
| sector | 0.104 | 0.008 |
| adopt | n | effective_n | max_weight |
|---|---|---|---|
| 0 | 550 | 508.469 | 5.265 |
| 1 | 350 | 295.283 | 9.089 |
The largest absolute weighted SMD is 0.008. Values around 0.10 are a review convention, not a proof of exchangeability. Check distributional overlap and nonlinear terms, not only means. The largest weight is 9.09. Extreme weights signal limited support; trimming changes the target population and should be prespecified and reported. No automatic trimming is applied here.
The Hájek estimate normalizes weights within adoption groups. The bootstrap resamples independent firms and refits the propensity model every time, capturing estimated-weight uncertainty. If firms are clustered in markets or firms repeat, resample the independent assignment clusters instead. Bootstrap intervals do not repair confounding or poor overlap.
| estimate | std_error | lower | upper | replicates |
|---|---|---|---|---|
| 4.091 | 0.36 | 3.367 | 4.798 | 1999 |
Results text: In 900 firms, the weighted productivity difference was 4.09 points, 95% bootstrap CI [3.37, 4.80], based on 1999 resamples. Interpretation as the ATE requires adequate overlap, the stated measured-confounding model, and no unmeasured confounding. Check interval stability across seeds or additional replicates before final reporting.
Firm fixed effects absorb stable differences; year effects absorb common shocks. Treatment is assigned at the firm level, so both the aggregate model and event-study intervals cluster by firm. This example has one treatment date and never-treated controls. Do not reuse this two-way fixed-effects event study unchanged for staggered adoption with heterogeneous effects.
| model | term | estimate | std_error | p | lower | upper |
|---|---|---|---|---|---|---|
| Average post-adoption ATT | treated:post | 3.391 | 0.274 | p < .001 | 2.851 | 3.932 |
| test | F | p | df1 | df2 |
|---|---|---|---|---|
| Joint pre-adoption coefficients | 1.974 | p = .100 | 4 | 179 |
Results text: The estimated average post-adoption ATT was b = 3.39, 95% CI [2.85, 3.93], p < .001. The event-study preperiod test returned p = .100. A nonsignificant pretrend test does not prove parallel trends: consider power, economic plausibility, concurrent policies, and changing sample composition. The intervals are pointwise, not simultaneous. The design uses 180 firm clusters; few-cluster studies need different inference, such as an appropriate wild-cluster procedure or randomization inference.
At the .05 level, the joint preperiod test does not reject zero preperiod contrasts. This is an absence of detected joint differences, not affirmative evidence that untreated trends would remain parallel.
Encouragement increases adoption, but some firms adopt regardless and others never adopt. The synthetic instrument is randomized and affects productivity only through adoption; unobserved uptake propensity also predicts the outcome. These assumptions are known because this is a simulation. A real encouragement may directly change behavior and violate exclusion. LATE applies to compliers, not necessarily all firms.
| model | term | estimate | std_error | p | lower | upper |
|---|---|---|---|---|---|---|
| OLS association | adopt | 6.622 | 0.331 | p < .001 | 5.973 | 7.271 |
| 2SLS LATE | fit_adopt | 4.981 | 0.900 | p < .001 | 3.214 | 6.747 |
| model | term | estimate | std_error | p | lower | upper |
|---|---|---|---|---|---|---|
| First stage | encouraged | 0.364 | 0.029 | p < .001 | 0.307 | 0.421 |
| model | term | estimate | std_error | p | lower | upper |
|---|---|---|---|---|---|---|
| Reduced form / ITT | encouraged | 1.813 | 0.379 | p < .001 | 1.069 | 2.557 |
The single-instrument heteroskedasticity-robust first-stage Wald statistic is 159.45. A cutoff such as 10 is not a universal assurance of strong identification. Two-stage least squares standard errors come from the IV estimator, not an ordinary regression on fitted treatment. Here the 2SLS estimate is b = 4.98, 95% CI [3.21, 6.75], p < .001. Compare this with the OLS association to illustrate endogeneity, not to prove the IV assumptions.
For each hypothesized effect, the Anderson–Rubin-style test regresses outcome minus the hypothesized effect times adoption on encouragement and the exogenous covariate. The heteroskedasticity-robust test is asymptotic; inversion can yield wide, disconnected, or unbounded sets. This finite grid is a diagnostic, not an exact endpoint calculation.
Accepted grid values range from 3.25 to 6.50. Acceptance does not reach the tested grid boundaries. The complete accepted set is available in the exported grid; a range alone can hide disconnected regions. With one excluded instrument the model is exactly identified, so there is no overidentification test. No statistical diagnostic verifies exclusion or monotonicity.
Staggered adoption, heterogeneous effects with changing treatment cohorts, anticipation, treatment reversal, spillovers, time-varying confounding, weak instruments, few clusters, fuzzy or discrete-score RD, and endogenous sample selection require extensions. Do not interpret a firm fixed-effect regression as causal merely because fixed effects are included. For clustered assignment, the number of independent treatment units—not the number of firm-year rows—drives design strength. Power planning should simulate the intended assignment process and estimator.
All examples are synthetic and do not reproduce proprietary business databases or dissertation datasets.