Incomplete panels: what chaining and observation correction add
On this page
cdid connects the outcome changes your panel supplies and makes explicit when those changes represent your study population. It also provides observation weighting, actual-change imputation, and augmented reconstruction when observation depends on measured characteristics. These choices work separately from the adjustment that makes treated units comparable with controls.
The foundational chained-DiD framework
Bellégo, Benatia and Dortet-Bernadet’s chained DiD paper provides the foundational framework for chained difference-in-differences with incomplete outcome histories. It explicitly formalizes missing-trends assumptions, distinguishes representativeness of outcome changes from representativeness of levels, and establishes identification, estimation and inference using overlapping incomplete panels. Its conditional extensions address selection on observed characteristics. This framework underlies cdid’s treatment of missing outcomes. The Chained Difference-in-Differences, section 2.2 · Published paper
Consider households interviewed at dates 1 and 2 and other households interviewed at dates 2 and 3. Few households appear at both endpoints of the longer interval. Chaining uses the two shorter comparisons to recover a longer effect when the observation and parallel trends assumptions support them. Additional valid pairs can contribute too. The package combines their information by GMM, accounting for shared units and estimated models. Worked example
Several other DiD packages accept incomplete panels. The useful comparison concerns which observations enter each estimate, which population it describes, and which assumptions justify it. Six distinctions clarify what cdid adds.
1. R did can use an unbalanced panel
The default allow_unbalanced_panel = FALSE drops units not observed in every period. With unbalanced panels allowed, R did retains incomplete histories and uses repeated-cross-section computations. Its documentation distinguishes changes within a unit, conditional on earlier covariates, from period-specific outcome calculations with contemporaneous covariates. R did reference
This is a legitimate estimation route under its assumptions. It does not construct the same information as chaining within-unit changes across overlapping histories. In cdid, mode = "ordinary" delegates to did, including its repeated-cross-section route; chained mode performs the within-unit calculation. Moving from did
2. Pair balancing supplies a change, not an observation correction
Stata csdid 2.0.3 offers three choices. bal(full), the default, retains units observed throughout. bal(pair) requires both outcomes separately for each two-date comparison. bal(none) uses repeated-cross-section computations with available observations. Pair balancing reproduces the observation rule of legacy version 1.82. Official unbalanced-panel guide
Having two observed outcomes makes their difference available. It does not establish that its mean represents the original cohort. A baseline-to-follow-up comparison can also be unavailable while shorter, overlapping changes remain informative. Chaining can connect those changes; an observation correction addresses their selectivity. These solve different problems. Matched comparisons
3. Linear etwfe and jwdid already support incomplete panels
Their regression specifications can use incomplete outcome histories. The distinction between unit and cohort fixed effects matters: their balanced-panel equivalence need not survive when observation dates differ within cohorts. R etwfe uses cohort effects when ivar is omitted; its unit-effects specification needs the identifier. Stata jwdid documents unit effects and its group, corr, and cre alternatives. etwfe reference · jwdid help
The cdid cohort-by-date preset reproduces the documented linear unit-effects comparisons on matched complete and incomplete panels. That match requires aligned samples, covariates, controls, effect evaluation, and covariance normalization. Event-study averages additionally require the same aggregation weights. Commands and verified scope
4. The connection to MTCAR in linear fixed-effects models
Wooldridge’s section 10.2 describes a related selection mechanism within linear fixed-effects models. It permits missingness related to additive individual heterogeneity and rules out systematic selection on idiosyncratic errors, using verbal restrictions and a regression construction that draws on Wooldridge (2019). That subsection does not separately state a missing-trends assumption or an identification theorem for incomplete panels. Wooldridge (2025), section 10.2
The argument below makes the connection to the chained-DiD framework explicit. Its full-history mean-exogeneity condition is a sufficient reconstruction, not an assumption displayed in that subsection. With suitable restrictions, the additive linear model implies mean-change representativeness. When covariates affect both trends and observation, the corresponding restriction is conditional on those covariates.
To see the connection, write an additive model as
Y(i,t) = c(i) + m(t, cohort, X(i)) + u(i,t).
Differencing removes c(i). If the residual has mean zero at every date conditional on cohort, covariates, the unit effect and the observation history, then
E[Y(i,t) - Y(i,s) | cohort, X(i), both dates observed]
= E[Y(i,t) - Y(i,s) | cohort, X(i)].
This is a sufficient-condition argument connecting the two formulations. Chained DiD states the restriction directly on outcome changes; it need not impose that stronger full-history restriction on the regression errors. Without covariate selection, this gives baseline mean-change representativeness. With covariate selection, recovering a population effect also requires averaging over the population’s covariate distribution.
The relevant changes include treatment gains. If observation selects heterogeneous gains not accounted for by the model, removing a unit effect does not resolve that selection. Thus the linear fixed-effects argument and the chained-DiD observation assumption share a mechanism, while their full assumptions and estimators need not coincide. The missing-outcome guide explains the additional adjustments available in cdid.
5. Chaining requires representative mean changes
In the chained adoption model, with selection = "none", units observed at both dates must be representative of their cohort in mean changes over each used interval:
E[change | cohort, both dates observed] = E[change | cohort]
This concerns population expectations. Individual changes need not be identical, and sample means need not coincide exactly. Observation can select outcome levels while preserving the needed mean changes. The original paper explains why mean-independence suffices even though it states a stronger independence condition for exposition. Chained DiD, discussion of Assumption 2
When representativeness holds only conditional on covariates, the adjustment must reconstruct population changes from those conditional comparisons. The restriction covers actual treated changes, including treatment effects, and every pair used. Connected observation windows alone do not establish identification. Missing-outcome assumptions
The static switching model instead uses switcher and stayer populations. With covariates, its baseline also requires representativeness of their reference-covariate composition. See the static-model assumptions.
6. Treatment DR does not automatically correct selective observation
A treatment propensity and an untreated-outcome regression address the treatment comparison. An observation propensity and an actual-change model address which changes are observed. The latter model includes treated changes; it is different from predicting the untreated counterfactual.
Under selection = "conditional_mtar", cdid offers three adjustments:
- Observation weighting: weight each observed change using its probability of being observed at both dates.
- Change imputation: predict actual changes and average them over the declared study population.
- Augmented correction: combine those predictions with weighted prediction errors from observed changes.
The paper’s Corollary 1 establishes the conditional observation-weighting argument. The package guide documents imputation and augmentation and their separate model requirements. Treatment IPW, OR, and DR remain distinct choices. Corollary 1 · Adjustment guide
Define the population before choosing the fit
For the cohort/adoption route, supply frame when the study includes units with no outcome rows. It retains their population information, with the cohort and covariates required by the selected adjustment. Probability-based corrections require observation support; imputation needs training and prediction support. Model restrictions, parallel trends, and no anticipation still matter. Binding probability safeguards can change the target, so the documented robustness claims retain their clipping qualifications. Prepare the unit table · Probability safeguards
Comparison reviewed November 16, 2026. The R syntax and comparison evidence on this page describe cdid 1.3.0.0. The native Stata package has its own guide and R correspondence. Upstream sources inspected were R did master 2.5.1.906, Stata csdid 2.0.3, R etwfe online documentation with a changelog labeled 0.6.2.99 development, and Stata jwdid source 2.201. The recorded cdid regression comparisons used etwfe 0.6.2; their exact settings remain in the linked guide. This source review does not rerun those numerical comparisons. did version · csdid release · etwfe changelog · jwdid source