Missingness

Incomplete records are not a single preprocessing step. CausalDynamics owns the Structural and Dynamical objects: response indicators, MAR/MNAR claims on a graph, observation masks, and generative dropout for simulation. Numerical fill or reweighting policies (:drop, :ipcw, …) belong in CausalTargeted and CausalMediation.

What this package does

Object	Role
`ObservationMask`	Binary (R=1) (observed) / (R=0) (missing) aligned to named columns
`MissingnessSpec`	Claimed regime (`:mcar`, `:mar`, `:mnar`) and optional conditioning set
`MissingnessCertificate`	Result of `certify_missingness`: identifiable status, `mar_set`, notes
`identify(...; missingness=)`	Attaches a certificate on `IdentificationResult`.missingness
`apply_missingness_mechanism`	Simulate MCAR/MAR masks on complete columns (MNAR refused)
`simulate_incomplete_panel`	Complete CDM panel, then apply a mechanism
`require_complete_values` / `require_complete_matrix`	Refuse silent `Missing` → `Float64`

Causal backdoor / frontdoor identification and missingness identification are separate: an MNAR claim can leave result.missingness.identifiable == false while result.identifiable remains true for the interventional query under complete observation.

Assignment form

For each substantive column (V) (e.g. (Y)),

\begin{aligned} V &\coloneqq fV\bigl(\mathrm{pa}G(V),\, U^V\bigr), \ RV &\coloneqq f{RV}\bigl(\mathrm{pa}G(RV),\, U^{RV}\bigr), \ V^{\mathrm{rec}} &\coloneqq \begin{cases} V & RV = 1, \ \texttt{missing} & RV = 0. \end{cases} \end{aligned} $

Latent complete values stay in the model; Julia missing is only the recorded token when (RV=0). Figure labels may write (V^) for (V^{\mathrm{rec}}); counterfactuals remain (V^{do(\cdot)}(\mathbf{u})), not (V^). Imputed fills (\tilde{V}) are Observable policies and are not new nodes on (G). Dynamics never invents a float fill inside CDM solvers or [`encodeto_panel`](@ref).

Regimes

Regime	Certificate	Simulation
`:mcar`	Identified (no conditioning set required)	Bernoulli missingness via `intercept`
`:mar`	Identified iff `conditioning_set` nonempty	Logistic in the conditioning covariates
`:mnar`	Unidentified without further assumptions	Throws in `apply_missingness_mechanism`

using CausalDynamics, Graphs

g = DiGraph(3)
add_edge!(g, 1, 2)  # W → A
add_edge!(g, 1, 3)  # W → Y
add_edge!(g, 2, 3)  # A → Y

miss = MissingnessSpec(:Y; regime = :mar, conditioning_set = [:W])
id = identify(g, TotalEffectQuery(:A, :Y);
    node_names = Dict(1 => :W, 2 => :A, 3 => :Y),
    missingness = miss,
)
id.missingness.identifiable  # true
id.missingness.mar_set       # [:W]

Masks and completeness guards

y = [1.0, missing, 3.0]
mask = observation_mask(Dict(:y => y); columns = [:y])
miss_rates(mask)  # Dict(:y => 1/3)

require_complete_values([1.0, 2.0]; context = "parents")  # ok
# require_complete_values([1.0, missing]; context = "parents")  # throws

High-dimensional assays must be completed or row-dropped under an Observable policy before encode_to_panel.

Hand-off to estimation

Estimators read the MAR set via CausalTargeted mar_set(id) or pass the IdentificationResult into impute_posterior(...; certificate=id). Do not treat survival censoring IPCW as the same object as MAR missing terminal outcomes; that distinction is owned downstream.

Stress ledgers for incomplete panels and encode guards live with the estimation stack (CausalTargeted missingness grid).