Missingness

Controlled injection of missing values into generated datasets.

Real-world tabular datasets are almost never complete. Missing data is endemic across domains, and the mechanism of missingness determines its statistical impact:

MCAR  (Missing Completely At Random)
      Each value is independently missing with some probability.
      Example: random sensor packet loss, database write failures.
      → Simplest mechanism.  Reduces sample size but does not bias estimates.

MAR   (Missing At Random)
      Missingness depends on *observed* values.
      Example: older survey respondents skip digital-interaction questions;
               patients with certain diagnoses are less likely to have lab X recorded.
      → Missingness patterns carry information about the observed features.
         A model that ignores this loses signal.

MNAR  (Missing Not At Random)
      Missingness depends on the *missing value itself*.
      Example: sicker patients have more missing lab results (the very values
               that would be most informative are the ones most likely absent).
      → Hardest mechanism.  Ignoring it introduces systematic bias.

A foundation model trained on a synthetic prior with no missing values has never seen incomplete inputs during pretraining. At inference, it must handle missingness zero-shot with no prior exposure. Incorporating missingness into synthetic priors has demonstrated performance gains on real datasets with incomplete features, and covering missingness regimes improves reliability in weak meta-feature settings.

Use missingness workflows to inject deterministic synthetic null patterns for robustness testing under MCAR, MAR, and MNAR regimes.

In dagzoo, missingness is an observation model applied after target generation. The default prior first samples a latent DAG, emits complete features from node-assigned converters, emits y from one selected latent target node, then samples a missingness process and emits X_obs = mask(X_complete, m). This keeps latent target derivation separate from the later censoring process.

When to use

Why it matters for your prior

  • Your synthetic prior should reflect the near-universal presence of missing data in real tabular tasks — without it, the model’s first encounter with incomplete inputs is at inference time.
  • You want to test whether MNAR-aware pretraining produces more robust in-context learning than MCAR-only priors — a key question for medical/clinical tabular applications where informative missingness is the norm.
  • You want to measure how effective diversity changes when missingness is present versus absent in the prior, as a targeted ablation of this axis.
  • You need to control the missingness mechanism (MCAR/MAR/MNAR) independently from noise family and mechanism complexity to isolate its contribution to downstream performance.

Operational triggers

  • You want realistic training/evaluation with incomplete tabular data.
  • You need controlled ablations across missingness mechanisms.
  • You want benchmark guardrails to include missingness-aware checks.

Preset workflows

Use presets for standard mechanism runs:

dagzoo generate --config configs/preset_missingness_mcar.yaml --num-datasets 25 --out data/run_missing_mcar
dagzoo generate --config configs/preset_missingness_mar.yaml --num-datasets 25 --out data/run_missing_mar
dagzoo generate --config configs/preset_missingness_mnar.yaml --num-datasets 25 --out data/run_missing_mnar

Targeted MAR calibration via CLI

dagzoo generate \
  --config configs/default.yaml \
  --num-datasets 25 \
  --device cpu \
  --set dataset.missing_rate=0.25 \
  --set dataset.missing_mechanism=mar \
  --set dataset.missing_mar_observed_fraction=0.6 \
  --set dataset.missing_mar_logit_scale=1.4 \
  --out data/run_missing_cli_mar

Key options

  • --set dataset.missing_rate=...: overall missingness probability (fraction of cells that are NaN in the output). Concrete examples:

    --set dataset.missing_rate=0.05  →  5% of cells missing  (light; common in clean survey data)
--set dataset.missing_rate=0.15  →  15% of cells missing (moderate; typical clinical datasets)
--set dataset.missing_rate=0.30  →  30% of cells missing (heavy; EHR or sensor-network data)

  • --set dataset.missing_mechanism=...: which statistical mechanism drives the missingness. mcar = independent coin flip per cell; mar = missingness depends on complete feature values through the observation model; mnar = missingness depends on the value being censored.

  • --set dataset.missing_mar_observed_fraction=...: fraction of features used to compute MAR logits (higher = more features influence which values go missing).

    --set dataset.missing_mar_observed_fraction=0.3  →  30% of features drive missingness
--set dataset.missing_mar_observed_fraction=0.8  →  80% of features drive missingness (strong MAR)

  • --set dataset.missing_mar_logit_scale=...: MAR logit sensitivity multiplier. Higher values make missingness more sharply dependent on the complete features used by the observation model.

    --set dataset.missing_mar_logit_scale=0.5  →  weak MAR signal (missingness is nearly random)
--set dataset.missing_mar_logit_scale=1.5  →  strong MAR signal (missingness is highly structured)

What to inspect

  • In-process DatasetBundle.metadata["missingness"] for resolved missingness configuration and realized rates.
  • In-process DatasetBundle.metadata["prior"] for the latent-node target semantics versus the later observation-model step.
  • Public dataset_catalog.parquet for the stable emitted dataset identity and schema surface.
  • Benchmark summaries for preset_results[*].scenarios.missingness (when enabled).

For output details, see output-format.md.