This stage is a decision-making workflow for diagnosing missingness at the dataset and column levels.

In real life, missing data is common. Although it should ideally receive ample time and attention, the reality is that scientists and analysts often face tight deadlines, and consequently may fall down a rabbit hole or fail to meet team and personal aspirations, especially in the fast-paced environment of a startup.

Definition: missing data (or missing values) is a data value that is not stored for a variable in the observation of interest. (Kang, 2013)

There is no single correct method for handling missing data, as it depends on the data sample, the modelling problem and the technique used to reach our end-of-day objective, not to mention how unclean the data is. But that doesn't stop the problem from being feasible. The main challenge in automating this workflow is settling on the optimal method for diagnosing the missing values, given these many layers of perception.

Using Bayes' theorem, the problem of choosing the optimal method, Methodᵢ, where i = 1, 2, 3, ... indexes the methods explored in the current session, simplifies to the following expression:

P(Methodᵢ | x_model_problem ∧ x_data_domain_description ∧ x_data_field)

where:

• x_model_problem : the target objective to achieve from this data sample
• x_data_domain_description : the type or origin of this dataset
• x_data_field : The data column containing the missing values.

From a programmer's point of view, the data persisted across states in this decision-making cycle are insights that highlight the currently observed sample. These insights also persist across the agent's episodes, and can be stored in BigQuery for easy access during data transformations.

Agent	Purpose	Output Format	Determinism Tactics
MissingMechanismClassifier	LLM classification of mechanism	JSON {mechanism, justification, risk_flags, confidence}	Constrained prompt + fallback parser
MissingImpactAssessor	Risk scoring across 4 domains	JSON	Defaults on parse failure
MissingStrategyRecommender	Ranked handling strategies	JSON array + rationale	Conservative fallback (mean_imputation)
(Heuristic Layer)	Pre-LLM signal extraction	Python dataclass	Deterministic numeric signals

Types of Missing Dataset

According to the source, the first step is distinguishing how the missing values occur in the data sample. These are the tests for telling each mechanism apart, by data field type:

Data Field Type	MAR vs MCAR	MNAR vs MAR	MNAR vs MCAR
Numeric (e.g., age, salary)	Little’s MCAR test; correlation tests between missing indicator & observed vars; logistic regression on missingness	Pattern-mixture models; Heckman selection models; sensitivity analysis with simulated unobserved values	Compare Little’s test with value-dependent dropout models; fit selection models predicting missing from true values
Binary classification	Chi-square test between missingness indicator & class; logistic regression of missingness ~ class	Logistic regression with observed covariates + EM imputation; examine residual dependence	Randomization/permutation tests vs fitted dropout models; check if imbalance remains after controlling for class
Multilabel classification	Missingness indicator regressed on known labels; random forests for feature importance	Pattern-mixture analysis across label sets; semi-supervised EM algorithms	Compare observed-label-predictive models with latent class models predicting missingness
Continuous (sensor data)	Time-series autocorrelation of missingness indicator; runs tests; logistic regression on device status	Joint modeling of outcome + missingness process (shared parameter models); survival analysis for dropout	Monte Carlo dropout simulation to test if missing aligns with extreme tails; likelihood ratio tests
Long text (speech, transcripts)	Logistic regression: missingness ~ metadata (speaker, duration, file size); clustering missing vs non-missing groups	Latent variable models on semantic embeddings; adversarial ML (train classifier to distinguish missing vs not)	Compare uniform random corruption vs topic-dependent dropout (e.g. TF-IDF clusters more missing)
Other text (name, email)	Chi-square test missingness ~ demographic groups; multiple imputation then compare fit metrics	Sensitivity analysis with reweighted likelihoods; Bayesian models with informative priors on unobserved values	Compare uniform random missing (MCAR) with models linking rare/unusual values to missingness (MNAR)

Decision-Making Tree in Distinguishing Missing Values

                          ┌───────────────────────────┐
                          │      Start: Missing?      │
                          └─────────────┬─────────────┘
                                        │
                          ┌─────────────┴─────────────┐
                          │                           │
                 ┌────────▼─────────┐       ┌─────────▼─────────┐
                 │ Test MCAR (Is     │       │   No missing data │
                 │ missingness       │       │ → Done            │
                 │ completely random?)│       └──────────────────┘
                 └─────────┬─────────┘
                           │
         ┌─────────────────┼──────────────────┐
         │                                    │
 ┌───────▼────────┐                 ┌─────────▼─────────┐
 │ MCAR confirmed │                 │ MCAR rejected     │
 │ (Little’s test │                 │ (missingness not  │
 │ not rejected,  │                 │ purely random)    │
 │ no predictors) │                 └─────────┬─────────┘
 └────────────────┘                           │
                                              │
                                 ┌────────────▼──────────────┐
                                 │ Test MAR (Dependence on   │
                                 │ observed variables?)      │
                                 └────────────┬──────────────┘
                                              │
                        ┌─────────────────────┼─────────────────────┐
                        │                                           │
             ┌──────────▼─────────┐                     ┌───────────▼──────────┐
             │ MAR confirmed      │                     │ Residual dependence   │
             │ (missingness can   │                     │ remains after all     │
             │ be explained by    │                     │ observed variables    │
             │ known covariates)  │                     │ tested                │
             └──────────┬─────────┘                     └───────────┬──────────┘
                        │                                           │
             ┌──────────▼─────────┐                     ┌───────────▼───────────┐
             │ Treat as MAR for   │                     │ Suspect MNAR          │
             │ imputation models  │                     │ (missingness depends  │
             │ and ML pipelines   │                     │ on unobserved value   │
             └────────────────────┘                     │ itself; use selection │
                                                        │ models, sensitivity   │
                                                        │ analysis)             │
                                                        └───────────────────────┘

Quantifying the Classes Attempt

However, this classification doesn't consider the context of the observed dataset, or of the data column where the missing values appear. The following are draft XOR / truth tables based on my understanding of Kang's paper. These matrix representations can guide an LLM's decision-making (online learning) processes, or serve as validation tests and rules.

MCAR

	Missing values	Non-Missing Values
Missing Values	0	0
Non-Missing values	0	0

MAR

	Missing values	Non-Missing Values
Missing Values	0	1
Non-Missing values	0	0

MNAR

	Missing values	Non-Missing Values
Missing Values	0.5	0.5
Non-Missing values	0.5	0.5

Worthy Considerations

In many applications, missing values may not appear in a consistent form. The most common forms in datasets are empty strings, pd.NA, None, np.nan, etc.

Explored Methods for Handling Missing Values

Common methods for handling missing values include:

• Expectation-Maximization
• Imputation Methods
• Sensitivity analysis
• Standardizing over the parameters from the column's approximated distribution

Concluding Actions

Actions taken on the dataset at the end of this stage:

• Dropping rows.
• Imputation strategies, e.g. mean, mode, etc., or adding an extra label such as 'UNKNOWN' for multi-label data types.

Resources

• Kang, H. (2013). The prevention and handling of the missing data. Korean Journal of Anesthesiology. https://pmc.ncbi.nlm.nih.gov/articles/PMC3668100/