A manufacturer builds a model to predict equipment failures from maintenance logs. The sensor table has blank readings whenever a device went offline, and technicians logged failure types only for repairs that went through the main ticket system. Some failure modes never appear at all, because crews fixed them on site and nobody wrote them down.
Missing data refers to values that are absent from a dataset where an observation was expected, such as blank fields, null values or records that were never collected.
Statisticians sort missing data by why it is missing. In the online book Flexible Imputation of Missing Data, Stef van Buuren summarizes the three categories Donald Rubin introduced in 1976. Data are missing completely at random (MCAR) when the probability of being missing is the same for all cases, missing at random (MAR) when that probability depends on information the team has observed, and missing not at random (MNAR) when it varies for reasons the team does not know. Van Buuren notes that most simple fixes only work under the restrictive MCAR assumption and can give biased estimates when it does not hold.
How are missing values different from missing cases?
- Missing values are blanks or nulls in records that exist. A data profiling report counts them by column.
- Missing cases are situations that never entered the dataset, such as a failure mode fixed on site and never logged. No cell is blank, so a null count cannot find them.
- Both affect AI results, but they need different responses. Missing values call for handling rules. Missing cases call for more data or a documented limit on what the model covers.
Where does missing data break in AI work?
- Biased fills: if a sensor drops out mostly under heavy load, filling the blanks with the column average understates the conditions that come before failures.
- Lost signal: dropping every row with a blank can remove the records where the system was under the most stress.
- Leakage through imputation: fitting an imputer on the full dataset before the train-test split lets evaluation data shape the training inputs.
- Silent gaps: a model trained without a failure mode will not flag it, and the evaluation scores will not reveal the gap if the test set lacks the same cases.
CUBIG Shorts
Plausible Rows Aren’t the Same as Useful Rows
From Ho Bae’s conversation with George Firican on the Lights On Data Show.
How do teams handle missing data?
Teams first measure where values are missing and ask why, by field, source system and time period. Common treatments include deleting incomplete rows, data imputation and indicator columns. The scikit-learn imputation guide notes that keeping track of which values had been missing can be informative, and its MissingIndicator transformer marks those positions so a model can use them.
For missing cases, teams compare the dataset against the situations the AI task has to handle, with input from the people who know the process. Where cases are scarce, teams may collect more records or add candidate data built under explicit conditions, as described in data augmentation, then check results on real evaluation data. Recording which gaps were filled and which remain makes it clear what the model was and was not tested on.
Related entries: data quality, golden dataset and AI-ready data.