AI Glossary / Statistics and probability / Intermediate
Deck imputation
Definition
Deck imputation fills a missing data item using information from a donor record. The donor is selected from a collection of available records, often by matching characteristics related to the missing item. Hot-deck methods generally use donors from the current survey or dataset; cold-deck methods use an earlier or otherwise external reference source. The supplied value remains an imputation rather than an observed response.
What it means
A record with a missing value is the recipient. A record supplying a replacement is the donor. The method tries to use information from relevant observed cases without requiring every missing value to be predicted by a fitted regression equation.
Choosing a donor is a modeling decision. Similarity in available fields is useful only insofar as it makes the donor's missing-item counterpart relevant to the recipient. Matching on arbitrary details does not guarantee a good replacement.
Terminology varies across statistical programs, especially for cold-deck defaults and reference sources. The distinction should be documented for the actual implementation. A historical donor can introduce problems when prices, behavior, definitions, or collection procedures have changed.
How it works
First define eligible donors with valid values for the required item and compatible record meanings. Establish matching classes or a distance measure using observed characteristics associated with the item and, where relevant, response patterns. Specify how to select among multiple eligible donors.
Some procedures select a nearest neighbor; others sample a donor within a class or use a sequential ordering. Rules are needed for ties, empty classes, donor reuse, and consistency between related variables. Copying several linked items together can preserve some relationships, but does not automatically resolve every constraint.
After imputation, validate the completed records and retain flags distinguishing reported and imputed values. Assess the consequences for the intended estimates. Treating one completed dataset as if every value had been observed can understate uncertainty. An appropriate variance or multiple-imputation procedure depends on the design and donor method, rather than being supplied automatically by copying values.
Three practical examples
- A business survey borrows a missing expenditure item from a responding organization in a comparable industry and size group.
- A customer research team uses a donor from the same study to fill a missing categorical answer while retaining an imputation indicator.
- A teaching exercise compares current-period and previous-period donor pools to show how changes over time can affect imputed amounts.
Think of it this way
Think of borrowing a replacement piece from a similar kit. The borrowed piece is not the missing original. You still need to check compatibility and remember which piece was substituted.
Worked example
Consider an original small business survey. One recipient has 12 employees and a missing monthly software expense. Three eligible donors in the same industry and region have 10, 14, and 25 employees, with expenses of 200, 360, and 700 currency units respectively.
Suppose the declared matching rule uses absolute employee-count difference. The distances are 2, 2, and 13. The first two donors tie for nearest. A documented random tie rule gives each probability one-half; the third donor is not selected under this rule.
The recipient's imputed expense is therefore either 200 or 360. Its expected imputed value under the tie rule is (200+360)/2 = 280. That expectation is not itself a copied donor value and should not be reported as the result of one random draw unless the method instead explicitly uses averaging.
If the three observed expenses total 1,260, the four-record completed total is either 1,460 or 1,620, with expected total 1,540. The donor-selection variance for the imputed item is [(200−280)²+(360−280)²]/2 = 6,400, giving a standard deviation of 80. This is only variation from the stipulated tie selection, not the total uncertainty from sampling, missingness, or an unsuitable donor model.
Illustrative case study
In this fictional case, Birch Services surveys small organizations about software costs before developing an AI budgeting assistant. Some firms omit the expense field. The analysts create donor groups using industry and a documented size measure, and examine groups with too few complete responses.
They preserve the original missing values, record the chosen donor identifiers internally, and store the imputation flags. A review finds that several matched firms use different accounting definitions, so those records cannot be exchanged without reconciliation.
The invented case shows why donor selection requires measurement consistency; imputation does not reveal the actual missing expenses.
Common misconceptions and mistakes
A donor's observed value is real for the donor, not necessarily correct for the recipient. Having no blanks does not mean that a dataset has no uncertainty or bias.
Matching on many variables can leave too few donors, while matching on too few can ignore important differences. Unlimited reuse of a small donor pool can make results depend heavily on a few respondents. Both problems deserve explicit diagnostics.
Hot-deck methods do not universally preserve correlations or multivariate consistency. Copying fields separately may create combinations that never occurred in any real record. Cold-deck values can also be outdated. Neither method automatically solves missingness driven by unobserved values after the matching information is considered.
Business use
Deck imputation can support consistent analysis when incomplete records would otherwise be discarded or handled differently by each analyst. Businesses should define the intended use, donor eligibility, matching variables, and fallback procedure before interpreting completed data.
Monitor imputation rates, donor reuse, empty matching groups, and changes in important summaries. Compare results under plausible alternative donor rules and retain the ability to analyze reported values separately. For AI development, fit any data-dependent imputation procedure within the training workflow and respect evaluation boundaries. Borrowing information from the evaluation set can create misleading performance estimates even when the imputation algorithm itself is reasonable.
People also ask
Is deck imputation always random?
No. Donor selection can be deterministic or stochastic, depending on the method.
Does hot-deck mean recent historical data?
Usually it refers to donors from the same current survey or dataset; document the program's convention.
Is one imputed file enough for uncertainty estimation?
Not automatically. The analysis needs an uncertainty procedure appropriate to its sampling and imputation design.
Check your understanding
Check your understanding: Why were two donors equally eligible in the example?
Answer: Both were two employees away under the chosen distance.
Is 280 a copied value?
Answer: No; it is the expected value over the two possible selections.
Does the standard deviation of 80 represent all uncertainty?
Answer: No.
Why retain imputation flags?
Answer: To distinguish replacement values from reported observations and support review.
Related terms
- Hot deck imputation
- Missing data
- Nearest neighbor
- cold-deck imputation
- donor record
- multiple imputation
- item nonresponse
Learn next
Compare deterministic and random donor selection. Study missing-data assumptions and variance estimation for imputed surveys.
More in Statistics and probability
- Data validation according to a list - Data validation according to a list is the checking of a data value against an explicitly defined set of acceptable values for a field.
- De Finetti's theorem - De Finetti's theorem connects exchangeability with conditional independence.
- De Moivre's law - De Moivre's law is a simple actuarial survival model in which lifetime is uniformly distributed between zero and a specified limiting age.
- Decoder models - Decoder models, in the language-modeling sense used here, generate sequences by predicting the next token from preceding context.
- Decoder-only model - A decoder-only model is a neural sequence model built around a decoder stack without a separate encoder stack for the text input.
- Decomposition of time series - Decomposition of time series separates observations recorded over time into components representing different patterns.
Sources
Sources: OECD Glossary of Statistical Terms, Deck imputation, https://www.oecd.org/content/dam/oecd/en/publications/reports/2008/09/oecd-glossary-of-statistical-terms_g1gh9ad7/9789264055087-en.pdf. US Census Bureau, SIPP Data Editing and Imputation, https://www.census.gov/programs-surveys/sipp/methodology/data-editing-and-imputation.html. The donor calculation and fictional business case are original.
Previous: Decision tree pruning Next: Declarative bias