Single Imputation and Missing-Value Uncertainty

One completed dataset is not the same as complete information

Replacing a missing value makes a dataset easier to analyze, but it does not reveal what that value actually was. A completed dataset contains a particular substitute; the uncertainty that the substitute might be wrong is not automatically carried into an ordinary analysis of that dataset. This distinction matters because standard errors and confidence intervals are meant to reflect uncertainty, not just the arithmetic consequences of the values currently in the data table.

Single imputation means filling each missing entry once and then analyzing the resulting completed dataset as if those entries had been observed. The replacement may be a fixed value, such as a mean or a prediction, or one random draw from a model for the missing value. These choices differ in how they select the replacement, but one completed dataset still represents only one particular completion.

An imputation is not an observation. It is a model-based substitute. A fixed imputation, such as a conditional mean, may summarize the model's expected missing value, but it does not show the range of plausible values around that expectation. A random imputation selects one plausible value, but one draw does not show how the analysis would change across other plausible draws.

For inference to be meaningful, the imputation model must be appropriate for the variable and its relationships with observed information. The assumptions about why data are missing also matter: an imputation based on observed predictors typically needs a missingness assumption such as MAR, along with a correctly specified conditional model. Under MNAR, the distribution of missing values may depend on values that are themselves unobserved, so additional assumptions are required. These assumptions cannot be established merely by completing the dataset.

How one imputation can conceal alternatives

Consider an estimate \(\hat{\theta}\), such as a sample mean or a treatment-effect estimate. If a missing value could plausibly take several values, then the estimate may differ across completed datasets. An analysis of only one completion calculates its estimate and usually its standard error conditional on that particular set of filled-in values. Unless the analysis separately accounts for uncertainty across plausible imputations, that standard error does not capture all the uncertainty created by missing data.

The idea can be expressed with the law of total variance. In a joint probability model, let \(\theta\) be the target parameter. Conditioning on the observed data and then considering the unknown values gives:

The first term describes average uncertainty remaining when a particular set of missing values is specified. The second describes how the parameter estimate or its conditional expectation varies as the missing values vary over plausible completions. The identity follows by adding and subtracting the conditional expectation:

The two bracketed terms have zero conditional cross-product in expectation, because the first has conditional expectation zero given \(D_{\mathrm{obs}}\) and \(Y_{\mathrm{mis}}\). Squaring, taking expectations given \(D_{\mathrm{obs}}\), and expanding therefore yields the variance decomposition above. A single completed dataset supplies just one set of \(Y_{\mathrm{mis}}\); it does not, by itself, estimate the variation represented by the second term.

This decomposition is a way to understand the issue, not a claim that every frequentist analysis uses this exact parameter-variance formula. In practice, the key distinction remains: uncertainty calculated conditional on one completion is not the same as uncertainty that also reflects alternatives for the missing values.

Why one random draw is not enough

Random imputation can represent uncertainty in the value being filled in better than a deterministic prediction: the imputed value is allowed to vary according to a specified predictive distribution. Yet after drawing once, the analyst still has only one completed dataset. The draw's randomness is not the same as an analysis that incorporates the spread of results across possible draws. In addition, a random draw from a model with estimated parameters may fail to account for uncertainty in those parameters unless that uncertainty is included in the draw procedure.

Assumptions, diagnostics, and alternatives

A single-imputation result depends on both the imputation rule and its assumptions. A fixed substitute can be especially misleading when the missing values differ systematically from the observed values after accounting for the predictors used. A random substitute can also be misleading if the predictive distribution is poorly specified, relevant predictors are omitted, or uncertainty about model parameters is ignored. No random number generator can correct a misspecified model or identify an untestable missingness mechanism.

Useful checks include comparing the observed data with the values the imputation model predicts, examining whether predictions are plausible on the variable's scale, and checking whether conclusions change under reasonable alternative imputation assumptions. For random imputations, rerunning the draw and noting how estimates change can reveal sensitivity to the particular draw. These checks do not prove the assumptions are right; they can expose instability or implausible imputations.

Single imputation also has practical limitations. A completed dataset may make later analyses convenient, but reporting an ordinary standard error or confidence interval from it as if every value had been observed can understate uncertainty. The point estimate can be distorted as well: the expected substitute need not preserve the target relationship or estimate when missingness and outcome structure are more complex. Neither a fixed substitute nor a random draw is automatically unbiased.

The central comparison is not that single imputation is always worse for every descriptive purpose. A filled dataset can be useful for a clearly limited task, and a fixed value may be a transparent convention when its purpose is not inferential. The problem arises when one imputed dataset is treated as though it contains the same information as the unobserved complete data, particularly when uncertainty statements are then based only on that one completion.

Common mistake

• Wrong reasoning: “I used a random draw, so the uncertainty from missing values is included in my standard error.”
• Correction: A random draw reflects one realization from the imputation model. An analysis of that single completed dataset still cannot show how estimates vary over other plausible draws. Accounting for missing-value uncertainty requires an inferential method that includes that variation, not merely a randomized replacement.