Mean Imputation's Point Mass and Distribution Shape

Why the Completed Histogram Can Mislead

A histogram of a completed data set can look reassuringly full: every row has a value, and no observations are left out. But if the missing values were all replaced by the same observed mean, the completed data set contains a feature that may not reflect the population at all: a pile of identical values at the imputation point. That pile can change the apparent shape of the distribution, including how concentrated or skewed it looks.

This tutorial focuses on the distributional footprint of mean imputation. It does not require a model for the missingness mechanism to describe the arithmetic effect. Whether the completed histogram is a reasonable representation of the underlying data is a different question, and depends on information that the observed histogram alone cannot supply.

The basic operation is deterministic: for a fixed observed data set and fixed missingness pattern, every imputed value is the same. If \(m>0\), at least \(m\) rows in the completed data set are tied at the observed mean. There may be more than \(m\) such rows if some observed values were already equal to that mean.

The Point Mass in the Completed Distribution

An empirical distribution assigns probability \(1/N\) to each of the \(N\) entries in a data set. After mean imputation, its mass consists of two parts: the original observed values, each with weight \(1/N\), and the \(m\) imputed copies at the mean. The completed-data empirical distribution can be written as

where \(\widehat{F}_{\mathrm{obs}}\) is the empirical distribution of the \(r\) observed values, and \(\mathbf{1}\{\bar{y}_{\mathrm{obs}}\le t\}\) equals 1 when the condition in braces holds and 0 otherwise. The formula says that a fraction \(r/N\) of the completed distribution comes from observed values, while a fraction \(m/N\) is placed at one location.

In particular, the total fraction of completed values equal to the imputation point is

where \(\#\) denotes the number of observed values satisfying the condition. The imputed share alone is \(m/N\). These are exact features of the constructed data set, not estimates of how often the true variable takes the mean.

How the Point Mass Changes the Shape

A histogram groups values into intervals, so adding a point mass changes the count in the bin containing the observed mean. The effect on the picture depends both on the fraction \(m/N\) and on the chosen bin boundaries. A large missing fraction can make the pile visually prominent; a small one may be difficult to notice. Neither visibility nor invisibility changes the underlying fact that all imputed values are tied.

The pile can also conceal features that matter for describing the variable. For instance, if missing values would have included large values, replacing them by a central mean can make a right tail appear less substantial. The observed values may still show some tail, but the completed histogram mixes that observed evidence with imputed values whose locations were assigned by rule. The completed shape is not, by itself, evidence about the unknown values.

A point mass can also merge with natural ties. If the observed mean equals one or more observed values, the completed histogram displays a combined group. A plot alone may not reveal which entries were imputed. Conversely, changing bin boundaries can split or combine the visual impression around the mean, even though the number of records imputed there remains unchanged.

Assumptions, Checks, and Alternatives

The calculation requires a quantitative variable, at least one observed value so the observed mean is defined, and a specified set of missing rows. It assumes only that the imputation rule is to replace each missing value by that mean. It does not assume that missingness is completely at random, at random conditional on observed variables, or not at random. Those mechanisms concern why values are missing; the pile occurs mechanically under this rule regardless.

Useful checks when examining a mean-imputed distribution include:

• Report \(r\), \(m\), \(N\), the observed mean, and the imputed fraction \(m/N\), rather than showing only the histogram.
• Mark the observed mean on the plot and check whether an unusually tall bin is centered on or contains it.
• Compare the observed-value display with the completed-data display, clearly labelling the different sample sizes and what each includes.
• Check sensitivity to reasonable bin widths and boundaries. If the visual story changes sharply, describe that as a display limitation, not as a change in the underlying completed values.
• Look for features that a single central fill could mask, such as an observed tail or distinct groups. Do not infer the unseen values from the imputed histogram.

A histogram of observed values avoids adding an artificial pile, but it displays only the observed cases and may therefore give an incomplete picture. Mean imputation creates a complete data set but imposes its point mass. Multiple imputation instead produces multiple completed data sets so that uncertainty about missing values can be represented across imputations; it does not make any one histogram a direct view of the unknown full distribution. The appropriate comparison depends on the descriptive or inferential question, not just on which plot looks smoother.

The key limitation is that the completed histogram records the chosen replacement rule as well as the observed data. It can be useful for documenting what a single-imputation data set contains, but its shape should not be treated as if every plotted value had been measured. A smooth-looking histogram is not evidence that the missing-value uncertainty has disappeared.