Why an Imputation Chain Needs Starting Values
A chained-equations algorithm begins with incomplete data, but its first conditional update needs a current value for the other variables in its conditioning set. A simple way to make that first update possible is to assign temporary values to missing cells. These values form a starting completed data set; they are computational starting points, not the final imputations and not additional observations.
Let \(Y_{ij}\) denote the value of variable \(j\) for case \(i\), where \(i=1,\ldots,n\) and \(j=1,\ldots,p\). Let \(R_{ij}=1\) when \(Y_{ij}\) is observed and \(R_{ij}=0\) when it is missing. Write \(Y_{ij}^{(0)}\) for the value in the initial completed data set.
Initialization does not, by itself, make an assumption about why values are missing or establish that an imputation model is appropriate. The observed-data assumptions and the conditional models matter for the subsequent imputation procedure; a convenient starting rule cannot repair a misspecified model. Its immediate job is narrower: provide valid values in the missing cells so the algorithm can begin.
Simple Initialization Strategies
The simplest options operate separately within each incomplete variable. For a numeric variable, a common deterministic start fills every missing cell with the mean of that variable's observed values. For a categorical variable, a common choice is its observed mode, the most frequent observed category. These starts are easy to reproduce and do not require fitting a model.
A second option is to assign each missing cell a value sampled from that variable's observed values, usually with replacement. This is a random start: two runs can begin with different completed data sets. Sampling observed values can retain the observed marginal range and categories, unlike a mean that might be outside a variable's allowable values. However, independently sampling each column does not generally preserve the relationships among variables.
| Starting rule | Typical use | Strength | Limitation |
|---|---|---|---|
| Observed mean | Numeric variables | Simple and reproducible | Can create ties, unrealistic values, or values outside the valid range |
| Observed mode | Categorical variables | Produces an allowed category | Can overrepresent the modal category in the initial data |
| Random observed value | Numeric or categorical variables | Uses observed support and avoids identical fills in many cases | Can distort relationships and adds random variation to the start |
For any rule, use only observed values in the relevant column to calculate its summary or form its sampling pool. The rule must also respect the variable's measurement scale and allowable values. For example, a mean is not a valid category for a binary variable, and it may be an invalid value for an integer count or a bounded measurement.
Worked Example: Mean Initialization for a Numeric Variable
Suppose a hypothetical data set records a numeric measurement, in millimetres, for five cases. The entries are \(12\), \(14\), missing, \(18\), and \(16\). Use the observed mean to initialize the missing cell.
There are four observed values, so the observed mean is
Set the third entry to \(15\) mm. The initialized column is therefore \(12,14,15,18,16\) mm. The first, second, fourth, and fifth entries remain exactly \(12,14,18,\) and \(16\) mm, respectively. The value \(15\) mm is a starting placeholder; this operation does not establish that the unobserved measurement actually equalled \(15\) mm.
This start is valid if the variable allows real-valued measurements across this range. If it were instead a count restricted to whole numbers, the mean could be unsuitable as a cell value, and another permissible starting rule would be needed.
Worked Example: Mode Initialization for a Categorical Variable
In a hypothetical survey, a categorical response has observed values Yes, No, Yes, Yes, No, with two additional responses missing. Initialize both missing cells using the observed mode.
Among the five observed responses, Yes occurs three times and No occurs twice. Therefore, the observed mode is Yes. Assign Yes to each missing cell, giving the initial sequence Yes, No, Yes, Yes, No, Yes, Yes.
The original five observed responses have not changed. The two added Yes values are placeholders, not evidence that the missing respondents answered Yes. In particular, the completed starting column now contains five Yes values and two No values, but that frequency distribution should not be reported as if it were the observed survey result.
Worked Example: Random Observed-Value Initialization
Consider a hypothetical numeric variable with observed values \(2,3,3,7\), and two missing cells. Initialize by sampling independently with replacement from the observed values. Suppose one possible random draw assigns \(3\) to the first missing cell and \(7\) to the second.
The resulting initialized column is \(2,3,3,7,3,7\). Both assigned values are in the observed support, and duplicates are allowed because sampling is with replacement. For each missing cell, the probability of drawing \(3\) is \(2/4=0.5\), while the probability of drawing \(2\) or \(7\) is \(1/4=0.25\) each.
This is one possible start, not a claim that these are the missing values. It also does not preserve associations with other variables if the draws are made separately by column. The purpose is only to provide permissible initial values before the conditional updates.
Why Observed Values Must Stay Fixed
The defining constraint can be implemented directly using the observation indicator. First copy the data, then assign starting values only at cells where \(R_{ij}=0\). In particular, a fill operation must not replace all entries in a column with its mean or mode.
Here \(a_{ij}\) is the value selected by the chosen initialization rule for a missing cell. The rule can vary by variable or by missing cell, but it must not override the observed-data branch.
The preservation property follows directly from the definition. For any cell with \(R_{ij}=1\), the first branch gives \(Y_{ij}^{(0)}=Y_{ij}\). Therefore, every observed cell is unchanged, regardless of how the missing-cell values \(a_{ij}\) are chosen. This is an exact cell-by-cell guarantee, not an approximate property of a particular mean or sampling procedure.
Identify which entries have \(R_{ij}=1\) and which have \(R_{ij}=0\). Keep this pattern fixed while creating the start.
Use a summary or sampling rule appropriate to its scale and valid values, calculated from that variable's observed entries.
Assign starting values where \(R_{ij}=0\), leaving every cell with \(R_{ij}=1\) untouched.
Verify that no cells remain missing, observed entries match the original data, and each assigned value is valid for its variable.
Assumptions, Diagnostics, and Limitations
A simple initialization requires at least one observed value in any variable initialized from that variable's observed distribution. A mean cannot be calculated from a column with no observed entries, and a mode or observed-value sample cannot be obtained from it either. Such a variable needs an externally specified starting value or a different modeling decision; a data-derived column rule is unavailable.
Check the start before proceeding. Useful checks include comparing every observed entry with the original data, confirming that no missing cells remain in the computational starting data, and validating ranges, categories, and integer constraints. For random starts, also check that the random-number procedure did not inadvertently sample missing entries or overwrite observed ones.
These strategies are deliberately crude. Column means and modes ignore other variables, while independent random draws ignore their joint structure. Consequently, an initialized data set can have distorted associations even when every individual value is allowable. The starting values should not be used as completed data for substantive analysis. They are temporary inputs that later conditional updates are intended to replace at missing cells.
Starting values can affect early stages of an iterative procedure, particularly when a variable has many missing entries or the initialization is far from plausible values. That is a reason to inspect the start and later assess the behavior of the imputation procedure, not a reason to interpret the placeholders as estimates. A simple initialization is not a substitute for suitable conditional models, adequate observed information, or a defensible missing-data analysis.
Common Mistakes
- Overwriting observed entries. Filling an entire column with its mean changes actual recorded data. Apply the assignment only where \(R_{ij}=0\).
- Treating placeholders as imputations for analysis. A mean-filled or mode-filled data set does not represent the intended multiple-imputation uncertainty. Use these values only as starts.
- Using a rule that violates the variable's scale. A mean can be fractional for a count, and a numerical average of category codes may not be a meaningful category. Choose a valid start for the variable.
- Assuming marginally plausible values preserve relationships. Drawing from each column separately may create unrealistic combinations across variables. Valid individual values do not guarantee a realistic joint data set.
- Trying to estimate a start from an entirely missing column. With no observed values, a column mean, mode, or observed-value sample is undefined. Additional information or a different modeling choice is required.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A numeric variable has observed values \(4,6,8\) and one missing value. What is its observed-mean starting value, and which entries must remain unchanged?
- A binary variable has observed categories A, A, B, and one missing cell. What does mode initialization assign, and what does that assignment not establish?
- A variable has observed values \(1,2,5\) and two missing cells. Under independent sampling with replacement from observed values, what is the probability that a particular missing cell receives \(2\)?
- Why can separate random observed-value starts for two variables produce a poor joint starting data set even when every assigned value is individually valid?
Solutions. (1) The mean is \((4+6+8)/3=6\), so the missing cell starts at \(6\); the observed entries \(4,6,8\) remain unchanged. (2) The mode is A, so the missing cell starts at A. This does not show that the missing response was actually A. (3) There is one occurrence of \(2\) among three equally likely observed entries, so the probability is \(1/3\). (4) Separate column draws do not condition on one another, so they can pair values in combinations that are rare or inconsistent in the observed data; marginal validity alone does not preserve relationships.