What an Imputed-Mean Trace Represents
A MICE run updates missing values through a sequence of conditional imputation steps. For a continuous variable, one useful diagnostic summary is the mean of the values currently imputed for that variable. Plotting that summary against iteration gives a trace: the horizontal coordinate records the order of the updates, and the vertical coordinate records the imputed mean. Reading the path can reveal whether the summary changes level, moves persistently in one direction, or fluctuates around a comparatively steady level.
The trace concerns the missing entries of one variable, not the mean of all observed and imputed values. The observed values of that variable do not change from iteration to iteration; including them would dilute or obscure movement in the imputed part. This distinction is essential to interpreting the vertical axis.
The set \(\mathcal{M}_j\) is fixed because the pattern of missingness in the original data is fixed. The individual imputed values can change as MICE proceeds, so their mean can change too. The index \(c\) identifies a chain, while \(t\) orders iterations within that chain. A trace is therefore a sequence of summaries, not a sequence of new observed-data means.
Reading Shifts, Trends, and Fluctuation
A shift is a noticeable change in the trace’s level between portions of the sequence. A trend is a sustained movement in one direction over a run of iterations. Stable fluctuation describes variation up and down without an evident persistent change in level across the displayed sequence. These are descriptive readings of a plot, not formal pass-or-fail rules. The size and meaning of a movement depend on the variable’s units and context, and there is no universal numerical cutoff that makes a trace “stable.”
A trace can fluctuate even when its broad level appears steady: the imputed values are updated stochastically, and their average need not be identical at every iteration. Conversely, a trace that looks smooth or nearly flat does not by itself prove that the full imputation procedure has settled into an appropriate distribution. The mean is only one summary; changes in other features of the imputed values may not be visible in it.
Worked Example: Calculating a Mean Trace for a Continuous Measure
Suppose a hypothetical clinical dataset has four missing values for a continuous laboratory measure, recorded in mg/dL. In one chain, the four imputed values at iteration 1 are 8, 10, 12, and 14 mg/dL. At iteration 2 they are 9, 11, 13, and 15 mg/dL; at iteration 3 they are 8.5, 10.5, 12.5, and 14.5 mg/dL. Calculate and interpret the three trace points.
At iteration 1, the mean of the four imputed values is
At iteration 2, the mean is
At iteration 3, it is
The plotted sequence is therefore 11, 12, and 11.5 mg/dL at iterations 1, 2, and 3. It rises by 1 mg/dL and then falls by 0.5 mg/dL. With only three points, these values do not establish a persistent trend or a stable long-run level. The appropriate description is a short sequence with a rise followed by a partial return, not evidence that the chain has converged.
Comparing Chains and Describing the Trace Carefully
When a MICE run has multiple chains, each chain has its own sequence of imputed means. The chains can be shown separately or in one plot, provided the chain identity remains clear. Comparing their paths helps distinguish movement within a chain from differences in level between chains. A trace may show similar fluctuation in every chain, or one chain may remain at a different level; neither pattern should be judged from a single point alone.
Worked Example: Stable Fluctuation Across Two Chains
In a hypothetical dataset, two chains produce imputed means for a continuous measure, in mg/dL, at five successive iterations:
| Chain | Iteration 1 | Iteration 2 | Iteration 3 | Iteration 4 | Iteration 5 |
|---|---|---|---|---|---|
| A | 11.8 | 12.0 | 11.9 | 12.1 | 12.0 |
| B | 12.1 | 11.9 | 12.0 | 11.8 | 12.1 |
For Chain A, the average of its five trace points is
For Chain B, the corresponding average is
Both paths move up and down within a narrow range around 12 mg/dL, and their five-point averages differ by 0.02 mg/dL. In this short illustration, the traces show stable-looking fluctuation and similar levels. That is a description of these plotted means, not a guarantee that the chains are adequate: five iterations provide little information about behavior over a longer run, and a mean trace cannot show every feature of the imputed values.
Keep the vertical scale and units visible when interpreting changes. A movement of 1 unit could be negligible or consequential depending on the measure. Also inspect the sequence in iteration order: a collection of values without their ordering can show a range, but cannot reveal whether values moved steadily, shifted suddenly, or fluctuated around a level.
Why the Mean Moves—and What It Can Hide
The change in the imputed mean between two successive iterations can be written directly in terms of the changes in the individual imputations. With the same \(r_j\) cases included at both iterations,
This follows by subtracting the two sums and factoring out the common denominator \(r_j\). Thus, the trace point changes by the average change across imputed cases. Individual imputations can move substantially in opposite directions and cancel in this average. A nearly unchanged mean therefore does not imply that the case-level imputations barely changed.
For the same reason, a mean trace is a focused diagnostic rather than a complete assessment. It is useful for detecting movement in the average imputed level, but it does not report how spread out the imputations are, whether the tails change, or whether particular cases receive unusual values. Other diagnostics answer different questions; no single summary should be treated as a complete description of a chain.
Worked Example: A Stable Mean Can Mask Individual Changes
Suppose two missing cases for a continuous measure have imputed values of 10 and 20 units at one iteration. At the next iteration, their values are 12 and 18 units. Calculate the two means and explain what the trace does and does not show.
The first mean is
The second mean is
The mean trace is flat at 15 units, even though the first imputation increased by 2 units and the second decreased by 2 units. The trace correctly reports no change in the average, but it conceals the individual movements. A flat mean trace should therefore be read narrowly: the average imputed value did not change, not that every imputation stayed fixed.
Common Mistakes, Diagnostics, and Alternatives
- Including observed values in the trace mean. The target is the mean over the fixed set of missing cases, not the full variable mean. Mixing observed and imputed cells can make the plotted change smaller and answer a different question.
- Calling any up-and-down movement a trend. A trend requires sustained directional movement, not just two adjacent points that differ. Describe what is visible over the sequence and avoid inferring a persistent pattern from a few iterations.
- Treating a flat trace as proof of convergence. The mean may be stable while individual imputations or other distributional features still change. Use the trace as one diagnostic target among several.
- Ignoring the scale or units. A vertical movement has no context-free interpretation. Check the axis, units, and plausible range of the measure before judging its size.
- Comparing chains without preserving their identities. Combining values while discarding chain and iteration order removes the information needed to read separate paths.
A trace plot is most informative when its axes identify iteration and the imputed mean with units, and when chain identities are distinguishable. If the path has implausible jumps, first check that the correct variable, missing cells, scale, and iteration ordering were used to form the summary. A sudden movement may reflect legitimate changes in imputed values, but a coding or plotting error can produce a misleading trace as well.
The main alternative is not to replace the mean trace with one universally superior plot, but to choose a summary that matches the diagnostic question. The mean is suited to changes in average level. A trace of a different summary can reveal features the mean omits, while direct inspection of imputed values can reveal case-level movement hidden by averaging. These views are complementary, not interchangeable.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A variable has three missing cases, imputed as 4, 7, and 10 units at one iteration. What is the trace value?
- At the next iteration the three imputations are 5, 6, and 10 units. What is the new trace value, and by how much did the mean change?
- Two missing-case imputations change from 8 and 14 units to 10 and 12 units. What does the mean trace show, and what does it hide?
- Why should the observed values of the variable generally be excluded when calculating this trace summary?
- Does a flat imputed-mean trace prove that the MICE chains have converged? Explain.
Solutions. 1. The trace value is \((4+7+10)/3=21/3=7\) units. 2. The new value is \((5+6+10)/3=21/3=7\) units, so the mean changed by \(7-7=0\) units. 3. Both means are \((8+14)/2=11\) and \((10+12)/2=11\) units. The trace is flat, but it hides that both case-level imputations moved by 2 units in opposite directions. 4. The target is the average over the missing cases being imputed; including observed values would instead calculate a mixed full-data summary and can dilute movement in the imputations. 5. No. It shows only that the average of those imputations did not visibly change; it cannot establish stability of individual imputations or other features of the imputed distribution.