Why Early Iterations Need Special Attention
A MICE chain begins from an initial completed data set and proceeds by updating incomplete variables in turn. Early updates can depend strongly on that starting state. A trace of a chain summary may therefore show an initial transient before later values fluctuate around a more stable range. Including that transient in summaries of the retained imputations can distort those summaries.
The task here is to assess this early behavior, not to certify that the chain has reached the correct distribution. The earlier material on trace plots introduced summaries such as imputed means, dispersion measures, quantiles, and tail statistics. For any such summary, examine how its level and range change across successive iteration segments. The conclusion should be a reasoned, provisional choice about which iterations to retain.
This description depends on what is summarized. A chain might look settled for an imputed mean while a high quantile continues to change. It is therefore more defensible to check several substantively relevant summaries than to select a burn-in from one trace alone. The summaries should be chosen to reflect features of the imputed variables that matter for the analysis.
Comparing Iteration Segments
Divide the trace into consecutive segments of equal length, or into blocks chosen in advance for a particular diagnostic. For each segment, compare the level and the variability of the summary. A visible early shift followed by similar later segments is evidence that the startup transient has diminished. Continued movement across later segments means the burn-in decision remains uncertain; it does not, by itself, identify the cause.
A simple numerical comparison is the difference between the mean of an early segment and the mean of a later segment. For a segment \(A\) containing \(n_A\) iterations, define its average summary as \(\bar{S}_A\). When comparing segments, keep the units of \(S_t\): if \(S_t\) is an imputed mean measured in mmol/L, the difference is also in mmol/L. Such differences are descriptive. Chain summaries from successive iterations are dependent, so treating iteration-level values as independent observations and applying an ordinary two-sample test is generally not justified.
The comparison is not a hunt for exact equality. A settled chain still fluctuates, and two segment averages will rarely match exactly. Ask whether later segments show persistent directional change or a change in their range, and whether the size of the difference matters on the scale of the summary. No universal numerical cutoff defines stabilization for every variable, dataset, model, or analysis.
Worked Example: A Declining Imputed Mean
Suppose a hypothetical MICE run produces 100 iterations. For one continuous variable, the trace summary is the mean of its imputed values, measured in mmol/L. The averages over five consecutive 20-iteration segments are 12.4, 8.1, 6.3, 6.0, and 5.9 mmol/L.
Compare early segments. The first change is \(8.1-12.4=-4.3\) mmol/L. The next is \(6.3-8.1=-1.8\) mmol/L. The early trace is not stable: the segment average is falling substantially.
Compare the last two segments. Their difference is \(5.9-6.0=-0.1\) mmol/L. This is small relative to the earlier changes, and the last two segment averages are close. A reasonable provisional decision is to regard iterations 1 through 60 as a candidate burn-in for this summary, then retain iterations 61 through 100 for further assessment.
This decision is limited. Forty retained iterations provide only a short basis for judging later behavior, and one summary cannot represent every aspect of the imputations. The arithmetic supports a conclusion that this mean’s initial transient appears to have diminished; it does not establish convergence.
Why Removing a Transient Can Matter
A segment average that includes startup iterations can be shifted by those iterations. The following identity shows why a long enough stable portion does not automatically make the choice irrelevant. If the first \(B\) iterations form one segment and the remaining \(T-B\) iterations form another, the average across all iterations is a weighted average of the two segment averages:
This follows by splitting the sum of all \(T\) values into the first \(B\) and last \(T-B\) values, then dividing by \(T\). If the first segment differs sharply from the later segment, its contribution shifts the overall average in proportion to \(B/T\). Discarding an identified transient avoids giving it that contribution. But the identity says nothing about whether the later segment has reached the desired distribution; that must be assessed separately.
Worked Example: A Mean Appears Settled While a Tail Summary Moves
In a second hypothetical run, summaries for one imputed variable are recorded across five equal segments. The mean of imputed values is 10.1, 10.0, 10.2, 10.1, and 10.0 units. The 90th percentile is 18.0, 17.0, 16.0, 14.0, and 14.6 units.
Check the mean. Its largest difference between successive segment averages is \(10.2-10.0=0.2\) units in magnitude. Across the last three segments it changes from 10.2 to 10.0 units, a difference of \(-0.2\) units. On its own, the mean gives little indication of a continuing startup shift.
Check the tail summary. The 90th percentile falls by \(14.0-18.0=-4.0\) units from the first to the fourth segment. It then rises by \(14.6-14.0=0.6\) units. The tail trace has not simply repeated the mean’s appearance: its earlier movement is substantial, and the latest segment alone does not show whether the level will persist.
The appropriate conclusion is not that the mean’s apparent settling is false, but that it is insufficient for an overall burn-in decision. The analyst should continue the run or examine additional later segments and relevant summaries before deciding which iterations to retain. The mean and percentile have different units only if the original variable is not on a single common scale; here both are measured in the variable’s units.
Making a Defensible Burn-In Decision
A practical decision combines the trace evidence with a clear record of what was examined. The following sequence encourages a cautious choice without turning a visual judgment into a proof claim.
Include summaries that could reveal meaningful changes in the imputed values, not only a convenient mean.
Compare both their levels and their within-segment ranges. Look for a startup shift that becomes smaller or disappears in later segments.
Place it after the apparent transient, and check whether the segments after the cutoff support a similar interpretation across the selected summaries.
Compare retained summaries under plausible nearby burn-in choices. Large changes indicate that the choice still matters and should not be presented as settled.
Report the cutoff as a practical diagnostic choice supported by the traces, not as a guarantee that the chain has converged.
Worked Example: Checking Sensitivity to the Cutoff
Consider two hypothetical chains for the same imputed-variable mean, measured in score points. Their averages over five 20-iteration segments are shown below.
| Chain | Iterations 1–20 | 21–40 | 41–60 | 61–80 | 81–100 |
|---|---|---|---|---|---|
| A | 14.0 | 9.0 | 6.4 | 6.1 | 6.0 |
| B | 4.0 | 5.4 | 5.8 | 5.9 | 6.1 |
Chain A begins high and falls, while chain B begins low and rises. The segment averages are more alike in later blocks, but a numerical sensitivity check is still useful. With a 20-iteration burn-in, the retained average across both chains is the average of the eight remaining block averages:
With a 40-iteration burn-in, the six retained block averages give \( (6.4+6.1+6.0+5.8+5.9+6.1)/6=36.3/6=6.05 \) score points. With a 60-iteration burn-in, the four remaining blocks give \( (6.1+6.0+5.9+6.1)/4=24.1/4=6.025 \) score points.
The retained average changes by 0.2875 score points between the 20- and 40-iteration cutoffs, but by only 0.025 score points between the 40- and 60-iteration cutoffs. These results suggest that retaining the first 20 iterations has a noticeable effect and that later choices are less sensitive for this particular summary. A cautious report could use 60 as a provisional burn-in for this trace, while noting that only 40 iterations per chain remain and that other summaries and further iterations are needed for a broader assessment.
Assumptions, Alternatives, and Failure Modes
A burn-in decision assumes that the chain summaries are meaningful for the imputed variables being assessed and that the run is long enough to display both its early behavior and a usable later portion. The MICE algorithm’s successive updates generate dependent iterations; the observed trace is therefore a finite, noisy record, not a direct view of an underlying distribution. The interpretation also depends on the chosen summaries and on the starting state. A trace cannot by itself establish that the imputation models are appropriate or that all relevant aspects of the chain have stabilized.
Segment comparison is more informative than choosing a fixed number of iterations simply because it is familiar, but segment choice is not the only possible approach. A visual trace can reveal a pattern that a table of segment averages obscures; the table makes the magnitude of changes easier to compare. Examining retained summaries under more than one plausible cutoff shows how sensitive the result is to the decision. These approaches complement each other. None converts stabilization into proof, and a formal diagnostic used elsewhere should not be treated as a substitute for understanding the trace being assessed.
Several failures deserve attention. A plot with a narrow vertical axis may make small fluctuations look dramatic, while a wide axis can conceal a practically important shift. Very short segments produce noisy averages; very long segments can conceal a transient within a block. A trace that looks calm over its final few iterations may simply be too short to reveal continued movement. And settling in a single summary can conceal changes in another, as the preceding example illustrates. When the evidence is ambiguous, the defensible response is to extend the run or qualify the decision, rather than to claim a precise cutoff has been proven.
Common Mistake
Check Your Understanding
Answer each question using the definitions and reasoning in this tutorial. Brief solutions follow the questions.
- A chain-summary segment average changes from 7.8 to 7.7 units across two late segments. What can this support, and what can it not establish?
- Why is a trace of the imputed mean alone not always enough to choose a burn-in?
- For \(T=100\) iterations, suppose the first 20 have mean summary 12 and the remaining 80 have mean summary 7. Calculate the full-run average and explain the effect of retaining the first 20.
- A trace has a large early shift, but its final segment contains only five iterations. What is a sensible limitation to report or next step to take?
Solutions. (1) The 0.1-unit difference is consistent with similar late segment levels for this summary. It does not prove convergence or show that other summaries have stabilized. (2) Other summaries, such as a quantile or dispersion measure, may continue to move even when the mean appears stable; assess summaries relevant to the analysis. (3) The full-run average is \( (20\times12+80\times7)/100=(240+560)/100=8 \) units. Retaining the first 20 raises the average from the later-segment mean of 7 to 8 units because the early segment has a higher mean and contributes one fifth of the iterations. (4) Five iterations provide little evidence about the sustained behavior of the late trace. State that the stabilization assessment is uncertain and, if feasible, continue the run and compare additional later segments.