Why Look for Stickiness and Abrupt Switching?
A MICE trace records a summary of the imputations as the algorithm iterates. The summary might be an imputed mean, a dispersion measure, or a quantile; the choice depends on the variable and diagnostic question. A trace is useful not only for spotting a sustained trend, but also for seeing whether movement is continuous, nearly absent, or concentrated in sudden changes.
These patterns matter because a chain that barely changes may be exploring only a small part of the relevant range. A chain that jumps between distinct levels may be moving between regions, but it may also reflect instability in a conditional model or an initialization effect. The visual pattern is a reason to investigate, not a diagnosis by itself: the trace shows a summary, not the full state of the imputation process.
The definition is deliberately based on the chosen summary. A trace of imputed means can appear stable while a tail quantile moves, or vice versa. The summary, its scale, the iterations included, and any tolerance used to call a change “small” should therefore be made explicit. The interpretations below assume that the same summary is calculated consistently at every iteration and that the trace is not being distorted by inconsistent recording or changing definitions.
Describe the Pattern Before Explaining It
Start by describing what is visible: how long a nearly unchanged run lasts, how large a jump is compared with ordinary step-to-step changes, and whether movement is spread through the trace or concentrated in short periods. These descriptions are more defensible than immediately naming a cause.
For numerical summaries, define the consecutive change as \(\Delta_t=q_t-q_{t-1}\). Its absolute value, \(|\Delta_t|\), describes the size of a step without regard to direction. A plateau can be operationally defined by selecting a tolerance \(\delta\) and counting consecutive steps with \(|\Delta_t|\leq\delta\). The tolerance should reflect the scale and practical meaning of the summary, not be chosen after seeing a result simply to produce a desired label. There is no universal \(\delta\) that identifies a faulty chain for every variable.
A rough scale comparison can help flag a potential jump. For example, compare the largest \(|\Delta_t|\) with the typical absolute step size in the same trace. This comparison is a descriptive aid, not a formal test with a universal cutoff. If most steps are zero, the typical step may itself be zero; in that case, report the jump in the summary’s original units and inspect the sequence rather than relying on a ratio with a zero denominator.
Worked Examples
Worked Example: A Long Plateau in an Imputed Mean
Suppose a hypothetical trace records the imputed mean of a continuous variable, measured in milligrams per liter, for seven consecutive iterations: 12.0, 12.1, 12.0, 12.1, 12.0, 12.0, and 12.1 mg/L. For this illustration, call a step small when its absolute size is at most \(\delta=0.15\) mg/L.
The six consecutive changes are \(0.1,-0.1,0.1,-0.1,0.0,\) and \(0.1\) mg/L. For example, the first is \(12.1-12.0=0.1\) mg/L, and the fifth is \(12.0-12.0=0.0\) mg/L. Each satisfies \(|\Delta_t|\leq0.15\) mg/L, so the full seven-point run meets this chosen plateau rule.
This is a plateau in the imputed-mean trace, not proof that the entire imputation state is frozen. The mean may remain nearly constant while individual imputed values or another summary change. If many relevant summaries are similarly immobile, poor exploration becomes more plausible; if this mean alone is stable, the behavior may simply be consistent with a stable location.
Worked Example: An Abrupt and Persistent Level Change
Consider a hypothetical trace of an imputed mean in units of points: 4.8, 5.0, 4.9, 5.1, 8.7, 8.9, 8.8, and 8.7. The first three changes are \(0.2,-0.1,\) and \(0.2\) points. The change from 5.1 to 8.7 is \(8.7-5.1=3.6\) points, far larger than those preceding steps. The next three changes are \(0.2,-0.1,\) and \(-0.1\) points.
The sequence shows a large upward switch followed by values that remain near 8.8 points. That pattern differs from a single isolated spike, which would quickly return to the earlier level. Plausible explanations include a chain moving into a different region, a transition after initialization, or an instability in a conditional update. The trace alone cannot distinguish these explanations.
The appropriate next step is to check where the change falls in the iteration history and whether the affected variable’s update could reasonably produce it. A persistent shift after an initial transient may have a different interpretation from a shift that recurs unpredictably throughout the run. Neither observation alone establishes that the chain has adequately explored its target distribution.
Worked Example: Intermittent Movement in Short Bursts
A hypothetical trace of an imputed median, in hours, is 2.0, 2.0, 2.1, 2.0, 2.0, 2.0, 3.4, 3.5, 2.0, 2.1, 2.0, and 2.0 hours. Let a movement be “appreciable” for this example if \(|\Delta_t|>0.5\) hours.
There are eleven changes. The change from 2.0 to 3.4 hours is \(1.4\) hours, and the change from 3.5 back to 2.0 hours is \(-1.5\) hours. Both exceed the 0.5-hour threshold in absolute size. The other nine changes have absolute sizes of 0.1 hours or less. Thus, \(2/11\), or about 18.2%, of the steps are appreciable, and both occur around one short excursion.
This is intermittent movement: most of the trace is nearly flat, interrupted by a brief excursion. It differs from the persistent level change in the previous example because the summary returns to its earlier neighborhood. Possible explanations include occasional movement into a different region followed by return, an unusual update, or a feature of the summary that responds to a small number of imputed values. Inspect the affected update and other relevant summaries before deciding whether the excursion signals poor exploration.
Why a Flat Trace Can Mislead
There is an important distinction between a stable summary and a chain that is failing to explore. If the relevant distribution is concentrated in a narrow region, a summary may vary little even when the chain is behaving appropriately. Conversely, a constant summary can conceal movement in other parts of the imputed data. The same caution applies to jumps: a large change may represent genuine movement rather than a computational failure.
A brief justification for this caution follows from the fact that \(q_t\) is a function of the chain state, not the state itself. Write the state at iteration \(t\) as \(S_t\), and the recorded summary as \(q_t=g(S_t)\). Different states can have the same summary if \(g\) discards information; therefore, observing \(q_t=q_{t-1}\) does not imply \(S_t=S_{t-1}\). Likewise, a small change in \(q_t\) does not rule out a large change in features not captured by \(g\).
This is why diagnostic claims should be scoped to the trace being examined. A plateau in an imputed-mean trace means the imputed mean changed little over that run. It does not establish that every imputed value, every conditional model parameter, or every other variable’s trace was unchanged.
Assumptions, Diagnostics, and Follow-Up
Interpreting these patterns as evidence about exploration requires several practical conditions. Iterations must be in their proper order and recorded at equal intervals; the same summary must be computed in the same way throughout; and the scale must be interpretable. To treat a plateau as a warning about inadequate exploration, the chain should also be in a period where a stable target distribution is expected. A movement during an initial transient may instead reflect settling from initialization, so its timing matters.
Confirm the iteration order, the summary definition, and the units. Check that missing or duplicated records have not created artificial flat runs or jumps.
Record the plateau length, the sizes of the largest changes, and whether large changes persist at a new level or return to the earlier range.
Relate the pattern to the variable and its conditional update. Consider whether sparse information, rare categories, unstable fitting, or donor-selection behavior could make movement irregular.
Inspect other relevant summaries and the update behavior. If a plausible computational or model issue is found, address it and rerun the diagnostic rather than treating a longer trace alone as a guaranteed remedy.
The earlier material on detecting drift and slow mixing concerns sustained movement or sluggish exploration over time. Stickiness and abrupt switching add different questions: is movement unusually scarce for long stretches, or concentrated at a few large transitions? These descriptions can coexist. A chain can drift slowly and also contain plateaus, or switch abruptly and then remain sticky.
Observed-data diagnostics answer a related but different question. They can reveal implausible imputations or discrepancies between observed and imputed values, but a plausible-looking completed-data distribution does not demonstrate that the chain explored well. Likewise, a trace pattern does not by itself establish that the imputation model fits the observed data. Use the diagnostic suited to the question instead of treating one check as a substitute for all others.
Common Mistakes and Limitations
| Misleading conclusion | Correction |
|---|---|
| “The trace is flat, so the chain is converged.” | A flat trace can reflect stability or poor movement. It is not sufficient evidence of adequate exploration. |
| “Every large jump means the algorithm failed.” | A jump may be legitimate movement or a transient. Check its timing, persistence, and relation to the update. |
| “The mean barely changed, so the imputed data barely changed.” | A mean can hide changes in spread, tails, or individual values. Interpret only the summary actually plotted. |
| “A numerical threshold gives a universal pass/fail rule.” | Tolerances depend on scale and purpose. Use them to describe and compare behavior, not as universal convergence tests. |
The central limitation is that these are visual and descriptive diagnostics. They help identify where to investigate, but do not provide a complete mathematical guarantee of convergence or adequate exploration. A trace may look quiet because the summary is insensitive, and a noisy trace may be compatible with appropriate stochastic movement. Conclusions should remain proportionate to the evidence.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A trace summary changes by at most 0.02 units for 20 consecutive steps. What can you conclude, and what can you not conclude?
- A summary changes from 6.1 to 9.0 units and then remains near 9.0. Name the pattern and give two possible explanations.
- In a trace of 15 steps, three steps exceed a chosen movement threshold, all within one short interval. What pattern does this suggest?
- Why does an unchanged imputed mean not prove that the full chain state is unchanged?
- What should be checked before interpreting a plateau as a sign of poor exploration?
Solutions. (1) The trace has a long plateau relative to the stated scale or tolerance. This establishes only that the recorded summary changed little; it does not prove convergence or that the full imputed state was static. (2) This is an abrupt, persistent level change. Possible explanations include movement to a different region or an initial transient; instability in an update is another possibility to investigate. (3) This suggests intermittent movement concentrated in a burst, though the threshold is a descriptive choice rather than a universal test. (4) The mean is a function of the chain state and can discard information: distinct states can have the same mean. (5) Verify the iteration order, consistent summary calculation and units, the timing of the run, and whether the summary is appropriate for the variable; then inspect other relevant summaries and the conditional update for corroborating evidence.