Why Overlay Chains?
A trace plot follows a chosen summary of an imputed variable across iterations. Examining one chain can reveal changes over time, but it cannot show whether other chains visit a similar range of values. Overlaying chains makes that comparison visible: persistent separation may signal that chains have not reached comparable behavior, while overlap provides some evidence that they explore a shared range.
An overlay is only informative if the plotted quantities are comparable. Each line must represent the same summary, computed at corresponding iterations and expressed in the same units. All chains must also use the same vertical axis. Without these conditions, apparent agreement or disagreement can be an artifact of the plot rather than a feature of the chains.
For example, \(S_{m,t}\) might be the mean of the imputed values for a continuous variable at iteration \(t\), or the proportion imputed into a particular category. The earlier material on forming chain summaries provides the basic idea: the line is a sequence of summaries, not a display of every individual imputed value. Here the concern is how to compare those sequences across chains.
Conditions for a Fair Visual Comparison
Before interpreting an overlay, check what each line represents and how it was drawn. The chains should be generated under the same MICE specification: the same variables, conditional models, predictor setup, and imputation settings. Their initial states may differ, which is useful for seeing whether different starting points lead to similar behavior. The plotted iteration ranges should be aligned, and any omitted initial iterations should be omitted consistently.
The summary must also be meaningful for the variable. A mean can be informative for a continuous variable, but a proportion or category count may be more appropriate for a categorical variable. Whichever summary is chosen, the definition must not change between chains. If a mean is plotted for one chain and a median for another, their similar numerical values do not establish similar chain behavior.
Use one set of axes with the same limits and units for every chain. Do not rescale each chain separately to make its own fluctuations fill the plotting region. A common vertical scale lets the reader compare both the location and the range of the summaries. Shared axes do not require the plotted lines to occupy the same amount of space; large differences are precisely what the overlay should reveal.
State which statistic is plotted and what observations it summarizes, such as the mean of the imputed values for one variable.
Confirm that every chain uses the same MICE specification, summary definition, iteration indexing, and displayed iteration range.
Use distinct line styles or colors, a readable legend, and one vertical scale in the summary's units.
Look for persistent separation, convergence toward a shared range, or markedly different movement. Treat visual agreement as evidence to assess, not a guarantee.
What Common Axes Preserve
A shared vertical scale matters because it preserves the numerical relationship between chains. Suppose all chains are plotted using the same linear transformation \(Y=a+bS\), with \(b>0\). Then a difference in summaries is represented by
Thus, the sign of the difference and its size relative to other differences are preserved, apart from the same multiplier \(b\) for all chains. But if each chain is independently rescaled, chain \(m\) may instead be plotted as \(Y_{m,t}=a_m+b_mS_{m,t}\). Differences in location and spread can then be hidden. For example, separate vertical limits can make two far-apart chains appear to fluctuate over nearly identical heights. Common axes are therefore part of the comparison, not a cosmetic preference.
Worked Example: Detecting Persistent Separation
Suppose two hypothetical MICE chains produce the following successive means of the imputed values for a continuous outcome, measured in milligrams. The same mean is calculated at each iteration in both chains.
| Iteration | Chain 1 mean (mg) | Chain 2 mean (mg) |
|---|---|---|
| 1 | 10 | 18 |
| 2 | 12 | 17 |
| 3 | 11 | 19 |
| 4 | 13 | 18 |
| 5 | 12 | 20 |
The chain means are
Chain 1 ranges from 10 to 13 mg, whereas chain 2 ranges from 17 to 20 mg. The ranges do not overlap, and the difference between their average plotted summaries is \(18.4-11.6=6.8\) mg. With common axes, the lines would appear persistently separated. This is a reason to investigate the chains, their starting states, and the MICE specification; it is not by itself proof of a particular failure or explanation. The hypothetical values are illustrative, not results from a real study.
Reading Overlap Without Overclaiming
When lines overlap, the corresponding summaries occupy similar numerical ranges at many displayed iterations. This is useful visual evidence that the chains are not persistently separated on that summary. It does not establish that the chains have explored all relevant states, that every part of the joint distribution is represented, or that the plotted summary is adequate for the scientific analysis. A short trace can overlap by chance, and a summary can conceal differences in other features of the imputations.
Likewise, crossing lines do not automatically indicate good mixing. Chains can cross while remaining highly variable in location over time, or while each remains in a different region for long stretches. Interpret the overall pattern: whether chains occupy a broadly comparable range, whether one stays apart, and whether differences persist through the displayed iterations. The earlier material on drift, slow mixing, and sticky or abrupt switching addresses time patterns within chains; overlaying adds the comparison between chains.
The visual assessment assumes that each chain is being used as a plausible realization of the same intended MICE process and that the plotted summaries are calculated consistently. If chains use different conditional models or different summary definitions, a line-to-line comparison is not a diagnostic of mixing. Even with a common setup, an overlay is a qualitative diagnostic, not a formal convergence test.
Worked Example: Overlap as Limited Evidence
Consider two hypothetical chains whose successive means, in grams, are:
| Iteration | Chain 1 mean (g) | Chain 2 mean (g) |
|---|---|---|
| 1 | 4.8 | 5.1 |
| 2 | 5.2 | 4.9 |
| 3 | 4.9 | 5.2 |
| 4 | 5.1 | 4.8 |
| 5 | 5.0 | 5.0 |
Both chain means equal \(5.0\) g:
Each chain ranges from 4.8 to 5.2 g, so their plotted values overlap throughout this short example. This supports the limited conclusion that the two chains' means occupy a similar range over these five iterations. It does not show that five iterations are sufficient, rule out sticky behavior, or establish that other summaries behave similarly. Those questions require broader diagnostics and adequate chain runs.
Matched Summaries and Misleading Comparisons
A trace overlay compares like with like. In a hypothetical analysis, suppose one line shows the imputed-value mean in days and another shows the imputed-value median in days. Both use days, but they are different statistics. Their numerical proximity cannot be interpreted as agreement between chains on the same chain-level summary. Recalculate the same statistic for each chain before overlaying.
The same principle applies across variables: do not overlay summaries that refer to different variables merely because their values happen to be on similar numerical scales. If a variable is transformed, state whether the plotted summary is on the transformed or original scale, and apply the same transformation in every chain. These checks prevent a visually neat graph from making an invalid comparison.
Worked Example: Correcting an Inconsistent Overlay
Suppose an analyst reports that chain 1 has a mean of 12 days at a particular iteration, while chain 2 has a median of 11 days at that iteration. The analyst plots these as two points and concludes that the chains agree because the values differ by only 1 day. That conclusion is unsupported: the plotted points are different statistics.
Assume the analyst recalculates the mean for both chains at three matched iterations. The hypothetical values are 12, 13, and 11 days for chain 1 and 17, 16, and 18 days for chain 2. The matched averages are
The mean summaries differ by \(17-12=5\) days, and the ranges, 11–13 days and 16–18 days, do not overlap. A common-axis overlay of these matched means would show separation, rather than the misleading near-agreement implied by comparing a mean with a median. The hypothetical values illustrate the correction; they are not real study data.
Alternatives, Diagnostics, and Limits
| Display or approach | What it helps assess | Main limitation |
|---|---|---|
| One chain's trace plot | Changes in that chain's summary over iterations | Cannot compare chain locations or ranges |
| Overlaid traces with common axes | Between-chain overlap and separation, alongside time patterns | Visual and summary-specific; does not prove convergence |
| Separate panels with independently chosen vertical limits | Within-chain shapes when values differ greatly | Can conceal numerical differences between chains |
| Numerical convergence diagnostics | Quantitative evidence under their own assumptions | Cannot replace checking model specification and the plotted process |
Use separate panels when lines obscure one another or a large difference makes smaller movements hard to see, but preserve identical vertical limits if direct comparison is intended. If panels use different limits for readability, make that choice explicit and do not infer between-chain overlap from the visual heights alone. Numerical convergence diagnostics complement trace plots; they answer questions that a visual overlay cannot settle.
Several failure modes deserve attention. Persistent separation can reflect inadequate exploration, different initial states that have not yet become comparable, or incompatible chain settings. A trace that occupies a narrow range can reflect genuine stability or a chain that is barely moving. A common axis can also make small fluctuations hard to see when one chain is far from another; this calls for careful inspection, not independent rescaling that hides separation. Finally, agreement on one summary does not guarantee agreement in other aspects of the imputed data.
Common mistake: “The two traces overlap after I set each chain's vertical axis separately, so the chains mix well.” Independent scales remove the numerical basis for comparing locations and spreads. Replot the same summary for every chain using one shared axis, then describe the evidence cautiously. Overlap supports comparability on the displayed summary; it does not prove convergence or validate the imputation model.
Check Your Understanding
Answer each question, then check the concise solutions below.
- What must be held constant across chains for a matched overlay of an imputed-value mean?
- Two chains occupy nonoverlapping ranges on a common-axis trace plot. What is a defensible conclusion, and what conclusion would go too far?
- Why can independently rescaling each chain's vertical axis create a misleading impression of overlap?
- At one iteration, chain 1's imputed-value mean is 8 units and chain 2's is 10 units. What is the numerical difference, and what additional information is needed to interpret the full chain comparison?
Solutions. (1) Use the same mean definition, variable, units, corresponding iteration range, and common vertical scale, with chains generated under the same MICE specification. (2) Nonoverlapping ranges are evidence of persistent separation for the displayed summary and warrant investigation; they do not alone prove why separation occurred or establish a definitive convergence failure. (3) Separate limits apply different transformations to each chain, so the plotted heights no longer preserve between-chain numerical differences. (4) The summaries differ by \(10-8=2\) units, with chain 2 higher at that iteration. To assess the chain comparison, inspect matched summaries over iterations on common axes and consider the broader diagnostic evidence; one pair of values is insufficient.