What an FCS Trace Plot Shows
An FCS algorithm updates the missing values in one variable at a time, cycling through the variables. A trace plot displays a selected value or summary of the completed data after each full cycle. Its purpose is to show how that quantity changes as the chain proceeds: a sustained movement may indicate that the chain is still drifting, while continued fluctuation around a stable level is more consistent with apparent stabilization.
Let \(Z^{(k)}\) denote the completed data state after the \(k\)th full FCS iteration, and let \(k=1,\ldots,K\) index iterations. A trace is the sequence \(S_k=g(Z^{(k)})\), plotted against \(k\). The function \(g\) might select one imputed cell, count imputed events, or calculate a mean across a specified set of imputed cells. A trace is therefore not a plot of the original observed records against time; its horizontal axis is algorithm iteration.
The set being summarized should be explicit. For example, a mean can be calculated over the cells missing for one variable, holding that set of cells fixed across iterations. Observed values are not changing imputations and should not be mixed into that mean unless the diagnostic is deliberately defined to include them. A clear axis label gives both the summary and its units, such as “mean imputed systolic pressure (mmHg)” or “proportion of imputed outcomes equal to 1.”
Choosing What to Trace
A single imputed cell can reveal whether that particular chain component moves, but its path may be unusually noisy or unrepresentative. A summary across several missing cells can make broader movement easier to see, though it can also hide changes that cancel one another. These traces answer different questions, so a useful diagnostic often includes more than one quantity.
- Selected continuous imputed values: trace one or a small number of missing cells in meaningful units. Choose cells that are substantively important or that represent different patterns of observed predictors; do not assume one cell represents all imputations.
- Summary of continuous imputations: trace, for example, the mean or median across a fixed set of missing cells. A mean can respond to overall movement but may be affected by extreme imputations.
- Selected categorical imputation: trace a 0/1 indicator for whether a chosen cell is in a specified category. The path will jump between values, so repeated switching is not itself evidence of a problem.
- Summary of categorical imputations: trace the proportion in a specified category among a fixed set of missing cells. This gives a bounded, interpretable summary, but changes in different subgroups can be obscured by an overall proportion.
The algorithm and data should be held fixed while interpreting a trace: changes should reflect the sequence of FCS updates, not a changed model, predictor set, or summary definition. The plotting scale matters as well. A narrow vertical scale can make small fluctuations look dramatic; a wide scale can conceal modest but persistent movement. Label the scale and units, and consider the range in relation to meaningful changes in the variable.
Worked Example: A Rising Trace of an Imputed Mean
Suppose a hypothetical FCS run produces a mean imputed value for a fixed set of missing measurements, in millimetres of mercury (mmHg). The first eight trace values are:
| Iteration \(k\) | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 |
|---|---|---|---|---|---|---|---|---|
| Mean imputed value (mmHg) | 12.1 | 12.8 | 13.0 | 13.7 | 14.2 | 14.8 | 15.1 | 15.9 |
The mean of the first four values is \((12.1+12.8+13.0+13.7)/4=51.6/4=12.9\) mmHg. The mean of the last four is \((14.2+14.8+15.1+15.9)/4=60.0/4=15.0\) mmHg. The sequence rises at every iteration, and the later four-value average is 2.1 mmHg higher than the earlier one. This trace gives a clear warning of persistent upward movement over the displayed iterations.
It does not identify why the mean is rising, nor does it prove that the algorithm will never stabilize. The appropriate conclusion is limited: this summary has not shown apparent stabilization over these eight iterations, so a conclusion based on a stationary-looking chain would be premature. A trace over more iterations and other relevant summaries can help assess whether the pattern persists.
Reading Drift and Apparent Stabilization
A trace is mainly a visual diagnostic, not a formal pass-or-fail test. Drift is sustained movement in the level of a summary over iterations. It may appear as a gradual increase or decrease, a sequence of long runs on one side of a typical level, or a change in the range of values. A persistent trend is more concerning than a short excursion followed by a return to the previous range.
A trace may appear stabilized when it fluctuates around a broadly similar level and range without a sustained directional pattern. “Apparent” is essential: a finite plot cannot show that the process has reached its long-run behavior. In particular, the earlier material on Autocorrelation Between FCS Iterations is relevant because neighboring points can be related. A cluster of similar values is not necessarily a sequence of independent confirmations.
A short justification explains why traces can be informative. If the chain has reached a stationary regime and \(S_k=g(Z^{(k)})\) has a finite expectation, then the distribution of \(S_k\) does not change with \(k\). In particular, its expected value is constant across iterations. A sustained change in the level of the trace is therefore inconsistent with what one would expect from a stationary regime, although random fluctuation can imitate a trend in a finite run. Conversely, a roughly constant trace cannot prove stationarity: the chosen summary may fail to reveal movement elsewhere in the completed data.
This justification does not require that the FCS conditionals come from a compatible joint distribution. A chain can show stable behavior for selected summaries even when compatibility is not established. Such a trace cannot certify compatibility or the validity of a joint target distribution; it describes observed behavior of the specified algorithm.
Worked Example: Tracing a Binary Imputation Summary
In a hypothetical data set, 40 records have a missing binary outcome. For each iteration, define \(S_k\) as the proportion of those 40 imputed outcomes equal to 1. The first five iterations produce counts of 17, 21, 19, 20, and 18 outcomes equal to 1.
The corresponding trace values are \(17/40=0.425\), \(21/40=0.525\), \(19/40=0.475\), \(20/40=0.500\), and \(18/40=0.450\). Thus the plotted proportions are 0.425, 0.525, 0.475, 0.500, and 0.450. They move up and down rather than steadily increasing or decreasing over these five iterations.
This small sequence does not show an obvious sustained trend, but five points provide little basis for judging a longer chain. The summary also does not say which records changed category. A trace of a selected record can complement it: for a chosen missing outcome, plot \(I_k=1\) when its imputed value is 1 and \(I_k=0\) otherwise. That individual trace will be discrete and potentially jumpy; its interpretation should be paired with summaries across records rather than treated as a smooth-valued measurement.
Practical Construction and Checks
Constructing a useful trace requires more than plotting any available column of software output. Define the quantity first, calculate it consistently at each completed iteration, then inspect both its level and its range. The following sequence keeps the diagnostic interpretable.
Name the variable, cells or fixed set of cells, calculation, and units. For a categorical summary, identify the category whose proportion is plotted.
Use the completed-data state after the same point in each FCS cycle. Do not compare partial updates from one iteration with completed updates from another.
Put iteration on the horizontal axis and the summary, with units where applicable, on the vertical axis. Make sure the scale does not exaggerate or hide meaningful changes.
Look for direction, persistent runs, shifts in level, and changes in spread. Compare selected-cell traces with summaries when both are informative.
Describe what the plotted summary did over the displayed iterations. Do not turn a visually calm trace into proof of convergence, compatibility, or adequate exploration of every feature.
Worked Example: Fluctuation Around a Similar Level
A hypothetical chain produces the following mean imputed ages, in years, for the same set of missing cells:
| Iteration \(k\) | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Mean age (years) | 6.1 | 5.7 | 6.4 | 5.9 | 6.2 | 6.0 | 6.3 | 5.8 | 6.1 | 6.0 | 6.2 | 5.9 |
The first four values average \((6.1+5.7+6.4+5.9)/4=24.1/4=6.025\) years. The last four average \((6.1+6.0+6.2+5.9)/4=24.2/4=6.05\) years. The values fluctuate between 5.7 and 6.4 years, while these two block averages differ by only 0.025 years. The displayed trace has no clear sustained increase or decrease and appears to fluctuate around a similar level.
The warranted conclusion is not that the chain has converged. It is that this particular mean, on this particular run and over these 12 iterations, shows no obvious persistent trend. A single average could conceal movement in some imputations offset by movement in others, so other selected values or summaries may be useful.
Alternatives, Failure Modes, and Limits
Trace plots are most useful as a direct view of change over iteration, but they are not the only diagnostic. A numerical summary by iteration blocks can make a level shift easier to describe, though reducing a trace to a few averages discards its pattern. A marginal distribution plot can show the range and shape of imputed values but removes their iteration order, so it may conceal drift. Autocorrelation summaries address dependence between neighboring iterations, a different feature from a persistent trend. These approaches complement rather than replace one another.
Several failures deserve particular attention:
- Tracing only one convenient quantity. A calm trace for one mean does not show that every variable, subgroup, or conditional relationship is stable. Use summaries chosen to reflect the features relevant to the analysis.
- Changing the summary as the chain proceeds. For example, using a different set of missing records in later iterations can create apparent movement unrelated to changes in those same imputations. Keep the definition fixed.
- Over-interpreting noise. An isolated spike is not automatically persistent drift. Check whether movement continues and whether other relevant summaries show a similar pattern.
- Over-interpreting smoothness. A visually quiet trace can reflect a summary that is insensitive to important changes. Stability of a plotted mean does not establish stability of the full imputed data state.
- Treating iteration as study time. Iteration counts algorithm steps, not days, follow-up visits, or a natural time variable. A rising trace describes the chain’s behavior, not a real-world increase in the outcome.
The principal limitation is that trace plots are qualitative and summary-dependent. They can reveal obvious problems, but they cannot prove that a chain has reached a stationary distribution, certify compatibility of the conditional models, or guarantee that the imputations are substantively appropriate. A stable-looking trace should be considered alongside the model specification and other diagnostics, not used as a stand-alone certificate.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- In an FCS trace plot, what does the horizontal axis represent, and what should a plotted point represent?
- A trace summary for 30 missing binary outcomes is the proportion imputed as 1. At one iteration, 12 outcomes are imputed as 1. Calculate the trace value.
- A selected continuous imputation rises for three iterations, then returns to its previous range. Is this by itself evidence of persistent drift? Explain.
- Why might a trace of the mean across all missing cells fail to reveal movement in particular records?
- What is the strongest conclusion justified by a trace that fluctuates around a similar level without an obvious trend?
Solutions. (1) The horizontal axis is FCS iteration; each point is the consistently defined imputed value or summary after that iteration. (2) The proportion is \(12/30=0.400\), or 40.0%. (3) Not by itself. A brief rise followed by a return is an excursion, not necessarily sustained movement; inspect the continuing pattern and other relevant traces. (4) Individual values can move in opposite directions and offset in the mean, leaving the overall average nearly unchanged. (5) The chosen summary shows no obvious persistent trend over the iterations displayed and appears stable there; this does not prove convergence, compatibility, or stability of all chain features.