Why Track Category Proportions?
For a multilevel categorical variable, a single summary such as the proportion in one category can reveal changes in the imputed data as a MICE chain progresses. But a multilevel variable has several categories, and their proportions are linked: an increase in one category must be balanced by decreases elsewhere. Tracking the categories together can therefore show patterns that a single binary summary would miss, such as a rare category appearing intermittently or different chains favoring different parts of the category distribution.
The purpose is diagnostic. A trace plot records how summaries of the current imputed values move across iterations and chains. A steady-looking set of category proportions is evidence to examine, not proof that the chains have converged, that the conditional models are appropriate, or that the imputations are valid.
Constructing the Trace Correctly
The definition requires a consistent denominator and category scheme. For one variable, the denominator is ordinarily the number of its missing values being imputed, not the full sample size and not the number of observed values. Keep the same set of missing records, category labels, and category ordering for every iteration and chain being compared. If these change, a movement in the plotted proportion may reflect a changed calculation rather than chain behavior.
At each iteration, every imputed record belongs to exactly one of the \(K\) categories. Consequently, the category counts add to \(m\), and the proportions add to one:
This constraint matters for interpretation. If one category’s proportion rises, one or more other categories’ proportions must fall in total. The traces are not independent measurements. For a nominal variable, categories have no natural ordering, so do not average numerical codes such as 1, 2, and 3 as though they were measurements. For an ordinal variable, the order is meaningful, but each category proportion still describes the fraction assigned to that level; it is not a numeric mean.
Worked Example: Computing a Three-Category Trace
Suppose a hypothetical survey variable has categories A, B, and C, and \(m=40\) records are missing this variable. In one chain, the category counts at three successive iterations are as follows.
| Iteration | A count | B count | C count |
|---|---|---|---|
| 1 | 18 | 14 | 8 |
| 2 | 16 | 16 | 8 |
| 3 | 17 | 15 | 8 |
At iteration 1, the proportions are \(18/40=0.450\), \(14/40=0.350\), and \(8/40=0.200\). They sum to \(1.000\). At iteration 2 they are \(16/40=0.400\), \(16/40=0.400\), and \(8/40=0.200\), again summing to \(1.000\). At iteration 3 they are \(17/40=0.425\), \(15/40=0.375\), and \(8/40=0.200\).
A trace plot would show A moving from 0.450 to 0.400 to 0.425, B moving from 0.350 to 0.400 to 0.375, and C remaining at 0.200. These numbers describe the current category allocation among the 40 imputed records. They do not reveal which particular records changed categories between iterations.
Resolution and a Short Justification
With \(m\) imputed values, a category proportion can change only in increments of \(1/m\), because its count is an integer. This discrete resolution explains why traces can look jumpy even when the underlying changes in counts are small.
To see why, write the count in category \(k\) as an integer \(N_{ctk}=j\). Substitution in the definition gives \(\hat{p}_{ctk}=j/m\). If the count changes by one, the difference between proportions is
Thus, when only a few values are imputed, each count change produces a relatively large visible step. With many imputed values the steps are smaller, though the trace may still be noisy. The formula describes the granularity of the summary; it does not give the probability that a chain has converged.
Worked Example: Reading a Rare-Category Trace
In a hypothetical data set, \(m=20\) records are missing a four-category variable. Category D is rare in the imputed values. Its counts over five iterations in one chain are 0, 1, 2, 1, and 1.
The corresponding proportions are \(0/20=0.000\), \(1/20=0.050\), \(2/20=0.100\), \(1/20=0.050\), and \(1/20=0.050\). Each one-count step is \(1/20=0.050\), or 5 percentage points. The move from 0.050 to 0.100 is a net increase of one in the category count. The two counts alone do not establish how many individual records changed categories: records could have entered and left D between iterations.
This short trace does not establish a problem by itself. It does show that the estimated proportion for D is coarse at this denominator and that the category is intermittently absent. The appropriate next checks include whether other chains show similar behavior, whether the pattern persists over a longer run, and whether the rare category is supported and represented appropriately in the imputation model. A trace with only a few iterations cannot distinguish ordinary stochastic fluctuation from a persistent chain problem.
Reading Patterns Across Chains
Use the same categories and denominator definition when overlaying chains. First consider whether each chain’s traces settle into a broadly stable range after any initial transient behavior. Then compare those ranges and their movement across chains. Persistent separation in one or more category proportions may signal that chains are exploring different regions or mixing poorly. It warrants investigation rather than an automatic conclusion that a particular chain is wrong.
Because proportions sum to one, inspect the category traces jointly. For example, if one chain has a higher proportion in category A, it must have a lower combined proportion in the remaining categories. A difference in A alone does not identify which other category accounts for the balance. For ordinal variables, adjacent category traces may move in compensating ways; retain the category labels and their order rather than replacing the whole distribution with an average category code.
A useful diagnostic sequence is:
Confirm the same missing records, denominator, category coding, and iteration definition are used throughout.
At every chain and iteration, counts must total \(m\) and category proportions must total one, subject only to displayed rounding.
Look for continuing drift, long periods at a boundary, abrupt changes, or a rare-category trace that repeatedly disappears and reappears.
Look for sustained differences in ranges or patterns, not merely a momentary separation at one iteration.
Use other diagnostics of chain behavior and review model specification and category support before deciding what action is needed.
This sequence is consistent with the earlier material on trace summaries, stabilization, mixing, and sparse-variable diagnostics. A proportion trace can be autocorrelated: neighboring iterations are not independent observations. Therefore, a visually dense trace is not equivalent to a large number of independent draws, and a stable-looking trace should not be treated as a formal guarantee.
Worked Example: Persistent Differences Between Chains
Suppose three hypothetical MICE chains impute a nominal variable with categories A, B, and C. There are \(m=50\) missing values. Over several later iterations, the category counts remain near the following values in each chain:
| Chain | A count | B count | C count | A proportion | B proportion | C proportion |
|---|---|---|---|---|---|---|
| 1 | 25 | 15 | 10 | 0.500 | 0.300 | 0.200 |
| 2 | 24 | 16 | 10 | 0.480 | 0.320 | 0.200 |
| 3 | 15 | 20 | 15 | 0.300 | 0.400 | 0.300 |
For chain 1, for example, the proportions are \(25/50=0.500\), \(15/50=0.300\), and \(10/50=0.200\). The corresponding proportions for chains 2 and 3 are calculated by dividing each count by 50; each row sums to 1.000. Chains 1 and 2 are close in this summary, while chain 3 has a lower A proportion and higher B and C proportions.
If the trace plot shows this separation persisting across many iterations, it is a warning to investigate mixing and the imputation setup. The table alone does not prove nonconvergence: these are illustrative chain summaries, and a stable difference could have several explanations. Check that chains use the same coding and missing records, inspect the full traces rather than selected iterations, and consult other diagnostics. Do not pool the chains’ proportions into one number to make the separation disappear.
Common Mistakes, Alternatives, and Limits
- Reading a count difference as a record-transition count. A category count rising by two is a net increase of two. The before-and-after counts do not reveal how many records entered or left that category.
- Ignoring the denominator. A 5-percentage-point jump is one count when \(m=20\), but it represents a different count change when the denominator differs. Report or inspect \(m\) alongside the trace.
- Plotting only one category. This can hide compensating movements among the other categories. Inspect the full set of proportions, remembering their sum constraint.
- Calling a flat trace proof of convergence. A category can remain stable while other aspects of the chain or model are problematic. Use the trace with other chain diagnostics and substantive review.
- Treating category codes as quantitative values. Numerical labels for nominal categories have no meaningful spacing. For ordinal variables, the order matters, but category proportions should not be replaced by an unexamined mean of the codes.
The main alternative is not another single summary that solves every diagnostic problem. Other chain-level summaries and diagnostics provide complementary evidence; a category-proportion trace is especially direct for shifts in category allocation. It is less informative about which records changed, relationships among variables, or whether the imputation model is scientifically adequate. Rare categories may also yield highly discrete traces, so apparent jumps need to be interpreted in relation to \(m\) and the surrounding chain behavior.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A categorical variable has 25 missing values. What is the smallest possible positive step in a category’s imputed proportion, in proportion units and percentage points?
- For three categories, the counts at one iteration are 9, 11, and 5. Compute the proportions and check their sum.
- A category count is 2 at one iteration and 4 at the next. What can be concluded about the net count change, and what cannot be inferred about individual records?
- Two chains have stable but separated traces for one category over many iterations. Give a cautious interpretation and name two checks to make.
Solutions. (1) The increment is \(1/25=0.040\), or 4 percentage points. (2) The denominator is \(9+11+5=25\); the proportions are \(9/25=0.360\), \(11/25=0.440\), and \(5/25=0.200\), which sum to 1.000. (3) The category count has a net increase of two. The two counts do not show how many individual records entered or left; both movements could have occurred. (4) Persistent separation is a warning that chains may not be mixing similarly, but it does not prove nonconvergence by itself. Check that the chains use identical category definitions and missing records, then inspect full traces and other chain diagnostics for continuing drift or poor mixing.