Potential Scale Reduction for Multiple MICE Chains

Why Compare Variation Across MICE Chains?

A MICE run can produce several chains: repeated sequences of imputations generated from different starting values and, ordinarily, separate random-number streams. A diagnostic should ask whether those chains appear to be exploring a common distribution for a quantity of interest. The potential scale reduction factor compares variation between chains with variation within chains. A substantial difference can signal that the chains have not yet settled into similar behavior.

The calculation is applied to a particular scalar summary recorded at each iteration, not to “the chain” in the abstract. For example, at every retained iteration one might record the mean of the imputed values for one variable. The result concerns that summary and that MICE run; it does not certify convergence of every imputed variable or every feature of the distribution.

Definition: Suppose \(m\) chains each provide \(n\) matched post-burn-in values of the same scalar summary. Write \(\theta_{ij}\) for the value from chain \(j\) at retained iteration \(i\), where \(i=1,\ldots,n\) and \(j=1,\ldots,m\). Let \(\bar{\theta}_j\) and \(s_j^2\) be the sample mean and sample variance within chain \(j\), and let \(\bar{\theta}_{..}\) be the mean of all \(mn\) values. Define the within-chain variance \(W\), between-chain variance \(B\), estimated marginal variance \(\hat{V}\), and classical potential scale reduction factor \(\hat{R}\) by the formulas below.

W=\frac{1}{m}\sum_{j=1}^{m}s_j^2,\qquad B=\frac{n}{m-1}\sum_{j=1}^{m}(\bar{\theta}_j-\bar{\theta}_{..})^2

\hat{V}=\frac{n-1}{n}W+\frac{1}{n}B,\qquad \hat{R}=\sqrt{\frac{\hat{V}}{W}}

The chain means must describe the same quantity, on the same scale, at corresponding retained iterations. The formula also requires \(m\geq2\), \(n\geq2\), and \(W>0\). If \(W=0\), the ratio is undefined: it is not evidence of perfect convergence.

What the Calculation Measures

The within-chain variance \(W\) summarizes how much the summary varies inside chains. The between-chain variance \(B\) scales the dispersion of the chain means to the scale of a variance for individual iterations. If all chains have settled into the same target distribution, their means should not remain systematically separated, and the two sources of variation should become broadly compatible as chains grow.

The estimate \(\hat{V}\) combines the within-chain estimate with the between-chain estimate. The factor \(\hat{R}\) is the square root of the ratio of that combined estimate to \(W\). Thus, a value appreciably above one indicates that between-chain differences are inflating the estimated variance relative to within-chain variation. It is a warning signal, not a probability that the chains have converged, and not a direct estimate of the probability that any conclusion is correct.

Worked Calculations

Worked Example: Three Chains With Noticeable Separation

Suppose the summary is the mean of the imputed values for a hypothetical continuous variable, recorded for four retained iterations in each of three MICE chains. The recorded summaries are:

Chain 1: 8, 10, 9, 9. Chain 2: 10, 11, 9, 10. Chain 3: 11, 10, 12, 11.

The chain means are 9, 10, and 11. Each chain has sample variance \(2/3\): for chain 1, for example, the squared deviations from 9 sum to \(1+1+0+0=2\), and division by \(n-1=3\) gives \(s_1^2=2/3\). The same calculation gives \(s_2^2=s_3^2=2/3\). Therefore \(W=2/3\), and the overall mean of the three chain means is 10.

With \(m=3\) and \(n=4\), the between-chain variance is

B=\frac{4}{3-1}\left[(9-10)^2+(10-10)^2+(11-10)^2\right] =2(1+0+1)=4.

Then

\hat{V}=\frac{3}{4}\left(\frac{2}{3}\right)+\frac{1}{4}(4) =\frac{1}{2}+1=1.5,\qquad \hat{R}=\sqrt{\frac{1.5}{2/3}}=\sqrt{2.25}=1.5.

A factor of 1.5 is a substantial indication that the chain means remain separated relative to the within-chain variation. These short sequences do not support treating the summary as well mixed. The calculation does not identify why they differ; trace plots and checks of the run are needed to investigate.

Worked Example: A More Modest Difference Between Two Chains

For a hypothetical imputed-variable mean, suppose two chains yield four retained values each: chain 1 has 4, 5, 5, 6, and chain 2 has 5, 6, 6, 7. Their means are 5 and 6. Each sample variance is \(2/3\), since the squared deviations from its chain mean sum to 2 and are divided by 3. Hence \(W=2/3\), and the overall mean is 5.5.

Here \(m=2\) and \(n=4\), so

B=\frac{4}{2-1}\left[(5-5.5)^2+(6-5.5)^2\right] =4(0.25+0.25)=2.

The estimated variance and factor are

\hat{V}=\frac{3}{4}\left(\frac{2}{3}\right)+\frac{2}{4} =0.5+0.5=1,\qquad \hat{R}=\sqrt{\frac{1}{2/3}}=\sqrt{1.5}\approx1.225.

The factor is closer to one than in the preceding example, but the chains still have only four retained values each. It is not a guarantee that the chains are adequately mixed: more iterations and inspection of their trajectories would be needed before drawing a reassuring conclusion.

Worked Example: Similar Chain Means Can Hide a Shared Drift

Consider three hypothetical chains of a scalar imputed-variable summary, each recorded for only three retained iterations. All three have the same values: 0, 1, 2. Each chain mean is 1 and each sample variance is 1. Thus \(W=1\) and \(B=0\), because the chain means are identical. The formula gives

\hat{V}=\frac{2}{3}(1)+\frac{1}{3}(0)=\frac{2}{3},\qquad \hat{R}=\sqrt{\frac{2/3}{1}}\approx0.8165.

All chains show a rising sequence; the matching chain means have not exposed that shared pattern. In this finite-sample calculation, the raw factor is below one. That is possible with the classical formula and should not be read as “better than converged.” Some software reports a version constrained to be at least one, but such a convention does not resolve the underlying diagnostic limitation. A trace plot would make the common rise visible.

Assumptions, Matching, and Practical Use

The interpretation of the factor relies on conditions that may be only approximate in an applied MICE run. The retained chains should be generated under the same imputation specification and target the same stationary distribution for the summary. Their starting points should be sufficiently dispersed to make between-chain differences informative. The retained period should be reasonably stationary, and the summary should have a finite variance. The chains must be long enough for within-chain variation and chain means to be meaningful estimates.

For MICE, these are diagnostic conditions, not automatic guarantees. Chained-equation updates can have slow transitions, and a chain-level summary can drift or move between regions over time. A numerical \(\hat{R}\) calculated during such behavior may not have the usual interpretation. The chain-level values are also typically autocorrelated; a long sequence of highly dependent values is not equivalent to the same number of independent draws.

Before calculating, verify that every chain contributes values of the same summary, using the same variable and definition, after a comparable burn-in period. If one chain has fewer retained iterations, do not silently pair unrelated observations or treat unequal sequences as matched. Use a common retained length and document any truncation, while recognizing that discarding observations does not fix nonstationarity.

Choose one summary.
For example, select the mean of the imputed values for one variable at each iteration. Keep the definition and units fixed across chains.

Check the retained sequences.
Use comparable post-burn-in iterations and inspect trace plots for trends, abrupt changes, or chains that remain separated.

Calculate and interpret.
Compute each chain mean and variance, then \(W\), \(B\), \(\hat{V}\), and \(\hat{R}\). Treat an elevated value as a prompt to investigate, not as a stand-alone verdict.

Diagnostics, Alternatives, and Limitations

A large factor is useful because it draws attention to disagreement between chains that may not be apparent from a single chain. The appropriate response is to inspect trace plots, consider whether additional iterations or a revised imputation specification are needed, and assess other relevant summaries. The earlier material on effective sample size and Monte Carlo error addresses a different question: how much independent information a correlated chain supplies and how precisely a Monte Carlo estimate is computed. Those quantities complement rather than replace between-chain comparison.

A factor close to one is not proof of convergence. Chains can share the same drift, have too few iterations to reveal separation, or agree on a mean while differing in other features of the distribution. A summary can also be insensitive to behavior that matters for the analysis. Calculate factors for substantively relevant summaries, and examine the corresponding traces; no single scalar diagnostic covers every feature of a MICE run.

There is no universal cutoff that turns this classical factor into a pass/fail test for MICE. Its reliability is particularly limited when \(n\) is small, when initial chains were not meaningfully dispersed, or when the retained transitions are nonstationary. The formula assumes equal chain lengths and has finite-sample behavior that can produce a raw value below one. It also summarizes a particular scalar quantity rather than providing a general measure of imputation quality or a check that the imputation model is appropriate.

Common Mistakes

Using mismatched summaries. Comparing, for example, one chain’s mean with another chain’s variance does not produce a meaningful factor. Use the same scalar summary in every chain.
Calling a value near one proof of convergence. A shared drift or a short run can yield reassuring-looking agreement. Inspect the chain histories and other diagnostics.
Treating the factor as an imputation-quality score. It assesses between- and within-chain variation for a chosen summary; it does not establish that the imputation model is appropriate.
Interpreting a raw value below one as exceptional convergence. Finite-sample variation can produce it under the classical formula. Check the calculation and the traces rather than assigning it a special success meaning.

Key takeaway: A potential scale reduction factor compares the estimated variation across chains with variation within chains for one matched summary. A large value warns of chain disagreement; a small value is reassuring only when chain length, stationarity, matching, and trace behavior also support that interpretation.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

Three chains each have 20 retained values for the same summary. What must be matched across chains before calculating the factor?
If \(W=2\) and \(B=8\) for \(n=4\), calculate \(\hat{V}\) and \(\hat{R}\).
Why does a factor close to one fail to establish that a MICE run is stationary?
What does a factor substantially above one suggest, and what is a sensible next diagnostic step?
What should be concluded if \(W=0\) in the classical formula?

Solutions. (1) The same scalar summary, definition, scale, and comparable post-burn-in iterations must be used in all chains. (2) \(\hat{V}=(3/4)(2)+(1/4)(8)=1.5+2=3.5\), so \(\hat{R}=\sqrt{3.5/2}=\sqrt{1.75}\approx1.323\). (3) Chains may share a drift, the retained run may be too short, or the selected summary may miss relevant differences; inspect traces and other summaries. (4) It suggests between-chain variation is large relative to within-chain variation; inspect trace plots and investigate the run rather than treating the number as a verdict. (5) The ratio is undefined, so it supplies no valid evidence of convergence.