Why Run More Than One FCS Chain?
An FCS run repeatedly imputes missing values using a sequence of conditional models. A single run can show how its summaries behave over iterations, but it cannot show whether a different independently generated trajectory would behave similarly. Multiple chains provide that replication: they help reveal whether the observed behavior is reproducible across runs, or whether a particular run has an unusual trajectory or summary.
The number of chains is a design choice, not a universal constant. More chains give more opportunities to see between-run discrepancies, but using a fixed computational budget for more chains may leave fewer iterations in each. The useful number therefore depends on the diagnostic purpose, the time required per iteration, and how much chain-to-chain variation needs to be assessed.
For notation, let \(c=1,\ldots,C\) index the chains and \(t=1,\ldots,T_c\) index iterations within chain \(c\). A summary of interest at iteration \(t\) of chain \(c\) can be written \(S_{ct}\), such as the mean of an imputed variable or a coefficient from an analysis model. The chains share the data and modeling plan, but their stochastic trajectories are generated independently. As in the earlier material on comparing independent FCS chains, comparisons are most informative when they use the same summary on the same iteration scale.
Independent chains are diagnostic replicates, not independent copies of the study participants. The observed data are identical across chains. Also, FCS trajectories should not automatically be described as draws from a joint posterior distribution: fully conditional specifications need not correspond to a compatible joint model. Several similar-looking chains cannot establish that the conditional models are correctly specified or compatible.
What Additional Chains Can Reveal
A second or subsequent chain can expose behavior that a single trace cannot. For example, one chain may remain in a range that other chains do not visit, or its summaries may drift while the others appear relatively stable. More chains can also make an overall pattern of between-run variation easier to see. These are warnings to investigate, not automatic diagnoses of a particular cause.
A chain count should be chosen to serve a stated diagnostic goal. Possible goals include checking whether several independently run trajectories show comparable summaries after the selected burn-in, or increasing the chance that at least one run reveals a specified unusual feature. The second goal can be expressed with a simple planning calculation, provided its assumptions are made explicit.
The formula follows by considering the complementary event. If one chain misses the feature with probability \(1-r\), then, under independence, all \(C\) chains miss it with probability \((1-r)^C\). Subtracting this probability from one gives the chance that at least one chain detects it:
Worked Example: Planning for a Chance of Detection
Suppose, for planning purposes, a chain has an estimated \(r=0.35\) probability of revealing a particular unexpected behavior during the proposed run. How many independent chains are needed for the probability of at least one detection to reach 0.90?
Under the stated equal-probability and independence assumptions, choose \(C\) so that \(1-(1-0.35)^C \geq 0.90\). Rearranging gives \((0.65)^C \leq 0.10\), and therefore \(C \geq \log(0.10)/\log(0.65) \approx 5.345\). The smallest whole-number choice is 6 chains.
Checking the boundary makes the decision clear. Five chains give \(1-0.65^5=0.8840\), while six give \(1-0.65^6=0.9246\). Thus five do not reach 0.90, and six do, under this illustrative planning model.
The value \(r=0.35\) is hypothetical and may be difficult to estimate in practice. If the chance of detection differs across chains, or if chains share random-number streams, the calculation does not apply as written. The result does not certify the FCS algorithm; it only quantifies the stated detection objective under its assumptions.
Choosing Chain Count Alongside Run Length
The earlier material on choosing the number of FCS iterations addresses how long to run a chain. Chain count is a related but distinct decision. At a fixed total number of updates, adding chains generally means fewer updates per chain. This can improve replication across trajectories while making each trajectory less informative about behavior that unfolds slowly. Conversely, a single long run can provide a detailed trace but no independent trajectory for comparison.
There is no chain-count threshold that guarantees a reliable diagnosis. A small number of chains may be adequate for an initial check when computation is expensive and the traces are consistent, but limited replication makes unusual behavior easier to miss. More chains are useful when computing resources allow, when between-run differences are a concern, or when the purpose is explicitly to increase the chance of detecting a specified feature. The decision should also consider whether each chain is long enough, after burn-in, to show interpretable behavior.
A practical plan states both quantities: the number of independent chains and the number of iterations in each. For example, “four chains, each run for 1,000 iterations” is more informative than “4,000 iterations,” which could mean one trajectory or several. If the total computational budget is fixed, document how it was divided and inspect whether the resulting per-chain runs are adequate for the diagnostic task.
Worked Example: Comparing Two Allocations of the Same Budget
A hypothetical FCS analysis has a budget of 4,000 post-initialization update sweeps. The analyst considers either four chains of 1,000 sweeps each or eight chains of 500 sweeps each.
Both plans use 4,000 sweeps in total: \(4\times1{,}000=4{,}000\) and \(8\times500=4{,}000\). The eight-chain plan doubles the number of independent trajectories, which can provide more between-chain comparisons. But every trajectory is half as long, so it may be less able to display behavior that requires many updates to become visible.
If preliminary runs suggest that 500 sweeps per chain are too short to assess the chosen summaries, the eight-chain allocation would not meet the diagnostic objective merely because it has more chains. If 500 sweeps are sufficient for that objective and replication across runs is especially valuable, it may be a reasonable allocation. The budget alone does not decide between the plans; the adequacy of each chain’s length and the purpose of replication do.
This comparison assumes a sweep has comparable computational meaning across chains. In practice, record the actual update schedule and run time as well as the nominal sweep count, particularly if the runs differ in settings or computational conditions.
Reading Chain-to-Chain Differences
For each chain, inspect the same summaries after the chosen burn-in, and compare their traces and distributions. Useful questions include: Are the summaries occupying comparable ranges? Is one chain showing a sustained shift or markedly different spread? Do apparent differences persist when more iterations are run? These checks complement within-chain trace assessment, autocorrelation, effective sample size, and Monte Carlo error; they do not replace them.
A numerical comparison can help make the scale of a discrepancy concrete, but chain-level averages are summaries, not independent observations from the original study. The following example illustrates both the comparison and its limits.
Worked Example: Interpreting Unequal Chain Summaries
Suppose four hypothetical chains have post-burn-in averages of a diagnostic summary equal to 5.00, 5.04, 4.97, and 5.39 units. The average across these four chain summaries is \((5.00+5.04+4.97+5.39)/4=5.10\) units.
Their deviations from 5.10 are \(-0.10\), \(-0.06\), \(-0.13\), and \(0.29\) units. The sample variance of the four chain averages is \[ \frac{(-0.10)^2+(-0.06)^2+(-0.13)^2+(0.29)^2}{4-1} =\frac{0.1146}{3}=0.0382 \] square units, giving a sample standard deviation of \(\sqrt{0.0382}=0.1954\) units.
The fourth chain’s average is noticeably higher than the others in this hypothetical example. That is a reason to examine its trace and run details, and possibly to run further chains or iterations. It is not proof that the fourth chain is wrong, nor does the average across chains establish a scientifically valid estimate. The summary’s scale, within-chain behavior, burn-in choice, and conditional models all matter.
In particular, the sample standard deviation across four chain averages is not the standard error for an analysis coefficient or a treatment effect. It describes dispersion of these four run-level summaries. Even dividing it by \(\sqrt{4}\) would only describe the Monte Carlo precision of their average under additional assumptions about the summaries; it would not be the sampling uncertainty of the scientific estimand.
Chains, Iterations, and Imputed Data Sets
The three counts answer different questions. The chain count \(C\) counts independent trajectories. The iteration count describes the number of successive FCS update cycles within a trajectory. The number of completed data sets retained for downstream analysis is a separate choice. Depending on the workflow, completed data sets may be saved at selected iterations from one or more chains, but increasing \(C\) does not automatically provide the desired number of imputations or make retained data sets independent.
Worked Example: Keeping the Three Counts Separate
An analyst plans four independently run chains and intends to retain five completed data sets from each chain after burn-in. This produces \(4\times5=20\) retained completed data sets. The plan still has four chains, not 20 chains; and it must separately specify how many update iterations each chain runs before and between those saved states.
Whether the 20 data sets provide adequate information for the intended analysis depends on the imputation and analysis plan, the behavior of the chains, and the spacing of retained states. If successive states are strongly related, retaining many closely spaced states does not make them independent. The chain count is useful for diagnostic replication, while the retained-data-set count serves a different purpose.
Assumptions, Diagnostics, and Limitations
Comparisons across chains are interpretable only when the runs are comparable. Use the same incomplete data, conditional models, predictor sets, update schedule, and summary definitions. Use independent random-number streams and record initialization and software settings. If one chain uses a different specification, a difference between summaries cannot be attributed simply to chain-to-chain stochastic variation.
Look for more than agreement in averages. Similar averages can coexist with differences in spread, drift, or ranges visited. Likewise, chains that look alike over a short run may all fail to display longer-term behavior. Flat or overlapping traces are useful evidence, but they do not prove convergence, identify a unique target distribution, or establish compatibility of the conditional models. The earlier material on why convergence does not prove compatibility is especially relevant: stability and compatibility are distinct questions.
The main limitations of adding chains are computational cost and the possibility of false reassurance. If the same modeling problem affects every run, all chains may agree while sharing the same deficiency. If a behavior is very unlikely to appear during any individual run, a modest increase in chain count may still leave it unseen. The planning formula applies only when a meaningful per-chain detection probability can be posited and the chain-level detection events are independent.
More chains also cannot compensate automatically for inadequate iterations per chain, poorly specified conditional models, or summaries that do not reveal the behavior of interest. Conversely, a large difference among chains is a diagnostic signal to investigate—not a reason to select whichever chain gives the preferred result. FCS may exhibit stable behavior even when its conditionals are incompatible, so agreement across chains should not be interpreted as proof of a joint model.
A Practical Planning Sequence
Name the summary or behavior to compare across runs. A chain count without a stated purpose is difficult to justify.
Hold the data and modeling settings fixed, and use independent initialization and random-number streams so chain differences are meaningful.
Allocate enough updates per chain for the planned checks; do not maximize the chain count at the expense of runs too short to interpret.
Inspect the same post-burn-in summaries across chains. If discrepancies remain, investigate traces, run settings, and model behavior rather than averaging away the disagreement.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In a plan with six chains, each run for 800 iterations, what does the number six count? Does it specify the number of completed data sets retained?
- If a chain has probability \(r=0.20\) of revealing a specified diagnostic feature, what is the probability that at least one of four independent chains reveals it?
- Two chains have similar post-burn-in averages, but one shows a sustained upward drift. Is agreement of their averages enough to conclude that the behavior is satisfactory?
- A fixed budget allows either three chains of 1,200 sweeps or six chains of 600 sweeps. What trade-off should guide the choice?
Solutions. (1) Six is the number of independently run trajectories. The plan does not state how many completed data sets will be retained; that count must be specified separately. (2) Using the planning result, \(1-(1-0.20)^4=1-0.8^4=1-0.4096=0.5904\). Under the stated assumptions, the probability is 0.5904, or 59.04%. (3) No. A similar average can hide a trend within one chain. Inspect the trace and other relevant summaries; average agreement alone does not establish satisfactory behavior. (4) Six chains offer more independent replication, while three chains provide twice as many sweeps per trajectory. Choose based on whether 600 sweeps are adequate for the intended checks and whether additional between-chain comparisons are worth the shorter runs; the total budget alone does not determine the better plan.