Sensitivity to FCS Initialization

Why Starting Values Matter

In fully conditional specification (FCS), missing values are updated variable by variable using conditional models. The first update needs values for the other variables, including their missing entries. An algorithm therefore begins with an initialization: provisional values are placed in missing cells, while observed values remain fixed. These provisional values are starting points for the chain, not final imputations.

A practical question follows: if the same FCS procedure begins from different plausible initial values, do the chains eventually give similar summaries? Early differences are expected because the chains begin in different states. If disagreement persists over the iterations used for inference, however, the run may not have explored its relevant distribution adequately. This is a warning to investigate, not a diagnosis of the specific cause.

“Plausible” means consistent with the variable’s support and reasonable given the observed data and imputation models. For example, a binary variable must start with zeros and ones, and a nonnegative measurement should not be initialized with negative values unless the variable itself allows them. Plausible starts need not be identical across chains; deliberately differing starts make the check more informative.

What the Comparison Can Establish

Consider a common FCS setup with fixed observed data and a fixed set of conditional models. At iteration \(t\), the algorithm updates the missing entries for each variable in turn, drawing from the corresponding conditional model given the current values of the other variables. A complete pass through the variables is one iteration, or sweep. Each initialization produces a starting state for this process.

For a summary of interest, write \(G_r^{(t)}\) for its value at iteration \(t\) in chain \(r\). The summary might be the mean of an imputed variable, a regression coefficient calculated from the completed data, or a proportion. The initialization check asks whether the trajectories and retained-run summaries of \(G_r^{(t)}\) depend substantially on chain \(r\)’s starting values.

The underlying justification is conditional: if the FCS process defines the same transition rule for every chain and that rule has a common stationary distribution to which it converges from the chosen starts, the influence of initialization should diminish. For example, let \(K^t(z,\cdot)\) denote the distribution after \(t\) updates from state \(z\), and let \(\pi\) be a stationary distribution. Convergence in total variation means that \(d_{\mathrm{TV}}(K^t(z,\cdot),\pi)\) tends to zero, where \(d_{\mathrm{TV}}\) is the largest difference in probability assigned to any event. For a bounded summary \(g\), the difference in its expectations from two starting states is bounded by

Here \(\|g\|_{\infty}\) is the largest absolute value of \(g\). If both distances on the right approach zero, the difference in expected summaries also approaches zero. This explains why loss of dependence on the start is a useful diagnostic. It is not a guarantee that every FCS specification has a suitable stationary distribution or converges to it. In particular, conditional models need not define a compatible joint distribution. Agreement between chains is evidence about their observed behavior under this run, not proof of compatibility or convergence.

Designing an Informative Initialization Check

To attribute differences to starting values, keep the data, conditional models, update procedure, and settings the same across chains. Use distinct plausible initial fills and independent stochastic draws. Compare trajectories for summaries relevant to the intended analysis, as well as summaries of the imputed variables themselves. A chain can have similar averages for one variable while differing in another variable or in a substantive analysis estimate.

Persistent Differences and Monte Carlo Error

A visible difference between chains should be interpreted in relation to its size, its persistence, and the uncertainty in estimating each chain’s summary. Monte Carlo error (MCSE) quantifies simulation uncertainty in a chain summary; it is not the same as the variability of individual imputations. When two chain summaries estimate the same quantity and their Monte Carlo errors are estimated separately, a useful descriptive scale for their difference is

This comparison assumes the MCSEs are meaningful estimates for the summaries being compared. The denominator combines the two simulation errors under the approximation that their estimation errors are independent. The resulting value is a diagnostic scale, not a formal test with a universal cutoff. A large value, especially alongside a persistent visual offset, is a reason to investigate. Even a small value does not establish that all parts of the imputed-data distribution have been explored.

What Counts as a Warning?

Potential warning signs include chain summaries that remain separated after the chosen burn-in, persistent differences in the level or variability of imputed values, and analysis estimates that depend noticeably on which initialization was used. A difference confined to the first few iterations is less concerning than a stable offset across a long retained segment, but the distinction depends on the summaries and the intended analysis.

Initialization sensitivity is most informative when the starting values probe plausible differences. Starting every chain with the same constant fill can make chains look similar initially without testing how they respond to other reasonable states. Conversely, deliberately extreme but implausible fills can produce long transients that say little about behavior from realistic starts. Check that the initial values are valid and explain the rationale for the range of starts.

The following points help avoid overinterpreting the check:

• Agreement is not proof. Chains can agree on a few monitored summaries while failing to explore other features. It also does not prove that the conditional models are compatible.
• Different early values are expected. A large gap at initialization is not, by itself, evidence of failure. The important question is whether differences persist in the relevant parts of the run.
• A longer run is not automatically a remedy. If the chains remain separated, simply adding iterations may or may not resolve the issue. Reassess the run and its summaries rather than assuming more updates guarantee adequate exploration.
• Do not confound causes. If both the initial values and the conditional models change between chains, the comparison cannot isolate sensitivity to initialization.

How This Check Differs From Alternatives

Using multiple chains with the same initialization can help assess variation between runs, but it may provide a weaker test of sensitivity to starting values because the chains do not begin from deliberately different states. The initialization check complements comparisons of independent chains by varying the initial fills intentionally. Trace plots show how summaries evolve and can reveal startup effects; comparing chain summaries quantifies disagreement in selected quantities. Neither diagnostic alone proves convergence.

Changing a conditional model or the update procedure answers a different question: whether results depend on the modeling or updating choices. To assess initialization specifically, keep those choices fixed. If persistent disagreement remains, other checks may help identify the problem, but do not attribute it to a particular cause based on initialization sensitivity alone. The earlier material on FCS compatibility is also relevant: chains agreeing on monitored summaries does not establish that the conditional models specify a coherent joint distribution.

There are limitations to this diagnostic. It can miss differences in aspects of the distribution that were not monitored; Monte Carlo error estimates can themselves be unreliable when exploration is poor; and a modest number of starts cannot represent every possible initial state. In addition, FCS need not always have the simple stationary target assumed in the convergence justification above. The check is best treated as one piece of evidence, reported with the initializations, summaries, and retained iterations used.