Why Consider Randomizing the Update Order?
In fully conditional specification (FCS), a sweep updates the variables with missing values one after another, using a conditional model for each variable. The earlier material on choosing the variable update order concerns selecting an order for this sequence. This tutorial considers a different question: should the order remain fixed, or should it be randomized between sweeps?
The question matters because updates within a sweep are sequential. When a variable is updated, its conditional model uses the current values of the other variables in its conditioning set. Values for variables already updated in that sweep may therefore be more recent than values for variables not yet updated. Changing the schedule can change which version of those values is available at each step.
Randomizing the order is best viewed as a robustness check on the schedule. It can reveal whether results depend noticeably on the chosen order. It does not, by itself, establish that the imputation models are appropriate, that the chain has stabilized, or that the resulting imputations have desirable statistical properties.
What Changes When the Order Changes?
Suppose the variables to be updated are \(X_1,\ldots,X_p\). An order can be written as a permutation \(\pi=(\pi_1,\ldots,\pi_p)\), where \(\pi_k\) is the variable updated in position \(k\). In a randomized-order scheme, one possible convention is to choose each of the \(p!\) permutations with equal probability at the start of each sweep. Other probability rules are possible, but the rule should be stated and applied consistently.
At the \(k\)th step of a sweep, the conditional update for \(X_{\pi_k}\) uses the current values of the variables in its conditioning set. If a conditioning variable has already been updated, its newly updated value is available. If it has not yet been updated, its value is the value carried into the sweep. The update order therefore affects the sequence of inputs to the conditional models, even when the models themselves are unchanged.
The following algebra isolates that scheduling effect. It uses deterministic update rules only to make the order dependence visible. It is not a description of the full stochastic imputation procedure: the way conditional updates generate replacements is a separate part of FCS.
Worked Example: Two Variables Updated in Opposite Orders
Consider two hypothetical variables, \(A\) and \(B\), with current values \(A=2\) points and \(B=10\) points. For this scheduling illustration only, suppose their update rules are \(A \leftarrow 1+0.5B\) and \(B \leftarrow 4+A\), with both variables measured in points. Compare the two orders.
Order \(A\), then \(B\). First update \(A\), using the current \(B=10\): \(A=1+0.5(10)=6\) points. Then update \(B\), using the newly updated \(A=6\): \(B=4+6=10\) points. The end-of-sweep values are \(A=6\), \(B=10\).
Order \(B\), then \(A\). First update \(B\), using the current \(A=2\): \(B=4+2=6\) points. Then update \(A\), using the newly updated \(B=6\): \(A=1+0.5(6)=4\) points. The end-of-sweep values are \(A=4\), \(B=6\).
The same two rules and the same starting values produced different end-of-sweep values because the second update in each order used a different newly updated value. This arithmetic demonstrates why schedule sensitivity is possible; it does not show that any particular order is preferable.
Ways to Randomize the Schedule
Two procedures are sometimes confused. In the first, one random permutation is drawn at the beginning of a run and then held fixed for every sweep. This makes the order random across runs, but it does not randomize the order between sweeps within a run. In the second, a new permutation is drawn at the beginning of every sweep. This second procedure is the direct form of randomized variable update orders described here.
For \(p\) variables, there are \(p!\) possible orders. With three variables, for example, the six possible orders are \(ABC\), \(ACB\), \(BAC\), \(BCA\), \(CAB\), and \(CBA\). A uniform scheme assigns probability \(1/6\) to each. In practice, the schedule may be generated by a random shuffle of the variable list. The randomization concerns only the sequence of updates; it does not mean that variables, conditioning sets, or conditional model specifications are being randomly selected.
Worked Example: A Three-Variable Randomized Schedule
Suppose a hypothetical FCS procedure updates \(A\), \(B\), and \(C\). Each sweep draws uniformly from the six possible permutations. A realized sequence of three orders is \(CAB\), then \(ABC\), then \(BCA\).
There are \(3!=3\times2\times1=6\) possible orders, so each has probability \(1/6\) on any sweep under this uniform rule. The probability of the particular three-order sequence shown is \( (1/6)(1/6)(1/6)=1/216\), assuming the sweep orders are drawn independently. This is the probability of that schedule sequence, not a probability about the values imputed.
In the first sweep, \(C\) is updated before \(A\) and \(B\); in the second, \(A\) is first; in the third, \(B\) is first. Each variable appears once in each sweep. Thus randomizing order changes sequencing, not the requirement to update every variable once per sweep.
A run that uses \(CAB\) for all three sweeps would instead have selected one order and kept it fixed. Although that order might have been chosen randomly before the run, it would not test the effect of varying the order from sweep to sweep.
Using Randomized Order as a Robustness Check
A useful check asks whether scientifically important summaries change materially when the update schedule changes, while other choices are held as comparable as possible. The target is not exact numerical equality: different schedules can lead to ordinary Monte Carlo variation as well as schedule-related differences. The practical question is whether the pattern of estimates and conclusions is sensitive enough to affect interpretation.
For a fair comparison, keep the data, variable-specific conditional models, conditioning sets, initialization strategy, number of sweeps, and downstream analysis fixed. Compare a chosen fixed-order run with several runs that use randomized orders, or compare multiple fixed orders with multiple randomized schedules. Use comparable saved portions of the runs. When possible, repeat the comparison so that an unusual result from one schedule is not mistaken for a general order effect.
Inspect estimates of the quantities that matter for the analysis, along with simple summaries of their variation across runs. Look for consistent shifts, wider between-run variation, or conclusions that change direction or practical interpretation. Trace plots of relevant summaries can help reveal patterns over sweeps, but visual inspection is not a formal guarantee of stability. A lack of visible differences is evidence only about the schedules and runs examined; it is not proof that the imputation procedure is valid.
Worked Example: Comparing Estimates Across Schedules
Suppose a hypothetical analysis estimates a regression coefficient in log-odds units. To assess schedule sensitivity, the analyst runs four comparable fixed-order analyses and four analyses with randomized orders, holding the conditional models and analysis plan constant. For illustration, suppose the resulting summaries are:
Fixed-order estimates: \(0.31,\ 0.35,\ 0.30,\ 0.32\). Their mean is \((0.31+0.35+0.30+0.32)/4=1.28/4=0.320\) log-odds units, and their range is \(0.35-0.30=0.05\) log-odds units.
Randomized-order estimates: \(0.30,\ 0.33,\ 0.34,\ 0.31\). Their mean is \((0.30+0.33+0.34+0.31)/4=1.28/4=0.320\) log-odds units, and their range is \(0.34-0.30=0.04\) log-odds units.
In this small illustration, the two groups have the same mean estimate, and their ranges are similar. That is reassuring with respect to this particular coefficient and these runs, but it does not establish equivalence of the schedules or rule out sensitivity in another quantity. The four values are diagnostic summaries, not four independent estimates to combine as though they were a new inferential procedure.
State the fixed order and how randomized orders are generated, including whether a fresh permutation is drawn each sweep.
Keep the data and conditional-model specifications the same so that a difference is interpretable as schedule sensitivity rather than a model change.
Compare relevant estimates and their patterns across comparable runs; do not base the assessment on one randomized schedule alone.
Describe which schedules and summaries were compared. Treat agreement as a limited robustness check, not as proof of validity or convergence.
What the Check Can and Cannot Tell You
Randomizing order is useful when a fixed order is conventional but there is concern that the sequence of updates could materially influence the results. It can also make the dependence on a particular hand-chosen order visible. If estimates differ systematically across schedules, that is a reason to investigate further: check that the same variables and conditional models were used, confirm that the schedule was implemented as intended, and examine whether the result is reproducible across runs.
There are important limits. First, random order does not repair a misspecified conditional model or an unsuitable conditioning set. Second, it does not guarantee that a chain has stabilized. Third, agreement between schedules cannot establish that the assumed statistical structure is appropriate. Finally, because different schedules can produce different sequences of updates, randomization itself may add variation; a single randomized run is an especially weak basis for a robustness claim.
This check also does not replace a principled choice of order where one is needed. If there is a reason to use a particular sequence, it should be stated and justified. Randomization is a way to examine sensitivity to scheduling, not a universal default that is automatically superior to a fixed order.
Common Mistake
Wrong reasoning: “I randomized the order, so the imputations are now unbiased and the chain must have converged.”
Correction: Randomization changes the schedule and can expose sensitivity to that schedule. It does not establish unbiasedness, convergence, or adequacy of the conditional models. Those claims require separate justification and assessment. Similarly, if two schedules give similar estimates in a few runs, the defensible conclusion is that no important difference was evident in the examined comparison—not that the schedules are mathematically equivalent.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- For four variables, how many possible update orders are there? If each order is equally likely, what is the probability of one specified order in a sweep?
- Explain why updating \(A\) before \(B\) can produce different end-of-sweep values from updating \(B\) before \(A\).
- A run draws one random order at its start and uses it for every sweep. Is this a randomized order between sweeps? Explain.
- Name two choices that should be held constant when comparing fixed-order and randomized-order runs, and state why.
- Two schedules give similar estimates in a few runs. What is a valid conclusion, and what conclusion would go beyond the evidence?
Solutions. (1) There are \(4!=4\times3\times2\times1=24\) orders; under a uniform rule, a specified order has probability \(1/24\). (2) The second variable updated can use the first variable's new value, whereas the first update uses the other variable's value from before its update in that sweep; the inputs to the conditional updates can therefore differ by order. (3) No. The order was selected randomly across possible runs, but remains fixed within this run; a new order must be drawn for each sweep to randomize between sweeps. (4) For example, hold the conditional models and conditioning sets constant. Otherwise, a difference could reflect a changed model specification rather than the schedule. Also keep the data and analysis plan constant for the same reason. (5) One may conclude that no important difference was evident for the examined schedules, runs, and estimates. One may not conclude that the schedules are equivalent or that the imputation procedure is valid or converged solely from that agreement.