Feedback Loops in Chained Updates

How an Update Becomes Feedback

In fully conditional specification (FCS), an imputed value is drawn from a conditional model, then may be used as a predictor when another variable is updated. If that later value is used as a predictor in the next sweep, the influence has travelled around a loop: a draw changes a later draw, and that later draw can change the first variable’s next draw. This is feedback in the algorithm. It describes how the chain’s state is updated, not a causal effect among the variables.

The key to tracing this process is to distinguish the state before a sweep from the values drawn during it. In the example below, \(X_t,Y_t,Z_t\) denote the values at the end of sweep \(t\), and therefore the starting state for sweep \(t+1\). The initial state is labelled with subscript 0. This convention prevents an initial value and a new draw from being assigned the same label.

For clarity, we will use a deliberately small hypothetical example with three binary variables and one record for which all three values are missing. Treat the conditional probabilities below as fixed and known for this illustration. A real analysis would estimate or draw conditional-model parameters, and would update only missing cells; observed cells remain fixed. This example isolates feedback through the imputed values rather than parameter uncertainty or other parts of a fitted MICE analysis.

A Small FCS System

Use the fixed update order \(X\), then \(Y\), then \(Z\). Each update uses the most recently available values: the new \(X\) is used to update \(Y\), and the new \(X\) and new \(Y\) are used to update \(Z\). The hypothetical conditional models are Bernoulli distributions:

Each right-hand side lies between 0 and 1 for binary predictors, so each defines a valid Bernoulli conditional distribution. A value of 0 has the complementary probability. In this toy system, each draw uses a fresh independent uniform random number, conditional on the current predictors. Those assumptions make it possible to calculate a path probability directly. They are simplifying assumptions for the example, not a claim that fitted MICE conditionals in general have fixed parameters or independent draws across the full analysis.

For example, when \(Y=0,Z=0\), the probability of drawing \(X=1\) is 0.10. If \(Y=1,Z=0\), that probability is 0.50. Thus the current values of \(Y\) and \(Z\) affect the conditional distribution used for \(X\). The same kind of dependence operates within a sweep and carries into the next one.

Transition Probabilities and Their Justification

A sweep can be represented as a transition from an old state \(s=(x,y,z)\) to a new state \(s'=(x',y',z')\). Write \(q_X,q_Y,q_Z\) for the three Bernoulli conditional probability functions above. Under the stated fixed-model and fresh-draw assumptions, the probability of that particular transition is

Here \(q_X(x'\mid y,z)\) means the probability assigned to the specific value \(x'\), not just to 1: it is the specified probability if \(x'=1\), and its complement if \(x'=0\). The second factor conditions on the newly drawn \(x'\), while the third conditions on both newly drawn values. This follows by applying the conditional probability multiplication rule in update order: first the probability of the \(X\) draw, then the probability of the \(Y\) draw given the information available at that update, then the probability of the \(Z\) draw given its available values.

The product also normalizes over all possible new states. Summing first over \(z'\) makes the last factor sum to 1 for every \(x',y'\); summing next over \(y'\) makes the middle factor sum to 1 for every \(x'\); and summing over \(x'\) makes the first factor sum to 1. Therefore \(\sum_{s'}K(s,s')=1\). The chain’s next-state distribution is determined by the current state in this fixed-parameter illustration.

Reading a Feedback Trace Carefully

A useful trace records more than the final completed value. For each sweep, retain the state at its start, the conditional probability used for each update, the uniform or random draw if available, and the updated value. This makes it possible to distinguish a changed conditional probability from a changed random outcome. It also exposes whether the implementation used the correct version of each predictor.

Several checks help locate failures. Verify that each conditional probability is between 0 and 1, that complements are used for zero-valued draws, and that every predictor is the latest value available at that update. Check that observed values remain unchanged and that the recorded end state of one sweep is the next sweep’s start state. If manually recalculated probabilities disagree with stored probabilities, suspect a stale predictor, incorrect indexing, or a coding error before attributing the discrepancy to statistical behavior.

Trace inspection can also reveal that the chosen conditional probabilities are highly sensitive to particular predictor values: small changes in one state component may produce large differences in later draw probabilities. That is a reason to scrutinize how the models and predictors are specified and to examine the chain’s behavior with appropriate diagnostics. A short path alone cannot establish whether the chain has explored its relevant state space adequately. In particular, a few consecutive feedback steps do not establish convergence or compatibility.

What Feedback Does—and Does Not—Tell You

Repeated chained updates differ from a single pass through the variables. A single pass produces one sequence of updates; repeating sweeps allows the completed values from one pass to influence the next. It therefore includes feedback through successive states, but a single traced path does not reveal the long-run distribution. The introductory material on autocorrelation between FCS iterations provides related context for dependence across iterations.

FCS specifies a collection of conditional models. If those conditionals are compatible with a joint distribution, the joint structure can provide one way to understand the target of the updates, subject to the relevant conditions. If they are incompatible, the sequence still defines update steps, but the existence of a common joint distribution cannot be assumed. A calculation such as \(K(s,s')\) describes the transition for the specified fixed conditionals; it does not prove compatibility or guarantee good exploration.

In a practical MICE run, conditional-model parameters may themselves be estimated or drawn, and the update system can be more complex than the fixed-probability illustration. Then the value state alone may not determine the next transition probabilities unless parameter state is also accounted for. The example’s product formula remains a transparent way to understand how value updates feed into one another, but its numerical path probabilities should not be transferred to a fitted analysis.

Common Mistakes and Practical Limits

• Reusing an index for the initial and updated state. Calling the initial state \(X_1=0\) and the first sweep’s new draw \(X_1=1\) makes the trace contradictory. Label the initial state with subscript 0 and the first sweep’s draws with subscript 1.
• Treating a probability increase as a guaranteed outcome. Raising \(P(Y=1)\) from 0.10 to 0.60 makes a draw of 1 more likely, not certain. The random draw must still be shown when tracing a realized path.
• Calling algorithmic dependence a causal effect. A conditional update makes one value computationally relevant to another draw. Causal interpretation requires scientific design and assumptions beyond this update mechanism.
• Assuming a loop proves a joint model exists. Chained dependencies describe the order and inputs of the updates. They do not establish compatibility among the conditional distributions.
• Overinterpreting a short trace. One path illustrates the mechanism but cannot by itself show that long-run behavior is satisfactory or representative.

The main alternative is to specify and sample from a joint model directly, rather than update a set of conditionals in sequence. A joint model makes the joint probability structure explicit, whereas FCS offers separate conditional specifications. Neither approach is automatically preferable in every application: the choice depends on the substantive model, its assumptions, and whether the chosen conditionals are suitable. In either approach, a feedback trace is a diagnostic aid, not a substitute for assessing the model and the chain.