Updating Imputations With Latent Cluster Effects

Why the Cluster Effect and the Imputation Must Be Updated Together

In multilevel chained equations, a missing value in one cluster is informed partly by that cluster’s observed and currently imputed values. A latent cluster effect summarizes some of this shared information. The dependence runs both ways: the current values inform the effect, and the updated effect informs the next imputation. Treating either part as fixed throughout the iterations can therefore give a misleading picture of the update.

Consider a continuous target \(Y_{ij}\), where \(i\) indexes individuals and \(j\) indexes clusters. A useful working model for one conditional update is a random-intercept model:

Y_{ij} = \beta_0 + \boldsymbol{x}_{ij}^{\mathsf T}\boldsymbol{\beta} + b_j + \varepsilon_{ij}, \qquad b_j \sim N(0,\tau^2), \qquad \varepsilon_{ij} \sim N(0,\sigma^2).

Here \(\boldsymbol{x}_{ij}\) contains the predictors in the conditional model, \(b_j\) is the latent cluster effect, and \(\varepsilon_{ij}\) is an individual-level residual. For this model, assume the random effects are independent across clusters, the residuals are independent across individuals conditional on the effects and predictors, and the random effects and residuals are independent. The model also assumes the conditional mean and variance structure are suitable for the target variable. The calculations below hold the regression coefficients and variance components fixed to isolate the update.

Definition: A latent cluster-effect update uses the current completed values and model parameters to update the cluster-specific effect \(b_j\). An imputation update then uses the current effect, predictors, and conditional model to update a missing \(Y_{ij}\). Repeating these updates lets the current imputed values and cluster effects inform one another.

In an actual multilevel imputation algorithm, updates may include stochastic draws and changes to model parameters. Here, conditional means make the feedback transparent; iterating conditional means alone is not a substitute for proper stochastic multiple imputation. The subsequent material on drawing random effects treats that additional uncertainty in detail.

Conditional Update for a Gaussian Random Intercept

Suppose the current completed values in cluster \(j\) are \(y_{ij}\), and temporarily define the residual after removing the fixed part of the model as \(r_{ij}=y_{ij}-\beta_0-\boldsymbol{x}_{ij}^{\mathsf T}\boldsymbol{\beta}\). Conditional on these values and fixed parameters, the density for \(b_j\) is proportional to its normal prior density multiplied by the residual likelihood:

p(b_j \mid \boldsymbol{y}_j,\boldsymbol{x}_j) \propto \exp\left\{-\frac{b_j^2}{2\tau^2} -\frac{1}{2\sigma^2}\sum_i(r_{ij}-b_j)^2\right\}.

Collecting the terms involving \(b_j\) gives a normal conditional distribution. Its variance is the reciprocal of the total precision, and its mean is that variance times the likelihood-weighted residual sum:

b_j \mid \boldsymbol{y}_j,\boldsymbol{x}_j \sim N\left( \frac{\tau^2\sum_i r_{ij}}{\sigma^2+n_j\tau^2}, \frac{\sigma^2\tau^2}{\sigma^2+n_j\tau^2} \right).

The sum and \(n_j\) refer to the current completed cluster data used in this update. In particular, if a missing response currently has an imputed value, that value contributes a residual to the sum. The conditional mean is pulled toward zero by the prior and is more responsive to the cluster’s residuals when the data provide more information relative to the prior variance.

Justification: The exponent’s quadratic terms in \(b_j\) are \(-\tfrac{1}{2}(1/\tau^2+n_j/\sigma^2)b_j^2 + b_j\sum_i r_{ij}/\sigma^2\), plus terms not involving \(b_j\). Completing the square gives variance \((1/\tau^2+n_j/\sigma^2)^{-1}=\sigma^2\tau^2/(\sigma^2+n_j\tau^2)\) and mean \(\sum_i r_{ij}/\sigma^2\) times that variance, which is \(\tau^2\sum_i r_{ij}/(\sigma^2+n_j\tau^2)\).

Once a current effect is available, the conditional distribution for an individual response is \(Y_{ij}\mid b_j,\boldsymbol{x}_{ij}\sim N(\beta_0+\boldsymbol{x}_{ij}^{\mathsf T}\boldsymbol{\beta}+b_j,\sigma^2)\). Its conditional mean supplies a simple way to trace the direction of the next update. A stochastic imputation would also account for the conditional residual variation rather than always replacing a missing value by its mean.

Worked Example: Updating an Effect From Current Cluster Values

Suppose a hypothetical cluster has a random-intercept model with \(\beta_0=10\), no varying predictors, \(\sigma^2=4\), and \(\tau^2=9\). Two observed responses are 13 and 15, and a third response is missing. Its current imputed value is 10. Find the conditional mean of the cluster effect using all three current values.

The residuals relative to the fixed intercept are \(13-10=3\), \(15-10=5\), and \(10-10=0\). Thus \(\sum_i r_{ij}=8\) and \(n_j=3\). The conditional mean is

E(b_j\mid \boldsymbol{y}_j) =\frac{\tau^2\sum_i r_{ij}}{\sigma^2+n_j\tau^2} =\frac{9(8)}{4+3(9)} =\frac{72}{31} \approx 2.323.

The conditional variance is \(\sigma^2\tau^2/(\sigma^2+n_j\tau^2)=36/31\approx1.161\), so the conditional standard deviation is about \(1.078\). The positive conditional mean reflects the cluster’s positive residual sum, while its magnitude is moderated by the zero-centered random-effect distribution. These values are in the units of \(Y\).

Using the Updated Effect to Impute a Missing Response

With the conditional mean \(2.323\) used as the current effect for this illustration, the missing response’s conditional mean is the intercept plus the effect. The predictor contribution is zero in this example:

Worked Example: Updating the Missing Value From the Effect

Continue with the hypothetical cluster above. Calculate the conditional mean for its missing response, using the updated effect’s conditional mean as the effect value.

E(Y_{\mathrm{mis},j}\mid b_j) =\beta_0+b_j =10+2.323 =12.323.

Thus the value informing the next effect update would be about 12.323 if this conditional-mean illustration were used. In a stochastic imputation, the missing response would not generally be set deterministically to 12.323: the conditional residual variance, as well as uncertainty in model quantities and latent effects, matters. The number here isolates the direction and scale of the feedback.

This update is conditional on the current effect and predictors. For a model with predictors, the same calculation adds \(\boldsymbol{x}_{ij}^{\mathsf T}\boldsymbol{\beta}\). The predictors and their values must be appropriate for the current chained-equation step: if a predictor is itself missing, the current imputation of that predictor is used.

Tracing One Feedback Cycle

The next example starts from the initial imputed value of 10, updates the effect, updates the missing value, and then recomputes the effect. Holding parameters fixed makes it possible to see exactly how the current imputation feeds back. It does not represent a full stochastic MICE chain.

Worked Example: Following the Effect–Imputation Feedback

Use the same hypothetical data: observed responses 13 and 15, intercept 10, \(\sigma^2=4\), and \(\tau^2=9\). Begin with the missing response set to 10. The first effect update is the calculation above, \(b_j=72/31\approx2.323\). The corresponding updated missing value is \(m=10+2.323=12.323\).

Now use this new current value in the next effect update. The three residuals are 3, 5, and \(12.323-10=2.323\). Their sum is approximately 10.323, so

b_j^{\mathrm{new}} =\frac{9(3+5+2.323)}{4+3(9)} =\frac{92.907}{31} \approx 2.997.

The next conditional-mean imputation is \(10+2.997=12.997\). The updated imputation has moved the cluster effect upward because its residual is positive; the higher effect, in turn, raises the next imputed value. For these fixed parameters, the conditional-mean iteration has a stable fixed point. At that point \(m=10+b_j\) and \(b_j=9(3+5+(m-10))/31\), so

b_j=\frac{9(8+b_j)}{31}, \qquad 22b_j=72, \qquad b_j=\frac{72}{22}\approx3.273, \qquad m\approx13.273.

This fixed point describes repeated conditional-mean substitution only. It is not the distribution of the missing value and should not be reported as if it were a stochastic multiple imputation. A proper chain has variation in updates and, in general, uncertainty in parameters and random effects.

Assumptions, Diagnostics, and Alternatives

The Gaussian formulas require the stated normal random-intercept model and known fixed parameters for the particular conditional calculation. Applying the model to imputation also requires the conditional specification to be suitable for the data and analysis, including relevant cluster structure and predictors. For missing-data validity, the missingness assumptions must be defensible given the variables included in the imputation models; the random-intercept equation alone does not establish them. If slopes vary across clusters, variances differ materially by cluster, or the outcome is binary or a count, this normal random-intercept update is not the appropriate conditional model without modification.

Diagnostics should examine the iterative system rather than only the final filled-in table. In particular:

Track cluster-effect summaries and imputed-value summaries across iterations and chains. Persistent drift, poor mixing, or chains that remain separated can indicate unstable updating.
Compare observed and imputed values by cluster and by important predictors. Systematic discrepancies, implausible values, or a changed within-cluster pattern can signal a poorly specified conditional model.
Check whether the effect updates are dominated by a few clusters or change sharply with initial imputations. Sparse information within clusters can make conditional effects sensitive to assumptions.
Inspect whether the implied within- and between-cluster variation is plausible. A model that omits important cluster structure may produce misleadingly narrow or poorly calibrated imputations.

The main alternatives make different trade-offs. A pooled model without cluster effects is simpler but can miss within-cluster dependence and between-cluster variation. Cluster fixed effects allow separate cluster intercepts but can be unstable in small or sparse clusters and do not directly represent a distribution of effects for imputation. A latent random-effect model partially pools cluster information through its distributional assumptions. These choices are not interchangeable diagnostics: the structure should follow the data and the intended analysis.

Approach	How cluster information enters	Important limitation
Pooled conditional model	No cluster-specific latent effect	May fail to preserve dependence and cluster differences
Cluster fixed effects	A separate intercept for each cluster	Can be unstable with sparse data and many clusters
Latent random intercept	Current values update a cluster effect under a shared distribution	Depends on the random-effect and conditional-model assumptions

A common mistake is to compute a cluster effect from observed responses once, then keep it fixed while updating missing responses. That omits the contribution of current imputations to the effect and breaks the feedback the model specifies. The correction is to use the current completed data for the effect update, then use the updated effect in the missing-value update, repeating the appropriate steps within the chain. A related mistake is to treat the conditional mean as the imputation itself; conditional-mean substitution omits residual and parameter uncertainty and can understate variability.

Key takeaway: In a multilevel chained-equation update, current completed values inform latent cluster effects, and those effects inform missing values. Trace this dependence explicitly, while distinguishing a conditional-mean illustration from the stochastic updates needed for multiple imputation.

Check Your Understanding

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

In the random-intercept model, what quantities from the current completed cluster data determine the conditional mean of \(b_j\)?
For a cluster with \(\beta_0=20\), \(\sigma^2=9\), \(\tau^2=4\), two observed responses 18 and 20, and one missing response currently imputed as 20, calculate the conditional mean of the effect.
Using your answer to question 2 as the effect value, what is the conditional-mean update for the missing response?
Why is repeatedly setting missing values to conditional means not, by itself, proper multiple imputation?

Solutions. (1) The conditional mean depends on the residual sum \(\sum_i r_{ij}\), the number of current completed values \(n_j\), and the variance components \(\sigma^2\) and \(\tau^2\). (2) The residuals are \(-2,0,0\), so the sum is \(-2\) and \(n_j=3\). The effect mean is \(4(-2)/(9+3(4))=-8/21\approx-0.3810\). (3) The missing response mean is \(20-0.3810=19.619\), in the outcome’s units. (4) Deterministic conditional-mean substitution omits the conditional residual variation and uncertainty in model quantities and latent effects; it does not generate the stochastic variation required for proper multiple imputation.