Singleton Clusters In Multilevel Imputation

Why Singleton Clusters Need a Separate Interpretation

A cluster with one observed member is not an empty cluster: its record may inform population means, associations between cluster means and outcomes, and the marginal variability of the response. But it cannot provide a comparison between two members of that cluster. This distinction matters in multilevel imputation because a model may use the same record for some parts of the conditional distribution while having no within-cluster contrast to learn from.

The key question is therefore not simply whether to retain a singleton. It is which model quantities the singleton can inform, and whether the data contain enough replicated clusters to distinguish those quantities. The discussion below uses the random-intercept normal model; it does not assume random slopes.

What Within-Between Decomposition Shows

For a level-1 predictor \(x_{ij}\), write its value as the cluster mean plus its deviation from that mean:

For a singleton, the only observed value is also the observed cluster mean, so \(x_{ij}-\bar{x}_{j}=0\). Thus that record supplies no nonzero within-cluster predictor deviation. It can still contribute to an association involving \(\bar{x}_{j}\), provided the between-cluster model and estimand make that association relevant. The same logic applies to the outcome: a single observed outcome can contribute to the cluster mean, but there is no within-cluster outcome contrast for that cluster.

This statement concerns observed data and the decomposition being used. If some cluster members are missing, the mean calculated from observed members may not equal the cluster’s underlying mean. Imputation and the model for the mean must account for that uncertainty rather than treating an observed-subset mean as known without qualification.

What a Singleton Says About a Random Intercept

Suppose the outcome follows a random-intercept model with known covariate contribution \(\mu_{ij}\):

where the cluster effects \(u_j\) are independent across clusters, the residuals \(e_{ij}\) are independent across members and clusters, and the two sets of effects are mutually independent. The covariate contribution \(\mu_{ij}\) may include fixed effects. For the calculations that follow, assume the variance components and fixed effects are known. In practice they are estimated, and their uncertainty also matters for proper imputation.

For a singleton with observed outcome \(y_j\), let \(r_j=y_j-\mu_j\). The observation is \(r_j=u_j+e_j\). It gives information about \(u_j\), but that information is noisy: the one residual combines the cluster effect and individual residual. Under the normal model, the conditional distribution is

This is a shrunken estimate of the cluster effect, not a direct observation of it. The conditional variance remains positive when both variance components are positive. A singleton therefore can update the distribution of its cluster effect, while leaving substantial uncertainty about how much of its residual is cluster-level and how much is individual-level.

Singletons and Variance Components

The covariance structure makes a second limitation precise. Under the random-intercept model, a singleton’s marginal outcome variance is \(\tau^2+\sigma^2\). A cluster with two members has covariance matrix

The off-diagonal covariance between two different members is \(\tau^2\), because they share the same random intercept but have independent residuals. A singleton has no such pairwise covariance. Consequently, if every cluster is a singleton and there are no other sources of information that distinguish the components, the outcome data identify only the sum \(\tau^2+\sigma^2\), not the two variances separately. Replicated clusters can provide the within-cluster covariance needed for that separation under the stated assumptions.

Implications for Imputation and Model Choice

In multilevel chained equations, the conditional imputation model should preserve the distinctions relevant to the analysis. A singleton’s observed predictors and outcome can contribute to fixed-effect and between-cluster relationships. If its outcome is missing, however, that missing value cannot itself update the cluster random intercept. Under the model above, with no other observed variable informing that effect, the random intercept remains drawn from its prior distribution \(N(0,\tau^2)\) before the missing outcome is generated. Any observed auxiliary variable can change this conclusion only if the imputation model specifies how it relates to the cluster effect.

A cluster-indicator model gives a different view. A singleton has its own indicator, but there is only one observation associated with that indicator; it cannot create a within-cluster comparison. Treating each singleton’s effect as an unrestricted fixed parameter is not equivalent to drawing a random effect from a population distribution. Conversely, excluding all singleton clusters removes valid observations that may inform population means and between-cluster associations. Neither choice is automatically appropriate: the decision should follow the target estimand and the assumptions of the imputation and analysis models.

The assumptions behind the random-intercept calculations include a correctly specified conditional mean, normally distributed independent random intercepts and residuals, and positive variance components. The imputation model must also be appropriate for the missingness and analysis goals. These assumptions are not verified merely because a software routine converges. In particular, the normal random-intercept calculation above does not establish that a singleton’s effect is known accurately, or that a model with very few replicated clusters can estimate the variance components reliably.

Common Mistakes and Limits

• Wrong reasoning: “A singleton contributes nothing to a multilevel model.” Correction: It supplies no within-cluster contrast, but may inform the fixed mean, a between-cluster association, or the marginal outcome distribution.
• Wrong reasoning: “The singleton’s observed residual is its random intercept.” Correction: The residual is the sum of the random intercept and individual error. The conditional random-intercept distribution reflects uncertainty in that separation.
• Wrong reasoning: “One singleton can identify both variance components because its outcome is observed.” Correction: Its marginal variance is \(\tau^2+\sigma^2\); separating the components requires information beyond that singleton variance, such as replicated clusters under the model.
• Wrong reasoning: “Dropping singletons is always safer.” Correction: Dropping them can discard useful between-cluster information and may change the target population. Assess the estimand, missingness, and model behavior instead.

Singletons have a real but limited role. They can add information about the distribution of observed records across clusters, yet cannot replace replication for learning within-cluster relationships or shared covariance. Their contribution to an imputation depends on what is observed, how the model represents cluster effects, and whether enough non-singleton clusters support the quantities being estimated.