Standard Errors for Pattern-Weighted Summaries

Why a Weighted Summary Needs Its Own Standard Error

Pattern-specific estimates answer questions about outcomes within dropout patterns. A population-level summary combines those estimates using the pattern proportions. Because both ingredients may be estimated, the uncertainty of the combined result is not generally the uncertainty of a pattern-specific mean, nor is it obtained by treating the weights as known.

Let \(G\) denote a participant’s dropout pattern, with \(K\) possible patterns. For pattern \(k\), let \(\pi_k=P(G=k)\) be its population proportion and let \(\mu_k\) be the outcome mean for that pattern under the analysis being performed. The pattern-weighted population mean is

In practice, the analysis uses estimates \(\hat{\pi}_k\) and \(\hat{\mu}_k\), giving \(\hat{\mu}=\sum_k\hat{\pi}_k\hat{\mu}_k\). This tutorial concerns the sampling uncertainty of this plug-in summary, conditional on the model and assumptions used to obtain its components. It does not determine whether a pattern-specific mean is identified or whether a particular missing-data assumption is credible.

Delta-Method Variance

The delta method approximates the variance of a smooth function of estimated parameters by linearizing that function near the estimates. Since the proportions sum to one, use \(K-1\) free weight parameters and define \(\pi_K=1-\sum_{j=1}^{K-1}\pi_j\). Write the parameter vector as \(\theta=(\pi_1,\ldots,\pi_{K-1},\mu_1,\ldots,\mu_K)\), and define \(h(\theta)=\sum_k\pi_k\mu_k\).

The derivative with respect to a free proportion \(\pi_j\) accounts for the fact that increasing it decreases \(\pi_K\). Thus \(\partial h/\partial\pi_j=\mu_j-\mu_K\), while \(\partial h/\partial\mu_k=\pi_k\). The gradient is therefore

If \(\hat{V}\) estimates the joint covariance matrix of \(\hat{\theta}\), the delta-method variance estimate is the gradient evaluated at the estimates, multiplied on both sides of \(\hat{V}\). The standard error is the square root of that variance estimate.

The joint covariance matrix matters. Estimating pattern proportions and means from the same participants can make them dependent. Even if they are estimated separately, their estimated means may be correlated through a shared model or shared data. Dropping covariance terms is justified only when the corresponding covariances are zero under the estimation design or an appropriate approximation.

Worked Example: Combining Joint Covariance Information

Why the Covariances and Weights Matter

For two patterns, the estimate can be written \(\hat{\mu}=\hat{\pi}_1\hat{\mu}_1+(1-\hat{\pi}_1)\hat{\mu}_2\). A first-order expansion around the true values shows the distinct sources of error:

The first term is error from estimating the pattern proportion; the remaining terms are error from estimating the pattern means. Taking the variance of this sum produces variance terms and covariances. This also explains why weight uncertainty has little first-order effect when the pattern means are nearly equal: its multiplier, \(\mu_1-\mu_2\), is then small. When the pattern means differ substantially, uncertainty about the pattern mix can matter considerably.

In a simple independent sample with observed outcomes in every pattern, there is a useful special case. If the pattern proportions are estimated by sample proportions and pattern means by within-pattern sample means, the weighted estimate equals the overall sample mean. Its sampling variance includes variation both within and between patterns. If instead pattern counts are fixed by design and inference conditions on those counts, the pattern proportions are fixed for that calculation, and only the within-pattern mean uncertainty contributes. These are different sampling setups, not competing formulas for the same setup.

Worked Example: Conditional and Unconditional Sampling

Bootstrap Standard Errors

A bootstrap estimates sampling variability by repeatedly resampling the observational units and recomputing the statistic. For each of \(B\) replicates, calculate the pattern proportions, refit or recompute the pattern-specific means, and form the weighted summary \(\hat{\mu}^{*(b)}\). The bootstrap standard error is the sample standard deviation of those replicate summaries:

Resampling must match the sampling design and repeat the steps that produced the original estimate. With independent participants sampled from a population, resample participants and recalculate both the weights and means. If the design fixes pattern counts, a stratified resampling scheme within patterns may be appropriate for inference conditional on those counts. Resampling pattern-specific estimates while holding estimated weights fixed omits weight uncertainty; resampling only fitted means may also miss their dependence on the weights.

Choosing and Checking an Approach

The delta method is often convenient when the estimator and a reliable joint covariance matrix are available. It is fast and makes the roles of the weights, means, and covariances explicit. The bootstrap is useful when the estimator is a complicated function of fitted pattern-specific models or when deriving the joint covariance is difficult. It can reflect nonlinearity in the estimator, but it cannot repair an inappropriate model, an invalid resampling design, or a lack of information in a sparse pattern.

Before interpreting either standard error, check the following:

• Weight definition and sum. Confirm that the estimated pattern proportions correspond to the target population and sum to one. If one pattern is omitted, the remaining weights must be normalized only when that is part of the intended estimand.
• Joint uncertainty. Check whether the covariance calculation includes estimated weights, pattern means, and their dependence. A covariance matrix for the means alone is not enough when the weights are also estimated.
• Pattern support. Very small pattern samples can produce unstable means or model fits. In bootstrap replicates, some patterns may be absent, causing a fit or the weighted summary to fail.
• Resampling design. Decide whether the target sampling scheme treats pattern counts as random or fixed. The resampling scheme must reproduce that choice and any clustering or stratification in the design.
• Approximation quality. Strong skewness, boundary estimates, or a highly nonlinear summary can make a symmetric normal approximation based on the delta method unreliable. Bootstrap replicates should also be inspected for outliers, failures, or a distribution very different from the original estimate.

A standard error describes sampling variation under the stated model and design. It does not measure uncertainty about the assumptions that justify the pattern-specific means, nor does it account automatically for misspecification. A small standard error can therefore accompany a result that is highly sensitive to the assumptions used to estimate an unobserved outcome.

Common Pitfall and Practical Limitations

Wrong reasoning: “The pattern means have standard errors, so I can multiply each by its weight, combine those standard errors, and ignore the weights because they add to one.” This is not generally correct. The weights may be estimated, the means may be correlated, and the uncertainty of a sum depends on covariance as well as on marginal variances.

Correction: Define the complete estimator first, then use a joint covariance calculation or a bootstrap that recomputes the complete estimator. Treat weights as fixed only when they are fixed by the design or when the inferential target explicitly conditions on them. In a fitted pattern-mixture analysis, the resulting uncertainty remains conditional on the model and assumptions used to estimate the pattern-specific quantities.