Root Mean Squared Error Across Trial Scenarios

Why Measure Error Across Trial Replicates?

A trial analysis can produce a treatment-effect estimate that differs from the value used to generate the trial. That difference may reflect systematic bias, the ordinary variability of estimates across repeated trials, or both. Root mean squared error (RMSE) summarizes the typical size of the estimation error while retaining the treatment effect’s units. It is therefore useful when assessing the accuracy of a point estimator across simulated trials.

In a Bayesian trial, the point estimate might be a posterior mean or another prespecified summary of the posterior distribution. RMSE evaluates that chosen estimator over repeated simulated data sets; it is not a summary of the posterior uncertainty within a single trial. The target must be the same estimand that the estimator is intended to estimate, expressed on the same scale and in the same units.

In simulation, the expectation is usually approximated by the average over the \(R\) replicates. Define the replicate mean estimate \(\bar{\hat{θ}}=R^{-1}\sum_{r=1}^{R}\hat{θ}_r\), the empirical bias \(b_R=\bar{\hat{θ}}-θ\), and the empirical variance with divisor \(R\), \(v_R=R^{-1}\sum_{r=1}^{R}(\hat{θ}_r-\bar{\hat{θ}})^2\). Then the empirical MSE and RMSE are

These are simulation estimates of the scenario’s MSE and RMSE, not exact population quantities. The convention of dividing by \(R\) in \(v_R\) makes the finite-replicate decomposition exact. If software instead reports the usual sample variance \(s^2\), with divisor \(R-1\), use \(v_R=(R-1)s^2/R\) in that decomposition.

Why RMSE Splits Into Bias and Variance

The population identity explains what contributes to RMSE. Write the error as \(\hat{θ}-θ=(\hat{θ}-E[\hat{θ}])+(E[\hat{θ}]-θ)\). The first term has expectation zero, so the cross-product vanishes when the squared error is averaged. Consequently,

This requires a finite second moment and a fixed target \(θ\). It does not require the estimator to be unbiased. Bias is the difference between the estimator’s repeated-sampling mean and the target; variance describes how estimates fluctuate around that repeated-sampling mean. Taking square roots gives \(RMSE=\sqrt{Var(\hat{θ})+Bias(\hat{θ})^2}\), not the sum of the standard deviation and the absolute bias.

For the simulated replicates, the same algebra gives an exact identity. Add and subtract \(\bar{\hat{θ}}\) inside each squared error. The cross-product sums to zero because \(\sum_r(\hat{θ}_r-\bar{\hat{θ}})=0\):

Thus, in a finite simulation, the MSE is exactly the empirical variance with divisor \(R\) plus squared empirical bias. The population identity is a property of the estimator’s sampling distribution; the empirical identity is an algebraic fact about the replicate values. A small number of replicates can make the empirical components unstable even though their identity holds exactly.

Compare the Components, Not Just the Total

Two scenarios can have the same RMSE while presenting different calibration concerns. A larger bias points to systematic displacement from the target; a larger variance points to less stable estimates across repeated trials. Reporting the components alongside RMSE makes that distinction visible. It also avoids interpreting a single summary as a complete account of estimator performance.

Worked Calculation on a Proportion-Difference Scale

The same calculations apply when the treatment effect is expressed as a risk difference. Keep the scale consistent: if the target is a difference in proportions, use estimates and errors in proportion units, or convert all of them to percentage points before calculating.

Diagnostics, Comparisons, and Limits

RMSE is interpretable only when the simulation design, estimator, and target are specified consistently. For every scenario, record the true estimand, the estimator used in each replicate, the number of attempted and analyzed replicates, and the units. Calculate each scenario’s metrics separately before making comparisons. Pooling estimates generated under different true effects can add between-scenario differences to the apparent spread and bias, so the resulting RMSE may not describe performance at any one scenario’s target.

Also inspect the replicate errors and their distribution. A few extreme estimates can dominate squared errors, making RMSE much larger than a typical absolute error. Check for coding or scale mistakes, numerical failures, and unexpectedly extreme estimates against the simulation’s analytic benchmarks and unit tests. If an analysis fails in some replicates, do not silently remove them: state how failures are handled and assess whether excluding them changes the estimand of the reported performance summary. A summary based only on successful fits describes that selected set of replicates unless the handling rule justifies a broader interpretation.

RMSE complements rather than replaces other summaries. Bias isolates average displacement but can be near zero when positive and negative errors cancel. Standard deviation measures spread around the replicate mean but ignores displacement from the target. Mean absolute error averages absolute errors and is less dominated by large errors than squared error, but it has different mathematical properties and is not the bias–variance decomposition above. Interval width and precision describe uncertainty summaries, not the repeated-simulation squared error of a point estimate. None of these quantities alone establishes whether a decision rule meets its operating requirements.

Monte Carlo error matters as well. With few replicates, the simulated RMSE can fluctuate simply because a different set of trials would produce different estimates. Increasing \(R\) generally stabilizes the Monte Carlo approximation, but the empirical decomposition does not itself measure that uncertainty. Report the replicate count and, when simulation precision matters, assess Monte Carlo uncertainty for the reported performance measures. Avoid treating extra decimal places as evidence of accuracy.

• Wrong: Add the squared empirical bias to the usual sample variance \(s^2\) and call the result the finite-replicate MSE. Correction: For the exact identity with \(R\) replicates, use \(v_R=(R-1)s^2/R\).
• Wrong: Pool estimates across true treatment effects and interpret their overall spread as sampling variance at one effect. Correction: calculate bias, variance, and RMSE within each fixed scenario, using its own true target.
• Wrong: Call RMSE “bias” or describe it as a posterior uncertainty interval. Correction: RMSE combines squared bias and sampling variance for a point estimator across repeated simulated trials; report the components to distinguish them.
• Wrong: Compare RMSE values on different scales without conversion. Correction: express the treatment effects and RMSEs in consistent units, and remember that MSE and variance use squared units.