Variance Reduction

Variance Reduction: Concept Architecture

Orientation and learning roadmap

The page first identifies the random error that remains when a simulation is run for a finite number of replications. It then develops methods for estimating a treatment difference more precisely, gives a checkable paired example, and shows how to test whether a technique helped. The closing sections separate computational precision from uncertainty about evidence and model assumptions.

Locate the variation that can be reduced

A patient-level model may produce different average costs or health outcomes each time it is run with a new random stream, even though the inputs and decision problem are unchanged. This variation in a finite-simulation estimate is Monte Carlo error. A reduction technique changes the sampling or coupling of simulated observations so that the estimator of the chosen target has less variance for a specified computational budget while retaining its intended validity.

For independent and identically distributed replications $Y_1,\ldots,Y_n$ with finite variance $\sigma^2$, the ordinary sample mean has variance $\sigma^2/n$. This result provides a benchmark for the number of replications needed to improve precision. Its independent-replication assumption does not apply unchanged to deliberately paired or stratified observations.

$$\widehat{\mu}=\frac{1}{n}\sum_{i=1}^{n}Y_i,\qquad \operatorname{Var}(\widehat{\mu})=\frac{\sigma^2}{n}.$$

The relevant target in economic evaluation is often an incremental cost, incremental QALY, or incremental net benefit rather than an outcome for one strategy alone. Variance reduction should therefore be judged against the uncertainty of that incremental estimator. A method that makes each arm separately look smoother need not improve the treatment contrast.

Pair simulated patients across strategies

Common random numbers (CRN) use matched random inputs for comparable uncertainty sources in competing strategies. This induces a covariance between simulated outcomes in the two arms. When that covariance is positive, subtracting paired outcomes can remove shared noise; the result must be checked because a poorly aligned model can instead produce little benefit or even a loss of precision.

Let $A_i$ and $B_i$ be simulated outcomes under strategies A and B for paired replication $i$, and let $D_i=B_i-A_i$. If pairs are independent across $i$, the variance of the mean paired difference is

$$\operatorname{Var}(\overline D)=\frac{\operatorname{Var}(B)+\operatorname{Var}(A)-2\operatorname{Cov}(A,B)}{n}.$$

A positive covariance decreases the variance of the difference compared with independently generated arm outcomes of the same marginal variances and sample size. A negative covariance increases it. Keep distinct random streams for different stochastic mechanisms, align clinically comparable events across strategies, and document what happens when pathways diverge or an intervention creates or removes an event.

A paired cost comparison

Suppose the per-person cost variances for A and B are each $100$ currency units squared, and the covariance under matched streams is $80$ currency units squared. With $n=100$ independent pairs, the variance of the estimated mean cost difference is $(100+100-2\times80)/100=0.4$ currency units squared. Independent arm streams with zero covariance would yield $2.0$ under the same marginal variances and 100 observations per arm.

The corresponding Monte Carlo standard errors are approximately $0.632$ and $1.414$ currency units, respectively. Thus pairing improves precision for this illustrative contrast by a variance factor of five; it does not change the expected treatment difference merely because the estimated difference is less noisy. The example assumes each pair is independent of the other pairs and the paired costs have the stated covariance.

Use other techniques only when their conditions fit

Different techniques alter the simulation estimator in different ways. Antithetic sampling creates deliberately complementary draws, while a control variate uses an auxiliary quantity whose expectation is known. Stratified and importance sampling alter which parts of the input distribution are sampled and require the correct estimator or weights to preserve the target.

• Antithetic variates: Pair a uniform draw $U$ with $1-U$ and average the paired outputs; improvement requires sufficiently negative covariance between paired outputs and cannot be assumed merely from complementary inputs.
• Control variates: If $X$ has known expectation $\mu_X$, use $\widehat\mu_{\mathrm{CV}}=\overline Y-\beta(\overline X-\mu_X)$; a correlated control and an appropriate coefficient $\beta$ can improve precision without changing the target expectation.
• Stratified sampling: Allocate draws across predefined portions of a distribution and combine stratum estimates with their probability weights; unweighted aggregation can estimate the wrong quantity.
• Importance sampling: Oversample rare but consequential events under a proposal distribution and reweight observations by the target-to-proposal density ratio; unstable weights can worsen precision or hide inadequate tail coverage.

Fixed seeds are valuable for reproducibility and for constructing matched runs, but merely repeating a run with the same seed does not create new independent evidence. Quasi-Monte Carlo sequences have their own error assessment requirements; their deterministic spacing should not be treated as if it were a sample of independent random replications.

Check precision against the decision question

The first check is whether the estimator still targets the specified expected cost, QALY, or net benefit under the model. The second is whether the Monte Carlo variance of the relevant comparison falls at a fixed computational cost, including any setup or additional model evaluations. Estimate Monte Carlo error using independent batches or another uncertainty calculation that respects pairing and the dependence introduced by the method.

For a paired incremental net benefit estimate, retain the within-pair differences and compute the standard error from the variation between independent pairs, not from treating both arms' observations as unrelated. Record the random-stream scheme, seed management, number of independent replications or batches, wall time, and estimated precision. A pilot comparing candidate methods at a common computational budget can reveal when a theoretically plausible method performs poorly in the actual model.

In a probabilistic sensitivity analysis (PSA), parameter draws represent uncertainty about inputs, whereas random patient histories conditional on a parameter draw introduce inner simulation noise. Pair strategies within each parameter draw if the model supports it, and ensure enough inner replications that Monte Carlo error does not obscure the outer distribution of decision outcomes. Reusing a single parameter draw across strategies is part of a fair comparison; it does not eliminate parameter uncertainty.

Interpret the gain without overstating certainty

Lower Monte Carlo error makes a finite simulation more stable, but it does not repair misspecified transition probabilities, missing costs, biased clinical evidence, or structural uncertainty. Likewise, a narrower interval for the simulation estimate alone is not a confidence or credible interval for the real-world intervention effect. Report which uncertainty has been reduced and which remains.

A useful efficiency measure is the ratio of the baseline estimator variance to the new estimator variance at equal cost, or the ratio of variance times computational cost when methods require different effort. Gains depend on the output and model: common random numbers can substantially sharpen incremental results while offering no comparable benefit for the stand-alone mean of either strategy. Interpret the precision gain for the decision quantity that the analysis is meant to support.

Evidence used for this explanation

These primary sources support the distinction between simulation error and substantive uncertainty and the treatment of common random numbers in health economic models. They also document the practical importance of stream alignment and simulation validation. Source descriptions are provided so the methodological claims can be checked without treating a citation as proof of a model's own numerical outputs.

• Weinstein MC et al. “Principles of Good Practice for Decision Analytic Modeling in Health-Care Evaluation”. Value in Health (2003).
• “Using Common Random Numbers in Health Care Cost-Effectiveness Simulation Modeling”. Medical Decision Making (2013).
• “Keeping the Noise Down: Common Random Numbers for Disease Simulation Modeling”. Medical Decision Making (2009).