---
title: 'Mean STD Sampling: Estimation & Variance Design'
url: https://www.emergentmind.com/topics/mean-std-sampling
type: topic
---

# Mean STD Sampling: Estimation & Variance Design

Mean STD Sampling denotes a family of procedures in which the mean and the standard deviation jointly determine estimation, inference, or sampling design. In systematic reviews and meta-analysis, it most commonly refers to reconstructing a study’s sample mean \(\bar{X}\) and sample standard deviation \(S\) from medians, ranges, and quartiles when primary studies report incomplete summaries rather than \((\bar{X},S)\) directly [1407.8038]. In a different Monte Carlo literature, it refers to estimating a mean \(\mu\) from i.i.d. samples when the relative standard deviation is bounded by a known constant \(c\), so that sample complexity is governed by \(c^2\epsilon^{-2}\ln(1/\delta)\) [1706.01478]. Related usages also appear in variance-aware sampling design, where standard deviations determine optimal allocation, ranked-set selection, or confidence-interval width [2408.15454, 2603.22598].

## 1. Core meta-analytic formulation

In continuous-outcome meta-analysis, the target inputs are the sample size \(n\), sample mean \(\bar{X}\), and sample standard deviation \(S\). A recurrent difficulty is that many trials instead report the median \(m\), minimum \(a\), maximum \(b\), and sometimes the first and third quartiles \(q_1,q_3\). Wan et al. formalized three reporting scenarios [1407.8038].

| Scenario | Reported summary | Principal information source |
|---|---|---|
| C1 | \(\{a,m,b;n\}\) | median and range |
| C2 | \(\{a,q_1,m,q_3,b;n\}\) | full five-number summary |
| C3 | \(\{q_1,m,q_3;n\}\) | quartiles and median |

The reconstruction target is
\[
\hat{\mu}\approx \bar{X}, \qquad \hat{\sigma}\approx S.
\]
In this setting, the main technical issue is that range-based and interquartile-range-based estimators depend strongly on \(n\). Wan et al. therefore replaced ad hoc range rules with order-statistic approximations that incorporate sample size continuously through normal quantiles [1407.8038].

For C1, the retained mean estimator is
\[
\hat{\mu}_{C1}=\frac{a+2m+b}{4},
\]
and the recommended standard deviation estimator is
\[
\hat{\sigma}_{C1}\approx \frac{b-a}{2\,\Phi^{-1}\!\left(\frac{n-0.375}{n+0.25}\right)}.
\]
For C2, the preferred mean estimator is
\[
\hat{\mu}_{C2}=\frac{a+2q_1+2m+2q_3+b}{8},
\]
while the standard deviation estimator combines range and interquartile range,
\[
\hat{\sigma}_{C2}\approx
\frac{b-a}{4\,\Phi^{-1}\!\left(\frac{n-0.375}{n+0.25}\right)}
+
\frac{q_3-q_1}{4\,\Phi^{-1}\!\left(\frac{0.75n-0.125}{n+0.25}\right)}.
\]
For C3, the mean and standard deviation estimators are
\[
\hat{\mu}_{C3}=\frac{q_1+m+q_3}{3},
\qquad
\hat{\sigma}_{C3}\approx
\frac{q_3-q_1}{2\,\Phi^{-1}\!\left(\frac{0.75n-0.125}{n+0.25}\right)}.
\]

These formulas are derived from expected normal order statistics, using Blom’s approximation
\[
E(Z_{(n)})\approx \Phi^{-1}\!\left(\frac{n-0.375}{n+0.25}\right),
\]
and analogous approximations for quartile order statistics. The practical significance is that the divisor in a range- or IQR-based SD estimator is no longer fixed; it varies with \(n\), which removes the arbitrary cutoffs used in earlier rules.

## 2. From ad hoc rules to order-statistic estimators

Earlier methods were simpler but less stable. Hozo et al. proposed the C1 mean approximation \((a+2m+b)/4\), together with range rules such as
\[
\hat{\sigma}\approx \frac{b-a}{4},
\]
or the piecewise rule
\[
\hat{\sigma}_{\text{Hozo}} \approx
\begin{cases}
\dfrac{b-a}{\sqrt{12}}, & n\le 15,\\[4pt]
\dfrac{b-a}{4}, & 15<n\le 70,\\[4pt]
\dfrac{b-a}{6}, & n>70.
\end{cases}
\]
Wan et al. identified several limitations: the cutoffs \(n=15\) and \(n=70\) are somewhat arbitrary, the divisor does not depend on \(n\) continuously, the range increases without bound for unbounded distributions as \(n\) grows, and the original derivation assumes non-negative data [1407.8038].

Bland extended the reconstruction problem to the full five-number summary and proposed
\[
\hat{\mu}_{C2,\text{Bland}} \approx \frac{a+2q_1+2m+2q_3+b}{8}.
\]
For SD, Bland used a formula independent of \(n\), which Wan et al. reported to underestimate for small \(n\) and overestimate for large \(n\), especially under skewness [1407.8038].

The main conceptual advance in Wan et al. is therefore not only the specific formulas but the replacement of fixed divisors by functions of \(n\). In their simulation study, the proposed C1, C2, and C3 SD estimators were nearly unbiased under normal data and had smaller bias than Hozo’s and Bland’s estimators under log-normal, beta, exponential, and Weibull data. The paper also concluded with a practical ranking of reporting formats: C2 was recommended as the ideal scenario, C1 was preferable for small samples, and C3 was preferable for large samples [1407.8038].

A common misconception is that all “median-to-mean” conversions are interchangeable. The reported results suggest the opposite: the available summary statistics determine which reconstruction is appropriate, and the sampling behavior of the range and IQR changes materially with \(n\).

## 3. Optimal weighting and non-normal extensions

The next stage of the literature sharpened the weighting of order-statistic summaries. Luo et al. treated the sample mean itself as an optimally weighted combination of the mid-range, mid-quartile, and median, with weights that vary smoothly with \(n\) [1505.05687]. For the full five-number summary, their recommended estimator is
\[
\bar{X}\approx
w_1\left(\frac{a+b}{2}\right)+
w_2\left(\frac{q_1+q_3}{2}\right)+
(1-w_1-w_2)m,
\]
with
\[
w_1=\frac{2.2}{2.2+n^{0.75}}, \qquad
w_2=0.7-\frac{0.72}{n^{0.55}}.
\]
This replaces fixed weights with a smooth function of sample size and was proposed as the optimal mean estimator under the five-number summary [1505.05687].

Shi et al. then argued that Wan et al.’s equal weighting of the range-based and IQR-based SD estimators in the five-number-summary case is also suboptimal, because the range is more informative for small \(n\) whereas the IQR is more stable for large \(n\) [2003.02130]. They proposed
\[
S(w)=
w\,\frac{b-a}{\xi} + (1-w)\,\frac{q_3-q_1}{\eta},
\]
with
\[
\xi(n)=2\,\Phi^{-1}\!\left(\frac{n-0.375}{n+0.25}\right),
\qquad
\eta(n)=2\,\Phi^{-1}\!\left(\frac{0.75n-0.125}{n+0.25}\right),
\]
and derived an optimal smooth weight. Their practical approximation is
\[
\tilde{w}_{\text{opt}}(n)\approx \frac{1}{1+0.07\,n^{0.6}},
\]
which yields the shortcut SD estimator
\[
S(n)\approx \frac{b-a}{\theta_1(n)}+\frac{q_3-q_1}{\theta_2(n)},
\]
where
\[
\theta_1(n)=
(2+0.14\,n^{0.6})
\Phi^{-1}\!\left(\frac{n-0.375}{n+0.25}\right),
\]
\[
\theta_2(n)=
\left(2+\frac{2}{0.07\,n^{0.6}}\right)
\Phi^{-1}\!\left(\frac{0.75n-0.125}{n+0.25}\right).
\]
Their numerical results showed a more accurate estimate for normal data and favorable performance for non-normal data [2003.02130].

A further limitation of the normal-order-statistic approach is that studies reporting medians often do so because the data are non-normal. McGrath et al. therefore proposed two explicitly non-normal approaches: a quantile estimation method that fits one of several candidate parametric distributions by minimizing squared discrepancies between observed and theoretical quantiles, and a Box-Cox method that transforms the reported quantiles to approximate symmetry before applying Luo/Wan-type formulas [1903.10498]. Their simulations and empirical assessments showed that the proposed methods often perform better than the existing methods when applied to non-normal data [1903.10498].

Taken together, these developments define the most technical sense of Mean STD Sampling in evidence synthesis: a sequence of increasingly refined estimators that use order-statistic theory, smooth weighting, and, where necessary, explicit non-normal modeling to recover \((\bar{X},S)\) from partial summaries.

## 4. Mean estimation with bounded relative standard deviation

In randomized approximation algorithms, “Mean STD Sampling” refers to a different problem: estimating a mean \(\mu\) from i.i.d. samples \(X_1,X_2,\ldots\) when the relative standard deviation is bounded,
\[
\left|\frac{\sigma}{\mu}\right|\le c,
\]
or equivalently \(\operatorname{SD}(X)\le c\mu\) when \(\mu>0\). The goal is to construct \(\hat{\mu}\) such that
\[
\mathbb{P}\bigl(|\hat{\mu}-\mu|>\epsilon\mu\bigr)\le \delta
\]
with as few samples as possible [1706.01478].

The classical median-of-means approach achieves leading-order sample complexity
\[
19.35\left(\frac{c}{\epsilon}\right)^2\ln(\delta^{-1}),
\]
and Huber’s 2014 smoothing construction reduced this to
\[
6.96\left(\frac{c}{\epsilon}\right)^2\ln(\delta^{-1})
\]
by averaging blocks, multiplying each block average by an independent uniform factor on \([1-\epsilon,1+\epsilon]\), and taking the median of the smoothed block means [1411.4074]. The same line of work proved that the optimal order is \(\Theta((c/\epsilon)^2\ln(1/\delta))\).

Huber’s 2017 estimator then achieved the optimal leading constant:
\[
(2+o(1))\,c^2\epsilon^{-2}\ln(1/\delta),
\]
matching the lower bound up to lower-order terms and doing so without requiring extra conditions such as bounds on third or fourth moments [1706.01478]. A 2019 refinement balanced the upper and lower tails of a robust \(M\)-estimator and obtained the explicit sample bound
\[
\left\lceil
\frac{2(c^2\epsilon^{-2}+1)}{1-\epsilon^2}\ln\left(\frac{2}{\delta}\right)
\right\rceil,
\]
again under a bounded relative standard deviation assumption [1908.05386].

This literature matters because many #P-complete randomized approximation schemes reduce to mean estimation with controlled relative variance. The listed applications include the volume of a convex body, the permanent of a nonnegative matrix or a \(\{0,1\}\)-matrix, the number of linear extensions of a poset, and the partition function of the Ising model [1411.4074, 1706.01478, 1908.05386]. Here the standard deviation is not reconstructed from quantiles; it is an a priori complexity parameter that determines the number of samples needed for a relative-error guarantee.

## 5. Variance-aware sampling designs

A third use of the phrase arises when the standard deviation determines how samples should be allocated or selected. In BayesSRW, the population is partitioned into \(k\) groups with group sizes \(N_i\), means \(\mu_i\), and standard deviations \(\sigma_i\), and the target is the overall population mean
\[
\mu=\frac{1}{N}\sum_{i=1}^k N_i\mu_i.
\]
With inverse-probability weighting, the variance of the estimator is
\[
\mathbb{V}(\hat{\mu})=
\frac{1}{N^2}\sum_{i=1}^k \frac{N_i^2\sigma_i^2}{n_i},
\]
and minimizing this variance subject to \(\sum_i n_i=n\) yields the optimal allocation
\[
n_i^*=
n\,\frac{N_i\sigma_i}{\sum_{l=1}^k N_l\sigma_l}.
\]
The rule is therefore to allocate in proportion to group size times group standard deviation, not merely group size [2408.15454]. The paper then extends this to a two-stage Bayesian procedure, BayesSRW, where an initial sample is used to estimate the \(\sigma_i\), and the remaining budget is allocated according to \(N_i\hat{\sigma}_i\) [2408.15454].

In computer architecture simulation, the same mean/SD logic appears in another form. For a sample of regions with CPI values \(X_1,\dots,X_n\), the paper uses the classical confidence interval for the mean,
\[
\text{CI}= y \pm z_{\alpha/2}\frac{s}{\sqrt{n}},
\]
so that the margin of error is proportional to \(s/\sqrt{n}\) [2603.22598]. Ranked set sampling is used to reduce the standard deviation of the mean estimator at fixed sample size by exploiting a cheap ranking proxy derived from a baseline configuration. Across the reported experiments, ranked set sampling reduces the confidence interval width by up to \(50\%\); repeated subsampling reduces the maximum observed error from \(35\%\) to \(10\%\); and a multi-configuration Chebyshev-distance selection criterion yields an average error below \(2\%\) and a maximum error of \(3.5\%\) across SPEC CPU 2017 Integer applications when simulating 30 regions of 1 million instructions each [2603.22598].

These papers emphasize a different principle from the meta-analysis literature. The goal is not to recover an unreported SD, but to use observed or estimated standard deviations to minimize the variance of a mean estimator through allocation, ranking, or representative subsample selection.

## 6. Related extensions, assumptions, and recurring limitations

Several adjacent literatures extend the same mean/dispersion logic to dependent data and interacting-particle systems. For stationary time series, self-normalized block sampling replaces explicit long-run variance estimation by a path-based self-normalizer
\[
D_n=
\left\{
n^{-1}\sum_{k=1}^n
\left(
S_{1,k}-\frac{k}{n}S_{1,n}
\right)^2
\right\}^{1/2},
\]
and uses overlapping block statistics to estimate the sampling distribution of the mean without a priori estimation of auxiliary parameters such as long-run variance, tail index, or memory parameters [1512.00820]. In consensus-based sampling and optimization, the key state variables are a weighted mean \(m_\beta(\mu)\) and a weighted covariance \(C_\beta(\mu)\), and quantitative mean-field limits show \(J^{-1/2}\) convergence of the particle system to its mean-field dynamics [2312.07373]. In a related mean-field Langevin setting, sampling from the mean-field stationary distribution is decoupled into propagation of chaos for the finite-particle approximation and log-concave sampling for the \(N\)-particle stationary law [2402.07355].

Across these domains, the main limitations are consistent. In the quantile-to-mean/SD setting, Wan et al.’s formulas rely on approximate normality or at least symmetry and moderate skewness, and accuracy degrades for very small \(n\), extreme skewness, heavy tails, or truncation [1407.8038]. McGrath et al. made the same point from the opposite direction: a major limitation of widely used conversion rules is their normality assumption, which is unlikely to be tenable precisely in studies that report medians [1903.10498]. In bounded-relative-variance Monte Carlo, the guarantees require a correct upper bound on \(\sigma/\mu\); underestimating that constant invalidates the sample-complexity guarantee [1908.05386]. In variance-aware allocation methods, gains are substantial only when heterogeneity in \(\sigma_i\) is real rather than negligible [2408.15454].

A recurring misunderstanding is to treat all of these problems as the same object. They are not. One line of work reconstructs missing \((\bar{X},S)\) from quantiles; another derives optimal sample complexity when a relative standard deviation bound is known; another uses standard deviations to allocate sampling effort or reduce variance. What unifies them is more limited but precise: the mean is the target quantity, the standard deviation governs either identifiability, accuracy, or cost, and the best sampling scheme depends on how that relationship is formalized in the underlying model.

Source: https://www.emergentmind.com/topics/mean-std-sampling