Standard Errors

Estimating the repeated-sampling precision of an estimator

Statistics

Inference

Uncertainty

Standard errors as estimated sampling standard deviations, with plug-in, information, delta-method, bootstrap, robust, and cluster-aware calculations.

Author

Jiahong Sun

Published

September 26, 2026

Modified

September 26, 2026

1 The Quantity a Standard Error Estimates

Let $\hat\theta=T(X)$ estimate an estimand $\theta(P)$. Under a stated data-generating process $P$, the estimator has sampling variance

\[ V_P(T)=\operatorname{Var}_P\{T(X)\}. \]

Its sampling standard deviation is

\[ \operatorname{SD}_P(\hat\theta)=\sqrt{V_P(T)}. \]

Because $P$ is unknown, an analysis usually reports a standard error:

\[ \boxed{ \widehat{\operatorname{SE}}(\hat\theta) =\sqrt{\widehat V(\hat\theta)}, } \]

an estimate of that repeated-sampling standard deviation. The standard error is itself a statistic and therefore has sampling variability. It summarizes precision under a model or design; it is not the observed estimation error $|\hat\theta-\theta|$, which cannot be calculated when $\theta$ is unknown.

2 Standard Deviation, Standard Error, and Bias

These quantities answer different questions:

Quantity	What varies?	Interpretation
Standard deviation of $X$	Individual observations	Population heterogeneity or measurement spread
Standard deviation of $\hat\theta$	Estimates across repeated samples	True repeated-sampling precision of the estimator
Estimated standard error	A function of the observed sample	Estimate of the preceding quantity
Bias $E(\hat\theta)-\theta$	Center of the estimator’s sampling distribution	Systematic displacement from the estimand

A small standard error does not imply small bias. With squared-error loss,

\[ \operatorname{MSE}(\hat\theta) =\operatorname{Var}(\hat\theta) +\operatorname{Bias}(\hat\theta)^2, \]

so a precise estimator may be precisely wrong because of confounding, selection, measurement error, or model misspecification.

3 Direct and Plug-in Standard Errors

For iid observations with variance $\sigma^2$, the sample mean satisfies

\[ \operatorname{Var}(\bar X)=\frac{\sigma^2}{n}. \]

Replacing the unknown $\sigma$ with the sample standard deviation $S$ gives

\[ \widehat{\operatorname{SE}}(\bar X)=\frac{S}{\sqrt n}. \]

This familiar formula is not a universal definition. It follows from equal weighting and independence:

\[ \operatorname{Var}\left(\frac{1}{n}\sum_iX_i\right) =\frac{1}{n^2}\left\{ \sum_i\operatorname{Var}(X_i) +2\sum_{i<j}\operatorname{Cov}(X_i,X_j) \right}. \]

The $\sigma^2/n$ result appears only when the variances are common and all covariance terms vanish. Positive within-cluster correlation slows the gain in precision even when the raw observation count is large.

For a Bernoulli proportion,

\[ \operatorname{SE}_p(\hat p) =\sqrt{\frac{p(1-p)}{n}}, \qquad \widehat{\operatorname{SE}}(\hat p) =\sqrt{\frac{\hat p(1-\hat p)}{n}}. \]

The second is a plug-in estimate. It becomes zero when all observed outcomes agree, even though uncertainty about $p$ is not generally zero. Boundary data and small samples therefore require more than mechanically applying the plug-in formula.

4 Information-Based Standard Errors

In a regular parametric model with log-likelihood $\ell(\theta)$, a maximum-likelihood estimator often has

\[ \sqrt n(\hat\theta-\theta_0) \xrightarrow{d} N\{0,I_1(\theta_0)^{-1}\}. \]

For a scalar parameter, common estimates are

\[ \widehat{\operatorname{SE}}(\hat\theta) =\{I_n(\hat\theta)\}^{-1/2} \]

using expected information, or

\[ \widehat{\operatorname{SE}}(\hat\theta) =\{j_n(\hat\theta)\}^{-1/2}, \qquad j_n(\theta)=-\frac{\partial^2\ell(\theta)}{\partial\theta^2}, \]

using observed information. For vector parameters, the square roots of the diagonal entries of the inverse information matrix are marginal standard errors. Off-diagonal entries matter for contrasts and reveal parameter dependence.

These estimates rely on local quadratic curvature, correct likelihood specification, identifiability, interior parameters, and an adequate large-sample approximation. Near a boundary or a nearly singular information matrix, a finite curvature-based standard error can be unstable or misleading.

5 Transformations and the Delta Method

Suppose $g$ is differentiable and $\hat\theta$ has estimated covariance matrix $\widehat\Sigma$. The first-order delta method gives

\[ \widehat{\operatorname{Var}}\{g(\hat\theta)\} \approx \nabla g(\hat\theta)^\mathsf T \widehat\Sigma \nabla g(\hat\theta). \]

For a scalar estimator,

\[ \widehat{\operatorname{SE}}\{g(\hat\theta)\} \approx |g'(\hat\theta)|\, \widehat{\operatorname{SE}}(\hat\theta). \]

If $g(\theta)=\log\theta$ for $\theta>0$, this becomes $\widehat{\operatorname{SE}}(\log\hat\theta)\approx\widehat{\operatorname{SE}}(\hat\theta)/\hat\theta$. Transforming the endpoints of an interval and constructing a symmetric interval from a transformed standard error are related but distinct operations.

The approximation can fail when the derivative is zero, undefined, or highly variable over the estimator’s plausible range. In those cases, higher-order theory, profile likelihood, or a design-faithful resampling distribution may better represent uncertainty.

6 Sandwich and Cluster-Robust Standard Errors

For an estimator defined by estimating equations

\[ \sum_{i=1}^n\psi_i(\theta)=0, \]

the asymptotic covariance often has the sandwich form

\[ \operatorname{Var}(\hat\theta) \approx A^{-1}BA^{-\mathsf T}, \]

where $A$ describes the sensitivity of the estimating equations to $\theta$ and $B$ describes their variability. The model-based information formula is a special case when the likelihood is correctly specified and the information identity holds.

A heteroskedasticity-consistent sandwich estimate permits observation-level variance misspecification while retaining assumptions about independent units and the mean model. With clusters $c=1,\ldots,G$, the score contributions must instead be summed within clusters before forming the middle matrix:

\[ \widehat B_{\text{cluster}} =\sum_{c=1}^G \left(\sum_{i\in c}\hat\psi_i\right) \left(\sum_{i\in c}\hat\psi_i\right)^\mathsf T. \]

Cluster-robust standard errors allow arbitrary within-cluster dependence under suitable conditions, but they rely on independent clusters and asymptotics in the number of clusters. Thousands of cells from eight donors do not provide a large-$G$ justification. Small-sample corrections or methods tailored to the design may be necessary.

“Robust” refers to a limited class of variance misspecification. It does not fix a biased coefficient, omitted confounding, dependence across clusters, severe leverage, or the wrong estimand.

7 Resampling Standard Errors

Given bootstrap estimates $\hat\theta^{*1},\ldots,\hat\theta^{*B}$ from $B$ resamples, the bootstrap standard error is

\[ \widehat{\operatorname{SE}}_{\mathrm{boot}}(\hat\theta) =\left[ \frac{1}{B-1} \sum_{b=1}^B \left(\hat\theta^{*b}-\bar{\hat\theta}^{*}\right)^2 \right]^{1/2}. \]

Here $B$ controls Monte Carlo error, not the amount of scientific information. Increasing $B$ stabilizes the numerical estimate of bootstrap spread; it cannot compensate for few independent experimental units.

The resampling scheme must reproduce the original design. Resample independent specimens for specimen-level inference; use cluster, block, or stratified resampling when the design requires it. Refit normalization, selection, and tuning steps inside every replicate if those would vary in a repeated study. Ordinary bootstrap methods can fail for maxima, boundary parameters, model selection, very small samples, and other nonregular statistics.

8 Worked Example: Correlated Technical Measurements

Suppose $m$ biological specimens each contribute $k$ technical measurements,

\[ Y_{ij}=\mu+b_i+\varepsilon_{ij}, \]

where $\operatorname{Var}(b_i)=\tau^2$, $\operatorname{Var}(\varepsilon_{ij})=\sigma^2$, and all random terms are independent across specimens. The grand mean is

\[ \bar Y=\frac{1}{mk}\sum_{i=1}^m\sum_{j=1}^kY_{ij}. \]

Its variance is

\[ \boxed{ \operatorname{Var}(\bar Y) =\frac{\tau^2}{m} +\frac{\sigma^2}{mk}. } \]

Adding technical replicates reduces only the second term. Adding biological specimens reduces both sources of uncertainty in the population mean. If one incorrectly treats all $mk$ measurements as iid with marginal variance $\tau^2+\sigma^2$, the reported variance becomes $(\tau^2+\sigma^2)/(mk)$, dividing the biological heterogeneity by $k$ when it should be divided only by $m$.

Equivalently, with intraclass correlation

\[ \rho=\frac{\tau^2}{\tau^2+\sigma^2}, \]

the variance inflation relative to $mk$ independent measurements is the design effect

\[ 1+(k-1)\rho. \]

This calculation shows why the experimental unit is part of a standard-error formula, not merely a reporting detail.

9 From Standard Errors to Inference

A common large-sample interval has the Wald form

\[ \hat\theta\pm z_{1-\alpha/2} \widehat{\operatorname{SE}}(\hat\theta). \]

Its validity needs more than a standard error: the centered, studentized statistic must have an appropriate reference distribution, bias must be negligible on the standard-error scale, and the variance estimate must be reliable. With normal iid data and an estimated variance, the exact reference for a mean is Student’s $t$, not standard normal. With bounded, skewed, or weakly identified targets, a symmetric Wald interval can be inappropriate even when the standard error itself is correctly estimated.

A Bayesian posterior standard deviation is not automatically a frequentist standard error. The two can agree approximately in regular large samples, but they arise from different probability statements and can differ under informative priors, weak identification, boundaries, or misspecification.

10 Reporting and Diagnostic Checklist

Report enough information to understand what uncertainty the number represents:

the estimand, estimate, and scale;
the independent experimental unit and number of such units;
the variance estimator: model-based, robust, cluster-robust, design-based, or resampling;
clustering, stratification, weighting, repeated measures, and small-sample corrections;
whether preprocessing and model selection are included;
the reference distribution or interval method; and
sensitivity to plausible dependence and model assumptions.

Useful warning signs include a singular Hessian, very large or implausibly tiny standard errors, strong disagreement between model-based and robust estimates, estimates near parameter boundaries, few clusters, high leverage, or bootstrap replicates that frequently fail. Each points to a specific assumption or identification problem rather than a reason to choose whichever standard error is smallest.

11 Bioinformatics Interpretation

For a differential-expression coefficient, the estimand might be a population log fold change among biological specimens under a specified normalization and regression model. The coefficient is the estimate; its standard error summarizes repeated-sampling precision for that fitted procedure. Read depth and technical replication can reduce measurement noise, while the number, diversity, and allocation of independent specimens control population-level information.

Pseudoreplicating reads or cells produces artificially small standard errors. Treating batch as independent noise when it is confounded with condition can produce a precise coefficient for an unidentified scientific contrast. Shrinkage and empirical-Bayes procedures may improve estimation by borrowing information across features, but the resulting posterior or moderated uncertainty should be described according to the method rather than mislabeled as a plain iid standard error.

Statistical precision does not establish biological importance. Report the effect estimate and uncertainty on an interpretable scale, then discuss whether the range is compatible with effects that matter biologically.

12 Connections and Take-home Messages

Sampling distributions defines the repeated-sample distribution whose spread is being estimated.
Point estimators: bias, variance, and mean squared error distinguishes precision from bias and total risk.
Fisher information and the Cramér–Rao bound develops inverse-information precision benchmarks.
Maximum-likelihood estimation explains Hessian-based covariance estimates.
Delta method derives uncertainty propagation through smooth transformations.
Slutsky’s theorem and continuous mapping justifies many plug-in and studentized approximations.
Independence and conditional independence clarifies the assumptions behind variance additivity.

A standard error estimates an estimator’s repeated-sampling standard deviation under a defined model or design. Its formula must reflect dependence, clustering, nuisance estimation, transformations, and every variable analysis step. A small standard error says nothing by itself about bias, causality, biological importance, or generalizability. The number of independent biological units—not the count of technical observations—sets the foundation for population-level precision.

13 References

Eicker, F. “Limit Theorems for Regressions with Unequal and Dependent Errors.” In Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, Volume 1, 1967, pp. 59–82. Article.
Huber, P. J. “The Behavior of Maximum Likelihood Estimates under Nonstandard Conditions.” In Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, Volume 1, 1967, pp. 221–233. Article.
White, H. “A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity.” Econometrica 48 (1980): 817–838. Article.
Liang, K.-Y. and Zeger, S. L. “Longitudinal Data Analysis Using Generalized Linear Models.” Biometrika 73 (1986): 13–22. Article.
Efron, B. and Tibshirani, R. J. An Introduction to the Bootstrap. Chapman & Hall/CRC, 1993. Publisher page.
Casella, G. and Berger, R. L. Statistical Inference, 2nd ed. Duxbury, 2002. Publisher page.

--- title: "Standard Errors" subtitle: "Estimating the repeated-sampling precision of an estimator" description: "Standard errors as estimated sampling standard deviations, with plug-in, information, delta-method, bootstrap, robust, and cluster-aware calculations." author: "Jiahong Sun" date: "2026-09-26" date-modified: "2026-09-26" categories: [Statistics, Inference, Uncertainty] toc: true number-sections: true draft: false --- # The Quantity a Standard Error Estimates Let $\hat\theta=T(X)$ estimate an estimand $\theta(P)$. Under a stated data-generating process $P$, the estimator has sampling variance $$ V_P(T)=\operatorname{Var}_P\{T(X)\}. $$ Its sampling standard deviation is $$ \operatorname{SD}_P(\hat\theta)=\sqrt{V_P(T)}. $$ Because $P$ is unknown, an analysis usually reports a **standard error**: $$ \boxed{ \widehat{\operatorname{SE}}(\hat\theta) =\sqrt{\widehat V(\hat\theta)}, } $$ an estimate of that repeated-sampling standard deviation. The standard error is itself a statistic and therefore has sampling variability. It summarizes precision under a model or design; it is not the observed estimation error $|\hat\theta-\theta|$, which cannot be calculated when $\theta$ is unknown. # Standard Deviation, Standard Error, and Bias These quantities answer different questions: | Quantity | What varies? | Interpretation | |---|---|---| | Standard deviation of $X$ | Individual observations | Population heterogeneity or measurement spread | | Standard deviation of $\hat\theta$ | Estimates across repeated samples | True repeated-sampling precision of the estimator | | Estimated standard error | A function of the observed sample | Estimate of the preceding quantity | | Bias $E(\hat\theta)-\theta$ | Center of the estimator's sampling distribution | Systematic displacement from the estimand | A small standard error does not imply small bias. With squared-error loss, $$ \operatorname{MSE}(\hat\theta) =\operatorname{Var}(\hat\theta) +\operatorname{Bias}(\hat\theta)^2, $$ so a precise estimator may be precisely wrong because of confounding, selection, measurement error, or model misspecification. # Direct and Plug-in Standard Errors For iid observations with variance $\sigma^2$, the sample mean satisfies $$ \operatorname{Var}(\bar X)=\frac{\sigma^2}{n}. $$ Replacing the unknown $\sigma$ with the sample standard deviation $S$ gives $$ \widehat{\operatorname{SE}}(\bar X)=\frac{S}{\sqrt n}. $$ This familiar formula is not a universal definition. It follows from equal weighting and independence: $$ \operatorname{Var}\left(\frac{1}{n}\sum_iX_i\right) =\frac{1}{n^2}\left\{ \sum_i\operatorname{Var}(X_i) +2\sum_{i<j}\operatorname{Cov}(X_i,X_j) \right}. $$ The $\sigma^2/n$ result appears only when the variances are common and all covariance terms vanish. Positive within-cluster correlation slows the gain in precision even when the raw observation count is large. For a Bernoulli proportion, $$ \operatorname{SE}_p(\hat p) =\sqrt{\frac{p(1-p)}{n}}, \qquad \widehat{\operatorname{SE}}(\hat p) =\sqrt{\frac{\hat p(1-\hat p)}{n}}. $$ The second is a plug-in estimate. It becomes zero when all observed outcomes agree, even though uncertainty about $p$ is not generally zero. Boundary data and small samples therefore require more than mechanically applying the plug-in formula. # Information-Based Standard Errors In a regular parametric model with log-likelihood $\ell(\theta)$, a maximum-likelihood estimator often has $$ \sqrt n(\hat\theta-\theta_0) \xrightarrow{d} N\{0,I_1(\theta_0)^{-1}\}. $$ For a scalar parameter, common estimates are $$ \widehat{\operatorname{SE}}(\hat\theta) =\{I_n(\hat\theta)\}^{-1/2} $$ using expected information, or $$ \widehat{\operatorname{SE}}(\hat\theta) =\{j_n(\hat\theta)\}^{-1/2}, \qquad j_n(\theta)=-\frac{\partial^2\ell(\theta)}{\partial\theta^2}, $$ using observed information. For vector parameters, the square roots of the diagonal entries of the inverse information matrix are marginal standard errors. Off-diagonal entries matter for contrasts and reveal parameter dependence. These estimates rely on local quadratic curvature, correct likelihood specification, identifiability, interior parameters, and an adequate large-sample approximation. Near a boundary or a nearly singular information matrix, a finite curvature-based standard error can be unstable or misleading. # Transformations and the Delta Method Suppose $g$ is differentiable and $\hat\theta$ has estimated covariance matrix $\widehat\Sigma$. The first-order delta method gives $$ \widehat{\operatorname{Var}}\{g(\hat\theta)\} \approx \nabla g(\hat\theta)^\mathsf T \widehat\Sigma \nabla g(\hat\theta). $$ For a scalar estimator, $$ \widehat{\operatorname{SE}}\{g(\hat\theta)\} \approx |g'(\hat\theta)|\, \widehat{\operatorname{SE}}(\hat\theta). $$ If $g(\theta)=\log\theta$ for $\theta>0$, this becomes $\widehat{\operatorname{SE}}(\log\hat\theta)\approx\widehat{\operatorname{SE}}(\hat\theta)/\hat\theta$. Transforming the endpoints of an interval and constructing a symmetric interval from a transformed standard error are related but distinct operations. The approximation can fail when the derivative is zero, undefined, or highly variable over the estimator's plausible range. In those cases, higher-order theory, profile likelihood, or a design-faithful resampling distribution may better represent uncertainty. # Sandwich and Cluster-Robust Standard Errors For an estimator defined by estimating equations $$ \sum_{i=1}^n\psi_i(\theta)=0, $$ the asymptotic covariance often has the sandwich form $$ \operatorname{Var}(\hat\theta) \approx A^{-1}BA^{-\mathsf T}, $$ where $A$ describes the sensitivity of the estimating equations to $\theta$ and $B$ describes their variability. The model-based information formula is a special case when the likelihood is correctly specified and the information identity holds. A heteroskedasticity-consistent sandwich estimate permits observation-level variance misspecification while retaining assumptions about independent units and the mean model. With clusters $c=1,\ldots,G$, the score contributions must instead be summed within clusters before forming the middle matrix: $$ \widehat B_{\text{cluster}} =\sum_{c=1}^G \left(\sum_{i\in c}\hat\psi_i\right) \left(\sum_{i\in c}\hat\psi_i\right)^\mathsf T. $$ Cluster-robust standard errors allow arbitrary within-cluster dependence under suitable conditions, but they rely on independent clusters and asymptotics in the number of clusters. Thousands of cells from eight donors do not provide a large-$G$ justification. Small-sample corrections or methods tailored to the design may be necessary. “Robust” refers to a limited class of variance misspecification. It does not fix a biased coefficient, omitted confounding, dependence across clusters, severe leverage, or the wrong estimand. # Resampling Standard Errors Given bootstrap estimates $\hat\theta^{*1},\ldots,\hat\theta^{*B}$ from $B$ resamples, the bootstrap standard error is $$ \widehat{\operatorname{SE}}_{\mathrm{boot}}(\hat\theta) =\left[ \frac{1}{B-1} \sum_{b=1}^B \left(\hat\theta^{*b}-\bar{\hat\theta}^{*}\right)^2 \right]^{1/2}. $$ Here $B$ controls Monte Carlo error, not the amount of scientific information. Increasing $B$ stabilizes the numerical estimate of bootstrap spread; it cannot compensate for few independent experimental units. The resampling scheme must reproduce the original design. Resample independent specimens for specimen-level inference; use cluster, block, or stratified resampling when the design requires it. Refit normalization, selection, and tuning steps inside every replicate if those would vary in a repeated study. Ordinary bootstrap methods can fail for maxima, boundary parameters, model selection, very small samples, and other nonregular statistics. # Worked Example: Correlated Technical Measurements Suppose $m$ biological specimens each contribute $k$ technical measurements, $$ Y_{ij}=\mu+b_i+\varepsilon_{ij}, $$ where $\operatorname{Var}(b_i)=\tau^2$, $\operatorname{Var}(\varepsilon_{ij})=\sigma^2$, and all random terms are independent across specimens. The grand mean is $$ \bar Y=\frac{1}{mk}\sum_{i=1}^m\sum_{j=1}^kY_{ij}. $$ Its variance is $$ \boxed{ \operatorname{Var}(\bar Y) =\frac{\tau^2}{m} +\frac{\sigma^2}{mk}. } $$ Adding technical replicates reduces only the second term. Adding biological specimens reduces both sources of uncertainty in the population mean. If one incorrectly treats all $mk$ measurements as iid with marginal variance $\tau^2+\sigma^2$, the reported variance becomes $(\tau^2+\sigma^2)/(mk)$, dividing the biological heterogeneity by $k$ when it should be divided only by $m$. Equivalently, with intraclass correlation $$ \rho=\frac{\tau^2}{\tau^2+\sigma^2}, $$ the variance inflation relative to $mk$ independent measurements is the design effect $$ 1+(k-1)\rho. $$ This calculation shows why the experimental unit is part of a standard-error formula, not merely a reporting detail. # From Standard Errors to Inference A common large-sample interval has the Wald form $$ \hat\theta\pm z_{1-\alpha/2} \widehat{\operatorname{SE}}(\hat\theta). $$ Its validity needs more than a standard error: the centered, studentized statistic must have an appropriate reference distribution, bias must be negligible on the standard-error scale, and the variance estimate must be reliable. With normal iid data and an estimated variance, the exact reference for a mean is Student's $t$, not standard normal. With bounded, skewed, or weakly identified targets, a symmetric Wald interval can be inappropriate even when the standard error itself is correctly estimated. A Bayesian posterior standard deviation is not automatically a frequentist standard error. The two can agree approximately in regular large samples, but they arise from different probability statements and can differ under informative priors, weak identification, boundaries, or misspecification. # Reporting and Diagnostic Checklist Report enough information to understand what uncertainty the number represents: - the estimand, estimate, and scale; - the independent experimental unit and number of such units; - the variance estimator: model-based, robust, cluster-robust, design-based, or resampling; - clustering, stratification, weighting, repeated measures, and small-sample corrections; - whether preprocessing and model selection are included; - the reference distribution or interval method; and - sensitivity to plausible dependence and model assumptions. Useful warning signs include a singular Hessian, very large or implausibly tiny standard errors, strong disagreement between model-based and robust estimates, estimates near parameter boundaries, few clusters, high leverage, or bootstrap replicates that frequently fail. Each points to a specific assumption or identification problem rather than a reason to choose whichever standard error is smallest. # Bioinformatics Interpretation For a differential-expression coefficient, the estimand might be a population log fold change among biological specimens under a specified normalization and regression model. The coefficient is the estimate; its standard error summarizes repeated-sampling precision for that fitted procedure. Read depth and technical replication can reduce measurement noise, while the number, diversity, and allocation of independent specimens control population-level information. Pseudoreplicating reads or cells produces artificially small standard errors. Treating batch as independent noise when it is confounded with condition can produce a precise coefficient for an unidentified scientific contrast. Shrinkage and empirical-Bayes procedures may improve estimation by borrowing information across features, but the resulting posterior or moderated uncertainty should be described according to the method rather than mislabeled as a plain iid standard error. Statistical precision does not establish biological importance. Report the effect estimate and uncertainty on an interpretable scale, then discuss whether the range is compatible with effects that matter biologically. # Connections and Take-home Messages - [Sampling distributions](sampling-distributions.qmd) defines the repeated-sample distribution whose spread is being estimated. - [Point estimators: bias, variance, and mean squared error](point-estimators-bias-variance-and-mean-squared-error.qmd) distinguishes precision from bias and total risk. - [Fisher information and the Cramér--Rao bound](fisher-information-and-cramer-rao-bound.qmd) develops inverse-information precision benchmarks. - [Maximum-likelihood estimation](maximum-likelihood-estimation.qmd) explains Hessian-based covariance estimates. - [Delta method](../asymptotic-thinking/delta-method.qmd) derives uncertainty propagation through smooth transformations. - [Slutsky's theorem and continuous mapping](../asymptotic-thinking/slutskys-theorem-and-continuous-mapping.qmd) justifies many plug-in and studentized approximations. - [Independence and conditional independence](../probability-randomness/independence-and-conditional-independence.qmd) clarifies the assumptions behind variance additivity. A standard error estimates an estimator's repeated-sampling standard deviation under a defined model or design. Its formula must reflect dependence, clustering, nuisance estimation, transformations, and every variable analysis step. A small standard error says nothing by itself about bias, causality, biological importance, or generalizability. The number of independent biological units—not the count of technical observations—sets the foundation for population-level precision. # References - Eicker, F. “Limit Theorems for Regressions with Unequal and Dependent Errors.” In *Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability*, Volume 1, 1967, pp. 59--82. [Article](https://projecteuclid.org/euclid.bsmsp/1200512981). - Huber, P. J. “The Behavior of Maximum Likelihood Estimates under Nonstandard Conditions.” In *Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability*, Volume 1, 1967, pp. 221--233. [Article](https://projecteuclid.org/euclid.bsmsp/1200512988). - White, H. “A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity.” *Econometrica* 48 (1980): 817--838. [Article](https://doi.org/10.2307/1912934). - Liang, K.-Y. and Zeger, S. L. “Longitudinal Data Analysis Using Generalized Linear Models.” *Biometrika* 73 (1986): 13--22. [Article](https://doi.org/10.1093/biomet/73.1.13). - Efron, B. and Tibshirani, R. J. *An Introduction to the Bootstrap*. Chapman & Hall/CRC, 1993. [Publisher page](https://doi.org/10.1201/9780429246593). - Casella, G. and Berger, R. L. *Statistical Inference*, 2nd ed. Duxbury, 2002. [Publisher page](https://www.routledge.com/Statistical-Inference/Casella-Berger/p/book/9781032593036).