|
Size: 2897
Comment: Some attachments
|
← Revision 15 as of 2025-05-26 21:15:15 ⇥
Size: 5766
Comment: Typo
|
| Deletions are marked like this. | Additions are marked like this. |
| Line 13: | Line 13: |
| In the classical OLS model,, estimated coefficients are: | The standard error of some estimate is the variance of that estimate divided by the square root of the sample size. One common use of standard errors is to estimate [[Statistics/MarginsOfError|margins of error]]. For a [[Statistics/BernoulliDistribution|Bernoulli-distributed]] variable, the standard error is ''p(1-p)'' and is maximized at ''p=0.5''. Therefore a conservative standard error is a function of only the sample size. Standard errors are also used in interpreting the estimated coefficients of a regression model. As a reminder, by classical [[Statistics/OrdinaryLeastSquares|OLS]], estimated coefficients are: |
| Line 17: | Line 21: |
| Standard errors are the standard deviations of these coefficients. | But specific regressions methods require assumptions about variance. Standard errors in this context are much more complicated. |
| Line 29: | Line 33: |
| Var(\hat(\beta)|X_i) = \frac{\sum_{i=1}^n Var((X_i-\bar{X})\hat{\epsilon}_i)}{(\sum_{i=1}^n(X_i-\bar{X})^2)^2} | {{attachment:unispec1.svg}} |
| Line 31: | Line 35: |
| Supposing the population ''Var(β)'' is known, this can be simplified. | Supposing the population ''Var(ε)'' is known and errors are homoskedastic, i.e. they are constant across all cases, this can be simplified. |
| Line 33: | Line 37: |
| Var(\hat{\beta}|X_i) = \frac{Var(\beta)(\sum_{i=1}^n(x_i-\bar{X})^2)}{(\sum_{i=1}^n(X_i-\bar{X})^2)^2} = \frac{Var(\beta)}{\sum_{i=1}^n(X_i-\bar{X})^2} | {{attachment:unispec2.svg}} |
| Line 37: | Line 41: |
| Var(\hat{\beta}|X_i) = \frac{Var(\beta)}{n (\frac{1}{n}\sum_{i=1}^n(X_i-\bar{X})^2)} = \frac{Var(\beta)}{n Var(X)} | {{attachment:unispec3.svg}} |
| Line 39: | Line 43: |
| ''Var(β)'' is unknown, so this term is estimated as: | ''Var(ε)'' is unknown, so this term is estimated as: |
| Line 41: | Line 45: |
| \hat{\epsilon}_i = Y_i - \hat{Y}_i | {{attachment:uniest1.svg}}, {{attachment:uniest2.svg}} |
| Line 43: | Line 47: |
| Var(\hat{\epsilon}) = \frac{1}{n-1}\sum_{i=1}^n(\hat{\epsilon}_i^2) 1 degree of freedom is lost in assuming homoskedasticity of errors, i.e. {{attachment:homosked.svg}} ''k'' degrees of freedom are lost in assuming independence of errors and ''k'' independent variables, which is necessarily 1 in the univariate case, i.e.: \sum_{i=1}^nX_i\hat{\epsilon}_i = 0 |
1 degree of freedom is lost in assuming homoskedasticity of errors, i.e. {{attachment:homosked.svg}}; and ''k'' degrees of freedom are lost in assuming independence of errors and ''k'' independent variables, which is necessarily 1 in the univariate case, i.e.: {{attachment:ind.svg}} |
| Line 53: | Line 51: |
| \hat{Var}(\hat{\beta}|X_i) = \frac{\frac{1}{n-2}Var(\hat{\epsilon})}{n Var(X)} | {{attachment:uniest3.svg}} |
| Line 61: | Line 59: |
| Var(\mathbf{b} | \mathbf{X}) = E\Bigl[(\mathbf{b}-\mathbf{\beta})(\mathbf{b}-\mathbf{\beta})^T \Big| \mathbf{X}\Bigr] | {{attachment:multspec1.svg}} |
| Line 63: | Line 61: |
| That term is rewritten as ''('''X'''^T^'''X''')^-1^'''X'''u''. | That term is rewritten as ''('''X'''^T^'''X''')^-1^'''Xε'''''. |
| Line 65: | Line 63: |
| Var(\mathbf{b} | \mathbf{X}) = E\Bigl[\bigl((\mathbf{X}^T\mathbf{X})^{-1}\mathbf{X}^T\mathbf{u}\bigr)\bigl((\mathbf{X}^T\mathbf{X})^{-1}\mathbf{X}^T\mathbf{u}\bigr)^{T} \Big| \mathbf{X}\Bigr] = E\Bigl[(\mathbf{X}^T\mathbf{X})^{-1}\mathbf{X}^T\mathbf{u}\mathbf{u}^T\mathbf{X}(\mathbf{X}^T\mathbf{X})^{-1} \Big| \mathbf{X}\Bigr] | {{attachment:multspec2.svg}} |
| Line 67: | Line 65: |
| Var(\mathbf{b} | \mathbf{X}) = (\mathbf{X}^T\mathbf{X})^{-1}\mathbf{X}^T E\bigl[\mathbf{u}\mathbf{u}^T\big|\mathbf{X}\bigr]\mathbf{X}(\mathbf{X}^T\mathbf{X})^{-1} | {{attachment:multspec3.svg}} |
| Line 69: | Line 67: |
| Practically speaking, ''E['''uu'''^T^|'''X''']'' is never known. But if homoskedasticity and independence are assumed, i.e.: | ''E['''εε'''^T^|'''X''']'' is not a practical matrix to work with, even if known. But if homoskedasticity and independence are assumed, i.e.: {{attachment:homosked_ind.svg}}, then this simplifies to: |
| Line 71: | Line 69: |
| E\bigl[\mathbf{u}\mathbf{u}^T\big|\mathbf{X}\bigr] = Var(\mathbf{\beta})\mathbf{I}_n | {{attachment:multspec4.svg}} |
| Line 73: | Line 71: |
| ...then this simplifies to: | ''s^2^'' is unknown, so this term is estimated as: |
| Line 75: | Line 73: |
| Var(\mathbf{b} | \mathbf{X}) = Var(\mathbf{\beta}) (\mathbf{X}^T\mathbf{X})^{-1} | {{attachment:multspec5.svg}} |
| Line 77: | Line 75: |
| ''Var(β)'' is unknown, so the estimate is: | This arrives at estimation as: |
| Line 79: | Line 77: |
| \hat{Var}(\mathbf{b} | \mathbf{X}) = \frac{1}{1-k} \mathbf{u}^T\mathbf{u} (\mathbf{X}^T\mathbf{X})^{-1} | {{attachment:multspec6.svg}} ---- == Robust == In the presence of heteroskedasticity of errors, the above simplifications cannot apply. In the univariate case, use the original estimator. This is mostly interesting in the multivariate case, where ''E['''εε'''^T^|'''X''']'' is still not practical. The assumptions made, when incorrect, lead to... * OLS estimators are not BLUE * they are unbiased, but no longer most efficient in terms of MSE * nonlinear GLMs, such as logit, can be biased * even if the model's estimates are unbiased, statistics derived from those estimates (e.g., conditioned probability distributions) can be biased '''Eicker-Huber-White heterskedasticity consistent errors''' ('''HCE''') assume that errors are still independent but allowed to vary, i.e. '''''Σ''' = diag(ε,,1,,,...ε,,n,,)''. Importantly, this is not a function of '''''X''''', so the standard errors can be estimated as: {{attachment:robust.svg}} Robust errors are only appropriate with large sample sizes. [[SamplingWeightsAndRegressionAnalysis|When fitting a model using data with survey weights, if those weights are a function of predictors including the dependent variable, then heteroskedastic consistent errors should be used.]] [[HowRobustStandardErrorsExposeMethodologicalProblemsTheyDoNotFix|If a model significantly diverges after introducing robust errors, there is likely a specification error.]] ---- == Clustered == '''Liang-Zeger clustered robust standard errors''' assume that errors covary within clusters. {{attachment:cluster1.svg}} where '''''x''',,g,,'' is an ''n,,g,,'' by ''k'' matrix constructed by stacking '''''x''',,i,,'' for all ''i'' belonging to cluster ''g''; and '''''ε''',,g,,'' is an ''n,,g,,'' long vector holding the errors for each cluster ''g''. The estimator becomes: {{attachment:cluster2.svg}} Clustered standard errors should only be used if the sample design or experimental design call for it. * A complex survey sample design leads to differential sampling errors across strata. * A two-stage sample design leads to differential sampling errors for the SSU within each PSU. * Assignment of an experimental treatment at a grouped level often leads to differential errors across those groups. * For time series evaluation of an experimental treatment that is assigned at the individual level, it is generally recommended to cluster at the individual level. There are parallels between [[Statistics/FixedEffectsModel|fixed effects]] and clusters, but use of one does not mandate nor conflict with the other. ---- == Finite Population Correction == Most formulations of standard errors assume the population is unknown and/or infinite. If the population is finite and the sampling rate is high (above 5%), the standard error is too conservative. The '''finite population correction''' ('''FPC''') is an adjustment to correct this: {{attachment:fpc.svg}} Intuitively, the FPC is 0 when ''n = N'' because there is no sampling error in a census. FPC approaches 1 when ''n'' approaches 0, demonstrating that the factor is meaningless for low sampling rates. ---- CategoryRicottone |
Standard Errors
Standard errors are the standard deviations of estimated coefficients.
Contents
Description
The standard error of some estimate is the variance of that estimate divided by the square root of the sample size.
One common use of standard errors is to estimate margins of error. For a Bernoulli-distributed variable, the standard error is p(1-p) and is maximized at p=0.5. Therefore a conservative standard error is a function of only the sample size.
Standard errors are also used in interpreting the estimated coefficients of a regression model. As a reminder, by classical OLS, estimated coefficients are:
univariate case:
multivariate case:
But specific regressions methods require assumptions about variance. Standard errors in this context are much more complicated.
Classical
Univariate
In the univariate case, standard errors are classically specified as:
Supposing the population Var(ε) is known and errors are homoskedastic, i.e. they are constant across all cases, this can be simplified.
Lastly, rewrite the denominator in terms of Var(X).
Var(ε) is unknown, so this term is estimated as:
, 
1 degree of freedom is lost in assuming homoskedasticity of errors, i.e. 
; and k degrees of freedom are lost in assuming independence of errors and k independent variables, which is necessarily 1 in the univariate case, i.e.: 
This arrives at estimation as:
Multivariate
The classical multivariate specification is expressed in terms of (b-β), as:
That term is rewritten as (XTX)-1Xε.
E[εεT|X] is not a practical matrix to work with, even if known. But if homoskedasticity and independence are assumed, i.e.: 
, then this simplifies to:
s2 is unknown, so this term is estimated as:
This arrives at estimation as:
Robust
In the presence of heteroskedasticity of errors, the above simplifications cannot apply. In the univariate case, use the original estimator.
This is mostly interesting in the multivariate case, where E[εεT|X] is still not practical. The assumptions made, when incorrect, lead to...
- OLS estimators are not BLUE
- they are unbiased, but no longer most efficient in terms of MSE
- nonlinear GLMs, such as logit, can be biased
- even if the model's estimates are unbiased, statistics derived from those estimates (e.g., conditioned probability distributions) can be biased
Eicker-Huber-White heterskedasticity consistent errors (HCE) assume that errors are still independent but allowed to vary, i.e. Σ = diag(ε1,...εn). Importantly, this is not a function of X, so the standard errors can be estimated as:
Robust errors are only appropriate with large sample sizes.
Clustered
Liang-Zeger clustered robust standard errors assume that errors covary within clusters.
where xg is an ng by k matrix constructed by stacking xi for all i belonging to cluster g; and εg is an ng long vector holding the errors for each cluster g.
The estimator becomes:
Clustered standard errors should only be used if the sample design or experimental design call for it.
- A complex survey sample design leads to differential sampling errors across strata.
- A two-stage sample design leads to differential sampling errors for the SSU within each PSU.
- Assignment of an experimental treatment at a grouped level often leads to differential errors across those groups.
- For time series evaluation of an experimental treatment that is assigned at the individual level, it is generally recommended to cluster at the individual level.
There are parallels between fixed effects and clusters, but use of one does not mandate nor conflict with the other.
Finite Population Correction
Most formulations of standard errors assume the population is unknown and/or infinite. If the population is finite and the sampling rate is high (above 5%), the standard error is too conservative. The finite population correction (FPC) is an adjustment to correct this:
Intuitively, the FPC is 0 when n = N because there is no sampling error in a census. FPC approaches 1 when n approaches 0, demonstrating that the factor is meaningless for low sampling rates.