C Error in estimated regression coefficients and standard errors

Results from statistics on regression under measurement error can help us understand the effect of bias on DA regression analysis.

Consider Equation (3.2) for the simple linear regression of log density of a species \(i\) on covariate \(x\). Let \(y\) stand for the actual log density of a focal species \(i\), \(z\) stand for the log density that we’ve measured, and \(d = z - y\) equal the difference between the two, which here is the log efficiency of species \(i\) minus the log mean efficiency of the sample. Let \(s_{xy}\) denote the sample covariance between variables \(x\) and \(y\), \(s^{2}_{x} = s_{xx}\) and \(s_{x}\) denote the sample variance and standard deviation, and \(r_{xy} = s_{xy}/(s_{x}s_{y})\) denote the sample correlation.

The ordinary least squares (OLS) and maximum likelihood (MLE) estimates of the slope of \(z\) equals the sample covariance of \(z\) and \(x\) divided by the sample variance in \(x\) or, equivalently, the sample correlation of \(z\) and \(x\) multiplied by the ratio of their sample standard deviations, \[\begin{align} \hat \beta_z = \frac{s_{zx}}{s^2_x} = r_{zx} \cdot \frac{s_z}{s_x}. \end{align}\] From the (bi)linearity of sample covariances it follows that \[\begin{align} \hat \beta_z = \frac{s_{yx}}{s^2_x} + \frac{s_{dx}}{s^2_x} = \frac{r_{yx} s_y}{s_x} + \frac{r_{dx} s_d}{s_x} = \hat \beta_y + \hat \beta_d, \end{align}\] where \(\hat \beta_y\) and \(\hat \beta_d\) denote the slope estimates for \(y\) and \(d\) (were these values to be known). The absolute error in the estimate \(\hat \beta_{z}\) of \(\hat \beta_{y}\) is therefore \(\hat \beta_{d}\); it is large in a practical sense when \(\hat \beta_{d}\) is large (in absolute value) compared to \(\hat \beta_{y}\), which corresponds to the covariance of \(d\) with \(x\) being large compared to the covariance of \(y\) with \(x\).

In our case, the covariance of \(d\) equals the negative of the covariance of log mean efficiency with \(x\). The absolute error in \(\hat \beta\) equals the negative covariance of log mean efficiency scaled by the variance in \(x\). This absolute error is the same for all species; however, its practical significance varies depending on its magnitude relative to that of the slope of the actual log densities. For species that covary with \(x\) more strongly than the log mean efficiency, the error will be relatively small. This situation might occur either because the mean efficiency varies relatively little across samples or because its variation is relatively less correlated with \(x\) compared to the log density of the focal species.

We can similarly understand the impact of measurement error on the precision of our slope estimates. The OLS and MLE estimated standard error in \(\hat \beta\) are both approximately \[\begin{align} \hat{\text{se}}(\hat \beta) \approx \frac{s_{\hat \varepsilon}}{s_x \sqrt{n}}, \end{align}\] where \(s_{\hat \varepsilon}\) is the sample standard deviation of the residuals (Wasserman (2004) Chapter 13). The sample residuals of \(z\), \(y\), and \(d\) have a similar relationship to the regression coefficient estimates, \[\begin{align} \hat \varepsilon_z &\equiv z - \hat \beta_z x \\&= (y + d) - (\hat \beta_y + \hat \beta_d) x \\&= (y - \hat \beta_y x) + (d - \hat \beta_d x) \\&= \hat \varepsilon_y + \hat \varepsilon_d. \end{align}\] (Note, here I’ve omitted the subscript indicating the dependence on the sample.) It follows that the sample variances of the residuals of \(z\), \(y\), and \(d\) are related through \[\begin{align} s^2_{\hat \varepsilon_{z}} = s^2_{\hat \varepsilon_{y} + \hat \varepsilon_{d}} = s^2_{\hat \varepsilon_{y}} + s^2_{\hat \varepsilon_{d}} + 2 s_{\hat \varepsilon_{y} \hat \varepsilon_{d}}. \end{align}\] The standard deviation of the \(z\) residuals is increased above that of the \(y\) residuals when the \(d\) residuals are either uncorrelated or positively correlated with the \(y\) residuals, but may be decreased when the \(y\) and \(d\) residuals are negatively correlated.

In our case, the \(d\) residuals equal the negative residuals of the log mean efficiency. It is plausible that for most species, their residual variation will have a small covariance with log mean efficiency and the net effect of variation in the mean efficiency will be to increase the estimated standard errors, as occurs with most species in Figure 3.1. However, high-efficiency species that vary substantially in proportion across samples may be strongly positively correlated with log mean efficiency such that the estimated standard errors decrease, as we see with Species 9 in Figure 3.1.

References

Wasserman, Larry. 2004. All of Statistics. Springer Texts in Statistics. New York, NY: Springer New York. https://doi.org/10.1007/978-0-387-21736-9.