My notes say

$$B(\hat\theta) = E(\hat\theta) - \theta $$

And I understand that the bias is the difference between a parameter and the expectation of its estimator. What I don't understand is how to calulate the bias given only an estimator? My notes lack ANY examples of calculating the bias, so even if anyone could please give me an example I could understand it better!

  • 3
    $\begingroup$ You do it by calculating the expectation. $\endgroup$ Oct 24 '16 at 5:18
  • 1
    $\begingroup$ The concept becomes clearer with examples. Wikipedia has a few. You could also try Google. $\endgroup$ Oct 24 '16 at 5:20
  • $\begingroup$ So if i was given the estimator $\hat\p$ = $X/n$ (p hat, haven't quite figured out the editing for this yet sorry) and i want to find the bias of that, i start by finding the expectation of $\hat\p$ ? @MichaelHardy $\endgroup$
    – H4-math
    Oct 24 '16 at 6:16
  • $\begingroup$ Yes. $\qquad\qquad$ $\endgroup$ Oct 24 '16 at 17:24

The concept of bias is related to sampling distribution of the statistic. Consider, for example, a random sample $X_{1},X_{2},\cdots X_{n}$ from $N(\mu, \sigma^{2})$. Then, it is easy to observe that, the sampling distribution of the sample mean $\bar{X}$ is $N(\mu,\frac{1}{n}\sigma^{2})$. we note that, $E(\bar{X})=\mu$. That is, the center of the sampling distribution of $\bar{X}$ is also $\mu$. Now consider, the statistics, \begin{equation*} S_{1}^{2}=\frac{1}{n-1}\sum_{i=1}^{n}(X_{i}-\bar{X})^{2},\qquad\qquad S_{2}^{2}=\frac{1}{n}\sum_{i=1}^{n}(X_{i}-\bar{X})^{2} \end{equation*} as estimators of the parameter $\sigma^{2}$. It can be shown that \begin{equation*} E(S_{1}^{2})=\sigma^{2} \mbox{ and } E(S_{2}^{2})=\frac{n-1}{n} \sigma^{2} \end{equation*} The sampling distribution of $S_{1}^{2}$ is centered at $\sigma^{2}$, where as that of $S_{2}^{2}$ is not. We say that, the estimator $S_{2}^{2}$ is a biased estimator for $\sigma^{2}$. Now using the definition of bias, we get the amount of bias in $S_{2}^{2}$ in estimating $\sigma^{2}$.

  • 2
    $\begingroup$ Nice simple example (+1). $\endgroup$
    – BruceET
    Oct 24 '16 at 8:24

Roughly speaking there are two favorable attributes for an estimator $T$ of a parameter $\tau$, accuracy and precision. Accuracy is lack of bias and precision is small variance. If an estimator is unbiased, then we just look at its variance. If it is biased we sometimes look at 'mean squared error', which is $$MSE_\tau = E[(T - \tau)^2] = B^2(T) + Var(T).$$

As an example, consider data $X_1, X_2, \dots, X_n \stackrel{iid}{\sim} UNIF(0, \tau).$ The estimator $T_1 = 2\bar X$ is unbiased, and the estimator $T_2 = X_{(n)} = \max(X_i)$ is biased because $E(T_2) = \frac{n}{n+1}\tau.$

As a substitute for a (fairly easy) analytical proof, here is a simulation to show that $T_2$ is 'better' in the sense that its MSE is smaller. We look at a million samples of size $n = 5$ from $UNIF(0, \tau = 1).$

m = 10^6;  n = 5;  tau = 1
x = runif(m*n, 0, 1)
DTA = matrix(x, nrow=m)  # each row a sample of n
t1 = 2*rowMeans(DTA);  t2 = apply(DTA, 1, max)
mean(t1);  mean(t2)
## 0.9997444  # aprx E(T1) = 1   unbiased
## 0.8332033  # aprx E(T2) = 5/6 biased
## 0.8333333
var(t1);  var(t2)
## 0.06665655 # aprx Var(T1)
## 0.01983109 # aprx Var(T2) < Var(T1)
mse.t1 = mean((t1-tau)^2); mse.t2 = mean((t2-tau)^2)
mse.t1; mse.t2
## 0.06665655 # aprx MSE(T1)
## 0.04765219 # aprx MSE(T2) < MSE(T1)

We see that the smaller variance of $T_2$ is enough to overcome its bias to give it the smaller MSE. It is possible to 'unbias' $T_2$ by multiplying by $(n+1)/n$ to get $T_3 = \frac{6}{5}T_2,$ which is unbiased and still has smaller variance than $T_1:$ $Var(T_3) \approx 0.029 < Var(T_1) \approx 0.067.$ The simulated distributions of the three estimators are shown in the figure below.

enter image description here


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.