Statistical Sciences 3859a More Quiz 1 Practice · PDF file2 (a) Write down the least-squares...
Click here to load reader
Transcript of Statistical Sciences 3859a More Quiz 1 Practice · PDF file2 (a) Write down the least-squares...
1
Department of Statistical and Actuarial SciencesStatistical Sciences 3859a
More Quiz 1 Practice Solutions
1. Consider the modelyi = β1xi + εi, i = 1, 2, . . . , n
where ε1, . . . , εn are independent normal random variables with mean 0 and varianceσ2.
(a) Derive the least-squares estimator for β1, β1.The least-squares estimator for β1 minimizes L =
∑(yi − β1xi)
2. Set the derivative ∂L∂β1
= 0 to obtain∑xiyi = β1
∑x2
i
from which it follows that
β1 =
∑xiyi∑x2
i
.
(b) Show that the estimator obtained above is unbiased.
E[β1] =
∑xiE[yi]∑
x2i
=
∑xiβ1xi∑
x2i
= β1.
Therefore, β1 is an unbiased estimator of β1.
(c) Show that the variance of the estimator obtained above is σ2/∑
x2i .
Var(β1) =
∑x2
i Var(yi)∑x2
i
=σ2
∑x2
i(∑x2
i
)2=
σ2∑x2
i
.
(d) Write down an unbiased estimator for σ2.
σ2 =
∑(ei)
2
n− 1
where ei = yi − β1xi. The n− 1 appearing in the denominator follows from the fact that only one parameter
has been estimated.
(e) Derive a formula to compute a (1− α)100% confidence interval for β1.
β1 ± tn−1,α/2σ√∑
x2i
(f) Suppose a new response y0 is to be observed at the predictor value x0, using the point predictory0 = β1x0. Show that the variance of y0 is σ2(x2
0/∑
x2 + 1). (Wording is bad here. We want thevariance of the predicted response at x0.)We would predict the new response to be y0 + ε which has variance σ2x2
0/∑
x2i + σ2.
(g) Derive a formula to compute a (1− α)100% prediction interval for y0.Our best prediction of the new response is y0 (i.e. our estimate of its expected value.) Thus, a predictioninterval is given by
y0 ± tn−1,α/2
√σ2
(x2
0∑x2
i
+ 1
).
2. Consider the modelyi = β0 + β1xi + εi, i = 1, 2, . . . , n
where ε1, . . . , εn are independent normal random variables with mean 0 and varianceσ2.
2
(a) Write down the least-squares estimator for β0.
β0 = y − β1x
where β1 = Sxy/Sxx.
(b) Show that the above estimator is unbiased.
E[β0] = E[y]− xE[β1]
= β0 + β1x− xβ1 (since β1 is unbiased)
= β0.
Therefore, the estimator is unbiased for β0.
(c) Derive the variance formula for the above estimator.
β0 =1
n
∑yi −
x
Sxx
∑(xi − x)yi
=∑(
1
n− x(xi − x)
Sxx
)yi.
Since the yi’s are independent with variance σ2, we have
Var(β0) =∑(
1
n− x(xi − x)
Sxx
)2
σ2.
Expanding the quadratic term, and noting that∑
(xi − x) = 0 gives
Var(β0) = σ2
(1
n+
x2
Sxx
).
(d) Give a formula to compute a (1− α)100% confidence interval for β0.
β0 ± tn−2,α/2
√MSE
(1
n+
x2
Sxx
)(e) Find the expected value and variance of y1−y
x1−x . Is it unbiased for β1? Is there an estimator withsmaller variance? Explain briefly.
E [] =1
x1 − x(β0 + β1x1 − β0 − β1x) = β1.
Therefore, this estimator is unbiased for β1. It is also linear in the responses. By the Gauss-Markov theorem,the least-squares estimator for β1 has the smallest variance among all such estimators. Therefore, there is alinear unbiased estimator with a smaller variance than the one given here.