You are on page 1of 5

6.

4 Prediction
Once our regression model has been estimated, we can use the estimated model to predict or
forecast some future value of the dependent variable, Y , that is associated with a given level of
the regression, X.
There are two types of prediction:

1. Prediction of the conditional mean value of Y that is associated with a given level of X, that
lies on the population regression line.

2. Prediction of the individual Y value that is associated with a given level of X.

The first and second types are known as mean prediction and individual prediction respectively.
In what follows, we will assume that X0 is a given level of X and Y0 is the corresponding true
level of Y .

6.4.1 Mean Prediction

Example 6.11. For Yi = β0 + β1 Xi + Ui , let Ŷi = 2.5 + 0.1Xi . Given that X0 = 20, what is
E(Y | X0 )?

Solution:
Using the equation of the fitted line,

Ŷ0 = 2.5 + 0.1(X0 )


= 2.5 + 0.1(20)
= 2.5 + 2 = 4.5

where Ŷ0 is an estimator of E(Y | X0 ). That is, Ŷ0 = Ê(Y | X0 ).

Note that Ŷ0 is the point predictor of E(Y | Xi ) and is BLUE. Given that Ŷ0 is an estimator,
we can conduct hypothesis tests for this estimator. Recall that to construct a test statistic for
hypothesis testing we need to find variance of the relevant estimator. In this case we need to find
var(Ŷ0 ).
Before we compute var(Ŷ0 ), let us show that Ŷ0 is an unbiased estimator of E(Y0 | X0 ).
For unbiasedness:
Let
Y0 = β0 + β1 X0 + Ui (6.4)

35
taking conditional expectation of (6.4) at X0 , yields

E(Y0 | X0 ) = β0 + β1 X0

Also using (6.4), the fitted line is


Ŷ0 = β̂0 + β̂1 X0 (6.5)

Taking expectation of (6.5) we have:

E(Ŷ0 ) = E(β̂0 ) + E(β̂1 X0 )


= β0 + E(β̂1 )X0
= β0 + β1 X0 by unbiasedness of β̂0 , β̂1 .

N ow, since E(Ŷ0 ) = β0 + β1 X0 = E(Y0 | X0 ),

this implies that Ŷ0 is an unbiased estimator of E(Y0 | X0 ).

Question: What is var(Ŷ0 )?


We will now show that · ¸
2 1 (X0 − X̄)2
var(Ŷ0 ) = σ +P .
n (Xi − X̄)2

36
Ŷ0 = β̂0 + β̂1 X0
=⇒ var(Ŷ0 ) = var(β̂0 ) + var(β̂1 X0 ) + 2Cov(β̂0 , β̂1 X0 )
= var(β̂0 ) + var(β̂1 )X0 2 + 2X0 Cov(β̂0 , β̂1 )
P 2 · ¸
σ2 Xi 2 σ2 −σ 2 X̄
= P + X0 P + 2X0 P
n (Xi − X̄)2 (Xi − X̄) (Xi − X̄)2
·P 2 ¸
σ2 Xi
= P + X0 2 − 2X0 X̄
(Xi − X̄) 2 n
2
·P 2 ¸
σ Xi 2 2 2
= P − X̄ + X0 − 2X0 X̄ + X̄
(Xi − X̄)2 n
·P 2 ¸
σ2 Xi 2 2
= P − X̄ + (X0 − X̄)
(Xi − X̄)2 n
·P 2 ¸
σ2 Xi − nX̄ 2 2
= P + (X0 − X̄)
(Xi − X̄)2 n
·P ¸
σ2 (Xi − X̄)2 2
= P + (X0 − X̄)
(Xi − X̄)2 n
· ¸
2 1 (X0 − X̄)2
=⇒ var(Ŷ0 ) = σ +P as required.
n (Xi − X̄)2

Given Ŷ0 and var(Ŷ0 ), we can now conduct hypothesis testing.

1. To conduct a test statistic for the null hypothesis,

H0 : Y0 = β0 + β1 X0
q
Ŷ0 − (β0 + βX0 )
we use tŶ0 = , whereSE(Ŷ0 ) = var(Ŷ0 ).
SE(Ŷ0 )

Note that as before we can show that tŶ0 ∼ tn−2(df ) .

2. To construct a (1 − α)% confidence interval for the true E(Y0 | X0 ), we use

P r[β̂0 + β̂1 X0 − t α2 SE(Ŷ0 ) ≤ β̂0 + β̂1 X0 + t α2 SE(Ŷ0 )] = 1 − α

37
6.4.2 Individual Prediction

This is the prediction or forecasting of the individual Y value associated with X = X0


Consider the population equation evaluated a specific point (X0 , Y0 ) so that

Y0 = β0 + β1 X0 + U0 .
T hen Y0 − Ŷ0 = (β0 − β̂0 ) + (β1 − β̂1 )X0 + U0 (6.6)
E(Y0 − Ŷ0 ) = E(β0 − β̂0 ) + E(β1 − β̂1 )X0 + E(U0 )
= 0

This shows that Ŷ0 is also an unbiased estimator for Y0 !!!


Question: What is var(Y0 − Ŷ0 )?
Using (6.6):

var(Y0 − Ŷ0 ) = var(β̂0 ) + X0 2 var(β̂1 ) + 2X0 Cov(β̂0 , β̂1 ) + var(U0 )


· ¸
2 1 (X0 − X̄)2
= σ 1+ + P .
n (X1 − X̄)2

Note that: mean prediction has smaller variance than individual prediction.

(i) To construct a test statistic for the null hypothesis

H0 : Y0 − Ŷ0 = 0 or H0 : Y0 = Ŷ0
q
Y0 − Ŷ0
t(Y0 −Ŷ0 ) = where SE(Y0 − Ŷ0 ) = var(Y0 − Ŷ0 ).
SE(Y0 − Ŷ0 )

(ii) var(Y0 − Ŷ0 ) > var(Ŷ0 ).

(iii) var(Y0 − Ŷ0 ) and var(Ŷ0 ) are minimized when X0 = X̄0 . This is because the point (X̄, Ȳ )
lies on the sample regression line; thus, at (X̄, Ȳ ), the corresponding residual is zero. What
is the implication of this for points far away from the sample mean X̄?

38
6.5 Interpreting Regression Results under different functional forms
Consider the functional forms:

M odel 1 : Yi = β0 + β1 Xi + ui (6.7)
M odel 2 : Yi = β0 + β1 ln Xi + ui (6.8)
M odel 3 : ln Yi = β0 + β1 Xi + ui (6.9)
M odel 4 : ln Yi = β0 + β1 ln Xi + ui (6.10)

Question: how do we interpret the estimated coefficients in each model?

Question: how do we compute elasticities using these estimated coefficients?

6.6 Economic, or Practical vs. Statistical significance


Economic or practical significance of a regressor X is related to the magnitude and sign of its
estimated coefficient.

Statistical significance of a regressor X is related solely to the magnitude of its associated t-


statistic, say tβ̂1 . Recall that such a t-statistic can be large if β̂1 is large or SE(β̂1 ) is small. It is
important to distinguish these two sources of statistical significance.

39

You might also like