Lecture 14 Simple Linear Regression
Ordinary Least Squares (OLS)
Consider the following simple linear regression model
Yi = α + βXi + εi where, for each unit i,
• Yi is the dependent variable (response).
• Xi is the independent variable (predictor).
• εi is the error between the observed Yi and what the model predicts.
This model describes the relation between Xi and Yi using an intercept and a slope parameter. The error εi = Yi − α − βXi is also known as the residual because it measures the amount of variation in Yi not explained by the model.
ˆ 2 We saw last class that there existsα ˆ and β that minimize the sum of εi . Specifically, we wish to findα ˆ and βˆ such that n X 2 Yi − (ˆα + βXˆ i) i=1 is the minimum. Using calculus, it turns out that
P(X − X¯)(Y − Y¯ ) αˆ = Y¯ − βˆX¯ βˆ = i i . P 2 (Xi − X¯)
OLS and Transformation
If we center the predictor, X˜i = Xi − X¯, then X˜i has mean zero. Therefore,
P X˜ (Y − Y¯ ) αˆ∗ = Y¯ βˆ∗ = i i . P ˜ 2 Xi ∗ By horizontally shifting the value of Xi, note that β = β, but the intercept changed to the overall average of Yi
Consider the linear transformation Zi = a + bXi with Z¯ = a + bX¯. Consider the linear model
∗ ∗ Yi = α + β Zi + εi
1 Letα ˆ and βˆ be the estimates using X as the predictor. Then the intercept and slope using Z as the predictor are P(Z − Z¯)(Y − Y¯ ) βˆ∗ = i i P 2 (Zi − Z¯) P(a + bX − a − bX¯)(Y − Y¯ ) = i i P(a + bX¯ − a − bX¯)2 P(bX − bX¯)(Y − Y¯ ) = i i P(bX¯ − bX¯)2 b P(X − X¯)(Y − Y¯ ) 1 = i i = βˆ b2 P(X¯ − X¯)2 b and
αˆ∗ = Y¯ − βˆ∗Z¯ βˆ aβˆ a = Y¯ − (a + bX¯) = Y¯ − βˆX¯ − =α ˆ − βˆ b b b
Zi−a We can get the same results by substituting Xi = b into the linear model.
Yi = α + βXi + εi Z − a Y = α + β i + ε i b i aβ β Y = α − + Z + ε i b b i i ∗ ∗ Yi = α + β Zi + εi
Properties of OLS
Given the estimatesα ˆ and βˆ, we can define (1) the estimated predicted value Yˆi and (2) the estimated residualε ˆi.
Yˆi =α ˆ + βXˆ i εˆi = Yi − Yˆi = Yi − αˆ − βXˆ i
The least squared estimates have the following properties. P 1. i εˆi = 0 n n n n X X X X εˆi = (Yi − αˆ − βXˆ i) = Yi − nαˆ − βˆ Xi i=1 i=1 i=1 i=1 = nY¯ − nαˆ − nβˆX¯ = n(Y¯ − αˆ − βˆX¯) = n(Y¯ − (Y¯ − βˆX¯) − βˆX¯) = 0
2 2. X¯ and Y¯ is always on the fitted line.
αˆ + βˆX¯ = (Y¯ − βˆX¯) + βˆX¯ = Y¯
ˆ sY 3. β = rXY × , where sY and sX are the sample standard deviation of X and Y , and rXY is sX the correlation between X and Y . Note that the sample correlation is given by: P (Xi − X¯)(Yi − Y¯ ) rXY = . (n − 1)sX sY Therefore,
P(X − X¯)(Y − Y¯ ) n − 1 s s βˆ = i i × × Y × X P 2 (Xi − X¯) n − 1 sY sX Cov(X,Y ) sY sX = 2 × × sX sY sX Cov(X,Y ) s 1 = × Y × sX sY 1 sX sY = rXY × sX
4. βˆ is a weighted sum of Yi.
P(X − X¯)(Y − Y¯ ) βˆ = i i P 2 (Xi − X¯) P(X − X¯)Y P(X − X¯)Y¯ = i i − i P 2 P 2 (Xi − X¯) (Xi − X¯) P(X − X¯)Y Y¯ P(X − X¯) = i i − i P 2 P 2 (Xi − X¯) (Xi − X¯) P(X − X¯)Y = i i − 0 P 2 (Xi − X¯) ¯ X Xi − X = Y P 2 i (Xi − X¯) ¯ X Xi − X = w Y , w = i i i P 2 (Xi − X¯)
3 5.ˆα is a weighted sum of Yi.
αˆ = Y¯ − βˆX¯ X 1 X = Y − X¯ w Y n i i i ¯ ¯ X 1 X(Xi − X) = v Y , v = − i i i P 2 n (Xi − X¯)
P 6. Xiεˆi = 0. X X X Xiεˆi = (Yi − Yˆi)Xi = (Yi − αˆ − βXˆ i)Xi X = (Yi − (Y¯ − βˆX¯) − βXˆ i)Xi X = ((Yi − Y¯ ) − βˆ(Xi − X¯))Xi X X = Xi(Yi − Y¯ ) − βˆ Xi(Xi − X¯) X 2 X = βˆ (Xi − X¯) − βˆ Xi(Xi − X¯) X 2 = βˆ (Xi − X¯) − Xi(Xi − X¯) ˆ X 2 ¯ ¯ 2 2 ¯ = β Xi − 2XiX + X − Xi + XiX) X 2 = βˆ X¯ − XiX¯ X = βˆX¯ X¯ − Xi = 0
P(X − X¯)(Y − Y¯ ) P X (Y − Y¯ ) Note that βˆ = i i = i i P 2 P 2 (Xi − X¯) (Xi − X¯)
7. Partition of total variation.
(Yi − Y¯ ) = Yi − Y¯ + Yˆi − Yˆi = (Yˆi − Y¯ ) + (Yi − Yˆi)
• (Yi − Y¯ ): total deviation of the observed Yi from a null model (β = 0).
• (Yˆ − Y¯ ): deviation of the modeled value Yˆi (β 6= 0) from a null model (β = 0).
• (Yˆ − Yi): deviation of the modeled value from the observed value.
Total Variations (Sum of Squares)
P 2 • (Yi − Y¯ ) : total variation: (n − 1) × V (Y ). • P(Yˆ − Y¯ )2: variation explained by the model
4 P ˆ 2 P 2 • (Y − Yi) : variation not explained by the model; residual sum of squares (SSE), εˆi
Partition of Total Variation
X 2 X 2 X 2 (Yi − Y¯ ) = (Yˆi − Y¯ ) + (Yi − Yˆi)
Proof:
2 X 2 X h i (Yi − Y¯ ) = (Yˆi − Y¯ ) + (Yi − Yˆi)
X 2 X 2 X = (Yˆi − Y¯ ) + (Yi − Yˆi) + 2 (Yˆi − Y¯ )(Yi − Yˆi) X 2 X 2 = (Yˆi − Y¯ ) + (Yi − Yˆi)
where,
X X h i (Yˆi − Y¯ )(Yi − Yˆi) = (ˆα + βXˆ i) − (ˆα + βˆX¯) εˆi X h i = βˆ(Xi − X¯) εˆi X X = βˆ Xiεˆi − βˆX¯ εˆi = 0 + 0
P 2 8. Alternative formula for εˆi
X 2 X 2 X 2 (Yi − Yˆi) = (Yi − Y¯ ) − (Yˆi − Y¯ ) X 2 X 2 = (Yi − Y¯ ) − (ˆα + βXˆ i − Y¯ ) X 2 X 2 = (Yi − Y¯ ) − (Y¯ + βXX¯ + βXˆ i − Y¯ ) 2 X 2 X h i = (Yi − Y¯ ) − βˆ(Xi − X¯)
X 2 2 X 2 = (Yi − Y¯ ) − βˆ (Xi − X¯)
or
P ¯ ¯ 2 X X (Xi − X)(Yi − Y ) (Y − Yˆ )2 = (Y − Y¯ )2 − i i i P 2 (Xi − X¯)
9. Variation Explained by the Model:
5 P 2 An easy way to calculate (Yˆi − Y¯ ) is
X 2 X 2 X 2 (Yˆi − Y¯ ) = (Yi − Y¯ ) − (Yi − Yˆi) P ¯ ¯ 2 X (Xi − X)(Yi − Y ) = (Y − Y¯ )2 − i P 2 (Xi − X¯) P 2 (Xi − X¯)(Yi − Y¯ ) = P 2 (Xi − X¯)
10. Coefficient of Determination r2:
variation explained by the model P(Yˆ − Y¯ )2 r2 = = i P 2 total variation (Yi − Y¯ )
(a) r2 is related the residual variance.
X 2 X 2 X 2 (Yi − Yˆi) = (Yi − Y¯ ) − (Yˆi − Y¯ ) P ¯ 2 hX X i (Yi − Y ) = (Y − Y¯ )2 − (Yˆ − Y¯ )2 × i i P 2 (Yi − Y¯ ) 2 X 2 = (1 − r ) × (Yi − Y¯ )
Diving the right-hand side by n − 2 and the left-hand side by n − 1, for large n
s2 ≈ (1 − r2)Var(Y )
(b) r2 is the square of the sample correlation between X and Y .
P(Yˆ − Y¯ )2 r2 = i P 2 (Yi − Y¯ ) P 2 (Xi − X¯)(Yi − Y¯ ) = P 2 P 2 (Yi − Y¯ ) (Xi − X¯) " #2 P(X − X¯)(Y − Y¯ ) = i i = r2 pP 2pP 2 XY (Yi − Y¯ ) (Xi − X¯)
Statistical Inference for OLS Estimates
Parametersα ˆ and βˆ can be estimated for any given sample of data. Therefore, we also need to consider their sampling distributions because each sample of (Xi,Yi) pairs will result in different estimates of α and β.
6 Consider the following distribution assumption on the error,
iid 2 Yi = α + βXi + εi εi ∼ N (0, σ ) .
The above is now a statistical model that describes the distribution of Yi given Xi. Specifically, we assume the observed Yi is error-prone but centered around the linear model for each value of Xi.
iid 2 Yi ∼ N(α + βXi, σ )
The error distribution results in the following assumptions.
1. Homoscedasticity: variance of Yi is the same for any Xi.
2 V (Yi) = V (α + βXi + εi) = V (εi) = σ
2. Linearity: the mean of Yi is linear with respect to Xi
E(Yi) = E(α + βXi + εi) = α + βXi + E(εi) = α + βXi
3. Independence: Yi are independent of each other.
Cov(Yi,Yj) = Cov(α + βXi + εi, α + βXj + εj) = Cov(εi, εj) = 0
We can now investigate the bias and variance of OLS estimatorsα ˆ and βˆ. We assume that Xi are known constants.
• βˆ is unbiased. ¯ hX i Xi − X E[βˆ] = E w Y , w = i i i P 2 (Xi − X¯) X = wiE[Yi] X = wi(α + βXi) X X = α wi + β wiXi = β
7 Because
X 1 X w = (X − X¯) = 0 i P 2 i (Xi − X¯) X 1 X w X = X (X − X¯) i i P 2 i i (Xi − X¯) 1 X ¯ = P P Xi(Xi − X) Xi(Xi − X¯) − X¯(Xi − X¯) 1 X ¯ = P P Xi(Xi − X) Xi(Xi − X¯) − X¯ (Xi − X¯) 1 X ¯ = P Xi(Xi − X) Xi(Xi − X¯) − 0 = 1
• αˆ is unbiased.
E[ˆα] = E[Y¯ − βˆX¯] = E[Y¯ ] − XE¯ [βˆ] 1 X 1 X = E Y − βX¯ = E [α + βX ] − βX¯ n i n i 1 = (nα + nβX ) − βX¯ n i = α + βX¯ − βX¯ = α
σ2 σ2 • βˆ has variance = , where σ2 is the residual variance and s2 is the Pn ¯ 2 2 x i=1(Xi − X1) (n − 1)sx sample variance of x1, . . . , xn. ¯ hX i Xi − X V [βˆ] = V w Y , w = i i i P 2 (Xi − X¯) X 2 = wi V [Yi] 2 X 2 = σ wi
where
X 1 X w2 = (X − X¯)2 i P 2 2 i [ (Xi − X¯) ] 1 = P 2 (Xi − X¯)
8 1 X¯ 2 • αˆ has variance σ2 + Pn ¯ 2 n i=1(Xi − X1) ¯ ¯ hX i 1 X(Xi − X) V [ˆα] = V v Y , v = − i i i P 2 n (Xi − X¯) X 2 = vi V [Yi] ¯ ¯ 2 X 1 X(Xi − X) = σ2 − P 2 n (Xi − X¯) ¯ 2 ¯ 2 ¯ ¯ X 1 X X (Xi − X) X X(Xi − X) = σ2 − − 2 2 P 2 2 P 2 n [ (Xi − X¯) ] n (Xi − X¯) 1 X¯ 2 X 2X¯ X = σ2 − (X − X¯)2 − (X − X¯) P 2 2 i P 2 i n [ (Xi − X¯) ] n (Xi − X¯) 1 X¯ 2 = σ2 + Pn ¯ 2 n i=1(Xi − X1)
• αˆ and βˆ are negatively correlated. ¯ ¯ ¯ X X 1 X(Xi − X) Xi − X Cov(ˆα, βˆ) = Cov v Y , w Y , v = − , w = i i i i i P 2 i P 2 n (Xi − X¯) (Xi − X¯) X = viwiCov (Yi,Yi) X 2 = viwiσ ¯ ¯ ¯ X 1 X(Xi − X) Xi − X = σ2 − P 2 P 2 n (Xi − X¯) (Xi − X¯) ¯ ¯ ¯ ¯ X Xi − X X X(Xi − X) Xi − X = σ2 − P 2 P 2 P 2 n (Xi − X¯) (Xi − X¯) (Xi − X¯) P(X − X¯) P X¯(X − X¯)2 = σ2 i − i P 2 P 2 2 n (Xi − X¯) [ (Xi − X¯) ] X¯ P(X − X¯)2 = σ2 0 − i P 2 2 [ (Xi − X¯) ] σ2X¯ = − P 2 (Xi − X¯)
Therefore, the sampling distributions for the regression parameters are:
1 X¯ 2 σ2 αˆ ∼ N α, σ2 + βˆ ∼ N β, Pn ¯ 2 Pn ¯ 2 n i=1(Xi − X1) i=1(Xi − X1)
9 However, in most cases, we don’t know what σ2 and will need to estimate it from the data. One unbiased estimator for σ2 is obtained from the residuals.
N 1 X s2 = (Y − Yˆ )2 n − 2 i i i=1
By replacing σ2 with s2, we have the following sampling distributions forα ˆ and βˆ. 1 X¯ 2 s2 αˆ ∼ t α, s2 + βˆ ∼ t β, n−2 Pn ¯ 2 n−2 Pn ¯ 2 n i=1(Xi − X1) i=1(Xi − X1) Note that our degrees of freedom is n − 2 because we estimated two parameters (α and β) in order P to calculate the residuals. Similarly recall that (Yˆi − Yi) = 0.
Knowing thatα ˆ and βˆ follow a t-distribution with degrees of freedom n-2, we are able to construct confidence interval and carry out hypothesis test as before. For example, a 90% confidence interval for β is given by s s2 βˆ ± t × . 5%,n−2 Pn ¯ 2 i=1(Xi − X1)
Similarly, consider testing the hypotheses H0β = β0 versus HAβ 6= β0. The corresponding t-test statistic is given by βˆ − β0 s ∼ tn−2 . s2 Pn ¯ 2 i=1(Xi − X1)
Prediction
Based on the sampling distribution ofα ˆ and βˆ we can also investigate the uncertainty associated with Yˆ0 =α ˆ + βXˆ 0 on the fitted line for any X0. Specifically, 1 (X − X¯)2 Yˆ ∼ N α + X β, σ2 + 0 . 0 0 P 2 n (Xi − X¯)
Note that α + X0β is the mean of Y given X0. The textbook uses the notation µ0. Replacing the unknown with our estimates, we have 1 (X − X¯)2 Yˆ ∼ t αˆ + X β,ˆ s2 + 0 . 0 n−2 0 P 2 n (Xi − X¯)
1 (X − X¯)2 • The fitted value Yˆ =α ˆ + βXˆ has mean α + X β and variance σ2 + 0 . 0 0 0 P 2 n (Xi − X¯)
E(Yˆ0) = E(ˆα + βXˆ 0) = E(ˆα) + X0E(βˆ) = α + X0β
10 ˆ ˆ 2 ˆ ˆ V (Y0) = V (ˆα + βX0) = V (ˆα) + X0 V (β) + X0Cov(ˆα, β) 1 X¯ 2 X2 X X¯ = σ2 + + σ2 0 − 2σ2 0 Pn ¯ 2 Pn ¯ 2 P ¯ 2 n i=1(Xi − X1) i=1(Xi − X1) (Xi − X) 1 (X − X¯)2 = σ2 + 0 P 2 n (Xi − X¯)
∗ We can also consider the uncertainty of a new observation Y0 at X0. In other words, what values ∗ ∗ ˆ of Y0 do I expect to see at X0 ? Note that this is different from Y0 which is the fitted value that represents the mean of Y at X0. Since
∗ 2 Y0 = α + βX0 + ε, ε ∼ N (0, σ ) , by replacing the unknown parameters with our estimates, we obtain
1 (X − X¯)2 Y ∗ ∼ t αˆ + βXˆ , s2 + s2 + 0 . 0 n−2 0 P 2 n (Xi − X¯) Again, the above distribution describes the actual observable data, instead of model parameters; interval estimates are also known as prediction intervals, instead of confidence intervals.
11