Estadística II Chapter 5. Regression analysis (second part)

Transcription

1 Estadística II Chapter 5. Regression analysis (second part)

2 Chapter 5. Regression analysis (second part) Contents Diagnostic: Residual analysis The ANOVA (ANalysis Of VAriance) decomposition Nonlinear relationships and linearizing transformations Simple linear regression model in the matrix notation Introduction to the multiple linear regression

3 Lesson 5. Regression analysis (second part) Bibliography references Newbold, P. Estadística para los negocios y la economía (1997) Chapters 12, 13 and 14 Peña, D. Regresión y Análisis de Experimentos (2002) Chapters 5 and 6

4 Regression diagnostics Diagnostics The theoretical assumptions for the simple linear regression model with one response variable y and one explicative variable x are: Linearity: y i = β 0 + β 1 x i + u i, for i = 1,..., n Homogeneity: E[u i ] = 0, for i = 1,..., n Homoscedasticity: Var[ui ] = σ 2, for i = 1,..., n Independence: ui and u j are independent for i j Normality: ui N(0, σ 2 ), for i = 1,..., n We study how to apply diagnostic procedures to test if these assumptions are appropriate for the available data (x i, y i ) Based on the analysis of the residuals ei = y i ŷ i

5 Diagnostics: scatterplots Scatterplots The simplest diagnostic procedure is based on the visual examination of the scatterplot for (x i, y i ) Often this simple but powerful method reveals patterns suggesting whether the theoretical model might be appropriate or not We illustrate its application on a classical example. Consider the four following datasets

6 Diagnostics: scatterplots The Anscombe datasets TABLE 3-10 Four Data Sets DATASET 1 DATASET 2 X Y X Y ' DATASET 3 DATASET 4 X Y X Y SOURCE: F. J. Anscombe, op. cit.

7 Diagnostics: scatterplots The Anscombe example The estimated regression model for each of the four previous datasets is the same ŷ i = x i n = 11, x = 9.0, ȳ = 7.5, rxy = The estimated standard error of the estimator ˆβ 1, s 2 R (n 1)s 2 x takes the value in all four cases. The corresponding T statistic takes the value T = 0.5/0.118 = But the corresponding scatterplots show that the four datasets are quite different. Which conclusions could we reach from these diagrams?

8 Diagnostics: scatterplots Anscombe data scatterplots o Data Set 1 Data Set Data Set 3 Data Set FIGURE 3-29 Scatterplots for the four data sets of Table 3-10 SOURCE: F. J. Anscombe, op cit.

9 Residual analysis Further analysis of the residuals If the observation of the scatterplot is not sufficient to reject the model, a further step would be to use diagnostic methods based on the analysis of the residuals e i = y i ŷ i This analysis starts by standarizing the residuals, that is, dividing them by the residuals (quasi-)standard deviation s R. The resulting quantities are known as standardized esiduals: e i s R Under the assumptions of the linear regression model, the standardized residuals are approximately independent standard normal random variables A plot of these standardized residuals should show no clear pattern

10 Residual analysis Residual plots Several types of residual plots can be constructed. The most common ones are: Plot of standardized residuals vs. x Plot of standardized residuals vs. ŷ (the predicted responses) Deviations from the model hypotheses result in patterns on these plots, which should be visually recognizable

11 Residual plots examples Consistency 5.1: Ej: of the consistencia theoretical model con el modelo teórico

12 Residual plots examples Nonlinearity 5.1: Ej: No linealidad

13 Residual plots examples Heteroscedasticity 5.1: Ej: Heterocedasticidad

14 Residual analysis Outliers In a plot of the regression line we may observe outlier data, that is, data that show significant deviations from the regression line (or from the remaining data) The parameter estimators for the regression model, ˆβ 0 and ˆβ 1, are very sensitive to these outliers It is important to identify the outliers and make sure that they really are valid data

15 Residual analysis Normality of the errors One of the theoretical assumptions of the linear regression model is that the errors follow a normal distribution This assumption can be checked visually from the analysis of the residuals e i, using different approaches: By inspection of the frequency histogram for the residuals By inspection of the Normal Probability Plot of the residuals (significant deviations of the data from the straight line in the plot correspond to significant departures from the normality assumption)

16 The ANOVA decomposition Introduction ANOVA: ANalysis Of VAriance When fitting the simple linear regression model ŷ i = ˆβ 0 + ˆβ 1 x i to a data set (x i, y i ) for i = 1,..., n, we may identify three sources of variability in the responses variability associated to the model: SSM = n i=1 (ŷ i ȳ) 2, where the initials SS denote sum of squares and M refers to the model variability of the residuals: SSR = n i=1 (y i ŷ i ) 2 = n i=1 e2 i total variability: SST = n i=1 (y i ȳ) 2 The ANOVA decomposition states that: SST = SSM + SSR

17 The ANOVA decomposition The coefficient of determination R 2 The ANOVA decomposition states that SST = SSM + SSR Note that y i ȳ = (y i ŷ i ) + (ŷ i ȳ) SSM = n i=1 (ŷ i ȳ) 2 measures the variation in the responses due to the regression model (explained by the predicted values ŷ i ) Thus, the ratio SSR/SST is the proportion of the variation in the responses that is not explained by the regression model The ratio R 2 = SSM/SST = 1 SSR/SST is the proportion of the variation in the responses that is explained by the regression model. It is known as the coefficient of determination The value of the coefficient of determination satisfies R 2 = r 2 xy (the squared correlation coefficient) For example, if R 2 = 0.85 the variable x explains 85% of the variation in the response variable y

18 The ANOVA decomposition ANOVA table Source of variability SS DF Mean F ratio Model SSM 1 SSM/1 SSM/s 2 R Residuals/errors SSR n 2 SSR/(n 2) = s 2 R Total SST n 1

19 The ANOVA decomposition ANOVA hypothesis testing Hypothesis test, H 0 : β 1 = 0 vs. H 1 : β 1 0 Consider the ratio F = SSM/1 SSR/(n 2) = SSM sr 2 Under H 0, F follows an F 1,n 2 distribution Test at a significance level α: reject H 0 if F > F 1,n 2;α How does this result relate to the test based on the Student-t we saw in Lesson 4? They are equivalent

20 The ANOVA decomposition Excel output

21 Nonlinear relationships and linearizing transformations Introduction Consider the case when the deterministic part f (x i ; a, b) of the response in the model y i = f (x i ; a, b) + u i, i = 1,..., n is a nonlinear function of x that depends on two parameters a and b (for example, f (x; a, b) = ab x ) In some cases we may apply transformations to the data to linearize them. We are then able to apply the linear regression procedure From the original data (x i, y i ) we obtain the transformed data (x i, y i ) The parameters β 0 and β 1 corresponding to the linear relation between x i and y i are transformations of the parameters a and b

22 Nonlinear relationships and linearizing transformations Linearizing transformations Examples of linearizing transformations: If y = f (x; a, b) = ax b then log y = log a + b log x. We have y = log y, x = log x, β 0 = log a, β 1 = b If y = f (x; a, b) = ab x then log y = log a + (log b)x. We have y = log y, x = x, β 0 = log a, β 1 = log b If y = f (x; a, b) = 1/(a + bx) then 1/y = a + bx. We have y = 1/y, x = x, β 0 = a, β 1 = b If y = f (x; a, b) = log(ax b ) then y = log a + b(log x). We have y = y, x = log x, β 0 = log a, β 1 = b

23 A matrix treatment of linear regression Introduction Remember the simple linear regression model, y i = β 0 + β 1 x i + u i, i = 1,..., n If we write one equation for each one of the observations, we have y 1 = β 0 + β 1 x 1 + u 1 y 2 = β 0 + β 1 x 2 + u 2.. y n = β 0 + β 1 x n + u n

24 A matrix treatment of linear regression The model in matrix form We can write the preceding equations in matrix form as y 1 β 0 + β 1 x 1 u 1 y 2 β 0 + β 1 x 2 u 2. y n =. + β 0 + β 1 x n. u n And splitting the parameters β from the variables x i, y 1 1 x 1 u 1 y 2 1 x 2 ( ) β0 u 2. y n =.. 1 x n β 1 +. u n

25 A matrix treatment of linear regression The regression model in matrix form We can write the preceding matrix relationship y 1 1 x 1 y 2 1 x 2 ( ) β0 as. y n =.. 1 x n β 1 y = Xβ + u + y: response vector; X: explanatory variables matrix (or experimental design matrix); β: vector of parameters; u: error vector u 1 u 2. u n

26 The regression model in matrix form Covariance matrix for the errors We denote as Cov(u) the n n matrix of covariances for the errors. Its (i, j)-th element is given by cov(u i, u j ) = 0 if i j and cov(u i, u i ) = Var(u i ) = σ 2 Thus, Cov(u) is the identity matrix I n n multiplied by σ 2 : σ σ 2 0 Cov(u) = = σ2 I 0 0 σ 2

27 The regression model in matrix form Least-squares estimation The least-squares vector parameter estimate ˆβ is the unique solution of the 2 2 matrix equation (check the dimensions) (X T X) ˆβ = X T y, that is, ˆβ = (X T X) 1 X T y. The vector ŷ = (ŷ i ) of response estimates is given by ŷ = X ˆβ and the residual vector is defined as e = y ŷ

28 The multiple linear regression model Introduction Use of the simple linear regression model: predict the value of a response y from the value of an explanatory variable x In many applications we wish to predict the response y from the values of several explanatory variables x 1,..., x k For example: forecast the value of a house as a function of its size, location, layout and number of bathrooms forecast the size of a parliament as a function of the population, its rate of growth, the number of political parties with parliamentary representation, etc.

29 The multiple linear regression model The model Use of the multiple linear regression model: predict a response y from several explanatory variables x 1,..., x k If we have n observations, for i = 1,..., n, y i = β 0 + β 1 x i1 + β 2 x i2 + + β k x ik + u i We assume that the error variables u i are independent random variables following a N(0, σ 2 ) distribution

30 The multiple linear regression model The least-squares fit We have n observations, and for i = 1,..., n y i = β 0 + β 1 x i1 + β 2 x i2 + + β k x ik + u i We wish to fit to the data (x i1, x i2,..., x ik, y i ) a hyperplane of the form ŷ i = ˆβ 0 + ˆβ 1 x i1 + ˆβ 2 x i2 + + ˆβ k x ik The residual for observation i is defined as e i = y i ŷ i We obtain the parameter estimates ˆβ j as the values that minimize the sum of the squares of the residuals

31 The multiple linear regression model The model in matrix form We can write the model as a matrix relationship, β y 1 1 x 11 x 12 x 0 1k y 2. = 1 x 21 x 22 x 2k β β y n 1 x n1 x n2 x nk and in compact form as y = Xβ + u y: response vector; X: explanatory variables matrix (or experimental design matrix); β: vector of parameters; u: error vector β k u 1 u 2. u n

32 The multiple linear regression model Least-squares estimation The least-squares vector parameter estimate ˆβ is the unique solution of the (k + 1) (k + 1) matrix equation (check the dimensions) (X T X) ˆβ = X T y, and as in the k = 1 case (simple linear regression) we have ˆβ = (X T X) 1 X T y. The vector ŷ = (ŷ i ) of response estimates is given by ŷ = X ˆβ and the residual vector is defined as e = y ŷ

33 The multiple linear regression model Variance estimation For the multiple linear regression model, an estimator for the error variance σ 2 is the residual (quasi-)variance, s 2 R = n i=1 e2 i n k 1, and this estimator is unbiased Note that for the simple linear regression case we had n 2 in the denominator

34 The multiple linear regression model The sampling distribution of ˆβ Under the model assumptions, the least-squares estimator ˆβ for the parameter vector β follows a multivariate normal distribution E( ˆβ) = β (it is an unbiased estimator) The covariance matrix for ˆβ is Cov( ˆβ) = σ 2 (X T X) 1 We estimate Cov( ˆβ) using s 2 R (XT X) 1 The estimate of Cov( ˆβ) provides estimates s 2 ( ˆβ j ) for the variance Var( ˆβ j ). s( ˆβ j ) is the standard error of the estimator ˆβ j If we standardize ˆβ j we have ˆβ j β j s( ˆβ j ) t n k 1 (the Student-t distribution)

35 The multiple linear regression model Inference on the parameters β j Confidence interval for the partial coefficient β j at a confidence level 1 α ˆβ j ± t n k 1;α/2 s( ˆβ j ) Hypothesis testing for H 0 : β j = 0 vs. H 1 : β j 0 at a confidence level α Reject H0 if T > t n k 1;α/2, where is the test statistic T = ˆβ j s( ˆβ j )

36 The ANOVA decomposition The multivariate case ANOVA: ANalysis Of VAriance When fitting the multiple linear regression model ŷ i = ˆβ 0 + ˆβ 1 x i1 + + ˆβ k x ik to a data set (x i1,..., x ik, y i ) for i = 1,..., n, we may identify three sources of variability in the responses variability associated to the model: SSM = n i=1 (ŷ i ȳ) 2, where the initials SS denote sum of squares and M refers to the model variability of the residuals: SSR = n i=1 (y i ŷ i ) 2 = n i=1 e2 i total variability: SST = n i=1 (y i ȳ) 2 The ANOVA decomposition states that: SST = SSM + SSR

37 The ANOVA decomposition The coefficient of determination R 2 The ANOVA decomposition states that SST = SSM + SSR Note that y i ȳ = (y i ŷ i ) + (ŷ i ȳ) SSM = n i=1 (ŷ i ȳ) 2 measures the variation in the responses due to the regression model (explained by the predicted values ŷ i ) Thus, the ratio SSR/SST is the proportion of the variation in the responses that is not explained by the regression model The ratio R 2 = SSM/SST = 1 SSR/SST is the proportion of the variation in the responses that is explained by the explanatory variables. It is known as the coefficient of multiple determination For example, if R 2 = 0.85 the variables x 1,..., x k explain 85% of the variation in the response variable y

38 The ANOVA decomposition ANOVA table Source of variability SS DF Mean F ratio Model SSM k SSM/k (SSM/k)/ Residuals/errors SSR n k 1 Total SST n 1 SSR/(n k 1) = sr 2

39 The ANOVA decomposition ANOVA hypothesis testing Hypothesis test, H 0 : β 1 = β 2 = = β k = 0 vs. H 1 : β j 0 for some j = 1,..., k H 0 : the response does not depend on any x j H 1 : the response depends on at least one x j ˆβ j interpretation: keeping everything else the same, a one unit increase in x j will produce, on average, a change by ˆβ j in y Consider the ratio F = SSM/k SSR/(n k 1) = SSM sr 2 Under H 0, F follows an F k,n k 1 distribution Test at a significance level α: reject H 0 if F > F k,n k 1;α