2. Regression Analysis

3. Introduction to Regression Analysis (RA) Regression Analysis is used to estimate a function f ( ) that describes the relationship between a continuous dependent variable and one or more independent variables. Y = f(X 1, X 2, X 3,…, X n ) +  Note: f ( ) describes systematic variation in the relationship.  represents the unsystematic variation (or random error) in the relationship.

4. In other words, the observations that we have interest can be separated into two parts: Y = f(X 1, X 2, X 3,…, X n ) +  Observations = Model + Error Observations = Signal + Noise Ideally, the noise shall be very small, comparing to the model.

5. Signal to Noise What we observe can be divided into: what we see signal noise

6. Model specification y i =  0 +  1 X i +  2 Z i If the true function is: And we fit: y i =  0 +  1 X i +  2 Z i + e i Our model is exactly specified and we obtain an unbiased and efficient estimate.

7. Model specification y i =  0 +  1 X i +  2 Z i +  3 X i Z i +  4 Z i And finally, if the true function is: And we fit: y i =  0 +  1 X i +  2 Z i + e i Our model is underspecified, we excluded some necessary terms, and we obtain a biased estimate. 2

8. Model specification y i =  0 +  1 X i +  2 Z i On the other hand, if the true function is: And we fit: y i =  0 +  1 X i +  2 Z i +  3 X i Z i + e i Our model is overspecified, we included some unnecessary terms, and we obtain an inefficient estimate.

9. Model specification if specify the model exactly, there is no bias if you overspecify the model (add more terms than needed), result is unbiased, but inefficient if you underspecify the model (omit one or more necessary terms (the result is biased) Overall Strategy –best option is to exactly specify the true function –we would prefer to err by overspecifying our model because that only leads to inefficiency –Therefore, start with a likely overspecified model and reduce it

10. An Example Consider the relationship between advertising ( X 1 ) and sales ( Y ) for a company. There probably is a relationship......as advertising increases, sales should increase. But how would we measure and quantify this relationship?

11. A Scatter Plot of the Data Advertising (in $1,000s) Sales (in $1,000s)

12. The Nature of a Statistical Relationship Regression Curve Probability distributions for Y at different levels of X Y X

13. A Simple Linear Regression Model The scatter plot shows a linear relation between advertising and sales. So the following regression model is suggested by the data, This refers to the true relationship between the entire population of advertising and sales values. The estimated regression function (based on our sample) will be represented as,

14. Determining the Best Fit Numerical values must be assigned to b 0 and b 1 The method of “least squares” selects the values that minimize: If ESS=0 our estimated function fits the data perfectly. We could solve this problem using Solver...

15. Estimation – Linear Regressin Formula for a straight line y = b 0 + b 1 x + e x y want to solve for b0 = intercept b1 = slope y x  y  x = outcomeprogram

16. The Estimated Regression Function The estimated regression function is:

17. Evaluating the “Fit” R 2 = 0.9691 0.0 100.0 200.0 300.0 400.0 500.0 600.0 2030405060708090100 Advertising (in $000s) Sales (in $000s)

18. The R 2 Statistic The R 2 statistic indicates how well an estimated regression function fits the data. 0<= R 2 <=1 It measures the proportion of the total variation in Y around its mean that is accounted for by the estimated regression equation. To understand this better, consider the following graph...

19. Error Decomposition Y X Y Y = b 0 + b 1 X ^ * Y i (actual value) Y i - Y Y i (estimated value) ^ Y i - Y ^ Y i - YiYi ^

20. Partition of the Total Sum of Squares or, TSS = ESS + RSS

22. Degree of Linear Correlation R 2 = 1 = perfect linear correlation; R 2 = 0 = no correlation High R 2 = good fit only if linear model is appropriate; always check with a scatterplot Correlation does not prove causation; x and y may both be correlated to a third (possibly unidentified) variable A more popular (but less meaningful) measure is the “correlation coefficient”: R 2 = RSQ([y-range],[x-range] r = CORREL([y-range],[x-range])

23. R 2 = 0.67

24. Testing for Significance: F Test n Hypotheses H 0 :  1 = 0 H 0 :  1 = 0 H a :  1 = 0 H a :  1 = 0 n Test Statistic n Rejection Rule Reject H 0 if F > F  where F  is based on an F distribution with 1 d.f. in the numerator and n - 2 d.f. in the denominator.

25. Some Cautions about the Interpretation of Significance Tests Rejecting H 0 : b 1 = 0 and concluding that the relationship between x and y is significant does not enable us to conclude that a cause- and-effect relationship is present between x and y. Just because we are able to reject H 0 : b 1 = 0 and demonstrate statistical significance does not enable us to conclude that there is a linear relationship between x and y.

26. An Example of Inappropriate Interpretation A study shows that, in elementary schools, the ability of spelling is stronger for the students with larger feet.  Could we conclude that the size of foot can influence the ability of spelling?  Or there exists another factor that can influence the foot size and the spelling ability?

27. Making Predictions Estimated Sales= 36.342 + 5.550 * 65 = 397.092 So when $65,000 is spent on advertising, we expect the average sales level to be $397,092. Suppose we want to estimate the average levels of sales expected if $65,000 is spent on advertising.

28. The Standard Error The standard error measures the scatter in the actual data around the estimate regression line. where k = the number of independent variables For our example, S e = 20.421 This is helpful in making predictions...

29. An Approximate Prediction Interval An approximate 95% prediction interval for a new value of Y when X 1 =X 1 h is given by where : Example: If $65,000 is spent on advertising: 95% lower prediction interval = 397.092 - 2*20.421 = 356.250 95% upper prediction interval = 397.092 + 2*20.421 = 437.934 If we spend $65,000 on advertising we are approximately 95% confident actual sales will be between $356,250 and $437,934.

30. An Exact Prediction Interval A (1-  )% prediction interval for a new value of Y when X 1 =X 1 h is given by where :

31. Example If $65,000 is spent on advertising: 95% lower prediction interval = 397.092 - 2.306*21.489 = 347.556 95% upper prediction interval = 397.092 + 2.306*21.489 = 446.666 If we spend $65,000 on advertising we are 95% confident actual sales will be between $347,556 and $446,666. This interval is only about $20,000 wider than the approximate one calculated earlier but was much more difficult to create. The greater accuracy is not always worth the trouble.

32. Comparison of Prediction Interval Techniques 125 175 225 275 325 375 425 475 525 575 2535455565758595 Advertising Expenditures Sales Regression Line Prediction intervals created using standard error S e Prediction intervals created using standard prediction error S p

33. Confidence Intervals for the Mean A (1-  )% confidence interval for the true mean value of Y when X 1 =X 1 h is given by where :

34. A Note About Extrapolation Predictions made using an estimated regression function may have little or no validity for values of the independent variables that are substantially different from those represented in the sample.

35. What Does “Regression” Mean?

36. 1.Draw “best-fit” line free hand 2.Find mother’s height = 60”, find average daughter’s height 3.Repeat for mother’s height = 62”, 64”… 70”; draw “best-fit” line for these points 4.Draw line daughter’s height = mother’s height 5.For a given mother’s height, daughter’s height tends to be between mother’s height and mean height: “regression toward the mean”

37. What Does “Regression” Mean?

38. Residual for Observation i y i – y i Standardized Residual for Observation i where: Residual Analysis ^^^ ^

39.  

40. Residual Analysis Detecting Outliers – An outlier is an observation that is unusual in comparison with the other data. – Minitab classifies an observation as an outlier if its standardized residual value is +2. – This standardized residual rule sometimes fails to identify an unusually large observation as being an outlier. – This rule’s shortcoming can be circumvented by using studentized deleted residuals. – The |i th studentized deleted residual| will be larger than the |i th standardized residual|.

41. Multiple Regression Analysis Most regression problems involve more than one independent variable. If each independent variables varies in a linear manner with Y, the estimated regression function in this case is: The optimal values for the b i can again be found by minimizing the ESS. The resulting function fits a hyperplane to our sample data.

42. Example Regression Surface for Two Independent Variables Y X1X1 X2X2 * * * * * * * * * * * * * * * * * * * * * * *

43. Multiple Regression Example: Real Estate Appraisal A real estate appraiser wants to develop a model to help predict the fair market values of residential properties. Three independent variables will be used to estimate the selling price of a house: –total square footage –number of bedrooms –size of the garage

44. Selecting the Model We want to identify the simplest model that adequately accounts for the systematic variation in the Y variable. Arbitrarily using all the independent variables may result in overfitting. A sample reflects characteristics: –representative of the population –specific to the sample We want to avoid fitting sample specific characteristics -- or overfitting the data.

45. Models with One Independent Variable With simplicity in mind, suppose we fit three simple linear regression functions: VariablesAdjustedParameter in the Model R 2 R 2 S e Estimates X 1 0.8700.85510.299b 0 =9.503, b 1 =56.394 X 2 0.7590.73114.030b 0 =78.290, b 2 =28.382 X 3 0.7930.77012.982b 0 =16.250, b 3 =27.607 Key regression results are: The model using X 1 accounts for 87% of the variation in Y, leaving 13% unaccounted for.

46. Important Software Note When using more than one independent variable, all variables for the X-range must be in one contiguous block of cells (that is, in adjacent columns).

47. Models with Two Independent Variables Now suppose we fit the following models with two independent variables: VariablesAdjustedParameter in the Model R 2 R 2 S e Estimates X 1 0.8700.85510.299b 0 =9.503, b 1 =56.394 X 1 & X 2 0.9390.924 7.471b 0 =27.684, b 1 =38.576 b 2 =12.875 X 1 & X 3 0.8770.84710.609b 0 =8.311, b 1 =44.313 b 3 =6.743 Key regression results are: The model using X 1 and X 2 accounts for 93.9% of the variation in Y, leaving 6.1% unaccounted for.

48. The Adjusted R 2 Statistic As additional independent variables are added to a model: –The R 2 statistic can only increase. –The Adjusted-R 2 statistic can increase or decrease. The R 2 statistic can be artificially inflated by adding any independent variable to the model. We can compare adjusted-R 2 values as a heuristic to tell if adding an additional independent variable really helps.

49. A Comment On Multicollinearity It should not be surprising that adding X 3 (# of bedrooms) to the model with X 1 (total square footage) did not significantly improve the model. Both variables represent the same (or very similar) things -- the size of the house. These variables are highly correlated (or collinear). Multicollinearity should be avoided.

50. Testing for Significance: Multicollinearity The term multicollinearity refers to the correlation among the independent variables. When the independent variables are highly correlated (say, |r | >.7), it is not possible to determine the separate effect of any particular independent variable on the dependent variable. If the estimated regression equation is to be used only for predictive purposes, multicollinearity is usually not a serious problem. Every attempt should be made to avoid including independent variables that are highly correlated.

51. Model with Three Independent Variables Now suppose we fit the following model with three independent variables: VariablesAdjustedParameter in the Model R 2 R 2 S e Estimates X 1 0.8700.85510.299b 0 =9.503, b 1 =56.394 X 1 & X 2 0.9390.924 7.471b 0 =27.684, b 1 =38.576, b 2 =12.875 X 1, X 2 & X 3 0.9430.9187.762b 0 =26.440, b 1 =30.803, b 2 =12.567, b 3 =4.576 Key regression results are: The model using X 1 and X 2 appears to be best: –Highest adjusted-R 2 –Lowest S e (most precise prediction intervals)

52. Making Predictions Let’s estimate the avg selling price of a house with 2,100 square feet and a 2-car garage: The estimated average selling price is $134,444 A 95% prediction interval for the actual selling price is approximately: 95% lower prediction interval = 134.444 - 2*7.471 = $119,502 95% lower prediction interval = 134.444 + 2*7.471 = $149,386

53. Binary Independent Variables Other types of non-quantitative factors could independent variables could be included in the analysis using binary variables. Example: The presence (or absence) of a swimming pool, Example: Whether the roof is in good, average or poor condition,

54. Polynomial Regression Sometimes the relationship between a dependent and independent variable is not linear. This graph suggests a quadratic relationship between square footage (X) and selling price (Y).

55. The Regression Model An appropriate regression function in this case might be, or equivalently, where,

56. Implementing the Model

57. Graph of Estimated Quadratic Regression Function

58. Fitting a Third Order Polynomial Model We could also fit a third order polynomial model, or equivalently, where,

59. Graph of Estimated Third Order Polynomial Regression Function

60. Overfitting When fitting polynomial models, care must be taken to avoid overfitting. The adjusted-R 2 statistic can be used for this purpose here also.

61. Example: Programmer Salary Survey A software firm collected data for a sample of 20 computer programmers. A suggestion was made that regression analysis could be used to determine if salary was related to the years of experience and the score on the firm’s programmer aptitude test. The years of experience, score on the aptitude test, and corresponding annual salary ($1000s) for a sample of 20 programmers is shown on the next slide.

62. Example: Programmer Salary Survey Exper. Score Salary Exper. Score Salary 4782498838 71004327326.6 18623.7107536.2 58234.358131.6 88635.867429 10843888734 07522.247930.1 18023.169433.9 6833037028.2 6913338930

63. Example: Programmer Salary Survey Multiple Regression Model Suppose we believe that salary (y) is related to the years of experience (x 1 ) and the score on the programmer aptitude test (x 2 ) by the following regression model: y =  0 +  1 x 1 +  2 x 2 +   where y = annual salary ($000) x 1 = years of experience x 2 = score on programmer aptitude test

64. Example: Programmer Salary Survey Multiple Regression Equation Using the assumption E (  ) = 0, we obtain E(y ) =  0 +  1 x 1 +  2 x 2 Estimated Regression Equation b 0, b 1, b 2 are the least squares estimates of  0,  1,  2.  Thus y = b 0 + b 1 x 1 + b 2 x 2. ^

65. Example: Programmer Salary Survey Solving for the Estimates of  0,  1,  2 ComputerPackage for Solving MultipleRegressionProblemsComputerPackage MultipleRegressionProblems b 0 = b 1 = b 1 = b 2 = b 2 = R 2 = etc. b 0 = b 1 = b 1 = b 2 = b 2 = R 2 = etc. Input Data Least Squares Output x 1 x 2 y 4 78 24 4 78 24 7 100 43 7 100 43...... 3 89 30 3 89 30 x 1 x 2 y 4 78 24 4 78 24 7 100 43 7 100 43...... 3 89 30 3 89 30

66. Example: Programmer Salary Survey Data Analysis Output The regression is Salary = 3.17 + 1.40 Exper + 0.251 Score Predictor Coef Stdev t-ratio p Constant3.1746.156.52.613 Exper1.4039.19867.07.000 Score.25089.077353.24.005 s = 2.419 R-sq = 83.4% R-sq(adj) = 81.5%

67. Example: Programmer Salary Survey Computer Output (continued) Analysis of Variance SOURCE DF SS MS F P Regression2500.33250.1642.760.000 Error1799.465.85 Total19599.79