Linear Regression Models 264a Marketing Research 1 Answering the questions you have asked. • What is the influence of a policy variable (price, advertising, and etc.) on a market outcome (market shares, sales, overall satisfaction)? 264a Marketing Research 2 Applications of Linear Regression Forecasting Sales Response to Marketing Actions: Linear Model Yt price X price,t t Linear Relation Between Sales and Price 3500 Sales 3000 2500 2000 1500 $15.00 $17.00 $19.00 $21.00 $23.00 $25.00 Price 264a Marketing Research 3 SUMMARY OUTPUT Regression Statistics Multiple R 0.898 R Square 0.806 Adjusted R Square 0.799 Standard Error 144.74 Observations 29 ANOVA df SS 2346991.779 565650.3691 2912642.148 MS 2346992 20950.01 Coefficients Standard Error 5126 240 -123.24 11.64 t Stat 21.34 -10.58 Regression Residual Total Intercept Price 1 27 28 264a Marketing Research F Significance F 112.03 4.14101E-11 P-value 1.9601E-18 4.141E-11 Lower 95% Upper 95% 4633 5619 -147.13 -99.35 4 Log-Linear Model log Yt price X price,t t Sales Log Linear Relation Between Sales and Price 50000 45000 40000 35000 30000 25000 20000 15000 10000 5000 0 $15.00 $16.00 $17.00 $18.00 $19.00 $20.00 Price What if you try to fit a linear model? 264a Marketing Research 5 SUMMARY OUTPUT Regression Statistics Multiple R 0.807358 R Square 0.651827 Adjusted R 0.645719 Square Standard Error 5976.594 Observations 59 ANOVA df Regression Residual Total Intercept Price SS MS F Significance F 1 3.81E+09 3.81E+09 106.7118 1.13E-14 57 2.04E+09 35719681 58 5.85E+09 Coefficie Standard nts Error 118158.3 10644.3 -6249.31 604.959 t Stat P-value Lower Upper 95% 95% 11.10061 7.13E-16 96843.45 139473.2 -10.3301 1.13E-14 -7460.72 -5037.9 Good fit alone is no guarantee that the model is apt. 264a Marketing Research 6 Look at the residuals! Yt price X price,t t Price Residual Plot 25000 20000 Residuals 15000 10000 5000 0 $15.00 -5000 $16.00 $17.00 $18.00 $19.00 $20.00 -10000 Price The residuals should be random noise Homoscedastic, not heteroscedastic No pattern. 264a Marketing Research 7 Try the right model. log Yt price X price,t t Log of Sales and Price 11.00 Log of Sales 10.00 9.00 8.00 7.00 6.00 $15.00 $16.00 $17.00 $18.00 $19.00 $20.00 Price 264a Marketing Research 8 SUMMARY OUTPUT Regression Statistics Multiple R R Square Adjusted R Square Standard Error Observations 0.96 0.92 0.92 0.30 59 ANOVA df Regression Residual Total Intercept Price 1 57 58 SS 59.24 5.15 64.40 Coefficients Standard Error 22.17 0.54 -0.78 0.03 MS 59.24 0.09 t Stat 41.40 -25.59 F Significance F 655.09 6.09178E-33 P-value 3.15E-44 6.09E-33 Lower 95% 21.10 -0.84 Upper 95% 23.24 -0.72 It looks better, but is it apt? 264a Marketing Research 9 Look at the residuals from this model! Price Residual Plot 0.6 Residuals 0.4 0.2 0 $15.00 -0.2 $16.00 $17.00 $18.00 $19.00 $20.00 -0.4 -0.6 Price Homoscedastic. No pattern. 264a Marketing Research 10 If the prices have a time order, check for serial dependency. Serial Dependency Among Residuals? 0.6 0.4 Residual 0.2 0 -0.2 -0.4 -0.6 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 Week No. 264a Marketing Research 11 Most data show a log-log relation between prices and sales. log Yt price log X price,t t Log-Log Relation Between Sales and Prices Sales 1200 1100 1000 900 800 700 600 500 400 300 $15.00 $20.00 $25.00 $30.00 $35.00 $40.00 Prices 264a Marketing Research 12 What happens if you fit a linear model? SUMMARY OUTPUT Regression Statistics Multiple R 0.964864 R Square 0.930962 Adjusted 0.929751 R Square Standard 49.31677 Error Observati 59 ons ANOVA df Regressio n Residual Total Intercept Price 1 SS 1869427 MS F 1869427 768.6334 Significance F 8.93E-35 57 138632.2 2432.143 58 2008059 Coefficient Standard t Stat P-value Lower Upper s Error 95% 95% 1344.148 25.87085 51.95606 1.04E-49 1292.342 1395.953 -25.3197 0.913271 -27.7242 8.93E-35 -27.1485 -23.4909 264a Marketing Research 13 Look at the residuals! Price Residual Plot 150 Residuals 100 50 0 -50 -100 $15.00 $20.00 $25.00 $30.00 $35.00 $40.00 Price 264a Marketing Research 14 Try the right model. Plot of Log Sales and Log Prices 7.20 Log of Sales 7.00 6.80 6.60 6.40 6.20 6.00 2.70 2.90 3.10 3.30 3.50 3.70 Log of Prices 264a Marketing Research 15 SUMMARY OUTPUT Regression Statistics Multiple R 0.994507 R Square 0.989044 Adjusted R 0.988852 Square Standard Error 0.028329 Observations 59 ANOVA df Regression 1 Residual 57 Total 58 Coefficients Intercept Log Price SS 4.129669 0.045744 4.175412 Standard Error 9.55979 0.04366 -0.9527 0.013281 264a Marketing Research MS 4.129669 0.000803 t Stat 218.9615 -71.7346 F 5145.847 P-value 4.46E-85 1.43E-57 Significance F 1.43E-57 Lower 95% 9.472363 -0.9793 Upper 95% 9.647217 -0.92611 16 Log Price Residual Plot 0.06 Residuals 0.04 0.02 0 -0.02 -0.04 -0.06 2.70 2.90 3.10 3.30 3.50 3.70 Log Price 264a Marketing Research 17 Tracking the Components of Customer Satisfaction 16 SPECIFIC CRITERIA The value received for the price. Product reliability Product capability How well the software utilizes the available computer system resources. How easy the product is to use. How trouble free is the installation process. The quality of technical support from the system engineers in the field. The quality of technical support from the Customer Support Center. The quality of the technical documentation. How easy is it to acquire maintenance for the product. How easy is it to apply the required maintenance. An evaluation of the company's sales representative. An evaluation of the company's billing services. An evaluation of the company's contracts services. An evaluation of the company's complaint-resolution process. An evaluation of the company's handling of customers' phone calls. And a measure of overall satisfaction. 264a Marketing Research 18 MEASURING PERFORMANCE How well is the company doing on each aspect? How important is it to the customer that the company performs well on each aspect? 264a Marketing Research 19 WHAT DRIVES OVERALL SATISFACTION? Yi 0 1 X1i 2 X2i ... 16 X16i i where: Yi is the overall satisfaction rating given by customer i, X 1i is the first rating given by customer i, 1 is the statistical importance weight for the first rating, 0 is the intercept term that adjusts the mean rating on the criterion (overall satisfaction) to match the mean ratings on the linear combination of rating scales, i is error. 264a Marketing Research 20 so as to Multiple regression attempts to estimate the weights minimize the error (i.e. the sum of squared errors). 264a Marketing Research 21 Ta b le 2. Rela tio n o f Perfo rm a nc e Sc a les to Overa ll Sa tisfa c tio n De p e nd e nt Va ria b le : O VERALL O ve ra ll sa tisfa c tio n w ith C o m p a ny So urc e DF Sum o f Sq ua re s Mea n Sq ua re Mod el Erro r C To ta l 16 1821 1837 470.78 423.91 894.69 Ro o t M SE R-sq ua re Ad j R-sq F Va lue Pro b >F 29.42 0.23 126.40 0.0001 0.48 0.53 0.52 De p M e a n C .V. 4.04 11.94 Pa ra m eter Estim a tes Va ria b le DF INTERC EP VALUE RELIABLE C APABLE RESO URC E EASE INSTALL TS_FES TS_C SC DO C S M AIN_AC Q M AIN_APP SALESREP BILLING C O NTRAC T C O M PLAIN PHO NEC AL 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 Pa ra m e te r Sta nd a rd Estim a te Erro r 0.29 0.19 0.20 0.16 0.02 0.09 0.00 0.01 0.14 0.05 0.04 -0.05 0.06 -0.04 -0.01 0.02 0.07 264a Marketing Research 0.10 0.02 0.02 0.02 0.02 0.02 0.02 0.02 0.03 0.02 0.02 0.03 0.02 0.04 0.04 0.03 0.02 T fo r H0: Sta nd a rd ize d Pa ra m e te r=0 Pro b > | T| Estim a te To le ra nc e La b e l 2.93 9.66 9.77 7.25 0.96 5.79 -0.07 0.38 5.73 3.16 1.77 -1.76 2.94 -1.07 -0.17 0.64 2.85 0.01 0.01 0.01 0.01 0.34 0.01 0.95 0.71 0.01 0.01 0.08 0.08 0.01 0.29 0.87 0.52 0.01 0.00 0.20 0.20 0.15 0.02 0.12 0.00 0.01 0.15 0.06 0.05 -0.05 0.07 -0.03 -0.01 0.02 0.07 . 0.61 0.59 0.57 0.65 0.65 0.54 0.50 0.39 0.62 0.40 0.36 0.52 0.25 0.24 0.39 0.43 Inte rc e p t Va lue re c e ive d fo r p ric e Pro d uc t re lia b ility Pro d uc t c a p a b ility Use o f syste m re so urc e s Pro d uc t e a se o f use Pro d uc t insta lla tio n Te c h sup p o rt - Fie ld SEs Te c h sup p o rt - C usto m e r Sup p o rt C e nte r Te c hnic a l d o c um e nta tio n Ea se o f a c q uiring m a inte na nc e Ea se o f a p p lying m a inte na nc e C o m p a ny's sa le s re p re se nta tive Billing se rvic e C o ntra c ts se rvic e s C o m p la int re so lutio n Ha nd ling o f yo ur c a lls 22 Linear Regression Models are Powerful • Powerful models can be powerfully abused. • Check to see if the model not only fits, but is also apt. • Think about what validity means for your model. – Forecasting on new data. – Relates to other concepts as expected. 264a Marketing Research 23