Chapter 2. Linear model. Put some concreteness on problem. The Bivariate Regression Model. Sample of n observations, labeled as i=1,2,..

Transcription

1 Linear model Chapter 2 The Bivariate Regression Model Sample of n observations, labeled as i=1,2,..n y i = + x i 1 + i and 1 are population values represent the true relationship between x and y Unfortunately these values are unknown The job of the researcher is to estimate these values 1 2 Put some concreteness on problem Notice that if we differentiate y with respect to x, we obtain y/ x = 1 1 represents how much y will change for a fixed change in x Increase in income for more education Change in crime or bankruptcy when casinos are opened Increase in test score if you study more Suppose a state is experiencing a significant budget shortfall Short-term solution raise tax on cigarettes by 35 cents/pack Problem a tax hike will reduce consumption (theory of demand) Question for state as taxes are raised, how much will cigarette consumption fall 3 4 1

2 Benefits and Costs of Model Suppose y is a state s per capita consumption of cigarettes x represents taxes on cigarettes Question how much will y fall if x is increased by 35 cents/pack? Note there are many reasons why people smoke cost is but one of them Placed more structure on the model, therefore we can obtain precise statements about the relationship between x and y These statements will be true so long as the hypothesized relationship is true As you place more structure on any model, the chance that the assumptions of the model are correct declines. 5 6 Data Data on state consumption/taxes, states x 2 years = 12 observations Y = per capita consumption Per capita packs per year Scatter Plot: Real taxes per pack vs. Per capita cigarette consumption 51 states, , 12 observations X = tax (State + Federal) in real cents per pack 2 dollars Real tax per pack (State + Federal, 2 $) 8 2

3 3 Scatter Plot: Real taxes per pack vs. Per capita cigarette consumption 51 states, , 12 observations What is i? Per capita packs per year Real tax per pack (State + Federal, 2 $) 9 There are many factors that determine a state s level of cigarette consumption Some of these factors we can measure, but for what ever reason, we do not have data Education, age, income, etc. Some of these factors we cannot measure Dislike of cigarettes, anti-smoking sentiment of your friends/neighbors/relatives i identified what we cannot measure in our model 1 Current smoking rates By demographic group Think of a difference way draw a vertical line at any tax level (e.g., 4 cents). Notice that at this level, there are multiple values of Y that are present Therefore on average, higher taxes will reduce consumption, but it cannot explain all of consumption across states Adults 2.8% Gender Males 23.5% Females 18.1% Age group % % % Race White 2.7% AA 21.1% AI/AN 26.9% Asian 1.7% Hispanic origin Hispanic 14.5% Non hispanic 21.9% Education < HS 28.9% HS 26.4% Some col. 22.1% College+ 8.2% Family Income <$2K 3.3% $2-$35K 27.2% $35-$55K 22.1% $55-$75K 18.3% $>75K 13.5%

4 Performance in the Olympics We can however estimate values of i by estimating values of β and 1. Estimates have hats : ˆ and ˆ 1 Our goal, is to choose values for ˆ and ˆ 1in an optimal way. Requires minimizing some function of the estimated errors associated with the model 13 Medal count in the Olympics is a simple measure of output Countries vary by Size Resources How is performance once we control for these attributes? 14 Ranking by Total Medal Count Ranking by Medals/1 million People medals~k name country medals United States of America USA 14 2 People's Republic of China CHI 88 3 Russian Federation RUS 82 4 Great Britain GB 65 5 Germany GER Japan JAP 38 7 Australia AUS 35 8 France FRE 34 9 Republic of Korea SKOR 28 9 Italy ITA m~a_rank name country medals medals~a Grenada GREN Jamaica JAM Trinidad and Tobago T&T New Zealand NZEL Bahamas BAH Slovenia SLOVE Mongolia MONG Hungary HUN Montenegro MONT Denmark DEN

5 5 ln(medals per 1 Million) vs ln(gdp per capita) 5 ln(medals per 1 Million) vs ln(gdp per capita) Greneda 4 4 Jamaica NZ 3 3 Australia ln(medals per 1 million) ln(medals per 1 million) China GB USA India ln(gdp per capita) ln(gdp per capita) 18 5 ln(medals per 1 Million) vs ln(gdp per capita) Greneda ln(medals per 1 million) India Jamaica NZ Australia GB USA China ln(gdp per capita) 19 Given linear model yi xi 1 i We can predict an level of consumption given parameter values y ˆ ˆ ˆi x i 1 The predicted value will not always be accurate sometimes we will over or under predict the true value Because of the linear relationship between x and y, predictions will lie along a line 2 5

6 y Difference between the actual & predicted value y 1 y ˆ x ˆ ˆi i 1 y ˆ ˆ ˆ ˆ i yi yi xi 1 i if y ˆ ˆ i yi i you underpredict (you did better than expected) ŷ 2 ŷ 1 y 2 ˆ i ˆ i If y ˆ ˆ i yi i you overpredict (you did worse than expected) x 1 x 2 x ln(medals per 1 million) ln(medals per 1 Million) vs ln(gdp per capita) Greneda Jamaica NZ Australia GB USA China Estimation Estimated errors measure what we don t know Want to minimize these errors as much as possible There are N errors in each model Need to select a criteria to somehow minimize all these errors -3 India ln(gdp per capita)

7 Criteria: Least squares Y let ˆ and ˆ be candidate values for 1 the parameters. The estimated error is then y ˆ x ˆ ˆi i i 1 f(x) Objective :minthe sum of squared errors n n 2 ˆ ˆ 2 ˆ i ( i i 1) i 1 i 1 SSE y x 25 X 2 X 1 X 26 Cigarette example Data available on web page state_cig_data.dta Already in a Stata data file To use, Download to a folder Change directory to the folder type use state_cig_data. * dscribe data. desc Contains data from state_cig_data.dta obs: 1,2 vars: 7 28 Aug 28 15:39 size: 29,58 (99.8% of memory free) - storage display value variable name type format label variable label - state str2 %9s 2-digit state code year int %8.g year state_tax float %9.g state tax in cents per pack retail_price float %9.g average retail price, nominal federal_tax byte %8.g federal tax in cents per pack packs_pc float %9.g packs of cigarettes per capita cpi float %9.g consumer price index, 2=

8 . * generate real taxes. gen real_tax=(state_tax+federal_tax)/cpi.. * get means of real tax and per capita consumption. sum packs_pc real_tax Generate real tax By dividing by CPI Variable Obs Mean Std. Dev. Min Max packs_pc real_tax ˆ ˆ ˆ xy y.6115( ) ˆ x. * get correlation coefficient. corr packs_pc real_tax (obs=12) packs_pc real_tax packs_pc 1. real_tax Means Std deviations ˆ y x ˆ 16.6 (56.5)(.937) Correlation coefficient 29 3 Run regression in stata SSM, SSE, SST R 2. * get regression estimate. reg packs_pc real_tax Source SS df MS Number of obs = F( 1, 118) = Model Prob > F =. Residual R-squared = Adj R-squared =.3733 Total Root MSE = Notice that SSE + SSM = SST R 2 = SSM/SST = 351.4/ =.3739 packs_pc Coef. Std. Err. t P> t [95% Conf. Interval] real_tax _cons β and β

9 Scatter Plot: Real taxes per pack vs. Per capita cigarette consumption 51 states, , 12 observations Scatter Plot: Real taxes per pack vs. Per capita cigarette consumption 51 states, , 12 observations Per capita packs per year Per capita packs per year Real tax per pack (State + Federal, 2 $) Real tax per pack (State + Federal, 2 $) Using the results. * output residuals. predict error_packs_pc, residuals Output residuals ˆ For every penny increase in taxes, per capita consumption falls by.94 packs per year A 35 cent increase in taxes will reduce consumption by (35)(.94) = 32 packs per person per year.. * show that the means of the errors are zero. sum error_packs_pc Variable Obs Mean Std. Dev. Min Max error_pack~c e * get correlation between x and error. corr error_packs_pc real_tax (obs=12) error_~c real_tax error_pack~c 1. real_tax Predicted error Has zero mean Predicted error not Correlated with x

10 Example 2 Do better performing teams have higher attendance? Data on wins and average attendance/game for 24 baseball season 3 observations attendance.dta - storage display value variable name type format label variable label - team str13 %13s team city attendance long %12.g avg attendance per game wins int %8.g wins during year -. * get means of wins and payroll. sum wins attendance Variable Obs Mean Std. Dev. Min Max wins attendance * get detailed data for attendance. sum attendance, detail avg attendance per game Percentiles Smallest 1% % % Obs 3 25% Sum of Wgt. 3 5% Mean Largest Std. Dev % % Variance 8.68e+7 95% Skewness % Kurtosis * run simple regression. reg attendance wins Source SS df MS Number of obs = F( 1, 28) = 8.89 Model Prob > F =.59 Residual 1.911e R-squared = Adj R-squared =.2139 Total e Root MSE = attendance Coef. Std. Err. t P> t [95% Conf. Interval] wins _cons ˆ for every addition win, attendance increase by 31 / game. An addition1 wins put 31.5 more people in seats 4 1

11 . * output predicted attendance. predict pred_att, xb.. list team wins attendance pred_att team wins attend~e pred_att Seattle New York-AL Baltimore Boston Cleveland Texas Anaheim Oakland Minnesota Chicago--AL Construct a new variable, The predicted value of Y Predicted attendance Plot: Actual and Predicted Attendance 45 degree line Consider Seattle Y p = (31.5)(93) = Actual attendance 42 Example 3: Education and Earnings Stylized fact: log wages or earnings is linear in education (above a certain range) Interpreted as a return to education Theoretical models why this would be the case Linear model: y=ln(weekly wages) endogenous variable x=years of education exogenous factor y i = + x i 1 + i 43 Notice that 1 has a different interpretation 1 =dy/dx In this case, y=ln(wages) dln(wages)/dx = (1/wages)dWages/dX dwages/wages = % change in changes (change in wages over base wages) when the endogenous variable is a natural log, 1 =dy/dx is interpreted as % change in y for a unit change in x 44 11

12 Data cps87.dta 19,96 observations from 1987 Current Population Survey on Full time (>3 hours) Males Aged % of sample 8% 7% 6% 5% 4% 3% 2% Weekly Earnings, Adult Males, 2 Census 1% 45 % Weekly Earnings 46 6% ln(wekly Earnings), Adult Males, 2 Census Average ln(weekly Wages) vs Educ % of Sample 5% 4% 3% 2% 1% Average ln(weekly Wages) Years of Education % ln(weekly Earnings)

13 ln(weekly Wages) ln(weekly Wages) vs. Education Years of Education. * construct ln weekly earnings. gen ln_weekly_earn=ln(weekly_earn).. * get descriptive information. sum weekly_earn ln_weekly_earn years_educ Generate ln(weekly earnings) Variable Obs Mean Std. Dev. Min Max weekly_earn ln_weekly_~n years_educ Means of the variables * run simple regression. reg ln_weekly_earn years_educ Source SS df MS Number of obs = F( 1, 1994) = Model Prob > F =. Residual R-squared = Adj R-squared =.163 Total Root MSE = ln_weekly_~n Coef. Std. Err. t P> t [95% Conf. Interval] years_educ _cons Example 4:London Olympics * generate measure of medals per person; * divide by 1,, people; gen medals_capita=medals/(population/1); label var medals_capita "medals per 1,, people"; * take natural ln of medals_capita and gdp_capita; gen ln_medals_capita=ln(medals_capita); gen ln_gdp_capita=ln(gdp_capita); * regress ln(medals/population) on ln(gdp/population); reg ln_medals_capita ln_gdp_capita;

14 Example 4: London Olympics. * regress ln(medals/population) on ln(gdp/population);. reg ln_medals_capita ln_gdp_capita; Source SS df MS Number of obs = F( 1, 83) = Model Prob > =.1 F Residual R-squared = Adj R-squared =.1677 Total Root MSE = ln_medals_~a Coef. Std. Err. t P> t [95% Conf. Interval] ln_gdp_cap~a _cons ln medals per 1 million people ln(medals/population) vs ln(gdp/population) GREN JAM T&TBAH NZEL MONG MONT HUN SLOVE GEO LITH CROA EST DEN AUS CUB BEL AZE CYPGBIRE NETH QUA ARM LATV CZE KAZ SLOVA BAH SWE TAI NOR MOLD SERB GABRUS SKOR PR FRECAN GER FIN UKR BOTS ROM ITA SWI ESP USA SING TUN JAP NKOR KEN BULG POL BEL KUW DREP IRAN COL GRE TAJ UZE SAFR HK ARG ETH BRA POR GUAT CHI MALA MEX TUR AFGH THAI MORO VEN SAUD UGA EGY ALG IND ln of medals per 1,, Fitted values INDO ln of gdp per capita ln medals per 1 million people ETH ln(medals/population) vs ln(gdp/population) NKOR IND EGY INDO JAM CUB CHI GREN MEX VEN SKOR SAUD AUS GB QUA NOR CAN USA ln of gdp per capita ln of medals per 1,, Fitted values 55 14