Linear Regression

(ver. 6.0)

Oscar Torres-Reyna
Data Consultant
[email protected]

http://dss.princeton.edu/training/

PU/DSS/OTR

Regression: a practical approach (overview) We use regression to estimate the unknown effect of changing one variable over another (Stock and Watson, 2003, ch. 4) When running a regression we are making two assumptions, 1) there is a linear relationship between two variables (i.e. X and Y) and 2) this relationship is additive (i.e. Y= x1 + x2 + +xN). Technically, linear regression estimates how much Y changes when X changes one unit. In Stata use the command regress, type: regress [dependent variable] [independent variable(s)] regress y x In a multivariate setting we type: regress y x1 x2 x3 Before running a regression it is recommended to have a clear idea of what you are trying to estimate (i.e. which are your outcome and predictor variables). A regression makes sense only if there is a sound theory behind it.

2
PU/DSS/OTR

Regression: a practical approach (setting)

Example: Are SAT scores higher in states that spend more money on education controlling by other factors?*

Outcome (Y) variable SAT scores, variable csat in dataset Predictor (X) variables Per pupil expenditures primary & secondary (expense) % HS graduates taking SAT (percent) Median household income (income) % adults with HS diploma (high) % adults with college degree (college) Region (region)

*Source: Data and examples come from the book Statistics with Stata (updated for version 9) by Lawrence C. Hamilton (chapter 6). Click here to download the data or search for it at http://www.duxbury.com/highered/. 3 Use the file states.dta (educational data for the U.S.).

PU/DSS/OTR

Regression: variables
It is recommended first to examine the variables in the model to check for possible errors, type: use http://dss.princeton.edu/training/states.dta describe csat expense percent income high college region summarize csat expense percent income high college region
. describe csat expense percent income high college region display format %9.0g %9.0g %9.0g %10.0g %9.0g %9.0g %9.0g value label variable label Mean composite SAT score Per pup il expenditures prim&sec % HS graduates taking SAT Median household income, $1,000 % adults HS diploma % adults college degree Geographical region

storage variable name type csat expense percent income high college region int int byte double float float byte

region

. summarize Variable csat expense percent income high college region

csat expense percent income high college region Obs 51 51 51 51 51 51 50 Mean 944.098 5235.961 35.76471 33.95657 76.26078 20.02157 2.54 Std. Dev. 66.93497 1401.155 26.19281 6.423134 5.588741 4.16578 1.128662 Min 832 2960 4 23.465 64.3 12.3 1 Max 1093 9259 81 48.618 86.6 33.3 4

PU/DSS/OTR

Regression: what to look for

Lets run the regression: regress csat expense, robust
Outcome variable (Y) Predictor variable (X)
Robust standard errors (to control for heteroskedasticity)
1

. regress

csat expense, robust Number of obs = F( 1, 49) = Prob > F = R-squared = Root MSE = 51 36.80 0.0000 0.2174 59.814
2

This is the p-value of the model. It tests whether R2 is different from 0. Usually we need a p-value lower than 0.05 to show a statistically significant relationship between X and Y. R-square shows the amount of variance of Y explained by X. In this case expense explains 22% of the variance in SAT scores. Adj R2 (not shown here) shows the same as R2 but adjusted by the # of cases and # of variables. When the # of variables is small and the # of cases is very large then Adj R2 is closer to R2. This provides a more honest association between X and Y.

Linear regression

Root MSE: root mean squared error, is the sd of the regression. The closer to zero better the fit.
Robust Std. Err. .0036719 24.35468
6

csat expense _cons

Coef. -.0222756 1060.732

t -6.07 43.55

P>|t| 0.000 0.000

[95% Conf. Interval] -.0296547 1011.79 -.0148966 1109.675

3

csat = 1061 - 0.022*expense For each one-point increase in expense, SAT scores decrease by 0.022 points.
5

The t-values test the hypothesis that the coefficient is different from 0. To reject this, you need a t-value greater than 1.96 (for 95% confidence). You can get the t-values by dividing the coefficient by its standard error. The tvalues also show the importance of a variable in the model.

Two-tail p-values test the hypothesis that each coefficient is different from 0. To reject this, the p-value has to be lower than 0.05 (you could choose also an alpha of 0.10). In this case, expense is statistically significant in explaining SAT. 5
PU/DSS/OTR

Regression: what to look for

Robust standard errors (to control Adding the rest of predictor variables: for heteroskedasticity) regress csat expense percent income high college, robust

Output variable (Y)

Predictor variables (X)

. regress csat expense percent income high college, robust Linear regression Number of obs = F( 5, 45) = Prob > F = R-squared = Root MSE = 51 50.90 0.0000 0.8243 29.571

This is the p-value of the model. It indicates the reliability of X to predict Y. Usually we need a pvalue lower than 0.05 to show a statistically significant relationship between X and Y.
2

Root MSE: root mean squared error, is the sd of the regression. The closer to zero better the fit.
csat expense percent income high college _cons Coef. .0033528 -2.618177 .1055853 1.630841 2.030894 851.5649 Robust Std. Err. .004781 .2288594 1.207246 .943318 2.113792 57.28743 t 0.70 -11.44 0.09 1.73 0.96 14.86 P>|t| 0.487 0.000 0.931 0.091 0.342 0.000

[95% Conf. Interval] -.0062766 -3.079123 -2.325933 -.2690989 -2.226502 736.1821 .0129823 -2.15723 2.537104 3.530781 6.28829 966.9477
3

R-square shows the amount of variance of Y explained by X. In this case the model explains 82.43% of the variance in SAT scores. Adj R2 (not shown here) shows the same as R2 but adjusted by the # of cases and # of variables. When the # of variables is small and the # of cases is very large then Adj R2 is closer to R2. This provides a more honest association between X and Y.

csat = 851.56 + 0.003*expense 2.62*percent + 0.11*income + 1.63*high + 2.03*college

4 5

The t-values test the hypothesis that the coefficient is different from 0. To reject this, you need a t-value greater than 1.96 (at 0.05 confidence). You can get the t-values by dividing the coefficient by its standard error. The tvalues also show the importance of a variable in the model. In this case, percent is the most important.

Two-tail p-values test the hypothesis that each coefficient is different from 0. To reject this, the p-value has to be lower than 0.05 (you could choose also an alpha of 0.10). In this case, expense, income, and college are not statistically significant in explaining SAT; high is almost significant at 0.10. Percent is the only variable that has some significant impact on SAT (its coefficient is different 6 from 0)

PU/DSS/OTR

Regression: with dummies

Region is entered here as dummy variable. The easy way to add dummy variables to a regression is using xi and the prefix i. (interpretation is the same as before). The first category is always the reference:
xi: regress csat expense percent income high college i.region, robust
. xi: regress i.region csat expense percent income high college i.region, robust _Iregion_1-4 (naturally coded; _Iregion_1 omitted) Number of obs = F( 8, 41) = Prob > F = R-squared = Root MSE = Robust Std. Err. .0035883 .2358047 1.196409 1.02694 1.599798 17.99933 12.52558 9.44989 67.86418 50 69.82 0.0000 0.9111 21.492

Linear regression

csat expense percent income high college _Iregion_2 _Iregion_3 _Iregion_4 _cons

Coef. -.002021 -3.007647 -.1674421 1.814731 4.670564 69.45333 25.39701 34.57704 808.0206

t -0.56 -12.75 -0.14 1.77 2.92 3.86 2.03 3.66 11.91

P>|t| 0.576 0.000 0.889 0.085 0.006 0.000 0.049 0.001 0.000

[95% Conf. Interval] -.0092676 -3.483864 -2.583638 -.2592168 1.439705 33.10295 .101086 15.4926 670.9661 .0052256 -2.53143 2.248754 3.888679 7.901422 105.8037 50.69293 53.66149 945.0751

Regions: 1 West 2 N. East 3 South 4 Midwest

NOTE: By default xi excludes the first value, to select a different value, before running the regression type: char region[omit] 4 xi: regress csat expense percent income high college i.region, robust This will select Midwest (4) as the reference category for the dummy variables. NOTE: Another way to create dummy variables is to type: tab region, gen(region) This will create four new variables (or a many a categories in the variable), one for each region in this case.

7
PU/DSS/OTR

Regression: ANOVA table

If you run the regression without the robust option you get the ANOVA table
xi: regress csat expense percent income high college i.region
Source Model Residual Total SS df 9 40 49 MS 22252.2045 (D) 317.28849 (E) 4346.15061 (F) Number of obs F( 9, 40) Prob > F R-squared Adj R-squared Root MSE = = = = = = 50 70.13 0.0000 0.9404 0.9270 17.813

(A) 200269.84 (B) 12691.5396 (C) 212961.38

MSS 200269.84 22252.2045 D (k 1) 9 F= = = = = 70.13 12691.5396 317.28849 E RSS 40 nk MSS ei = 200269.84 = A = 0.9404 = 1 R = 2 TSS ( yi y ) 12691.5396 C
2 2

AdjR 2 = 1

317.28849 49 E n 1 (1 R 2 ) = 1 (1 0.9404) = 1 = 1 = 0.9270 4346.15061 40 F nk

RootMSE =

12691.5396 RSS B = = = 17.813 (n k ) 40 40

A = Model Sum of Squares (MSS). The closer to TSS the better fit. B = Residual Sum of Squares (RSS) C = Total Sum of Squares (TSS) D = Average Model Sum of Squares = MSS/(k-1) where k = # predictors E = Average Residual Sum of Squares = RSS/(n k) where n = # of observations F = Average Total Sum of Squares = TSS/(n 1) R2 shows the amount of observed variance explained by the model, in this case 94%. The F-statistic, F(9,40), tests whether R2 is different from zero. Root MSE shows the average distance of the estimator from the mean, in this case 18 points in estimating SAT scores. Source: Kohler, Ulrich, Frauke Kreuter, Data Analysis Using Stata, 2009 8
PU/DSS/OTR

Regression: estto/esttab
To show the models side-by-side you can use the commands estto and esttab: regress csat expense, robust eststo model1 regress csat expense percent income high college, robust eststo model2 xi: regress csat expense percent income high college i.region, robust eststo model3 . esttab, r2 ar2 se scalar(rmse) esttab, r2 ar2 se scalar(rmse)
(1) csat expense percent income high college _Iregion_2 _Iregion_3 _Iregion_4 percent2 _cons N R-sq adj. R-sq rmse 1060.7*** (24.35) 51 0.217 0.201 59.81 851.6*** (57.29) 51 0.824 0.805 29.57 808.0*** (67.86) 50 0.911 0.894 21.49 -0.0223*** (0.00367) (2) csat 0.00335 (0.00478) -2.618*** (0.229) 0.106 (1.207) 1.631 (0.943) 2.031 (2.114) (3) csat -0.00202 (0.00359) -3.008*** (0.236) -0.167 (1.196) 1.815 (1.027) 4.671** (1.600) 69.45*** (18.00) 25.40* (12.53) 34.58*** (9.450)

Type help eststo and help esttab for more options.

Standard errors in parentheses * p<0.05, ** p<0.01, *** p<0.001

9
PU/DSS/OTR

Regression: correlation matrix

Below is a correlation matrix for all variables in the model. Numbers are Pearson correlation coefficients, go from -1 to 1. Closer to 1 means strong correlation. A negative value indicates an inverse relationship (roughly, when one goes up the other goes down).

. pwcorr csat expense percent income high college, star(0.05) sig csat csat expense percent income high college 1.0000 -0.4663* 0.0006 -0.8758* 0.0000 -0.4713* 0.0005 0.0858 0.5495 -0.3729* 0.0070 1.0000 0.6509* 0.0000 0.6784* 0.0000 0.3133* 0.0252 0.6400* 0.0000 1.0000 0.6733* 0.0000 0.1413 0.3226 0.6091* 0.0000 1.0000 0.5099* 0.0001 0.7234* 0.0000 1.0000 0.5319* 0.0001 1.0000 expense percent income high college

10
PU/DSS/OTR

Regression: graph matrix

Command graph matrix produces a graphical representation of the correlation matrix by presenting a series of scatterplots for all variables. Type: graph matrix csat expense percent income high college, half maxis(ylabel(none) xlabel(none))

Y
Mean composite SAT score

Per pupil expenditures prim&sec

% HS graduates taking SAT Median household income, $1,000 % adults HS diploma

% adults college degree

11
PU/DSS/OTR

Regression: exploring relationships

scatter csat percent
1100

scatter csat high

1100

Mean composite SAT score 900 1000

800

20

40 % HS graduates taking SAT

60

80

800

Mean composite SAT score 900 1000

65

70

75 % adults HS diploma

80

85

There seem to be a curvilinear relationship between csat and percent, and slightly linear between csat and high. To deal with U-shaped curves we need to add a square version of the variable, in this case percent square generate percent2 = percent^2

12
PU/DSS/OTR

Regression: functional form/linearity

The command acprplot (augmented component-plus-residual plot) provides another graphical way to examine the relationship between variables. It does provide a good testing for linearity. Run this command after running a regression regress csat percent high /* Notice we do not include percent2 */ acprplot percent, lowess acprplot high, lowess acprplot percent, lowess
0 Augmented component plus residual -200 -150 -100 -50

acprplot high, lowess

Augmented component plus residual 500 550 450
65

-250

600

20

40 % HS graduates taking SAT

60

80

70

75 % adults HS diploma

80

85

The option lowess (locally weighted scatterplot smoothing) draw the observed pattern in the data to help identify nonlinearities. Percent shows a quadratic relation, it makes sense to add a square version of it. High shows a polynomial pattern as well but goes around the regression line (except on the right). We could keep it as is for now. The model is: xi: regress csat expense percent percent2 income high college i.region, robust
Form more details see http://www.ats.ucla.edu/stat/stata/webbooks/reg/chapter2/statareg2.htm, and/or type help acprplot and help lowess.

13
PU/DSS/OTR

Regression: models
xi: regress csat expense percent percent2 income high college i.region, robust eststo model4 esttab, r2 ar2 se scalar(rmse)
. esttab, r2 ar2 se scalar(rmse) (1) csat expense percent income high college _Iregion_2 _Iregion_3 _Iregion_4 percent2 _cons N R-sq adj. R-sq rmse 1060.7*** (24.35) 51 0.217 0.201 59.81 851.6*** (57.29) 51 0.824 0.805 29.57 808.0*** (67.86) 50 0.911 0.894 21.49 -0.0223*** (0.00367) (2) csat 0.00335 (0.00478) -2.618*** (0.229) 0.106 (1.207) 1.631 (0.943) 2.031 (2.114 ) (3) csat -0.00202 (0.00359) -3.008*** (0.236) -0.167 (1.196) 1.815 (1.027) 4.671** (1.600) 69.45*** (18.00) 25.40* (12.53) 34.58*** (9.450) (4) csat 0.00141 (0.00372) -5.945*** (0.641) -0.914 (0.973) 1.869 (0.931) 3.418** (1.145) 5.077 (20.75) 5.209 (10.42) 19.25* (8.110) 0.0460*** (0.0102) 874.0*** (58.13) 50 0.940 0.927 17.81

Standard errors in parentheses * p<0.05, ** p<0.01, *** p<0.001

14
PU/DSS/OTR

Regression: getting predicted values

How good the model is will depend on how well it predicts Y, the linearity of the model and the behavior of the residuals. There are two ways to generate the predicted values of Y (usually called Yhat) given the model: Option A, using generate after running the regression:
xi: regress csat expense percent percent2 income high college i.region, robust generate csat_predict = _b[_cons] + _b[percent]*percent + _b[percent]*percent + _b[percent2]*percent2 + _b[high]*high +

Option B, using predict immediately after running the regression: xi: regress csat expense percent percent2 income high college i.region, robust predict csat_predict label variable csat_predict "csat predicted"

. predict csat_predict (option xb assumed; fitted values) (1 missing value generated) . label variable csat_predict "csat predicted"

15
PU/DSS/OTR

Regression: observed vs. predicted values

For a quick assessment of the model run a scatter plot scatter csat csat_predict
1100 800 Mean composite SAT score 900 1000

850

900

950 csat predicted

1000

1050

We should expect a 45 degree pattern in the data. Y-axis is the observed data and x-axis the predicted data (Yhat). In this case the model seems to be doing a good job in predicting csat
16
PU/DSS/OTR

Regression: testing for homoskedasticity

An important assumption is that the variance in the residuals has to be homoskedastic or constant. Residuals cannot varied for lower of higher values of X (i.e. fitted values of Y since Y=Xb). A definition:
The error term [e] is homoskedastic if the variance of the conditional distribution of [ei] given Xi [var(ei|Xi)], is constant for i=1n, and in particular does not depend on x; otherwise, the error term is heteroskedastic (Stock and Watson, 2003, p.126)

When plotting residuals vs. predicted values (Yhat) we should not observe any pattern at all. In Stata we do this using rvfplot right after running the regression, it will automatically draw a scatterplot between residuals and predicted values.
rvfplot, yline(0)
40 Residuals 0 20

Residuals seem to slightly expand at higher levels of Yhat.

-40

-20

850

900

950 Fitted values

1000

1050

17
PU/DSS/OTR

Regression: testing for homoskedasticity

A non-graphical way to detect heteroskedasticiy is the Breusch-Pagan test. The null hypothesis is that residuals are homoskedastic. In the example below we fail to reject the null at 95% and concluded that residuals are homogeneous. However at 90% we reject the null and conclude that residuals are not homogeneous.
estat hettest
. estat hettest Breusch-Pagan / Cook-Weisberg test for heteroskedasticity Ho: Constant variance Variables: fitted values of csat chi2(1) Prob > chi2 = = 2.72 0.0993

The graphical and the Breush-Pagan test suggest the possible presence of heteroskedasticity in our model. The problem with this is that we may have the wrong estimates of the standard errors for the coefficients and therefore their t-values. There are two ways to deal with this problem, one is using heteroskedasticity-robust standard errors, the other one is using weighted least squares (see Stock and Watson, 2003, chapter 15). WLS requires knowledge of the conditional variance on which the weights are based, if this is known (rarely the case) then use WLS. In practice it is recommended to use heteroskedasticity-robust standard errors to deal with heteroskedasticity. By default Stata assumes homoskedastic standard errors, so we need to adjust our model to account for heteroskedasticity. To do this we use the option robust in the regress command. xi: regress csat expense percent percent2 income high college i.region, robust Following Stock and Watson, as a rule-of-thumb, you should always assume heteroskedasticiy in your model 18 (see Stock and Watson, 2003, chapter 4) .
PU/DSS/OTR

Regression: omitted-variable test

How do we know we have included all variables we need to explain Y? Testing for omitted variable bias is important for our model since it is related to the assumption that the error term and the independent variables in the model are not correlated (E(e|X) = 0) If we are missing variables in our model and is correlated with the included regressor and, the omitted variable is a determinant of the dependent variable (Stock and Watson, 2003, p.144), then our regression coefficients are inconsistent. In Stata we test for omitted-variable bias using the ovtest command: xi: regress csat expense percent percent2 income high college i.region, robust ovtest
. ovtest Ramsey RESET test using powers of the fitted values of csat Ho: model has no omitted variables F(3, 37) = 1.25 Prob > F = 0.3068

The null hypothesis is that the model does not have omitted-variables bias, the p-value is higher than the usual threshold of 0.05 (95% significance), so we fail to reject the null and conclude that we do not need more variables.
19
PU/DSS/OTR

Regression: specification error

Another command to test model specification is linktest. It basically checks whether we need more variables in our model by running a new regression with the observed Y (csat) against Yhat (csat_predicted or X) and Yhat-squared as independent variables1. The thing to look for here is the significance of _hatsq. The null hypothesis is that there is no specification error. If the p-value of _hatsq is not significant then we fail to reject the null and conclude that our model is correctly specified. Type:
xi: regress linktest csat expense percent percent2 income high college i.region, robust

. linktest Source Model Residual Total csat _hat _hatsq _cons SS 200272.359 12689.0209 212961.38 Coef. 1.144949 -.0000761 -68.69417 df 2 47 49 MS 100136.18 269.979169 4346.15061 t 0.76 -0.10 -0.10 P>|t| 0.450 0.923 0.924 Number of obs F( 2, 47) Prob > F R-squared Adj R-squared Root MSE = = = = = = 50 370.90 0.0000 0.9404 0.9379 16.431

Std. Err. 1.50184 .0007885 712.388

[95% Conf. Interval] -1.876362 -.0016623 -1501.834 4.166261 .0015101 1364.446

For more details see http://www.ats.ucla.edu/stat/stata/webbooks/reg/chapter2/statareg2.htm, and/or type help linktest.

20
PU/DSS/OTR

Regression: multicollinearity
An important assumption for the multiple regression model is that independent variables are not perfectly multicolinear. One regressor should not be a linear function of another. When multicollinearity is present standand errors may be inflated. Stata will drop one of the variables to avoid a division by zero in the OLS procedure (see Stock and Watson, 2003, chapter 5). The Stata command to check for multicollinearity is vif (variance inflation factor). Right after running the regression type:
. vif Variable percent2 percent _Iregion_2 income _Iregion_3 high college expense _Iregion_4 Mean VIF VIF 70.80 49.52 8.47 4.97 4.89 4.71 4.52 3.33 2.14 17.04 1/VIF 0.014124 0.020193 0.118063 0.201326 0.204445 0.212134 0.221348 0.300111 0.467506

A vif > 10 or a 1/vif < 0.10 indicates trouble. We know that percent and percent2 are related since one is the square of the other. They are ok since percent has a quadratic relationship with Y, but this would be an example of multicolinearity. The rest of the variables look ok.
Form more details see http://www.ats.ucla.edu/stat/stata/webbooks/reg/chapter2/statareg2.htm, and/or type help vif.

21
PU/DSS/OTR

Regression: outliers
To check for outliers we use the avplots command (added-variable plots). Outliers are data points with extreme values that could have a negative effect on our estimators. After running the regression type:

avplot percent
100

avplot expense
40 -40 -20 e( csat | X ) 0 20

-100

-50

e( csat | X ) 0

50

-10

-5

0 e( percent | X )

10

-2000

-1000

0 e( expense | X )

1000

2000

coef = -5.9452674, (robust) se = .64055286, t = -9.28

coef = .00141156, (robust) se = .00371641, t = .38

These plots regress each variable against all others, notice the coefficients on each. All data points seem to be in range, no outliers observed.
For more details and tests on this and influential and leverage variables please check http://www.ats.ucla.edu/stat/stata/webbooks/reg/chapter2/statareg2.htm Also type help diagplots in the Stata command window.

22
PU/DSS/OTR

Regression: outliers
avplots
e( csat X) -100-50 0 |50 100 e( csat |X ) 60 -40-20 0 20 40 e( csat X )40 -40-20 0 |20

-2000 -1000 0 1000 e( expense | X )

2000

-10

-5 0 5 e( percent | X )

10

-500

0 e( percent2 | X )

500

coef = .00141156, (robust) se = .00371641, t = .38

coef = -5.9452674, (robust) se = .64055286, t = -9.28

coef = .0460468, (robust) se = .01019105, t = 4.5

e( csat X )40 -40-20 0 |20

-10

-5 0 5 e( income | X )

10

-6

-4

-2 0 2 e( high | X )

e( csat X )40 -40-20 0 |20

e( csat | X40 ) -40-20 0 20

-5

0 e( college | X )

coef = -.9143708, (robust) se = .97326373, t = -.94

coef = 1.8691679, (robust) se = .93111302, t = 2.01

coef = 3.4175732, (robust) se = 1.1450333, t = 2.

e( csat X )40 -40-20 0 |20

e( csat | X ) -40-20 0 20 40

-.4

-.2 0 .2 e( _Iregion_2 | X )

.4

-.5

0 e( _Iregion_3 | X )

.5

e( csat | X 40 ) -40-20 0 20

-.5

0 e( _Iregion_4 | X )

.5

coef = 5.0765963, (robust) se = 20.753948, t = .24

coef = 5.2088169, (robust) se = 10.422781, t = .5

coef = 19.245404, (robust) se = 8.1097615, t = 2.

23
PU/DSS/OTR

Regression: summary of influence indicators

In SPSS: Analyze-RegressionLinear; click Save. Select under Influence Statistics to add as a reg y x1 x2 x3 new variable (DFB1_1) or in syntax type dfbeta x1 Measures the influence of REGRESSION A case is an influential outlier each observation on the /MISSING LISTWISE if Note: you could also type: coefficient of a particular /STATISTICS COEFF OUTS independent variable (for R ANOVA |DfBeta|> 2/SQRT(N) predict DFx1, dfbeta(x1) example, x1). This is in /CRITERIA=PIN(.05) POUT(.10) standard errors terms. Where N is the sample size. To estimate the dfbetas for all predictors just /NOORIGIN type: /DEPENDENT Y An observation is influential if Note: Stata estimates /METHOD=ENTER X1 X2 X3 it has a significant effect on dfbeta standardized DfBetas. /CASEWISE PLOT(ZRESID) the coefficient. OUTLIERS(3) DEFAULTS To flag the cutoff DFBETA /SAVE MAHAL COOK LEVER gen cutoffdfbeta = abs(DFx1) > DFBETA SDBETA DFFIT 2/sqrt(e(N)) & e(sample) SDFIT COVRATIO . In Stata type:

DfBeta

Indicator of leverage and high residuals.

High influence if

After running the regression type: predict dfits if e(sample), dfits To generate the flag for the cutoff type: gen cutoffdfit= abs(dfits)>2*sqrt((e(df_m) +1)/e(N)) & e(sample) Same as DfBeta above (DFF_1)

DfFit

Measures how much an |DfFIT| >2*SQRT(k/N) observation influences the regression model as a whole. Where k is the number of parameters (including the How much the predicted intercept) and N is the values change as a result of sample size. including and excluding a particular observation. High impact if |COVRATIO-1| 3*k/N Where k is the number of parameters (including the intercept) and N is the sample size.

Covariance ratio

Measures the impact of an observation on the standard errors

In Stata after running the regression type predict covratio if e(sample), covratio Same as DfBeta above (COV_1)

24

PU/DSS/OTR

Regression: summary of distance measures

Measures how much an observation influences the overall model or predicted values.
In SPSS: Analyze-Regression-Linear; click Save. Select under Distances to add as a new variable (COO_1) or in syntax type REGRESSION /MISSING LISTWISE /STATISTICS COEFF OUTS R ANOVA /CRITERIA=PIN(.05) POUT(.10) /NOORIGIN /DEPENDENT Y /METHOD=ENTER X1 X2 X3 /CASEWISE PLOT(ZRESID) OUTLIERS(3) DEFAULTS DFBETA /SAVE MAHAL COOK LEVER DFBETA SDBETA DFFIT SDFIT COVRATIO.

High influence if D > 4/N Where N is the sample size.

Cooks distance

In Stata after running the regression type: predict D, cooksd

It is a summary measure of A D>1 indicates big outlier leverage and high residuals. problem .

High influence if leverage h > 2*k/N

Leverage

Measures how much an observation influences regression coefficients.

Where k is the number of parameters (including the intercept) and N is the sample size. A rule-of-thumb: Leverage goes from 0 to 1. A value closer to 1 or over 0.5 may indicate problems. Higher levels indicate higher distance from average values. The M-distance follows a Chisquare distribution with k-1 df and alpha=0.001 (where k is the number of independent variables). Any value over this Chi-square value may indicate problems.

In Stata after running the regression type:

predict lev, leverage

Same as above (LEV_1)

It is rescaled measure of leverage. Mahalanobis distance M = leverage*(N-1) Where N is sample size.

Not available

Same as above (MAH_1) 25

PU/DSS/OTR

Sources for the summary tables: influence indicators and distance measures Statnotes: http://faculty.chass.ncsu.edu/garson/PA765/regress.htm#outlier2 An Introduction to Econometrics Using Stata/Christopher F. Baum, Stata Press, 2006 Statistics with Stata (updated for version 9) / Lawrence Hamilton, Thomson Books/Cole, 2006 UCLA http://www.ats.ucla.edu/stat/stata/webbooks/reg/chapter2/statareg2.htm

26
PU/DSS/OTR

Regression: testing for normality

Another assumption of the regression model (OLS) that impact the validity of all tests (p, t and F) is that residuals behave normal. Residuals (here indicated by the letter e) are the difference between the observed values (Y) and the predicted values (Yhat): e = Y Yhat. In Stata you type: predict e, resid. It will generate a variable called e (residuals). Three graphs will help us check for normality in the residuals: kdensity, pnorm and qnorm. kdensity e, normal
.025

Kernel density estimate

A kernel density plot produces a kind of histogram for the residuals, the option normal overlays a normal distribution to compare. Here residuals seem to follow a normal distribution. Below is an example using histogram.

Density .01 .015

.02

histogram e, kdensity normal

.03

.005

Residuals Kernel density estimate Normal density

kernel = epanechnikov, bandwidth = 6.4865

Density 0 .01

If residuals do not follow a normal pattern then you should check for omitted variables, model specification, linearity, functional forms. In sum, you may need to reassess your model/theory. In practice normality does not represent much of a problem when dealing with really big samples.

.02

-40

-20

20

40

60

-40

-20

0 Residuals

20

40

27
PU/DSS/OTR

Regression: testing for normality

Standardize normal probability plot (pnorm) checks for non-normality in the middle range of residuals. Again, slightly off the line but looks ok. pnorm e
1.00

Quintile-normal plots (qnorm) check for non-normality in the extremes of the data (tails). It plots quintiles of residuals vs quintiles of a normal distribution. Tails are a bit off the normal. qnorm e
40

Normal F[(e-m)/s] 0.50 0.75

0.25

0.00

0.00

0.25

0.50 Empirical P[i] = i/(N+1)

0.75

1.00

-40

-20

Residuals 0

20

-40

-20

0 Inverse Normal

20

40

A non-graphical test is the Shapiro-Wilk test for normality. It tests the hypothesis that the distribution is normal, in this case the null hypothesis is that the distribution of the residuals is normal. Type swilk e

. swilk e Shapiro-Wilk W test for normal data Variable e Obs 50 W 0.95566 V 2.085 z 1.567 Prob>z 0.05855

The null hypothesis is that the distribution of the residuals is normal, here the p-value is 0.06 we failed to reject the null (at 95%). We conclude then that residuals are normally distributed, with the caveat that they are not at 90%. 28
PU/DSS/OTR

Regression: joint test (F-test)

To test whether two coefficients are jointly different from 0 use the command test (see Hamilton, 2006, p.175).
xi: quietly regress csat expense percent percent2 income high college i.region, robust
Note quietly suppress the regression output

To test the null hypothesis that both coefficients do not have any effect on csat (high = 0 and college = 0), type: test high college
. test high college ( 1) ( 2) high = 0 college = 0 F( 2, 40) = Prob > F = 17.12 0.0000

The p-value is 0.0000, we reject the null and conclude that both variables have indeed a significant effect on SAT.

Some other possible tests are (see Hamilton, 2006, p.176): test income = 1 test high = college test income = (high + college)/100 29
Note: Not to be confused with ttest. Type help test and help ttest for more details
PU/DSS/OTR

Regression: saving regression coefficients

Stata temporarily stores the coefficients as _b[varname], so if you type: gen percent_b = _b[percent] gen constant_b = _b[_cons] You can also save the standard errors of the variables _se[varname] gen percent_se = _se[percent] gen constant_se = _se[_cons]

. summarize Variable percent_b percent_se constant_b constant_se

percent_b percent_se constant_b constant_se Obs 51 51 51 51 Mean -5.945267 .6405529 873.9537 58.12895 Std. Dev. 0 0 0 0 Min -5.945267 .6405529 873.9537 58.12895 Max -5.945267 .6405529 873.9537 58.12895

30
PU/DSS/OTR

Regression: saving regression coefficients/getting predicted values

You can see a list of stored results by typing after the regression ereturn list:
. xi: quietly regress csat expense percent percent2 income high college i.region, robust i.region _Iregion_1-4 (naturally coded; _Iregion_1 omitted) . ereturn list scalars: e(N) e(df_m) e(df_r) e(F) e(r2) e(rmse) e(mss) e(rss) e(r2_a) e(ll) e(ll_0) macros: e(cmdline) e(title) e(vce) e(depvar) e(cmd) e(properties) e(predict) e(model) e(estat_cmd) e(vcetype) matrices: e(b) : e(V) : functions: e(sample) 1 x 10 10 x 10 : : : : : : : : : : "regress csat expense percent percent2 income high college _Iregion_*, robust" "Linear regression" "robust" "csat" "regress" "b V" "regres_p" "ols" "regress_estat" "Robust" = = = = = = = = = = = 50 9 40 76.92400040408057 .9404045015031877 17.81259357987284 200269.8403983309 12691.53960166909 .9269955143414049 -209.3636234584767 -279.8680043669825

Type help return for more details

31
PU/DSS/OTR

Regression: general guidelines The following are general guidelines for building a regression model* 1. 2. 3. 4. Make sure all relevant predictors are included. These are based on your research question, theory and knowledge on the topic. Combine those predictors that tend to measure the same thing (i.e. as an index). Consider the possibility of adding interactions (mainly for those variables with large effects) Strategy to keep or drop variables:
1. 2. 3. 4. Predictor not significant and has the expected sign -> Keep it Predictor not significant and does not have the expected sign -> Drop it Predictor is significant and has the expected sign -> Keep it Predictor is significant but does not have the expected sign -> Review, you may need more variables, it may be interacting with another variable in the model or there may be an error in the data.

*Gelman, Andrew, Jennifer Hill, Data Analysis Using Regression and Multilevel/Hierarchical Models, 2007, p. 69

32
PU/DSS/OTR

Regression: publishing regression output (outreg2)

The command outreg2 gives you the type of presentation you see in published papers. If outreg2 is not available you need to install it by typing ssc install outreg2 Lets say the regression is regress csat percent percent2 high, robust The basic syntax for outreg2 is: outreg2 using [pick a name], [type either word or excel] After the regression type the following if you want to export the results to excel* outreg2 using results, excel
. outreg2 using results, excel "results.xml" Click here to see the seeout file

Or this if you want to export to word outreg2 using results, word

. outreg2 using results, word "results.rtf" seeout

In excel

In word

Click here to see the file

33
*See the following document for some additional info/tips http://www.fiu.edu/~tardanic/brianne.pdf PU/DSS/OTR

Regression: publishing regression output (outreg2)

You can add more models to compare. Lets say you want to add another model without percent2: regress csat percent high, robust Now type to export the results to word (notice we add the append option) outreg2 using results, word append

In excel

In word

NOTE: If you run logit/probit regression with odds ratios you need to add the option eform to export the odd ratios

Type help outreg2 for more details. If you do not see outreg2, you may have to install it by typing ssc install outreg2. If this does not work type findit outreg2, select from the list and click install. Note: If you get the following error message (when you use the option append or replace it means that you need to close the excel/word window.

34
PU/DSS/OTR

Regression: publishing regression output (outreg2) continue

For a customized look, here are some options:
*** Excel outreg2 using results, bdec(2) tdec(2) rdec(2) adec(2) alpha (0.01, 0.05, 0.10) addstat(Adj. R-squared, e(r2_a)) excel *** Word outreg2 using results, bdec(2) tdec(2) rdec(2) adec(2) alpha (0.01, 0.05, 0.10) addstat(Adj. R-squared, e(r2_a)) word

For excel

For word

Click here to see the output, a excel/word window will open

Name of the file for the output Click on seeout to browse the results Set # of decimals for coefficients

Set # of decimals for auxiliary statistics Set # of decimals for the R2

Set # of decimals for added statistics (addstat option) Levels of significance

Include some additional statistic, in this case adj. R-sqr. You can select any statistics on the return lists (e-class, rclass or s-class). After running the regression type ereturn list for a list of available statistics.

35
PU/DSS/OTR

Regression: interaction between dummies

Interaction terms are needed whenever there is reason to believe that the effect of one independent variable depends on the value of another independent variable. We will explore here the interaction between two dummy (binary) variables. In the example below there could be the case that the effect of student-teacher ratio on test scores may depend on the percent of English learners in the district*.
Dependent variable (Y) Average test score, variable testscr in dataset. Independent variables (X) Binary hi_str, where 0 if student-teacher ratio (str) is lower than 20, 1 equal to 20 or higher.
In Stata, first generate hi_str = 0 if str<20. Then replace hi_str=1 if str>=20. In Stata, first generate hi_el = 0 if el_pct<10. Then replace hi_el=1 if el_pct>=10.

Binary hi_el, where 0 if English learners (el_pct) is lower than 10%, 1 equal to 10% or higher

Interaction term str_el = hi_str * hi_el. In Stata: generate str_el = hi_str*hi_el

We run the regression regress testscr hi_el hi_str str_el, robust

The equation is testscr_hat = 664.1 18.1*hi_el 1.9*hi_str 3.5*str_el The effect of hi_str on the tests scores is -1.9 but given the interaction term (and assuming all coefficients are significant), the net effect is -1.9 -3.5*hi_el. If hi_el is 0 then the effect is -1.9 (which is hi_str coefficient), but if hi_el is 1 then the effect is -1.9 -3.5 = -5.4. In this case, the effect of student-teacher ratio is more negative in districts where the percent of English learners is higher. See the next slide for more detailed computations. 36 *The data used in this section is the California Test Score data set (caschool.dta) from chapter 6 of the book Introduction to Econometrics from Stock and Watson, 2003. Data can be downloaded from http://wps.aw.com/aw_stock_ie_2/50/13016/3332253.cw/index.html.For a detailed discussion please refer to the respective section in the book. PU/DSS/OTR

Regression: interaction between dummies (cont.)

You can compute the expected values of test scores given different values of hi_str and hi_el. To see the effect of hi_str given hi_el type the following right after running the regression in the previous slide. These are different scenarios holding constant hi_el and varying hi_str. Below we add some labels

We then obtain the average of the estimations for the test scores (for all four scenarios, notice same values for all cases).

Here we estimate the net effect of low/high student-teacher ratio holding constant the percent of English learners. When hi_el is 0 the effect of going from low to high student-teacher ratio goes from a score of 664.2 to 662.2, a difference of 1.9. From a policy perspective you could argue that moving from high str to low str improve test scores by 1.9 in low English learners districts. When hi_el is 1, the effect of going from low to high student-teacher ratio goes from a score of 645.9 down to 640.5, a decline of 5.4 points (1.9+3.5). From a policy perspective you could say that reducing the str in districts with high percentage of English learners could improve test scores by 5.4 points. 37 *The data used in this section is the California Test Score data set (caschool.dta) from chapter 6 of the book Introduction to Econometrics from Stock and Watson, 2003. Data can be downloaded from
http://wps.aw.com/aw_stock_ie_2/50/13016/3332253.cw/index.html.For a detailed discussion please refer to the respective section in the book. PU/DSS/OTR

Regression: interaction between a dummy and a continuous variable

Lets explore the same interaction as before but we keep student-teacher ratio continuous and the English learners variable as binary. The question remains the same*. Dependent variable (Y) Average test score, variable testscr in dataset. Independent variables (X) Continuous str, student-teacher ratio. Binary hi_el, where 0 if English learners (el_pct) is lower than 10%, 1 equal to 10% or higher Interaction term str_el2 = str * hi_el. In Stata: generate str_el2 = str*hi_el We will run the regression regress testscr str hi_el str_el2, robust

The equation is testscr_hat = 682.2 0.97*str + 5.6*hi_el 1.28*str_el2 The effect of str on testscr will be mediated by hi_el.
If hi_el is 0 (low) then the effect of str is 682.2 0.97*str. If hi_el is 1 (high) then the effect of str is 682.2 0.97*str + 5.6 1.28*str = 687.8 2.25*str

Notice that how hi_el changes both the intercept and the slope of str. Reducing str by one in low EL districts will increase test scores by 0.97 points, but it will have a higher impact (2.25 points) in high EL districts. The difference between these two effects is 1.28 which is the coefficient of the interaction (Stock and Watson, 2003, p.223). 38 *The data used in this section is the California Test Score data set (caschool.dta) from chapter 6 of the book Introduction to Econometrics from Stock and Watson, 2003. Data can be downloaded from http://wps.aw.com/aw_stock_ie_2/50/13016/3332253.cw/index.html.For a detailed discussion please refer to the respective section in the book. PU/DSS/OTR

Regression: interaction between two continuous variables

Lets keep now both variables continuous. The question remains the same*. Dependent variable (Y) Average test score, variable testscr in dataset. Independent variables (X) Continuous str, student-teacher ratio. Continuous el_pct, percent of English learners. Interaction term str_el3 = str * el_pct. In Stata: generate str_el3 = str*el_pct We will run the regression regress testscr str el_pct str_el3, robust

The equation is testscr_hat = 686.3 1.12*str - 0.67*el_pct + 0.0012*str_el3 The effect of the interaction term is very small. Following Stock and Watson (2003, p.229), algebraically the slope of str is 1.12 + 0.0012*el_pct (remember that str_el3 is equal to str*el_pct). So:
If el_pct = 10, the slope of str is -1.108 If el_pct = 20, the slope of str is -1.096. A difference in effect of 0.012 points.

In the continuous case there is an effect but is very small (and not significant). See Stock and Watson, 2003, for further details.

39 *The data used in this section is the California Test Score data set (caschool.dta) from chapter 6 of the book Introduction to Econometrics from Stock and Watson, 2003. Data can be downloaded from http://wps.aw.com/aw_stock_ie_2/50/13016/3332253.cw/index.html.For a detailed discussion please refer to the respective section in the book. PU/DSS/OTR

Creating dummies
You can create dummy variables by either using recode or using a combination of tab/gen commands: tab major, generate(major_dum)

Check the variables window, at the end you will see three new variables. Using tab1 (for multiple frequencies) you can check that they are all 0 and 1 values

40
PU/DSS/OTR

Creating dummies (cont.)

Here is another example: tab agregroups, generate(agegroups_dum)

Check the variables window, at the end you will see three new variables. Using tab1 (for multiple frequencies) you can check that they are all 0 and 1 values

41
PU/DSS/OTR

Frequently used Stata commands

Category Getting on-line help help search Operating-system interface pwd cd sysdir mkdir dir / ls erase copy type Using and saving data from disk use clear save append merge compress Inputting data into Stata input edit infile infix insheet The Internet and Updating Stata update net ado news Stata commands

Basic data reporting

describe codebook inspect list browse count assert summarize Table (tab) tabulate

Type help [command name] in the windows command for details

Source: http://www.ats.ucla.edu/stat/stata/notes2/commands.htm

Data manipulation

generate replace egen recode rename drop keep sort encode decode order by reshape

Formatting

format label

Keeping track of your work

log notes

42
PU/DSS/OTR

Convenience

display

Is my model OK? (links)

Regression diagnostics: A checklist http://www.ats.ucla.edu/stat/stata/webbooks/reg/chapter2/statareg2.htm Logistic regression diagnostics: A checklist http://www.ats.ucla.edu/stat/stata/webbooks/logistic/chapter3/statalog3.htm Times series diagnostics: A checklist (pdf) http://homepages.nyu.edu/~mrg217/timeseries.pdf Times series: dfueller test for unit roots (for R and Stata) http://www.econ.uiuc.edu/~econ472/tutorial9.html Panel data tests: heteroskedasticity and autocorrelation
http://www.stata.com/support/faqs/stat/panel.html http://www.stata.com/support/faqs/stat/xtreg.html http://www.stata.com/support/faqs/stat/xt.html http://dss.princeton.edu/online_help/analysis/panel.htm
43
PU/DSS/OTR

I cant read the output of my model!!! (links)

Data Analysis: Annotated Output http://www.ats.ucla.edu/stat/AnnotatedOutput/default.htm Data Analysis Examples http://www.ats.ucla.edu/stat/dae/ Regression with Stata http://www.ats.ucla.edu/STAT/stata/webbooks/reg/default.htm Regression http://www.ats.ucla.edu/stat/stata/topics/regression.htm How to interpret dummy variables in a regression http://www.ats.ucla.edu/stat/Stata/webbooks/reg/chapter3/statareg3.htm How to create dummies http://www.stata.com/support/faqs/data/dummy.html http://www.ats.ucla.edu/stat/stata/faq/dummy.htm Logit output: what are the odds ratios? http://www.ats.ucla.edu/stat/stata/library/odds_ratio_logistic.htm
44
PU/DSS/OTR

Topics in Statistics (links)

What statistical analysis should I use? http://www.ats.ucla.edu/stat/mult_pkg/whatstat/default.htm Statnotes: Topics in Multivariate Analysis, by G. David Garson http://www2.chass.ncsu.edu/garson/pa765/statnote.htm Elementary Concepts in Statistics http://www.statsoft.com/textbook/stathome.html Introductory Statistics: Concepts, Models, and Applications http://www.psychstat.missouristate.edu/introbook/sbk00.htm Statistical Data Analysis http://math.nicholls.edu/badie/statdataanalysis.html Stata Library. Graph Examples (some may not work with STATA 10) http://www.ats.ucla.edu/STAT/stata/library/GraphExamples/default.htm Comparing Group Means: The T-test and One-way ANOVA Using STATA, SAS, and SPSS http://www.indiana.edu/~statmath/stat/all/ttest/
45
PU/DSS/OTR

Useful links / Recommended books

DSS Online Training Section http://dss.princeton.edu/training/ UCLA Resources to learn and use STATA http://www.ats.ucla.edu/stat/stata/ DSS help-sheets for STATA http://dss/online_help/stats_packages/stata/stata.htm Introduction to Stata (PDF), Christopher F. Baum, Boston College, USA. A 67-page description of Stata, its key features and benefits, and other useful information. http://fmwww.bc.edu/GStat/docs/StataIntro.pdf STATA FAQ website http://stata.com/support/faqs/ Princeton DSS Libguides http://libguides.princeton.edu/dss

Books Introduction to econometrics / James H. Stock, Mark W. Watson. 2nd ed., Boston: Pearson Addison Wesley, 2007. Data analysis using regression and multilevel/hierarchical models / Andrew Gelman, Jennifer Hill. Cambridge ; New York : Cambridge University Press, 2007. Econometric analysis / William H. Greene. 6th ed., Upper Saddle River, N.J. : Prentice Hall, 2008. Designing Social Inquiry: Scientific Inference in Qualitative Research / Gary King, Robert O. Keohane, Sidney Verba, Princeton University Press, 1994. Unifying Political Methodology: The Likelihood Theory of Statistical Inference / Gary King, Cambridge University Press, 1989 Statistical Analysis: an interdisciplinary introduction to univariate & multivariate methods / Sam Kachigan, New York : Radius Press, c1986 Statistics with Stata (updated for version 9) / Lawrence Hamilton, Thomson Books/Cole, 2006
46
PU/DSS/OTR