Linear Regression Models

Final Project

Fall 2011

Professor: Thian Zheng

Student: Ezequiel Zambaglione

Abstract
The objective of this project is to develop a model to help Countrys B Government to identify the
factors that can predict the outbreaks of virus A infection for further policy on virus prevention.
The data set taken into account to develop the model is a sample of outbreaks of Virus A infection in
100 geographical areas of the Country. The response variable is going to be the percentage of the area
population infected during the winter.
First of all, it is important to make a preliminary analysis of the data to select the candidate explanatory
variables. Then the use of econometrics methods of model selection will give the model that best fit the
sample data.
Secondly, some diagnostics to the residuals of the model were perform in order to be sure that there
are not missing explanatory variables, presence of outliers that can influence in future prediction,
multicollinearity and other issues.
Finally, the final model is present together with the conclusions.

Preliminary Analisys
To start the preliminary analysis of the data, first the list of candidate explanatory variables are
presented and an analysis of the correlation between each other and with the response variable % of
the area population infected during the winter of this study are going to be perform.
List of candidates variables:
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.

perc.male: percent of the area population that are male.

perc.college: percent of the area population that are college degree holders.
perc.child: percent of the area population that are school-aged children.
num.bed: number of hospital beds.
num.phys: number of physicians.
perc.vac: percent of the area population that received virus A vaccine.
population: size of area population in thousands.
hoshold.size: average households size in the area.
household.income: average household income in the area.
perc.insured: percent of the area population that have medical insurance.
geo.size: geographical size of the area in km2.
urban: indicator of urban area (1=urban, 0=otherwise).
camp.virus: indicator of campaign for virus A prevention in the area (1= yes, 0= no).
"temp": average temperature of winter months in the area.
"public.per.cap": public area per capita.
"employ": employment rate.
"rain": average amount of rain during winter months.
"region": country B has four regions (1, 2, 3, and 4).

In the following table it shows the correlation between all the explanatory variables and the response
variable (V1). Is good to notice that the percentage of males, college and insured, the employment rate
and the average amount of rain during the winter show a correlation with V1 near to zero, so they are
candidates to get out of the model.
Correlation between explanatory and response variables
perc.male
-0.0538
population
0.4305
camp.virus
-0.3683

perc.college
-0.0162
household.size
0.1520
temp
-0.2149

perc.child
0.1099
household.income
0.1507
public.per.cap
-0.4556

num.bed
0.5042
perc.insured
-0.0267
employ
-0.0072

num.phys
0.5087
geo.size
-0.6037
rain
0.0118

perc.vac
-0.1311
urban
0.4378
region
-0.4230

On the other hand, the number of beds and physicians, the population, size of the area, the indicator
variable of urban, the public space and the region show a correlation against V1 greater than 40%. They
are natural candidates for the model.
To obtain some information about possible future problems of multicollinearity, is important check the
correlation matrix of the explanatory variables. This matrix shows high correlation between the percent
of male, people with college degree and children in school age. The same behavior shows the number of
2

bed and physicians between each other and against the total population and the indicator variable of
urban. Finally, the household size and the household income are perfect correlated1.
Some other candidate variables are considered for the model,
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.

hab.km2: Habitant per km2=population/geo.size

perc.public.area: percent of public area= public.per.cap*population/geo.size
phys.bed: Physicians per bed=num.phys/num.bed
households: Total Households=population/household.size
total.income: Total income=household.income*(population/household.size)
num.unemploy: Number of uneployed=(1-population*employ)
phys.per.hab: Physicians per habitant=num.phys/population
phys.per.house: Physicians per house=num.phys/households
phys.per.km2: Physicians per km2=num.phys/geo.size
beds.per.km2: Beds per km2=num.bed/geo.size
beds.per.hab: Beds per habitant=num.bed/population
phys.per.child: Physicians per child=num.phys/(perc.child*population)
vac.km2: People that receive vaccine per km2=(perc.vac*population)/geo.size
income.per.cap: Income per habitant=household.income/household.size

Again, is important to check the relation between these added variables against the response variable
with a scatter plot, if they show some relation we were include them in the set of candidate variables2.
The decision of add habitants per km2, percentage of public area, number of unemployed, vaccine per
km2, physicians per bed, physicians per km2 and beds per km2 has been taken.
All these variables seem to have some correlation to the response variable, so they are going to be
included. However, some of them show high correlation between each other so is important to take
care about multicolinearity, for example, physicians and beds per km2.

1
2

The table with all the correlations you can see the appendix A
The plot with the pairs relations of the variables and the added variables are in the appendix A

Model selection
The objective of this section is to find the model that best fit the sample data, using the selected
candidate variables.
The method that is going to be used for the selection is the Stepwise Regression Method, starting with
the model V1~13 and using the AIC criteria to add and delete variables.
Step: AIC=-1344.3 (V1 ~ hab.km2 + region2 + region1 + camp.virus + perc.vac + perc.child + employ +
public.per.cap + phys.bed + households + temp). And the coefficients are the followings,
Coefficients of Model 1
Estimate
Std. Error
t value
Pr(>|t|)
(Intercept)
0.035
0.004
7.925
0.0000
perc.child
0.036
0.017
2.144
0.0348
perc.vac
-0.004
0.001
-3.718
0.0004
camp.virus
-0.002
0.000
-6.381
0.0000
temp
0.000
0.000
1.610
0.1109
public.per.cap
-0.469
0.161
-2.921
0.0044
employ
-0.007
0.004
-1.697
0.0932
region1TRUE
0.002
0.000
6.095
0.0000
region2TRUE
0.002
0.000
6.775
0.0000
phys.bed
0.009
0.005
1.905
0.0600
hab.km2
0.000
0.000
10.923
0.0000
households
0.000
0.000
-1.692
0.0942

The individual t-test was performed to test the significance of the coefficients. As it can be seen, the pvalue of the coefficients related to the temperature, employment rate, physicians per bed and number
of households are greater than 5%. However, we use test of Lack of Fit to decide if we continue with this
model or another model without these variables.
Analysis of Variance Table
V1 ~ perc.child + perc.vac + camp.virus + public.per.cap + region1 + region2 +
hab.km2
V1 ~ perc.child + perc.vac + camp.virus + temp + public.per.cap + employ +
Model1
region1 + region2 + phys.bed + hab.km2 + households
Res.Df
RSS
Df
Sum of Sq F
Pr(>F)
1
92 0.00012897
2
88 0.00011417
4
1.48E-05
2.8531
0.02829
Model2

The reduced in the RSS is pretty small, so we could prefer to have 4 more degree of freedom. However,
the p-value suggests rejecting H0: b3=b5=b6=b8=0, so we stay with model 1.

The test was running starting with the model V1~1 because we arrive to a lower AIC value than starting from the
full model.

Residuals analysis
Once we have chosen the best model, we are going to check its residuals. The following graph shows
that the residuals seem to follow a quadratic function, which means that probably we need to perform a
transformation in the response variable.
The Box-Cox transformation was used to identify the power of the necessary transformation. On the
other hand, the residuals against the fitted values looks like a white noise.
Residuals

0.040
0.035

model1$fit

0.000
-0.001
-0.003

0.030

-0.002

model1$res

0.001

0.002

0.045

Residuals vs Fitted values

20

40

60
Index

80

100

-0.003

-0.002

-0.001

0.000

0.001

0.002

model1$res

To check the normality of the residuals we use the Normal Q-Q plot and histograms. Based on these two
graphs and the previous one, it is conclude that the residuals follow a normal distribution, maybe with
some skew to the right.

Histogram of model1$res

-0.003

-0.002

10

Frequency

15

0.001
0.000
-0.001

Sample Quantiles

0.002

20

Normal Q-Q Plot

-2

-1

-0.003

-0.002

-0.001

0.000

0.001

0.002

model1$res

Theoretical Quantiles

The quadratic behavior showed by the residual, suggest a transformation of the response variable V1.
The variable was transformed to Y=V1-1 and Y=(V1)-0.5, and now the residuals seems to follow a random
behavior. However, they seem to have less variability in the center than in the extremes.
Residuals transformed model Y=V1^(-1)

-0.1

0.0

model6$res

-0.2

-2

-1

model7$res

0.1

0.2

Residuals transformed model Y=V1^(-0.5)

20

40

60
Index

80

100

20

40

60

80

100

Index

The transformation Y1=V1-1 was selected because both arrive the same results and this one is more
simple. With this transformation we go back to the model selection and we obtain the new model with
the same variables but different coefficients.

Coefficients of Model 2
Estimate
Std. Error
t value
Pr(>|t|)
(Intercept)
29.12
3.90
7.46
0.00000
perc.child
-26.39
14.76
-1.79
0.07724
perc.vac
3.10
0.98
3.17
0.00210
camp.virus
1.30
0.22
5.87
0.00000
temp
-0.03
0.01
-2.07
0.04169
public.per.cap
559.30
142.80
3.92
0.00018
employ
5.38
3.64
1.48
0.14297
region1TRUE
-1.69
0.28
-5.94
0.00000
region2TRUE
-1.75
0.25
-6.89
0.00000
phys.bed
-9.64
4.16
-2.32
0.02282
hab.km2
-0.01
0.00
-8.59
0.00000
households
0.00
0.00
1.38
0.16982

Now a check on the residuals of the new model is needed. The histogram of the residuals shows a
distribution really similar to a normal one, however the residuals against the fitted values graph shows
some increase in the volatility while the fitted values increases, this problem could be related with some
multicollinearity between the explanatory variables. On the other hand, the graphs of the residuals
against the explanatory variables show a white noise behavior4.
Residuals vs Fitted values

15
10

Frequency

-2

-1

model8$res

20

25

Histogram of model8$res

22

24

26

28

30

32

model8$fit

-2

-1

model8$res

Finally, the square of the residuals and its absolute value against the fitted value was checked to be sure
that they follow a white noise distribution. It is possible to see in the following graphs that both, the
square and the absolute values of the residuals against the fitted values show some positive trend and
increase in the variability.

Check the appendix A to see this scatter plot.

Absolut value of the residuals vs Fitted values

0.0

0.5

model8$res^2

1.5
1.0

abs(model8$res)

2.0

2.5

Square of the residuals vs Fitted values

22

24

26

28

30

32

22

24

26

model8$fit

28

30

32

model8$fit

This behavior of the residuals could be related with the fact that we are missing some explanatory
variable. Because of that the square of the explanatory variables and some interaction terms are going
to be included. The final model with an AIC value of 6.21, better than the 13.9 of the previous model,
was selected. In this last model, the variables were centered to avoid some multicollinearity issue
between the variables and its squares.

(Intercept)
c.perchild
c.perc.vac
c.households
c.employ
camp.virus
region1TRUE
region2TRUE
c.hab.km2
c.publicpercap
c.publicpercap2
c.hab.km22
region2TRUE:c.hab.km2

Coefficients of Model 3
Estimate
Std. Error
28.07
0.36
-23.77
14.31
3.16
0.95
0.00
0.00
10.18
3.46
1.17
0.21
-1.70
0.27
-1.44
0.25
-0.03
0.00
-75.41
162.80
271300.00 92780.00
0.00
0.00
0.01
0.00

t value
78.09
-1.66
3.33
2.60
2.94
5.62
-6.18
-5.84
-6.43
-0.46
2.93
3.04
2.56

Pr(>|t|)
0.00
0.10
0.00
0.01
0.00
0.00
0.00
0.00
0.00
0.64
0.00
0.00
0.01

The residuals of this new model still show some trend in the variance, in order to correct this problem
we are going to estimate the model with weights least squares, using the hat values as the weights. The
residuals against the fitted values seems to are5 ok and the new coefficients are the following,

To check the new residuals go to appendix A

(Intercept)
c.perchild
c.perc.vac
c.households
c.employ
camp.virus
region1TRUE
region2TRUE
c.hab.km2
c.publicpercap
c.publicpercap2
c.hab.km22
region2TRUE:c.hab.km2

Coefficients of Model 4
Estimate
Std. Error
28.21
0.32
-23.27
13.22
2.62
0.91
0.00
0.00
13.26
3.26
1.12
0.21
-1.70
0.26
-1.44
0.25
-0.03
0.00
-28.51
152.30
284300.00 81390.00
0.00
0.00
0.01
0.00

t value
88.19
-1.76
2.89
2.51
4.07
5.42
-6.48
-5.68
-6.84
-0.19
3.49
3.31
3.22

Pr(>|t|)
0.00
0.10
0.00
0.01
0.00
0.00
0.00
0.00
0.00
0.64
0.00
0.00
0.01

It is important to check for outliers, so in the first place we check for outliers on the Y observations using
the Studentized deleted errors. There is not presence of outliers in Y observations. Then we check for
outliers in the X observations using the Hat matrix leverage values. There is not presence of outliers in X
observations6.
And the last diagnostic is the Variance Inflation Factor, in order to check if the explanatory variables
show multicollinearity. The model shows serious multicollinearity issues because the maximum VIF
value is greater than 10 and the average VIF value is greater than 1.

c.perchild
c.perc.vac
c.households
c.employ
1.122
1.249
1.990
1.131
camp.virus
region1
region2
c.hab.km2
1.196
1.511
1.481
37.252
c.publicpercap c.publicpercap2 c.hab.km22 region2:c.hab.km2
3.013
2.194
29.082
1.822

The tables with the result of both tests are in the appendix A.

Conclusions
Along this project a model to predict the percentage of the population infected by the Virus A in the
Country B was developed. To do that we use the variables included in the sample data, we create new
variables and at the end we use the square of this variables, the interaction between each other and we
also transformed the response variable in order to get the best results.
The best model to predict the transformed response variable (Y1=V1-1) was,
Y1 = 28.21-23.27*c.perchild+2.62*c.perc.var+0.41E-4*c.households+13.26*c.emply+1.12*camp.virus1.7*region1-1.44*region2-0.03*c.hab.km2-28.51*c.publicpercap-28.43E+4*c.publicpercap2+0.23E-4
*c.hab.km22 +0.01*region2*c.hab.km2
The model seems to fit really well, with an adjusted R2 of 89.24%, which means that the 89.24% of the
variation of the response variables is explain by the model.
One limitation of the model is that we find multicollinearity between the explanatory variables, which
means that these variables have high correlation between each other. This issue could increase the
mean square error of these variables. However, the multicollinearity issue doesnt affect the prediction
power of the model, so we conclude that this model is a good tool to make the best prediction of the
percentage of the population infected with the Virus A.

10

Apendix A
Correlation matrix of the sample variables
V1
V1
perc.male
perc.college
perc.child
num.bed
num.phys
perc.vac
population
household.size
household.income
perc.insured
geo.size
urban
camp.virus
temp
public.per.cap
employ
rain
region

1.0000
-0.0538
-0.0162
0.1099
0.5042
0.5087
-0.1311
0.4305
0.1520
0.1507
-0.0267
-0.6037
0.4378
-0.3683
-0.2149
-0.4556
-0.0072
0.0118
-0.4230

perc.male
perc.college
perc.child
num.bed
num.phys
perc.vac
population household.size household.income perc.insured geo.size
urban
camp.virus temp
public.per.cap employ
rain
region
-0.05379185
-0.016239036
0.109909523
0.504218733
0.508748195 -0.131081034 0.43048432
0.15201332
0.150738919 -0.026732281 -0.603651932 0.437805016 -0.36833661 -0.214920555 -0.455620944 -0.007221472 0.01176554 -0.422998072
1
0.962909313
-0.914793855
0.14996299
0.131697934 0.072201059 0.09659195
0.03245287
0.076018307 0.125211244 0.040532916 0.089086629 0.10062439 0.049526291 0.030377274 -0.231542409 -0.044748119 0.119368858
0.96290931
1
-0.87415802
0.151829575
0.129064518 0.059845935 0.11501937
0.08234228
0.127280056 0.101606782 0.03104878 0.075338931 0.07441531 0.021969768 -0.003727268 -0.250959554 -0.031603178 0.060832004
-0.91479385
-0.87415802
1
-0.090994145
-0.074138337 -0.007453921 -0.05871094
-0.02338423
-0.062674736 -0.129227114 -0.074004725 -0.021437323 -0.08682309 0.003911223 -0.029381878 0.190215232 -0.052637884 -0.126157206
0.14996299
0.151829575
-0.090994145
1
0.945603843 0.195119591 0.79944142
0.0355717
0.041892933 0.006545029 -0.336654044 0.842372441 -0.08353608 -0.028253979 -0.088867222 -0.044900966 0.187778643 -0.09640023
0.13169793
0.129064518
-0.074138337
0.945603843
1 0.133602216 0.67182316
0.03654678
0.042060555 0.022692612 -0.386901068 0.900935808 -0.04472267 -0.043452632 -0.003097691 -0.029899445 0.191047857 -0.070914348
0.07220106
0.059845935
-0.007453921
0.195119591
0.133602216
1 0.20443806
0.01547453
0.017222982 0.130679036 -0.075647759 0.103700763 0.10046444 0.08291165 -0.050041915 -0.014499702 0.091875188 0.100483597
0.09659195
0.115019371
-0.05871094
0.799441421
0.671823158 0.204438061
1
0.06885906
0.073964098 -0.023112172 -0.162222753 0.476397949 -0.12470366 -0.085647889 -0.38589852 0.058129231 0.208401526 -0.109526469
0.03245287
0.082342283
-0.023384231
0.035571704
0.03654678 0.015474526 0.06885906
1
0.998948416 -0.026522967 -0.027913535 -0.020886456 0.01964967 -0.135394383 -0.077483079 0.016666948 -0.133816336 -0.070302971
0.07601831
0.127280056
-0.062674736
0.041892933
0.042060555 0.017222982 0.0739641
0.99894842
1 -0.020799924 -0.026474332 -0.017860898 0.02248706 -0.133947121 -0.078446773 0.004915188 -0.135392072 -0.06714767
0.12521124
0.101606782
-0.129227114
0.006545029
0.022692612 0.130679036 -0.02311217
-0.02652297
-0.020799924
1 -0.03506215 0.02540019 0.13967916 0.025298329 0.028860324 0.065061261 0.10514793 0.086176471
0.04053292
0.03104878
-0.074004725
-0.336654044
-0.386901068 -0.075647759 -0.16222275
-0.02791354
-0.026474332 -0.03506215
1 -0.431631853 -0.04174145 0.098784403 0.493078933 -0.067108329 -0.005145648 0.085411298
0.08908663
0.075338931
-0.021437323
0.842372441
0.900935808 0.103700763 0.47639795
-0.02088646
-0.017860898 0.02540019 -0.431631853
1 -0.04467011 0.013929722 0.184959123 -0.037122816 0.102242036 -0.004807952
0.10062439
0.074415313
-0.086823089
-0.083536082
-0.044722665 0.10046444 -0.12470366
0.01964967
0.022487057 0.139679159 -0.041741451 -0.044670111
1 0.239928841 0.082487002 -0.017534417 0.028369109 0.022655664
0.04952629
0.021969768
0.003911223
-0.028253979
-0.043452632 0.08291165 -0.08564789
-0.13539438
-0.133947121 0.025298329 0.098784403 0.013929722 0.23992884
1 0.215311068 -0.232143626 -0.072323202 0.164453208
0.03037727
-0.003727268
-0.029381878
-0.088867222
-0.003097691 -0.050041915 -0.38589852
-0.07748308
-0.078446773 0.028860324 0.493078933 0.184959123
0.082487 0.215311068
1 -0.183543762 -0.035585086 0.126658913
-0.23154241
-0.250959554
0.190215232
-0.044900966
-0.029899445 -0.014499702 0.05812923
0.01666695
0.004915188 0.065061261 -0.067108329 -0.037122816 -0.01753442 -0.232143626 -0.183543762
1 -0.019797261 0.062310507
-0.04474812
-0.031603178
-0.052637884
0.187778643
0.191047857 0.091875188 0.20840153
-0.13381634
-0.135392072 0.10514793 -0.005145648 0.102242036 0.02836911 -0.072323202 -0.035585086 -0.019797261
1 0.071927922
0.11936886
0.060832004
-0.126157206
-0.09640023
-0.070914348 0.100483597 -0.10952647
-0.07030297
-0.06714767 0.086176471 0.085411298 -0.004807952 0.02265566 0.164453208 0.126658913 0.062310507 0.071927922
1

11

Scatter plot of the added variables

0.0

4.0e+08

0.01

0.02

0
0.030

0.08

phys.bed

100

0.08

V1

perc.public.area

10000

0.0

hab.km2

4.0e+08

households

-80000

total.income

num.unemploy

16800

vac.km2

0.01

income.per.cap

0.10

phys.per.hab

0.02

beds.per.hab

phys.per.house

phys.per.km2

phys.per.child

0.030

100

10000

-80000

16800

0.10

0.05

beds.per.km2

0.05

12

Scatter plot of the transformed variables and explanatory variables

0.000

0.86

0.01

0.0

0.0

-2 2

0.140

0.140

model9$res

perc.child

0.000

0.5

perc.vac

public.per.cap

0.86

30

temp

0.08

employ

0.01

phys.bed

0.02

phys.per.hab

phys.per.house

0.0

camp.virus

0.0

region1

hab.km2

0.5

30

0.08

0.02

0.0

100

-2

-1

Residuals vs Fitted values

model13$res

-2 2

100

0.0

region2

22

24

26

28

30

32

34

model13$fit

13

Studentized deleted errors test

1
FALSE

2
FALSE

15
FALSE

FALSE
16

FALSE
29

FALSE

57

58

71
FALSE

72
FALSE

85
FALSE

73

86

99

60
74
FALSE

87
FALSE

61
75
FALSE

88
FALSE

62
76
FALSE

89
FALSE

63
77
FALSE

90
FALSE

64
78
FALSE

91
FALSE

65
79
FALSE

92
FALSE

66
80
FALSE

93
FALSE

94
FALSE

69

82

95

70
FALSE

83
FALSE

96
FALSE

56
FALSE

FALSE

FALSE

FALSE

55

68

81

42
FALSE

FALSE

FALSE

FALSE

41

54

67

28
FALSE

FALSE

FALSE

FALSE

27

40

53

14
FALSE

FALSE

FALSE

FALSE

FALSE

26

39

52

13
FALSE

FALSE

FALSE

FALSE

FALSE

25

38

51

12
FALSE

FALSE

FALSE

FALSE

FALSE

24

37

50

11
FALSE

FALSE

FALSE

FALSE

FALSE

23

36

49

10
FALSE

FALSE

FALSE

FALSE

FALSE

22

35

48

9
FALSE

FALSE

FALSE

FALSE

FALSE

21

34

47

8
FALSE

FALSE

FALSE

FALSE

FALSE

FALSE

FALSE

FALSE

59

20

33

46

7
FALSE

FALSE

FALSE

FALSE

FALSE

19

32

45

6
FALSE

FALSE

FALSE

FALSE

FALSE

18

31

44

5
FALSE

FALSE

FALSE

FALSE

FALSE

17

30

43

4
FALSE

FALSE

FALSE

FALSE

84
FALSE

97
FALSE

98
FALSE

100
FALSE

Use hat matrix to indentified outliers

1
FALSE

2
FALSE

15
FALSE

16

29

57

58

71
FALSE

72
FALSE

85
FALSE

59
73
FALSE

86
FALSE

99

60
74
FALSE

87
FALSE

61
75
FALSE

88
FALSE

62
76
FALSE

89
FALSE

63
77
FALSE

90
FALSE

64
78
FALSE

91
FALSE

65
79
FALSE

92
FALSE

66
80
FALSE

93
FALSE

94
FALSE

69

82

70
FALSE

83
FALSE

96
FALSE

56
FALSE

FALSE

FALSE
95

FALSE

55

68

81

42
FALSE

FALSE

FALSE

FALSE

41

54

67

28
FALSE

FALSE

FALSE

FALSE

27

40

53

14
FALSE

FALSE

FALSE

FALSE

FALSE

26

39

52

13
FALSE

FALSE

FALSE

FALSE

FALSE

25

38

51

12
FALSE

FALSE

TRUE

FALSE

FALSE

24

37

50

11
FALSE

FALSE

FALSE

FALSE

FALSE

23

36

49

10
FALSE

FALSE

FALSE

FALSE

FALSE

22

35

48

9
FALSE

FALSE

FALSE

FALSE

FALSE

21

34

47

8
FALSE

FALSE

FALSE

FALSE

FALSE

20

33

46

7
FALSE

FALSE

FALSE

FALSE

FALSE

19

32

45

6
FALSE

FALSE

FALSE

FALSE

FALSE

18

31

44

5
FALSE

FALSE

FALSE

FALSE

FALSE

17

30

43

4
FALSE

FALSE

FALSE

FALSE

FALSE

FALSE

FALSE

FALSE

84
FALSE

97
FALSE

98
FALSE

100
FALSE

14