Advanced Statistics

Descriptive Statistics

Descriptive Statistics

Economics
University of Manchester

1
Numerical data summaries
LOCATION
‘central value’ - mean and median
DISPERSION/SPREAD
how concentrated are data values around central
location – variance and standard deviation
CORRELATION AND REGRESSION
Relationships between variables
Building block of analysis of “effects”

2
Notation for Variables
VARIABLE
A “variable” is simply a label, with description, for an event of
interest (X,Y,Z, etc)

For example:
“Let X = your weekly expenditure on food”
“Let P = the price of a litre of petrol”
“Let Y = your household income”
X 1
=
.......
; X 2
=
.......
; X 3
=
.......
; X 4
=
.......
;

etc ….

Xi : in general, denotes the i th observation on the variable X.

Note the use of “subscript i”
If we have X1 , X2 , ... , Xn
a (random) sample of n observations on X

3
Summation: “adding up”

We use the following definitions and notation

to signify summing up:

Greek “SIGMA”, S: =X1+X2+…+Xn

n n
Dummy SUBSCRIPT:  Xi =  X j
i =1 j =1

4
Measures of location
Suppose we have sample data: X1, X2, …, Xn

Sample (arithmetic) mean:

Weighted mean sometimes used: (eg, price

indices)
n

 w X = (w X
i =1
i i 1 1 + w2 X 2 ++ wn X n );
n
with w
i =1
i =1
5
Measures of dispersion I
(Mean) deviation: xi = X i − X
Mean Squared Deviation (MSD)

1 n 2 1 n
 i n
n i =1
x = ( X i)− X
2

i =1

=
1
n
( ) ( 2
X1 − X + X 2 − X ) 2
(
+ + X n − X )
2

1 n
=   Xi  − X( )
2 2

 n i =1 

• Note: MSD ≥ 0

• Manipulations in Worked Exercises Q1 6

Measures of dispersion II
Sample variance

( )
n n
1 1
  i
2
s2 = Xi − X = x 2

n − 1 i =1 n − 1 i =1

Sample standard deviation: s=+ s 2

Note:
1 Divisor of n – 1, not n, for statistical reasons
2 Both s 2 and s ≥ 0
3 Sample mean and s are measured in the units
of the original data
4 Variance measured in squared units
7
Regression
Summarise the data/scatter with line of best fit
Data (Xi, Yi), i = 1, 2, …, n
LINEAR relationship between Y and X?
What is the equation of the line of BEST FIT?

Salary against education

40
Salary (000s)

30
20

10
0
0 2 4 6
8
Years of FE
Line of best fit I

Consider the following scatter

(Xi, Yi), i = 1, 2, …, 6

9
Line of best fit II

Z=a+bX

Draw any straight line through the points:

depicted by the green line, Z=a+bX,
where a and b define the
intercept and slope, respectively.

10
Line of best fit III

Z=a+bX

Z
Choose a particular value of X in
the sample and get the
Y corresponding Z value, from the
line that has been drawn

Now compare this Z value with the actual Y value

in the sample associated with the chosen X

X
11
Line of best fit IV

Z=a+bX

Z
Residual = Y – Z
Y

The difference is called a residual.

Residual = Y – Z
= Y - a - bX

X
12
Line of best fit V

Z=a+bX

Can construct such a residual for each X value in the

sample.

Then square the residual values.

Sum these squares to get the “sum of squared

residuals” associated with the line Z=a+bX

13
Line of best fit VI

Z=a+bX

This sum of squared residuals is defined by:

n n

 (Yi − Zi ) =  (Yi − a − bX i )
2 2

i =1 i =1

14
Line of best fit VII

Z=a+bX

Z W=c+dX
W
Y
Could repeat this for any other line;
eg, W=c+dX

This will give rise to a different set

of residuals:

Residual = Y - W
= Y - c - dX

X
15
Line of best fit VIII

Z=a+bX

W=c+dX

And, therefore, a sum of squared residuals:

n n

 (Yi − Wi ) =  (Yi − c − dX i )
2 2

i =1 i =1

16
Line of best fit IX

Z=a+bX

W=c+dX

Which line is better?

Choose the line which has the smallest sum of squared

residuals. That is, choose Z = a+bX if
n n

 i
( − − )   i
( − − )
2 2
Y a bX i Y c dX i
i =1 i =1

17
Line of best fit: summary

Choose that line (i.e., choose a and b) which

minimises
n n n n n n

 (Y − a − bX ) =  Y +a +b X − 2a  Yi − 2b Yi X i + 2ab X i
2 2 2 2 2
i i i i
i =1 i =1 i =1 i =1 i =1 i =1

This is called Ordinary Least Squares

Regression
or, “regressing Y on X”

18
The regression equation
Regression line (line of best fit) has the mathematical
form:
Z i = Yˆi = a + bX i

The intercept and slope are given by (using calculus):

a = Y −bX ,

 (X )( ) x y
n n

i − X Yi − Y i i n
b= i =1
= i =1
=  wi yi
 (X )
n n

 i
2
i −X x 2 i =1

i =1 i =1

Derived Worked Exercises Q2. 19

Sample correlation
Gives an INDEX of the LINEAR relationship between
observed Y and X.
Defined as

 (X
n
)(
− X Yi − Y ) Y .
. . Y
..
r= i =1
i
, .. .
+
 (X
n

i −X )  (Y − Y )
2
n

i
2
. X . . X
i =1 i =1
0<r<1 r1
x y i i x 2
i
= i
=b i

y
Y
x y 2 2 2
i . .
.. . .
i i
i i i

• NB: -1  r  1 X

r=0
20
Regression and correlation:
calculations and interpretation
These calculations illustrated in the last four slides

But in practice use Excel, for example

EXAMPLE: Data on salary (£000) and education (years ofFE):

r = 0.83
Regression line has +ve slope (higher salary with more years FE);
regression line fits fairly well.
Yˆi = 16.3 + 2.86 X i
Line of best fit:

Salary (on average) estimated to be higher by £2,860 for each

additional year of FE;

Salary estimated to be £16,300 when no FE (Xi =0).

21
Example
67 industrial firms, cross-section data:
CEO salary (in 1990, thousand US$)
Firm sales (in 1990, million US$)

CEO Salary versus Firm Sales

5000

4000

3000
Salary

2000

1000

0
0 20000 40000 60000 80000
Sales

22
Example continued
r = 0.53
Regression line:

Yˆi = 930.3 + 0.025 X i

Xi : firm sales (million US$)
Yi : CEO salary (thousand US$)

CEO Salary versus Firm Sales

5000

4000

3000
Salary

2000

1000

0
0 20000 40000 60000 80000
Sales
23