0% found this document useful (0 votes)
115 views35 pagesRegression & Correlation
Uploaded by
Danish VohraCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
0% found this document useful (0 votes)
115 views35 pagesRegression & Correlation
Uploaded by
Danish VohraCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
/ 35
CORRELATI9;
AND REGRESsjo;
10.1. INTRODUCTION
Im the preceding chapters, we were dealing with the statistical analysis involving aunizig
data, i.e., a data related to a single measurement. Many statistical problems arise where we ue
Pairs, triplets or a higher number of measurements, Such data are called bivariate, tivareg
multivariate data, In the present chapter, we shall be dealing with a bivariate data in whic ee
will be pairs of values corresponding to two variables or characteristics for each unit uit
study, ¢.g., fertilizer used and yield of various plots, income and expenditure of dfs
households, volume and pressure of a gas, height and weighYof individuals, etc. In such acie
the interest lies in analysing whether there is any relatioriship between the two characteris
under study, c., whether changes in the value of one variable cause changes in the value oft
in Some systematic way or not. For example, one may be interested in studying whether wih
increase in the quantity of fertilizer used, the yield of a crop increases, decreases of ems
unaffected,
Jn the case of physical sciences, most of the times there exist a strong relationship b¥
id ef aisles, ¢.8., for a gas, at fixed temperature, a given value of pressure cores
TP yemune: In such a case, the relationship can be studied by ploting the pairs of vibe
Sates ts the other hand, in social and in natural sciences, the exact effect of change 3
different plots eel 1 wrely known, For example, if we add the same quantity ot wen
other natural teens Ot Bet the same yield in all the plots. The reason being om 4 ene
one wishes to a ecting the phenomenon which cannot be fully controlled. hh aici!
Which ditection, This cajentont UPt© Which yield is affected by the quantity of ferilize”
™ ‘his will form the subject matter of correlation analysis. aod
if the two variables are related, one may wish to estimate or predict HY Tsp
other known variable, in case the sell iA ii ‘
low the expected yield if a given quamttY © ios
the help of analysis which we shall discuss”
cot CORRELATION AND REGRESSION
10.2. CORRELATION ANALYSIS
127
Asalready pointed out, correlation analysis aims at fin ing the degree of relationshi ip and the
direction of relationship between two variables. As far as direction is concerned, if it is found that
ashe value of one vari s, the value of other variable also increases, on the average, the
two variable are said tq be positively correlated. On the other hand if tié increase in one variable is
followed by a decrease in the other, the Variables are sai to be negatively correlated: There may be
a third situation where with the change in one variable the value o
f the other remains constant on the
average. In such cases there is zerd or.no correlation and variables are then said to be uncorrelated.
Following are the methods commonly used for studying correlation between two variables :
\ 1. Scatter diagram
2. Karl Pearson’s coefficient of correlation
N 3. Rank correlation coefficient
= Of these, first method is based on the knowled,
ige of diagram whereas the other two are
mathematical procedure. Now, we shall discuss these methods in detail.
1. Scatter Diagram. For a bivariate data, if the pairs of values are plotted in the x-y plane,
sae De seam of dots so obtained is called a scatter diagram. From this scatter diagram, one can have
Tit) a fairly good, though vague, idea about the correlation between two variables. If the points (dots)
have) are concentrated about some curve the variables are expected to be correlated. If all the plotted
te t| points fall around a straight line, then we say that linear correlation exist between the two variables.
there | Formostof the practical situation:
Ss, We assume the relationship to be linear just to introduce. simplicity
sndet| inthe analysis,
fre On the other hand if the dots are widely scattered, a poor correlation is expected. Figure 10.1
cas | TSHesens some scatter diagrams, In (a) all the points lie on a straight line with positive slope. This
ristics Y Y Y
f oft 4 7 4
ith be
emits
ae Ql rtxo x0 x
ds © (@) PERFECT POSITIVE (b) LIMITED DEGREE OF (C) PERFECT NEGATIVE
“3 on? CORRELATION Positive conreLario (©) PERFECT NEGA
Pd
2 io
aos y
ioe y
oe }
oe
am iP
f
ve % oO +x 0 ox
ae (0) LIMITED DEGREE OF —_(E) ZERO CORRELATION
gor i NEGATIVE CORRELATION
ils 8+ 10.1, Scatter'diagram showing positive, negative and zero correlation.
jo? STATISTICAL METHODS FOR
RESEAR(
ICH Woy
128
iti ion. the dots are
is £ perfect positive correlation. In (6) S are scattered
is the case of M “4 degree of positive correlation, sig
i iti is indi limites a
th positive slope. This indicates a z i .
pce in indicate perfect negative correlation and limited dent thy
8 Ofte
(0) and (4) A ©
Dots in (e) are scattered in.an irregular manner. Corre
atic
10n in By
correlation respectively.
case is said to be zero
ibjective meth
This is quite a sut : :
value to the extent of relationship that exists betwee!
‘ber of observations is fairly large.
a scatter diagram if the num!
cient of Correlation. To measure the degree of linear,
i
od of measuring correlation. It does riot give
a
two variables. Moreover, itis ditt
9
2. Karl Pearson Coeff
h Biometrician, Karl Pearson, developed a formula called
Comte
between two variables, @ Britisl
coefficient. It is a numerical measure of linear relationship and is defined below :
If (Kp Y)s = Ay Qronera is a bivariate data then Karl Pearson’s Linoou"
Correlation Coefficient r (X, Y) is given as (@.
“Cov (% Y) Aeghientn)
r®& Y= Cov {e, *) |
iD i
rg Pa = 2 <<
where, Cov (X,Y) = ye, -X) %-Y) -t > XY, - XY
i= = i=1 -
is called the Covariance of X and-y, in analogy with the term variance that is used inca
univariate data,
:
Thus substituting the expressions for Cov (X, Y), 0, and gy in (1), we get
(XY) = —2& -)%-¥) S EX 0X! 4
& Ex, x ayy ex?-e en!
also called pot
¢ correlation coeffici i
Heanor ficient calculated by the above formula is also (8 ~ >
2
is variance of X, and
1
n
5 _y= 2_¥
> (y%- ¥)’ = B¥/ - Y” is variance of Y.
Properties of C .
wearing Colton one hein rp
Th sir proofs.
{G) The value of comelation coefficient lies between | and J a ond)
"roof. Let us put 2
x; ay
i and ¥
ey
Then, r= 1 pyr :
° XY}, and Lex? = 1
n = 2X Lye = Lay? 1
Since, 1 xx
_ Since, 77 2K + ¥, = 0
oof
me
or used inc
CORRELATION AND REGRESSION 129
ot dexp+t Lyy? +2 sxvie0 or 20+nz0 or orz-l
agin +E (Xj -¥)? = 0 or = %1-Hz0 o rei,
Which proves the other part of property.
Ths-1srstl.
The value of r will be —1 if
=o =.
x= -¥i, or Y= ¥->-K -X), for each i.
Similarly, the value of r will be 1 if
=X or ¥,-¥+ 2 (x,-%), for each &
Thus, in these cases each variable is an exact linear combination of the other and the variables
are said to be perfectly correlated. ae
Atif The correlation coefficient is a/pure number which is independent of the units of
measurement of the two vari
Proof. Let for each pair of values (X;, Y,), we have the pair of values (U;, V) given as
Xi-A Yi-B
Ua, and WG
X=A+cU, since X,=A+cU;
or X,-X=c(;-W)
Similarly, =d(V,-V)
Hence, we gave variances and covariances of X and Y as
Cov (KY) = ed HEU: ~ H) (Vi - V) = ed Cov (U,V)
1 v aa
o2=c? peG -BP =co2
1 =
and 0} = d? SEM - VP? = dot
Thus, we get
edCov(U, V)
r%&Y)= _ ed CovU, V) Oy
“ay leaner =a ao
In case c and d are of same sign wil be equal to 1 and, thus, r (X, Y) and r Oy) will
Ta
be equal, both in magnitude and sign. Alternatively, if ¢ and d have opposite signs 7; a | wil
be~1 and r (X, ¥) and r (U, V) will be equal in magnitude but will have opposite signs. 130 Oe ESOS RESEARCH wy
(©) Let the following information for two groups are available :
Group
Group ay
Group size ma -
Group mean (X) xX G
Group mean (Y) a y |
Group variance (X) on @,
—_ 2 }
Group variance (Y) oy 3, |
Correlation coefficient r (X,Y) n a |
|
Then for the combined group, we have
=_mXi +m X2 m Yi +m Yo
m tn cu m+n
3 _ m2, + mor, + m(Xi ~ X)° +X - X)*
oa nm +m,
mo? +moz, + m(Vi~ YF? + m(% ~ ¥)?
oy mm +12
MADy Oy + Min Opt m HH ~ HH - ¥) +m HG -H-H
nm +m
‘Substituting these values in (1) we can get the correlation coefficient for the combined sanpe
and Cov (X, ¥) =
Illustration 1. In an agricultural field experiment on the wheat crop, following dat wt
obtained :
Nitrogen content in soil (kg/ha) | 69.6 | 93.7 | 81.5 96.2 | 95.9 | 8.5 | 708 8
[Wheat yield (qvha) 12.2 | 19.6 | 16.9 | 27.8 | 19.2 | 17.0 | 154 [18
Calculate the product moment correlation coefficient and interpret the result.
Solution. To calculate r, we obtained for the following table :
Nitrogen Content
—~_
~ 69.6
93.7
81.5
96.2
95.9
82.5
70.8
73.8
=X? = $5982.68 | SY2= 2739.06
e& - 131
CORRELATION AND REGRESSION \o
| x: $0 $8 = 93, % = M4 a 18, ond 4,
| XY >
I) ZXY -nxX¥ NO
ny r®&Y)= Se .
Y (EX? ~ nk") (ey? — ny?)
% 1224.87 ~ 8 x 83 x 18,
: ~ ¥55982.68— 8383 x83) (2739.06 - 8 x 18x 18)
;
: oe - 0.8185,
n 870.68 X 147.06
r Alternative Method : By taking deviations of X and Y from arbitrary numbers 83 and 18
a respectively, we get the following table,
x y | exe
69.6 12.2 | 13.4
93.7 19.6 107
815 169 “1s
96.2 278 132
95.9 19.2 ao
82,5 170 ~05
708 x4 | a2
Y-H . 5.9 -9.2
h-i Bs} 1
0
30
. Sol
be formu
2UV-~nUV
Sr Q&Yy=
a zi
Thus,
X= 135,
Later on at
Wrong values
apy] x
Obtain the
lution,
ao
>
s We observe a posi
by, thatthe soil rich in
Mhustration 2,
Pairs of observatio:
To obtain comect value of FweTee
fr,
_ 292.87 - 0 Sota
(20? ~ nd?) (av? nv) VETS oy (147.060) ~
itive correlation of the order 0.8185 by
etween
nitrogen contents gives more wheat
the two variables, indicating
yield,
While calculating correlation coefficient between two variables X and Y from
ns, following results were obtained,
=X= 540, ZY
= 120, ZY? = 530 and ZXY = 575
the time of che
cking , following two mistakes were noted :
Correct values
v | x Ty
2
9 7
Correct value of, Correlation coefficient, ae
, We need to find the Corrected summations to be used i 3
132 STATISTICAL METHODS FOR RESEARGy 1
Corrected EX = 135-7-6+6+7= 135 Wore of
Corrected BY = 120 - 15-9 + 16 +8 = 120 gf
| -
Corrected EX? = 640 — 7? - 62 + 62 + 72 = 640
Corrected EY? = 530 — 157-92 + 162 + 82 = 544
Corrected = 575-15 X 7-6 X 9+6X 16+7 x 8=568
x. 135 120
X= 37245, Y=s>=40
. 568 - 30x 45x4 as
and OY)" [(640 — 30x 45 44 — 3042)” Vang ye 06139,
Illustration 3. Calculate coeffici
nt of correlation from the followin;
1g data on marks
students in two subjects. a
Marks in Mathematics (x): | 38 | 70 | 40 [30 [90 | 80 | © | w [aw
[Marks in Statistics (Y) : 65 | 40 | 30 | so | 60 | 70 | 50 | so] @ |
Solution. Computation of r changing origin as well as scale
x x60 | soys=u] w | ¥ | 7-50) (r-soysev |( ul
55 -5 +1 -1 | 65 15 3 9} 3]
70 10 2 4} 40 | -10 2 4] 4 n
40 20 4 16 | 30 -20 4 16 | i
30 -30 6 36 | 50 0 0 oj}!
90 30 6 36 | 60 10 2 4) 2
80 20 4 16 | 70 20 4 16 '
60 0 0 0 | so 0 0 wae
80 20 4 16 | 50 0 0 an ¥
90 30 6 36 | 60 10 2 Aya
80 20 4 16 | 70 20 4 : |
6
15 179 9 ot ;
forctv# ‘
As one can see that mean for X as well as Y will not be a whole number. So we £0" ,
of origin by some convenient numbers,
if)
; these
Let us change the origin of X by 60 and Y by 50. Also it is convenient to cance |
Xand Y by 5. Denote (X-60)/5 as U and (Y-50)/5 as V. With these new variallc® "Day
n=10, TU=15, ZV=9, DU2 =177, ZV2
both
=69, and EUV =65
_ Gy eV) | ey
zuy -
Rank corelaton coefficient, r= 1-22 _ 1 _ OOD - ois
Sn? -1) Woy (99)
ranks vt
“SRepeated Rank It may happen that two or more individuals are assigned equal ms
respect to characteristics A or B. In such a case, common ranks equal to the average of a 7
which these individuals would have assumed that had been slightly different from each of
given to repeated items. Other items get the ranks as usual.
While calculating rank correlation coefficient in such a case, the foll
should be added to 3a2. ‘
should be added to 34°.
(mm? — 1)
12
it
lowing correction
Correction factor
Iie series 80 the
sexo wade been differ
ie 25isbe eommon rank
Ate valve 6 occurs {
5607, Simin
Tht
on
{Clore ig repea
be
tai Mey
where m is the number of times an item is repeated. ion fa
i
Ir there are more than one such group of items with repeated ranks then the 6o7= (gg
rly gait
should be added as many times as the number of such groups. ios] Ea,
2 ii *
Mlustration 7, Following data relate tothe marks often students in mathemati Xs were a
Marks in Math. to | «| 7] s4[ | so[ 75] ” ray) yo, ued
Marks in Biology 62] 57) «| a7] si | oo| 68 4 ‘a * No >»
&
PEL X RS,
coRRELATION AND REGRESSION 2 1) 437
Obtain the rank correlation coefficient.
Solution. Here while ranking the students we face the problem of repeated ranks.
‘WGateulation of Rank Correlation coefficient
Mortsin Marks in Rank in Rank in d=Ry—- Ry @
a Biology Math, Biology
x ¥ Ry R,
go } 2 4 5 al 1
"644 57 l6 7 -1 1
953 68 \ 25 3.5 4 1
ae 47 9 10 -1 1
64 F 81 6 1 5 25
80! 60 1 6 5 25
25> 68 “25 35 -1 !
40 ¢ 45 10 9 1 1
55 54 8 8 0 0
~ 6 70 6 2 4 16
Zd=0 Be =72
In the X series 80 the maximum value is given rank 1. next to this value 75 occurs twice. Had
these two values been different, they would have got ranks 2 and 3. Thus, the average of 2 and 3,
Sis the common rank given to both these values. The value 68 gets the next rank which is 4,
Again the value 64 occurs thrice and is given the common rank 6 which is the average of three
ranks 5, 6 and 7. Similarly in Y-series value 68 is given the common rank 3.5.
Thus, for the two repeated values in the X-series and one repeated value in the Y-series. the
‘otal correction is
? 7-1), 2027-1) _ 36
Comection factor -20 =) 30-0, @ =)
2 D2?
‘ 6 (2d? +3) 6(72+3) _
“Rank correlation coefficient r = 1 — aa - 0x99 0.545
Limitations of the Correlation Coefficient. It may be noted that a special care must be
taken while interpreting the value of correlation coefficient. Following are few situations where
Conelation coefficient may give misleading results.
( Sometimes correlation between two factors may occur due to the common effect of a
"d factor, eg, the positive correlation between the per hectare yield of jute and rice may be due
‘ote fact that both are related to the amount of rainfall. sicawn a
(i) On ation between two variables when no one exists in
eT nye ds tal ape owing ec
Ane shoukd assure that there is some theoretical reasoning behind this interdependence of the tw:
Variables,
di) As'stated earlier the correlation cvefficient measures the degree of linear rtationship. So
“low value of correlation coefficient rules out the possibility of a linear relationship, though @ 138 STATISTICAL METHODS FoR RESEARG
linear relationship may exist between the two variables. As an example, consi WOR
below: * Consider the dae by
[x ] 3 -2 | -l 0 1 ; |
ni
Ly [1s 10 7 6 7 10 1
It can be seen that r(X, Y) = 0 though here the two variables are related as Y 6
=6+y0
So, before using r (X, Y) as a measure of relationship one should see (withthe h
agra) whether the general relationship is linear or not. el of cay
court ows
et tha Vouragle
GRESSION ANALY: 2
caer
Tithe correlation analysis we have discussed the degree of relationship without consi
which is the cause and which is the effect. For example-in-case of field experiments the fet
used is a cause variable while the yield is effect of that cause. So depending of the circuiting
one can decide which is the cause apd which is the effect variable. The cause variable
denoted by X is also called the independent variable and the effect variable denoted by Y scale
\. dependent-variable. In regression analysis we find an algebraic function of the fom
Thus, TEs
[we express the dependent variable as a function of the indepéndent variable}
analysis makes possible to estimiate or predict the unknown values of dependent variables forkara
values of independent variable. In this chapter we shall discuss the case where the relation}
between two variables is linear. Non-linear relationships shall be discussed in the next chapter.
Principle of Least Squares. [itt ‘two variables are related, the points in the scatter aga
will cluster around certain curve called the curve of, regression in regression analysis ura it
find this curve i.e. to find the equation of a curve that gives the best representation of the relations
between the two variables.
In figure 10.2 the points are clustered around a straight line. One can draw a number
straight lines through this scatter. (According to the principle of least squares, a CUN®. that mim
Aooies 2
Y
Fig. 10.2 A scatter diagram for pur of values (x, Y)- Burnsf (RY)
CORRELATION AND REGRESSION 139
ependent variable) gives.a i Fore
depet ebe the distance of it point (say A) of the scatter from the regression line, parallel
fg Then afine that minimises the quantity Sez will be called the least square line. (Here
raion ‘sextended over all the paired observations).
Lines of Regression. If the scatter diagram indicates @ linear relationship between two
jae, ibn the next step iso find the equation ofa straight line that gives a best fit to the data.)
‘ae tin, thus obtained is called the line of regression. Let us suppose, in the bivariate data
YI LZ Y bbe the dependent variable and X be the independent variable, Then, the line
Y=at bX
js called the line of regression of Y on X. In this equation we want to estimate the values of
vmsants a and B from the given data, which determine the position of the line completely. The
parameters and b determine the intercept on Y-axis for X=0 and slope of the line respectively.
Using the principle of least squares, the values of a and b can be’ calculated by solving simultaneously
the following two equations called the normal equations :
a+ b=X «(2)
EX +b TX? (3)
After some simplificat je get the following expressions for a and b.
_ nEXY - EX3Y
~ n=X? — (2X)* (4)
XK sos
ad a= 2X p - 7-0 (5)
From (5) we observe that the line of regression passes through the point (X, ¥)
Further we know
Cov &yy-L sexy - 2% 2%
n no
we aetee (BY 7
x on n
Cov (X,Y)
o%
Again, we have
_ Cov (X,Y)
re oxay
Multiplying both sides by oy /ox, we get
Or COVKYY) gy _ CVOGY) 5
Ox 2
xy Ox ot
ltrepes , above regression ‘ ee
x ents the increase in the value of Y-variable corressponding to a umit jncrease in the value of
ra
axe For the sake of convenience we denote p by by, Thus, the regression coefficient of Y
'S given by eas 140 STATISTICAL METHODS FOR RESEARCH Wo
= Cov Y) _ oy
ibe = a ox
Similarly, if we take X as the dependent variable and Y as the independent vari
regression line of X on Y will be
X=a'—by Y.
Expressions for a’ and by, can be got from the expression (4) to (6) by interchanging Xa
Y. Thus we get
KE
(8)
able then, te
= MEXY ~ EXEY _ Cov (X,Y) _ | o5,
ay 2 2 A o »
nZY? ~ (ZY) o 'y ~@)
and ad’ =X-by¥ (lt
“Here by is called the coefficient of regression of X on Y and it indicates the increase in the
value of X variable if the value of Y-variable increases by one unit.
Remarks, (i) It may be noted that we always have two lines of regression. Line of regressin
of Y on X is to be used when we want to estimate the value of Y for a given value of X, ie, whe
Y is dependent and X is independent variable. On the other hand, if we want to estimate the value
of X for a given value of Y, we use line of regression of X on Y. In the case of perfect correlation,
the two lines coincide.
Gi) From (5) and (10) we conclude that the two lines of regression intersect each other at the
point (X, ¥).
(iii) We have two different lines of regression because in the regression of Y on X, the sumo
squares of the difference between the observed points and the regression line parallel to the Y-ats
are minimised. On the other hand in case of line regression of X on Y squares of these differences
Parallel to the X-axis are minimised.
Properties of Regression Coefficients. Few important properties of regressiot
coefficients are given below :
© Correlation coefficient (r) and the two regression coefficients (by, and ign e
Same sign.
Proof : we have
r= SKY), _ Cove%yy Cov (X, Y)
Geo, 7h EY pg = a
. is te
2 these three expressions o, ad o, are always positive. So the signs ofr by and by ®
same as the sign of Cov (X, Y),
ii) Correlation coefficient is equal to the geometric mean of the two regression coefiiet®
Proof : Multiplying (8) and (9); we get
Obese
, eo, oa, [Bxy X Byx
The sign of correlation coefficient will be the same as that of By and Byx- i
(i If one of the regression than um
Coefficient is greater than unity, the other must be less CORRELATION AND REGRESSION 141
Proof, Let us suppose that by, is greater than unity.
1
‘Then, bry >1 or ys
Also Psi or bay bye S 1
1 1
by Sz by S$=—<1
o he or SB
(jv) Arithmetic mean of the two regression coefficients is greater than the correlation coefficient.
1
Proof : We want to prove that z (bx + bye) Br
o.
y ox ay
r2z|>r 4M 5y
or | or oto 2,
3 +05 -2oxoy20 or ? > 0,
or o% +05 ~2oroy= (ox - ay)? 20,
Which is always true, since the square of a real number is always positive.
(©) Regression coefficients are independent of change of origin but not of scale.
Proof : Let U = Xoa and v-%t where a, b, and h and k are constants,
Then, it can be proved that
Cov (X, Y) = h k Cov (U, V)
=n o2 and 0} =k of
Cov (X,Y) = CV GY) ky
— and by = 2 — how
oF of
Hence the result.
_ Illustration 1. Following table gives the length of green jute plant (in cm) and the weight of
Ay jute fbre (in gm) for ten jute plants.
Length of green Weight of dry Length of green Weight of dry
Plant jute fibre plant jute fibre
mM 1.20 118 2.10
125 215 140 3.05
135 2.70 150 4.10
165 $5.25 160 5.70
1m 6.05 185 7.28
fing Calculate the linear regression equation of weight of dry fibre on length of green plant. Also
the expected dry weight of a 120 cm long green plant. N
STATISTICAL METHODS FOR RESEARCH wy
142 OR
Solution.
Weight of dry os Ps
Length of Plant (Y)
green plant (X) +
1.20 os 1332)
i 92.
an 2.10 13524 24189
25 oe 268.35
te 19600
140 3.05 42109
135 2.70 18225 36459 |
os 4.10 22500 615.09
165 5.25 27225 866.5
160 5.70 25600 912.09
im 6.05 29241 133455
185, io 34225 1341.25
X = 1460 EY = 39.55 =X? = 218486 EXY = 621030
We are to find the regression equation
Yeatb,.X
Now, we have
~ NEXY -=XzY = 10 x 6210.30 — 1460 x 39.55 = 0.082
*NEX? — (Ex? 10 x 218486 — (1460)?
2 39.55 1460 _ |
and a=Y — by “qo. ~ 0.082 x S5- = - 7.997
«. The reqired regression line is
— 7.997 + 0.082 X
Subsistuting X = 120, we get Y = — 7,997 + 0.082 x 120 = 1.827
Thus, expected dry weight of fibre from a 120 ems long plant will be 1.827 gms
,__ Illustration 2. The following data Pertains to marks in two subjects, statistics and m™
ina certain examination 1
Marks in Physics Marks in Stas
Mean = 8
Standard deviation 10.5 13
Correlation betw:
een marks in
f
, _ - ikely >=
gets 60 marke y Physics and statistics is 0.53. Find the most likely
in Statistics.
Physics if a student
Solution, Let
4)
we
me «og by ¥.9""
Tequired 10 find the enact Ute MAES in physics by X and marks in statsies bY Ys
ll
ed y 0
Tegtession line of X on Yin Value of the variable X for a given value of Y. So ™
wy
(0
of
d
we ir
a
ol
so, the
substi?
which |
Must
ses it
(Fin
(i Fi
(ii) F
(wit
heart weigh
Solut
solve the tv
Malti
Subst
fy
Ua to th
Xiang y .
8o,
Bu
een
ith Lug
Bay
XY = 62109
CORRELATION AND REGRESSION
143
X=atbyY
by = r2£ = 0:53
We know ny = 1G = 0.53. 775 = 0.492
and a=X~ by .¥ = 70 — 0.492 x 78 = 31,587
So, the line of regression of X on Y is X =
Substituting Y = 60, we get
X = 31.587 + 0.492 x 60 = 61.10.
which are the required marks of the student getting 60 marks in statistics.
Illustration 3. For a study related to. the weight of kidneys (Y) and weight of heart (X) both
measured in gms the two lines of regression are
4X —10¥ + 1725 = 0 and SX - 6Y + 325 =0.
(i Find the mean of two variables.
(ii) Find the correlation between the two variables.
1.587 + 0.492 Y.
(i) Find the expected weight of heart for an individual whose kidney weight is 250 gm.
() Ifthe standard deviation for kidney weight is 87.5 gm, find the standard deviation for the
heart weight.
Solution. (/) We know that the two lines of regression intersect at the point (X, ¥) . So we
solve the two equations, simultaneously.
4X — 10¥ + 1725 =0 aa)
3K - 6 +325=0 of)
Multiplying (i) by 5 and (ii) by 4 substracting we get
Y= 281.73
Substituting in (i), we get
X= 273.08
i To calculate the correlation coefficient, we use the fact that the correlation coefficient is
‘cual tothe geometric mean of two regression coefficients. So, let (i) be the regression equation of
Xand Y and (ii) be the regression equation Y on X. Then, we can write the two equations as
325.5
and Y="e+gx
So, r=+ a
Vom Bx ih:
Them the correlation coefficient cannot be greater than 1. So our supposition is ons
¥. wet, (iis the regression equation of Y on X while (i) is the regression equation o}
“With this we have on STATISTICAL METHODS FOR RESEang on
H i
Regression of Y on X as Y= 172.5 + 04X, Oi) gt i, oe
and Regression of X on ¥ as X= -65.0 + 1.2¥ | Oye 8
p= & Jb On = + VI2X04 = + 0.69 | v4 rw
‘Again, as the signs of the two regression coefficients are positive the comea
will alsebe positive. Thus, coreelation ‘coefficient = 0.69. lation g 7
(ii) We are to find the expected value of X for Y = 250. So we substitute y = so
mn Y to get the required value of X. ing 3 eo
regression equation of X 0
‘ X =-65 + (1.2) (250) = 235 gm.
ox . oslo
(iv) We know that by = 7.5% on substituting the values, we get pe
y
1.2 = 0.69 x 2 or oy = 152.2 gm
35
Remarks : 1, For each of the observed values of X if we estimate the value of Yx
from the lng of regression of Y on X and then calculate the correlation coefficient betwee tg) be te8
estimated values and the actually observed values of Y, then this correlation coefficient wiles rege kam
as the correlation coefficient between X and Y ioe
3 also it ean be observed thatthe square ofthe correlation coefficient is equaltohenig| & Se aty
the variance of the estimated values of Y to the variance of observed values of Y. men
Variance of estimated values of Y
2 -
1” (% Y) = “Variance of observed values of Y ;
tegand mean sg
Thus, the square of correlation coefficient may be interpreted as the proportion of thet
variance that is accounted for by the regression of Y on X.
3. The main point of difference between correlation and regression analysis, is ttt
correlation tells us the degree of linear relationship where as the regression analysis gives 8) Coreton rato (
expected value of dependent variable given the value of independent variable.
10.4. CORRELATION RATIO
Ithas been emphasized in the above discussion that the techniques of correlation andregest
analysis are useful only when the relationship between two variables is linear. ‘Sometimes wee!
low value of correlation coefficient mearly because there exists some sort of curvilinear (0%
linear) relationship between two variables. In such a case, correlation ratio denoted bY ot
appropriate measure to study the degree of relationship between two variables. Just asris a ea,
of concentration of points about the straight line of best fit, 7 is a measure of the concen @
points about a curve of best fit. ates
Suppose the pairs of . wenn) ta
ao pairs of values X and Y be arranged in arrays of Y according ‘0 ( nig “lin
it
- (iy, eben
1 Yn Yn = Yu ay,
= Que, Dat
7 Yu Yo ~ Yor, CT
he, ea
ty
~ : Vlg to,
ke :
Yur Ye ~ ae “ash
Ms a
lo S 145
apéLATION AND REGRESSION
oF
we observe that out of the total » individuals n, have the same value of X, say X; while
et onding values of Y are Yit, Yon Yet
the
el, fn
ation ce) erve define the mean of ith array as
lue of,
ent between
cient will
qual tothe
.
ortion of tea
alysis, i
ysis gv
Yi = DYy/mi
?
and the grand mean of Y as
Y= SY Yy/n = Yn Yin
v7 i
Then the correlation ratio of Y on X denoted by myx is the positive square root of
dni Y - YP?
oe
SEC - YP
Ifthe values are grouped into a k x 1 bivariate frequency table, the classes of X may be
‘supposed to give k arrays of Y-values. Let fj be the frequency of jth Y-values in the ith array, fi. be
the total frequency of the ith array and f,, be the total frequency of the jth Y-values. Then ith array
‘mean is given by
Yi= Diy Yy/Sio
5
and the grand mean is given is
Yi= Dfoj Yj/n
t
Conlation ratio (7,,) will, then, be given as positive square root of
Dla (Y; - ¥)?
ee
Shay (¥; - YY
7
yf Perties of Correlation Ratio. Few important properties of correlation ratio are given
(the value of correlation ratio lies between 0 and 1, ie.,0 < n2<1.
( the value of correlation ratio is always greater than the correlation coefficient, If the
ston perfect linear than r= 9,
(ti ni
i) r,
©) the
Mtervals by
the val
's independent of change of origin and scale.
‘and rye are same but ry and rYy~ are, in general different from each other
Value of correlation ratio is not indep
come NATOWE My,
® Of 7. approaches r,
lustrati,
Ue stration. Calculate correla
“ton 5 of section 10.2.
endent of the classification of data, As the class
approaches unity. On the other hand, if the grouping is very coarse,
tion ratio from the bivariate frequency data presented in nae STATISTICAL METHODS FOR RESEARG, |
Wi
Solution. Taking V = CY - 225)/50, we have the following table :
y-25 | _, i|l2
v-*3 1| 0
y [175 | 225] 275 | 325 | Total | 2h vy
x Sio
tas 9 | 10/7 | 0 | 2 2 om
17s] 5 | 5 | 12] 23 9 cast | 2m
12s} 1 | 6} 20] 2] 29 2B 0703 |
1875 | 0 ieee, 12 18 1500 i
Total foj 15 | 22| 43 | 10 | 90 48 lawl
an
353 | 62] 94] 21s] 724 |
‘The grand mean of V is
C1) (15) + (0) (22) + (1) (43) + 2010)
90
jis given by
2 _ 2 fio (Vi- VW)? _ 23.317
1p, = tLe NO - Bah = 322
Sew 72400
So that, "yx = V0.322 = 0.567
10.5. INTRACLASS CORRELATION
Intraclass correlation means correlation with in a class. As compared to product moe
correlation coefficient, in this case both the variables measure the same characteristic. Frew
wwe may be interested in studying the correlation between the height of brothers or bei
‘weight of cubs from the same maize plant. In such cases there is nothing to distinguish ow
from the other, so that one may be treated as X and the other as Y.
present
Let us have n classes Cj, C2, Cy and ky, ky «ky member of each may be
Cy: Xu Xi ~ Xikt
1
where Xij (i=, ng m3 = 1, 2, ny k) denote the value of jth member in the ith class
intraclass correlation coefficient is given as
BR - HP - BBKy- 3?
- 24
LTD Ch = 1) Ky - X*
a
File intaclss correlatior
Solution, We have the gran:
YIX=68 +
Wy
ti Bx = CORRELATION AND REGRESSION war
ere %; is the mean of the ith class and X is the grand mean of all the classes.
w
nine above formula if we put ki~ k, i. ifall the classes have equal member, then intraclass
coneation coefficient i given as
1 (i -}
here denotes the variances of X and 02 denote the variance of means of classes.
. 1
‘The value of intraclass correlation lies between €yD and 1. Also, it is independent of
change of origin and scale.
Illustration. In five families of three brothers each, the heights of brothers are :
Family Height in inches
1 68, ny, 2
2 72, 70, n
3 72, 69, 2
4 72, 22, 70
5 74, 22, B
Find the intraclass correlation coefficient.
Solution. We have the grand sum and sum of squares as,
DD X = 68 + 71 + 72 + 72 +..474 + 72 + 73 = 1071
id
and DD x? = (68)? + (71)? + (72)? ++ 72)” + (73)? = 76497
77
2
2 _ 16497 _ (1071 _
=> =ise 1.840
Again, the means of the heights of brothers of five families are
70.33, 71.33, 71.00, 71.33, 73.00
and the grand mean is
1071
xeap a4
(7033 — 71.4)? + (71.33 - 71.4)? + a, nay?
o= +(71.33 - 71.4)? + (73 - 71.47 =0.775
" 3
x 0.775
1.840
0.132
-enf2)-|
is requi, .
‘ited intraclass correlation coefficient.
jl 14
&
STATISTICAL METHODS FOR RESEARCH | WOE
PER
EXERCISES
What do you understand by the term correlation ? Discuss the various
measure it methods ty
‘Show that the correlation coefficient is independent of a change of origin and
the variable. State the limits between which r lies and give its proof. Scale
Comment on the following statements :
(@) Anegative correlation indicates that if the value of one variable decreases
value of the other also decreases. the
(b) Ifthe correlation coefficient between X and Vis positive then the correlation coef
between (i) - X and -Y is positive; (i) -X and Y is positive.
Explain the difference between product moment and rank correlation coefficients,
Write short notes on the following :
(a)Lines of regression _—_(b) Least squares principle
(c) Regression coefficient (4) Correlation ratio. (0) Intraclass
correlation.
Why do we have two lines of regression ? Prove that the correlation coefficients he
geometric mean of the two regression coefficients.
From the following data calculate the coefficient of correlation between age and boty
Weight of white leghorn chicks. Also estimate the weights of an 11 weeks old chick
‘Age in weeks : 1 2 3 4 5
Body wt. in gm: 60 100 160 240 320
Age in weeks : 6 7 8 9 10
Body wt in gm : 410 480 580 670 750
Ina field experiment on different varieties of raya crop, the following data are obtains
Variety No. + 2 3 4 5 6 7 8 9 1
Maturitydays 164 178 167 163 158 169 185 157 193 180
SeedYield 48 64 48 41 26 48 54 32 59 54
per plant (gm)
(a) Drawa scatter diagram.
jeen the maturiy pero!
(b) Calculate the product moment correlation coefficient betw
and yield. Comment on the results. aux!
The following data gives the number of blind persons per lakh of population inditfer
age groups. Find Pearsons correlation coefficient between age and blindness *
Age in yrs No. of blinds Age in yrs
perlakh
0-10 50 40-50
10-20 68 50-60
20-30 102 60-70
30-40 109 70-80 oRRELATION AND REGRESSION
149
40. The following table gives the age and daily wages of 50 workers ina factory :
Daily pay in Rs. (x)
Age(Y) 160-169 170-179 180-189 190-199 200-209
20-30 5 3 1 = =
30-40 2 6 2 1 _
40-50 1 2 4 2 5
50-60 - 1 3 6 2
60-70 = = 1 1 5
correlation ratio (1x)
11, Following are the ranks obtained by 12 students in mathematics and statistics. To what
extent the knowledge of students in two subjects is related.
Compute the correlation coefficient between age and daily wages. Also compute
Mathematics | 1
2
3 4 5
6|7
8 9
to [11] 12
Statistrics 1
6
3) 4/7
5 | 2
10| 8
[9 [2
12. Inamuscial competition ten participants were ranked by three judges A, B and C in the
following order :
RanksbyjudgeA ] 3] 5] 8] 4] 7] t0]2]1]6]o9
RanksbyjudgeB | 6 | 4/9 ]8)] 1 ]2{3] [5 ]7
RanksbyjudgeC | 1 | 6] 5] 0] 3| 2/4] 9/71] 8
Use the rank correlation coefficients to see which pair of judges
to. common liking in music.
13. The following table gives the heights in inches in 10 fathers and their eldest sons.
has the nearest approach
Calculate coefficient of rank correlation.
Ht. of father 63 | 67 | 64 | 68] 62] 66 | 68 | 67 | 69 | 71
Ht. of son ee | 68 | 65 | 69 | 66 | 65 | 68 | 69] 71 | 70
14. Fromthe following data
calculate the rank correlati
on height (in inches) and weight (in pounds) of eight individuals,
ion coefficient and the product moment correlation coefficient
Weight : a 7 90] 99] 108] 117 | 126 | 135 | 144
Height = oo | se} a] o | 8 | o | 70 | 7
Why the two coefficients are equal ?
15. Arandom sample of ten families had the following total expen
ture and food expenditure
in Rs. per day per person.
Total Exp. 766 20.1 262 30.4 22.5
|_Food Exp. 123 16.0 206 245 172
Total exp. 492 66.95 52.81 356 985
Food Exp. 364 2 a74 282 61.62 150
STATISTICAL METHODS FOR RESEARg
Ht
Estimate the regression line of food expenditure on total expenditure and
16. ne numbers of food (X) and clothing (Y) at eight Succes;
ive 1996 1997 1998 1999 2000 ang, ee
x 100 124 123 132 136 139 49 ay
Y 100 104120 192 143g Pg
(i) Fitthe lines of regression of X on Y and Y on X and hence
coefficient.
(i) Suggest what the value of Y will be when X is expected to be 150,
17. From the following data, estimate the most likely price at Calcutta Cotrespon,
price Rs. 68 at Bombay and that at Bombay corresponding to price Rs. 70; my
ra,
IVE Years ar
Calculate thee”
Calcutta (X) Boma
Average price in Rs. 7 6
Standard deviation of price 15 05
AX, Y) = 0.68
18. In correlation analysis the following two regression equations were Obtained:
12X+3Y=19, 3X+9Y=46
Obtain: () the value of correlation coefficient,
(i) mean values of X and Y, and
(ii) the ratio of coefficient of variability of X to that of Y.
19. (a) Ina regression analysis the following two regression coefficients were obiarl
by=3.5 and by,=0.5,
Comment on these values.
(b) Given the following two regression equations calculate the correlation cose
and comment on the results,
Regression of Y on X:
Regression of X on Y : X = 0.5 + 0.25Y.
20. Calculate the correlation coefficient between the following series {
Year 1991 1992 1988 1994 1905 1906 1967 108 1
Deaths fromcancer 612 583 671 692 689 635 601 600
Production of 66.6 84.9 886 780 968 1052 932 116 9!
steel (in millions)
What can you infer from the value of r? |
21. For five plants each containing five ears, the numberof grains per ea"a'®9
Calculate the intraclass Correlation coefficient.
Plant No. of Ears 5
1 2 3 7 ‘
1 Bs Z 3 2 4
2 33 34 7 » fl
3 58 a 50 51 #
4 4 7 8 2 7
5 35 44 7 “
“y
Oo
Oo
o
Oo
o
a
ii. INT
In the pre
volving t¥0 V
ie, data having
ne variable ma
‘qattyofseed,
canbe studied MULTIVARIATE
CORRELATION AND
REGRESSION
11.1. INTRODUCTION
In the preceding chapter, we were concerned with the correlation and regression studies
involving two variables only. But many a times we are required to work with a multivariate data,
ie, data having simultaneous information on more than two characteristics. In such a case value of
one variable may be influenced by many others, e.g., yield per acre of a crop depends upon the
quantity of seed, fertility of soil, fertilizer used, irrigation facilities and so on. This type of association
can be studied with the help of multiple correlation and multiple regression techniques. These
techniques help in investigating the extent of relationship and effect of a group of variables upon a
variable not included in the group.
Suppose in case of multivariate data we are interested in a relationship between two variables
only. Two altemative methods to study this relationship are : (i) consider only those data in which
allthe variables, other than the two under study, have a constant value or (ii) eliminate mathematically
the effect of other variables on the two variables of interest. The first method is rarely used since
the data in the required form are usually not available and if available, its size may be too small. In the
second method it is not possible to eliminate the entire influence of the other variables, but their
linear effect can be easily eliminated and these are called the partial correlation and partial regression
‘echniques, which are the subject matter of the present chapter.
discal” 13 section, some commonly used non linear regression equations have also been
scussed,
11.2, MULTIPLE REGRESSION
.__ Suppose we have a multivariate data involving p variables X), X2,..., Xp and we are interested
Onlvsing the effect of p — 1 independent variables X2, X3,-...., Xp on the dependent variable X;
ta bijective here is to build up a mathematical relationship of a form X; = f (Xz, X3,... Xp) with
‘dea of using it for predicting the value of variable X from a knowledge of the value of variables
en X,,. For example, we may be interested in estimating the yield of a crop in a year from
nena of rainfall, average temperature and average humidity, etc., during the period from
variate to harvesting of the crop. The common sense suggests that as more and more independent
“s are included in the regression, the prediction becomes better. 152 STATISTICAL METHODS FOR RESEARCH yy
1. . OR,
In the most general and simple case, the relationship between X and X3, x,y y
‘pis Ug |
|
to be linear one given by an equation of the form
X1 =a + by Xp t b3 Xs +. + by Xp
|
Sty |
i i Itiple regression equation of X; on Xp, X3,..,,
bs pace eased es cl reresion coefficients. These constants ai ie dam |
p-variate data consisting ofp values, corresponding to the p variables, for n (> p) distinct ing oe
Pit these values for the i-th individual be denoted by X1i Xap »» Xpi Then, the least du
catimates of the p constants 4, bp, b3,.--,B pcan be obtained by solving simultaneously the fog
pnormal equations. _
BX, = a + by UX py + 3 EX3i tt bp EXpi
DX) Xj = a EXyz + by EXof + bs EX2i Xai +--+ Bp EX2j Xpi
BX Xoj = a EXzi + bz EXri Xai + 43 X25, +... bp EX3; Xpi
Q)
EX, Xp= a EXpi + by EXaj Xpi + bs EXsy Xpi + + bp EXHi
Ths summations extend over i from | to 7. |
Aliter. Here we give an alternative method to determine the multiple regression equation (I)
case of two independent variables. However, the results can be extended for the regression equation.
having more than two independent variables.
We write the regression equation of X; on X2 and X3 as
X= 4 + by23 Xo + b13.2 Xs Q)
Here b12,3 is the partial regression coefficient of X on X when the linear effect of X;0nX,
and X; is eliminated. Simialrly one can define b;3,2. In order to determine the values of these patil
regression coefficents, we define the following terms :
n3
3
(@) Correlation matrix of X,, Xp and X3 :
ni 2 13 1 on
R=In m2 m3]=]ni2 1
3 12 [os ry = Land y=)
M1 132 133 ns 3 1 |
(b) Co-factors of r11, r12, r13 elements in the correlation matrix R :
%
Co-factor of ry = Ry = (-1)!+4 lei |
— 3
ny 1
n2 13
Corfactor of rip = Ry = C1)? =~ (ia — nisms)
ag!
fl
Co-factor of ‘, ® !
13 = Ry3 = (ays =
13 1) 73723 —N3
13 3 Say O
7 x, ia ‘a ul TWARIATE CORRELATION AND REGRESSION 153
{@ Also we denote by Xi, X2, X3, 01, 0 and o3 the means and standard deviations of X,,
Ey] ua one .
seri, ‘hen, br23 and br3a can be obtained from the following formulae:
isting i r
n, ing, 427323.
» the “ 4)
ugly sn a
“hay yg =— 2b Ra = 21 ta ts
32° sR 92 1h «5)
wd a= Xi ~ bias Xp ~ by32 Xs- 6)
Remakrs. (i) The partial regression coefficient by 3 determines the increase in the value of X
shen the value of Xs increased by one unit and the linear effect of Xs on X, and X; is eliminated.
(ip Interms of the notations of equation (3), by, 3 will be the partial regression coefficient of
yon X; when the linear effect or X3 on Xz and X, is eliminated and it will be different from bya.
(ii) It may be noted that in both the methods discussed above, we get the same values of the
constants of the regression equation, The same point has been verified for a particular case in the
z fallowingilustration,
s10N | tion Ij . *,
repeamadl Ilustration. From a field experiment the following data related to yield of wheat (X;),
sirogen contents in the soil at the time of sowing (Xq) and nitrogen applied (X;), are recorded.
Yield (qil/ha.) in the soil Nitrogen contents Nitrogen applied
4 Gs/ua) ig/ha)
r effect of Xs 162 69.6 0
ues of these 315 69.6 40
30.6 69.6 60
394 69.6 80
129 815 0
25.0 81.5 40
31.9 81.5 60
“|
yr! and 375 81.5 80
189 95.9 0
36.1 95.9 40
38.0 95.9 60
40.3 95.9 80
Find the regression of X, on Xz and X; and interpret the regression coefficients.
value elation, It is required to find the regression equation Xj = @ + by Xz + bs Xs. For this,
f various sums and sum of products, are calculated in table 11.1.
ex ing the values from table 11.1 in the normal equations (2), we get the following three
358.3 = 12a + 988b + 5406s i)
> -
156 STATISTICAL METHODS FoR
rB-rh- At na n3 5
Also -r3
The, substituting these expressions, we get
24,2
Ry oq = (12 23 = 2ririars
123 = >
%
Properties of Multiple Corelation Coefficient.
of multiple correlation coefficient.
@ Suppose, for all the given values of X», X5,., Xp We estimate the value
ES
oo,
We listbelow certain mponay
. . a t
1ultiple regression equation (1) and denote it by X1.25, .p. Then its product mones
coefficient with X; is equal to the multiple corelation coeficient. So we can say har
correlation coefficient measures the closeness of the association between the obserefa
expected values of the dependent variable obtained from the multiple regression equaton
(ii) Another way to interpret the multiple correlation coefficient is that its sua
Percentage of variation in the dependent variable explained by the independent variables tim
fitted linear regression equation.
(iii) While calculating the value of multiple correlation coefficient from (7) we take tie
square root of the expression on the right hand side. So the value of Ry 25). is take ai
Again as said above it is the simple correlation coefficient between X; and X95, is
will be less than 1. Therefore we conclude that 0 < Ry 93... 1.
(i) Ry23,-4 = 1 means that the association between the dependent and the inde
variables is perfect. In this case the observed and expected values will be equal, ie. Xi=Xu
and thus the multiple regression equation is said to be a perfect perdiction formula
() Ri 23... = 0 means that X; is completely uncorrelated with all the other variables
mliple regression equation fails to throw any light on the value of X; when Xp, Xs
own,
(i Ris, is always greater than all the simple correlation coefficients of the de
variable with the individual independent variables, i.e.,
Ri23e—p = rie M1,
. Tip: a
Ilustration, For a data related to three variables, Xy Xp and X; following conesion
were observed.
; 712 = 0.80, 743 =~ 0.56 and 795 =
Find Ry 25 and Rs 19 and interpret the results,
Solution, Ris = [2+ 7 ~ 2nnans
1-r
23,
(080)? + (0.56)? — 2 x (0.80) (-0.56) (-0.40)
(080)? + 0.56)? — 2 x (0.80) 0.56) -0.40)
0.40)
ni3 + 1B — Qnon
Rap= as = tis = 9 56g
1-1
0.40.
ein afer the linea ¢
eee Xy and Xp, dem
Ar some algebra yLTWARIATE CORRELATION AND REGRESSION
Ml 157
i 2s
interpretation. R?2; = (0.842)? = 0.709 means that 70 per cent of the total variation of
variable X; can be explained with the help of the multiple i
oe nly we can interpret the value of Rp iple regression of Xj on Xz and Xs.
{14 PARTIAL CORRELATION COEFFICIENT
‘Sometimes two variables X, and Xz may be correlated partly due to the fact that one of them
is dependent on the other and partly due to the correlation of a third variable, X3 with both X, and
Xp, Insuch a situation, one may wish to analyse the correlation between X; and X; after eliminating
the effect of X3. This type of correlation is called the partial correlation and the coefficient of
comelation between X, and X; obtained after eliminating the linear effect of X;, is called the partial
comelation coefficient.
‘The residual e.3 = Xj — a ~ by3 X3, may be regarded as the part of the variable X; obtained
after eliminating the linear effect of X3. Similarly the residual e 3 gives the part of variable X» that
remains after the linear effect of X3 has been eliminated. Thus the partial correlation coefficient
between X; and Xz, denotes by riz, is given by ri23 =r (e153, €23)
Cov (¢1,3, €23)
VV (413) V (23)
‘After some algebraic manipyulations we get the folloiwng expression
riz = "323
Ri
oye Rae = tas
pe VRuRa~ Ja-)d-B) 8)
Remarks : (i) As in case of simple correlation coefficient, partial correlation coefficient r2.3
is equal to the geometric mean of the partial regression coefficient of X, on Xp (b,23) and the
partial regression coefficient of X_ on X1 (bp1.3)-
Thus we have n23 = £ bi23 X bas
(ii) The expressions for 713.2 and r23,1 obtained similar to (8), are given as :
na Ris n3 = N23
3- Ra = ms
PS e RnR a7) - A)
73 = Mi Tig
Ra3
and 1 ty
VRnRss fa 13) (0-73)
(ii) Ifry2 = ry3 X rpg, then we have 712.3 = 0. Thus, one may ‘observe a correlation between
X, and X; which is due tro the effect of another variable X; but after eliminmating the effect of Xs
the two vaiables may be observed as uncorrelated.
__ (i) Partial correlation coefficient being the simple correlation coefficient between the two
‘esiduals, it lies between —1 and 1.
(0) Partial cortelation coefficient isa useful too! to decide whether a variable should be included
‘nthe regression analysis or not.
Ilustration, For a trivariate data following results are known.
%=42, 07=53, 03> 6.1 168 STATISTICAL METHODS FOR Reg,
ARCH
rr
r= 0.80, 2370.7, 13 = 06 ey
Find (i) rin3 (i) 723. (##) Ras and () bis bisa
Solution.
_ fin = 73 3 0.8 - 0.7) 0.6)
0 ast JEM - cose 7
Ja-Ba-m) — ¥0-049)0-036) 68s |
ngs = 9.458 | |
° |
@) RaI>
Jo-ma-m) |
1
(ip Roe (RE _ [EA FOH HOH OD OH |
1- 0.36 = 0.846, |
1-13
:
1 ng—n3ns 02
2° 1-r
eee
33°
pe second
a1 ng—nams _ 0 find
D3 N2BS = 0,54, | Wee y= bo”
and baa = SE -
O° =
| qaeqonding 0 norma
11.5. ESTIMATION OF NON-LINEAR RELATIONSHIP | sy = nb
Cur discussion of the correlation and regression analysis in tis es well is ree ee a
chapter is based on the assumption that the relationship between the dependent and theindepaia | a
1, one can well expectanotina | Siting the values of
variables is a linear one. But, given the complexity of the real worl
variables. The most common fms) we
Telationship between the dependent and the independent
vronlinear relationships can be expressed in the form of (®) polynomial equations ori) =) 937 <8
‘curves. These two forms are discussed below. oes
1. Polynomial relations. Consider the relationship betweet yield (¥) and fern | oe =
certain iveloft| =
‘Solving these equations, \
(0). As we increase the quantity of fertilizer the yield go on increasing, at
the yield attains a maximum value and with a further increase 19 the quantity of fet
decreases. Such a relationship can be adequately represented by a second degree polyno
of the form
Y = by + bX + bX?
In general, a polynomial of degree p can be written as
Y= by + 1X + XP + ack BpX?
0 (ad Py water
___Tocstimate this polynomial equation we set x
with the fitting on the linear multiple regression equation.
Y= by + by Xi + BaXa touchy aot]
Illustration. Fit a second degree polynomial curve to the following data ™
type chicks. :
‘Age in weeks (X) : 1 2 | 3 4 5 me
Gain in wt. ingm.(¥): | 62 [ 125 [ 150 | 150 130 | 120 P
11 as in the pre
and the indepat
11 expecta nora
st common fis)
ations or (2
RIATE CORRELATION AND REGRESSION
wi 159
olution. We need the type of table given below :
x x? x3 xé YX x?
1 1 1 1 62 62
2 4 8 16 250 500
3 27 81 450 1350
4 16 64 256 600 2400
130 5 25 125 625 650 3250
120 6 36 216 1296 720 4320
110 7 49 343 2041 70 5390
0 8 64 512 4096 720 $760
937 36 204 1296 8772 4222 25032
We want to find the second degree curve
Y = bo + DIX + DX?
Corresponding to the normal equations (2), we get the following three equations :
BY = ny + BiEX + EX,
EXY = boEX + BEX? + HEX?
ad EX2Y = boEX? + BEX! + DyEX4.
Substituting the values of sums, sum of squares and products from the table given above,
we get
937 = Bby + 36b; + 204,
4222 = 36by + 204bj + 1296,
23032 = 204by + 1296by + 8772b,
Solving these equations, we get
by = 35.206, by = 48.935, and by = ~ 5.423
Therefore, the required equation is
, Y = 35.206 + 48.935X — 5.423X?. ee
. i ations where with the incre
sche cures any snes ve ie tio i ne
cumple, a the income of an individual increases the demand forthe items of necessities #10
increases but at a decreasing rate and after a certain income level there is ni further increase in
with increase in income. Such relations can be adequately represented with the help 0!
Sowth function of the form
4
Y = by Xi"X2? Xp ; -
For the estimation of this function, if we take logarithms on ee sides, we
Tog, ¥ = log, by by loge Xi + Ba Hoke Xa tot Bp BX
Which 19 + by loge Ai * 92 . yew to the
chs tinear in log Y and log. X's. Thus by transforming the varabies eee be mar dr
Vaiabvies ¥# = Jog YX} = log Xtvus Xp" = 108 Xp» We can find the v8 »
troy .
7m the linear regression equation.
> STATISTICAL METHODS FY
160 i OR RESEAR GA,
Yt = bf + by Xi + bp Xft... by x} OR,
Remarks.The value of bi (i= 1, 2,-p) gives the percentage increase in y |
one percent increase in X;. Pini |
Illustration. The following table shows the demand (Y) for a commodity ang in
measured in arbitrary units. Pree
re 343] 580] 618 | 695 | 724] 812 | 887 | on 1 Tap
Zz: oi | 4 | so | 43 | 38 | 36 | 28 | 23 |) ~
Estimate the demand function Y = boX°.
Solution. We take logarithum of the variables Y and X.
Y x Y* = log.-¥ X= logy
543 61 6.2971 4.1109
580 54 6.3631 3.9890
618 50 6.4265 3.9129
695 43 6.5439 3.1612
724 38 6.5848 3.6316
812 36 6.6995 3.5835
887 28 6.7879 332 |
991 23 6.8987 3.1385
1186 19 7.0685 29445 |
1940 10 7.5705 23026 |
The readers can verify that the regression equation of Y* on X* will be given as Y*=9.0
0.69X*
“Y= Antilog (9.121) X-0-69 = 9145.4 X-0.69
The value of regression coefficient b =
of the commodity will result in a 0.69 per cel
3. Exponential Relations.
— 0.69 indicates that a one per cent increase ‘npn
nt decrease in the demand of that commodity
' n
: Non linear least squares regression is more general cs
ne, Mi deconrsineat curves can be put into a linear form by appropriate transforms
er the dependent variable , i snfineat
is of the form, Y or some (or all) of the independent variables. An
Y=fx;B)+e
: Exponential Functions.
size, in the spread of diseases,
‘pes of decline typified by radi
describe continuous growth or
y= aebt
Exponential functions are used to model changes in PPA
and in the growth of investments. They can also accurstel Pi,
active decay. Exponential functions with the base ¢ are of"
decay. Consider the nonlinear regression problem (ignoring
> taking a logarithm transformations of both sides, it becomes
In) = Ina) + by
The exponenti ‘ ji
: ial i : 0
increasingss an ino tnetion Sse more generally for functions when a series #7
State such that percent difference for observation to observation
g
we 61
rE CORRELATION AND REGRESSION 1
aT
on + form of the model is
y= ab, | 7
or growth, model when a > 0 and b> 1 and for decay when a > 0 and 0
