A C T A U N I V E R S I T A T I S L O D Z I E N S I S
FO L IA O EC O N O M IC A 141, 1997
Wiesław Wagner*
U SEFU L LN ESS O F D U R B IN M E T H O D F O R IN V E ST IG A T IN G N O R M A L IT Y IN O N E -D IM E N SIO N A L L IN E A R M O D E L
Abstract. In this paper the verification o f hypotheses o f univariate norm ality by D urbin random ized m ethod is presented.
The m ethod o f elimination o f the nuisance param eters by calculating the residual vector and connected residual vector is presented, too.
Key words: linear model, D urbin random ized m ethod, tests for norm ality o f residuals.
1. IN T R O D U C T IO N
Investigating the norm ality of one-dim ensional simple sample has been done by m any statisticians. Vast literature on the subject is given by M a r d i a (1980). M ost frequently the problem m ay be reduced to constructing tests statistics which are functions o f sample estim ators o f unknow n expectation ц and variance а г, w hich in the problem o f testing for norm ality are nuisance param eters. We eliminate them by transform ing observable random variables. One of the ways is given by D u r b i n (1961). We will extend it to the case of investigating the norm al distribution of random errors in a linear model, which perform ed on the transform ed residuals obtained from the LSM in order to get a simple sample. F o r such samples which comprise independent random variables with identical distributions we apply norm ality tests.
In the paper the D urbin m ethod for investigating norm ality o f random errors in a linear m odel is given. Its characteristic is supplem ented with assisting results.
2. LIN EA R M O D E L A N D R ESID U A LS
Let y\ nxl be a vector of n observable random variables given by a linear model
y = X ß + e, (2.1)
where: X: nxq is a known m atrix of order r (X ) = m < q < n, ß : qx 1 - a vector of unknow n constant param eters,
while e: лх1 is a vector o f unobserved random variables called random errors. The linear model is expressed by a triple (y, X ß , a 2l ) , where a 2 > 0 is unknow n and I: n x n is the identity m atrix. This m eans that E(e) = 0 and D(e) — a 2I.
The vector o f residuals from the LSM is expressed by r = фу, where Ф — I — ( X ( X ' X ) ~ X ' and ( X ' X y is the g-converse m atrix o f X' X.
F o r the vector of residuals r we have properties which proofs can be found in m any m onographies concerning linear models: r 'l = 0, r ' X = 0, r'r = у'фу = е'фе, E (r) = 0, D (r) = а 2 ф, E(rr') = Cov(y, r) = Cov(e, r) = а 2ф and s2 = r ' r / ( n - m ) is the BLUE estim ator o f a 2. The property D(r) = а 2ф states that the com ponents o f vector r are correlated and (usually) have different variances. F o r presenting the issue o f investigating norm ality of the distribution of vector e we give some assisting results which are used in the applications of the D urbin m ethod.
3. A SSISSTING R ESU LTS
Let X t , . . . , X n be a simple random sample, being a sequence o f n in-dependent values of random variable X which is assumed to have distribution N (fa a 2). By X and S 2 we denote the sample mean and unbiased estim ator o f variance from this sample, for which some properties hold: X ~ N ( p , a 2In), (n — 1)S2 ~ a 2x l - x,_ Cov(X, S 2) = 0, X and an a rb itra ry function g ( X v — X,..., X n — X ) are independent (n —1)S2 and are independent and ( X t — X)/(n — 1)S are independent.
Let us define standardized variables Ut = ( X t - X ) / S from sam ple X t , . . . , Xn. The variables t/„ i = 1...n, are not independent but we have:
C o v ( X !, Ui) — 0, Cov(X, [/,) = 0, Cov(S2, Ut) = 0.
The density function f ( u , ) = f ( u ) for random variable Ui is ( P e a r s o n , S e k a r 1936, C r a m e r 1958, p. 273).
/ ( u ) _ _ ^ . p . , ) ( n - l ) V n 9 n - z|_ ( " - 1) J
for — (n — 1)1 J n 4: и < (n — l)/\Jn, where g„ = Г(п/2).
Density (3.1) is a p articular case o f random variable U w ith the generalized beta distribution
f ( u ) = 1 + t _ ! • -Г О + q) (u - a)p Ч Ь - и ) 9 l , a < и < b, (b — a)p+q~ l Г(р)Г(Я)
when and a = — (n — 1 )/\Jn, b = ( n — 1 )Д /п and p = q — (n — 2)/2
Density (3.1) is a symmetric function, so all m om ents ab o u t the origin o f odd degree are equal 0, while the ones of even degrees are given by the form ula
E( Uk) = . (3.2)
V i! nkl2 9n + k - l
M om ents about the origin o f degree к o f standard deviation S are given by (e.g. P a w ł o w s k i 1976, p. 46)
2*/2 a
E(Sk) = k = 0’ 1’ 2> " (3 3>
4 . D U R B IN M E T H O D FO R A SIM PLE SA M PLE
Let us suppose that we verify a composite hypothesis H C th at X u ..., X„ is a simple sample from the distribution N(p, a 2), with unknow n p and a 2. Let X and S 2 be the sample m ean and unbised variance estim ator. M oreover, let us denote by ? and S'2, the sample m ean and variance from the population with the standardized norm al distribution N ( 0, 1). Thus, we have ? ~ N ( 0 , 1/n), ( n — l ) S ' 2 ~ X n ~ i and Y and S'2 are independent. W hat’s m ore, after replacing a by 1 and к by 2 we get E(S'2) = 1.
As we have already m entioned in the first paragraph the param eters p. and a 2 in the problem of investigating norm ality are nuisance param eters. D u r b i n (1961) suggests a random ization process to eliminate them. The idea of it is to consider two further random variables Y and S'2 which have the distributions m entioned earlier. According to this form ula, we determ ine such a random sample Yy,..., Yn that
We will show that sequence Y„ is a simple sample from the population with the distribution N( 0, 1). The relation (4.1) we write as
where Ui is given in point 3 and Y and S' are random variables generated independently o f X l t ...t X n.
Lemma 4.1. Random variables X , S 2, Ub ?, S'2' are pairwise independent. Proof. The independence o f X , S 2, Ut follows from the results given in point 3. O ther independences follow from assum ption that X ' and S'2 are independent o f ATl v .., X n.
Lemma 4.2. E ( ? + S ' U ,.) = 0 and D2( ? + S ' U i) = 1 Proof. From Lemma 4.1 we get
E ( ? + S' Ut) = E ( Y ) + E( S' UJ = E ( Y ) + E i S' UJ = В Д В Д ) = 0
which follows from disappearing o f the m om ents of odd orders of va-riable Ut. F rom the fact, that C o v ( Y + S ' U ^ = 0, we get for the variance
where we used E(S'2) = 1.
We have shown that the first two m om ents o f the left side expression Y, = Y + S' Ui are identical with those for the variable with distribution N( 0, 1). Now, we will give a lemma in which we will prove that variable S ’U is norm ally distributed and the first one has the chi distribution and the second has the symmetric beta distribution.
Lemma 4.3. R andom variable Z = S’U has the norm al distribution N ( 0, (n-l)/ri).
Proof. We use the result given by F i s z (1967, p . 71/ If S ’ and U are independent random variables w ith densities f ^ s ' ) and f 2(u) then the distribution o f Z = S' U is given by the density
Yt = ? + S V b i
D \ ? + S'U;) = D 2( ? ) + D 2(SU) = - + E(S'2Uf) - [E(S')E(Ui)]2 =
---j---n n
* п- 1),29п-1 2 ( n - l ) J n g n- 2
Using the density f t (sf) given, am ong others, by P a w ł o w s k i (1976, p. 45) and density (3.1) we get
m
= cic>ji.*->«p{—^ i } [ i
J " 4” * ’ =
We change the variable s' by consecutive substitutions t = (n— l ) s '2/2, v = 2(n — l ) t — nz2, w = /2(n — 1). Then we get
/ ( г ) » ' « ^ х р Г
-Com ing back to the constants C t and C2, from the definition o f the gam m a function, we get
f(z ) = 2(n — 1)("~ 1)/2 yfk g„~i 2("~4)/2(n — 1)<"~4)/2 2 (n - 1)/2^„_! ( n - 1 ) g„-2 ( n - 1 ) " -3
[ nz2 } 1 I n Г nz2
^ " - 2 ex p l 2(n 1) J ^/2ГГ\/ ЙГ-ТeXP j2 (n ^ T )
Finally, we get the density function of the distribution N ( 0, ( n - l ) / n ) i.e. random variable Z has the norm al distribution.
5. DURBIN METHOD FOR A LINEAR MODEL
In point 2 we gave the vector o f residuals from LSM for m odel (2.1). We are transform ing it further, using D u rb in ’s random ization procedure. Its use is connected with the elimination of the nuisance param eter a 2 on which the covariance m atrix o f the vector o f residuals depends. It requires the use o f a transform ation which expresses the quotient of two random variables with the chi-square distribution. On the other hand, the transfor-m ation of the vector o f residuals should give such randotransfor-m variables which are uncorrelated. Both problem s can be solved in two ways.
The idea o f the first one is to use some random variables, exactly as m any as there are unknow n nuisance param eters, in the problem s of
investigating norm ality. The problem was developed in th at direction by S a r k a d i (1960, 1967), S t o r m e r (1964), T h e i l (1968) and S a l l y and S a r k a d i (1982).
The other way considered is to extend the set o f random variables by the num ber o f them equal exactly to the num ber o f nuisance param eters. T hese additional variables are treated as generated random variables, independent o f the observed ones. Such reasoning was already presented in point 4. W hen the elements o f the sample are correlated, one should still generate a random vector with the и-dimensional N n(0, I) distribution in order to use it to eliminate correlated variables. This idea was presented by G o l u b et al. (1973) and W agner (1982, 1990).
In our considerations we apply the D urbin form ula. We have one nuisance param eter (variance a 2) and n correlated random variables being the com ponents o f the random vector r o f the residuals. We generate a random variable (n — \ )S' 2 ~ X n - i and independently of it random vector v with an arbitrary distribution with the m om ents E ( v ) = 0 and D ( v ) — I. We create a corrected vector of LSM residuals
r* = y + ( I - i t / ) v , (5.1)
where S2 = r'r/(n — m).
T o prove that the com ponents of vector r* are uncorrelated we use the result given in points 2, 3 and 4.
Lemma 5.1. E(r*) = 0 and D(r*) = I.
Proof. From independence of S'2 and vector v of vector y, we get
Cov(S', r) = Cov(S', v) = Cov(S, r) = Cov(S, v) = 0 and Cov(r, v). F o r the expectation we have E(r*) = E ^ - r^j + (I — ф)Е(у) — E ( S ' ) E ^ j . But every com ponent of vector r/S has the same distribution with expectation equal 0 according to Lem ma 4.3, i.e. E(r*) = 0. F urther, due to the earlier m entioned covariances, we get
D(r*) = + tp)D(v)(I — ф)' — E [ J rr'J
-«!■
+ 1 - Ф = ~ ^ S 2j £ (rr') + 1 - ф = ^ а 2Ф + 1 ~ Ф = 1, w hat follows from E(S ) = 1 and (3.3) at к = 2.Given lemma shows th at the com ponents o f vector r* are uncorrelated and get a simple sample.
6. TEST IN G FO R N O R M A L IT Y O F R A N D O M E R R O R S
The result o f Lem ma 5.1 will be used to test norm ality of random errors in m odel (2.1). Let N = <t2I): ц = X ß , a 2 > 0} be a class of и-dimensional norm al distributions with the given param eters. We set the null hypothesis th at the distribution o f vector e belongs to class N, which we write as H 0: e e N against the alternative
We verify the hypothesis H 0 with vector of r* o f corrected residuals. It is the sum of two vectors. The first is created from the transform ation of observable random vector у and generated random variable w ith the chi-square distribution with n-1 degrees of freedom. If the hypothesis H 0 is true then, according to Lem ma 4.3 each of its com ponents is norm ally distributed. A bout the second vector we can assume that, in particular, it is a random vector with the distribution N„(0, I). Thus, the sum o f the two independent vectors, each norm ally distributed, gives a random vector norm ally distributed. And conversly, with the help o f Cram er Lem ma, (see e.g. R a o 1982, p. 525) assuming that vector r* is norm ally distributed and, at the same tim e, it is com posed o f the two earlier m entioned independent random vectors, then each o f them is norm ally distributed. It implies th at the vector of random errors is norm ally distributed.
The verification o f H 0 is done with the help of a simple sample, which is created by the com ponents o f vector r* and w ith the help o f test for norm ality. A t n < 50 the S h a p i r o , W i l k (1965) test is recom mended, and at 5 0 < n < | 1 0 0 t he D ’ A g o s t i n o (1971) test. The tests m entioned are om nibus tests i.e. they are both sensitive to de-partures from the sym m etry and curtosis o f the norm al distribution. They are characterized by power and their critical values are known. T he Shapiro-W ilk and S hapiro-F rancia tests have left-side critical re-gions. This m eans that the big values o f the tests statistics do n ot lead to rejection of H 0. Furtherm ore, the D ’A gostino test has a two-sided critical region. The H 0 hypothesis is not rejected with this test when the value o f the test statistic lies between the upper and lower critical values.
R EFE R E N C E S
C r a m e r H. (1958): M etody m atematyczne w statystyce, PW N, W arszawa.
D ’ A g o s t i n o R. В. (1971): An omnibus test o f normality fo r moderate and large size samples, „B iom etrika’', 58, p. 341-348.
G o l u b G. H., G u t m a n I ., D u t e r R . (1973): Examination o f pseudo-residuals o f outliers fo r detecting suporsity in the general univariate linear model, [in:] (eds) D . G. K a b e , R. P. G u p t a , M ultivariate statistical inference, N orth-H olland Publishing Company, A m sterdam .
F i s z M . (1967): Rachunek prawdopodobieństwa i statystyka matematyczna, PW N , W arszawa. M a r d i a K.. V. (1980): Tests o f univariate and multivariate normality. Handbook o f statistics,
Vol. 1, p. 279-320.
P a w ł o w s k i Z. (1976): Statystyka matematyczna, PW N, Warszawa.
P e a r s o n E. S., S e k a r C. A. (1936): The efficiency o f statistical tools and a criterion fo r the rejection o f outlying observations, „B iom etrika” , 28, p. 308-320.
R a o C. R. (1982): Modele liniowe statystyki matematycznej, PW N, W arszawa.
S a l l y L., S a r k a d i K. (1982): Beurteilung der Normaliat an Handmehrerer Stichproben kleinen Umfangs, „Q ualität und Zuverlässigkeit” , 27, p . 197-199.
S a r k a d i K. (1960): On testing fo r normality, Publ. M ath. Inst. Hungar. Acad. Sd. 5, p. 269-275. S a r k a d i K. (1967): On testing fo r normality, Proc. 5th Berkeley Symp. M ath. Statist.
Prob. 1, p. 373-387.
S h a p i r o S. S., F r a n c i a R. S. (1972): Approximate analysis o f variance test fo r normality, JA SA , 67, p. 215-216.
S h a p i r o S. S., W i l k M . B. (1965): A n analysis o f variance test fo r normality (complete samples), „B iom etrika” , 52, p . 591-611.
S t o r m e r H. (1964): Ein Test zun Erkennen von Normalverteilungen, Z. f. W ahrscheinlich-keitstheorie, 2, p. 420-433.
T h e i l H. (1968): A simplification o f the B L U S procedure fo r analyzing regression distribuance, JASA, 63, p. 242-251.
W a g n e r W. (1982): Testing fo r normality o f errors in linear models, Studia Sei. M ath. H ungarica, 17, p . 393-401.
W a g n e r W. (1990): Test normalności wielowymiarowej Shapiro-W ilka i jego zastosowania w doświadczalnictwie rolniczym, Roczniki AR w Poznaniu, Rozprawy N aukowe, 197, Poznań.
Wiesław Wagner
PR ZY D A T N O ŚĆ M E T O D Y D U R B IN A D O B A D A N IA N O R M A LN O ŚC I W JED N O W Y M IA R O W Y M M O D E L U LIN IO W Y M
W pracy przedstaw iono metodę randam izacyjną D urbina do testow ania norm alności błędów losowych.
Zaprezentow ano również metodę eliminacji param etrów zakłócenia poprzez obliczanie w ektora resztowego i skorygowanego w ektora resztowego.