See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/274394043

An Application on Multinomial Logistic Regression Model

Article  in  Pakistan Journal of Statistics and Operation Research · March 2012

DOI: 10.18187/pjsor.v8i2.234

CITATIONS READS
40 16,890

1 author:

Abdalla M. El-Habil
Al-Azhar University - Gaza
32 PUBLICATIONS   70 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

name of paper is An Application on Multinomial Logistic Regression Model View project

All content following this page was uploaded by Abdalla M. El-Habil on 28 October 2016.

The user has requested enhancement of the downloaded file.

 An Application on Multinomial Logistic Regression Model

Abdalla M. EL-HABIL
Head of the Department of Applied Statistics
Faculty of Economics andAdministrative Sciences
Al-Azhar University, Gaza - Palestine
abdalla20022002@yahoo.com

Abstract
This study aims to identify an application of Multinomial Logistic Regression model which is one
of the important methods for categorical data analysis. This model deals with one nominal/ordinal
response variable that has more than two categories, whether nominal or ordinal variable. This
model has been applied in data analysis in many areas, for example health, social, behavioral,
and educational.To identify the model by practical way, we used real data on physical violence
against children, from a survey of Youth 2003 which was conducted by Palestinian Central
Bureau of Statistics (PCBS). Segment of the population of children in the age group (10-14 years)
for residents in Gaza governorate, size of 66,935 had been selected, and the response variable
consisted of four categories. Eighteen of explanatory variables were used for building the primary
multinomial logistic regression model. Model had been tested through a set of statistical tests to
ensure its appropriateness for the data. Also the model had been tested by selecting randomly of
two observations of the data used to predict the position of each observation in any classified
group it can be, by knowing the values of the explanatory variables used. We concluded by using
the multinomial logistic regression model that we can able to define accurately the relationship
between the group of explanatory variables and the response variable, identify the effect of each
of the variables, and we can predict the classification of any individual case.

Keywords: Multinomial logistic regression model - categorical data analysis -

maximum likelihood method - generalized linear models -classification.

I. Introduction
In recent years, specialized statistical methods for analyzed categorical data
have increased, particularly for application in biomedical and social science.
Regression analysis is one of these statistical tools that utilize the relationship
between two or more variables. The regression models can be divided into two
groups, the first related to linear relationship models, and the second related to
non-linear relationship models. The linear models, considered up to this point,
are satisfactory for most regression applications. Nonlinear model used when the
linear model is not suitable anyhow. Many of statisticians believe that the logistic
regression model is one of the important models can be applied to analyze a
categorical data; this model is a special case of generalized linear models (GLM).
The multinomial logistic regression (MLR) model used in generally effective
where the response variable is composed of more than two levels or categories.
The basic concept was generalized from binary logistic regression. Continuous
variables are not used as response variable in logistic regression, and only one
response variable can be used. The MLR model can be used to predict a
response variable on the basis of continuous and/or categorical explanatory
variables to determine the percent of variance in the response variable explained
by the explanatory variables, to rank the relative importance of independents, to

Pak.j.stat.oper.res. Vol.VIII No.2 2012 pp271-291

Abdalla M. EL-HABIL

assess interaction effects, and to understand the impact of covariate control

variables. The MLR model allows the simultaneous comparison of more than one
contrast, that is, the log odds of three or more contrasts are estimated
simultaneously, Garson (2009). The logistic regression model assumes that the
categorical response variable has only two values, in general, 1for success and 0
for failure. The logistic regression model can be extended to situations where the
response variable has more than two values, and there is no natural ordering of
the categories. Natural ordering can be treated as nominal scale, such data can
be analyzed by slightly modified methods used in dichotomous outcomes, and
this method is called the multinomial logistic. The impact of predictor variables is
usually explained in terms of odds ratios. Logistic regression applies maximum
likelihood estimation after transforming the dependent into a logit variable (the
natural log of the odds of the dependent occurring or not). Logistic regression
calculates changes in the log odds of the dependent, not changes in the
dependent itself as ordinary least square (OLS) regression does. Logistic
regression has many analogies to OLS regression: logit coefficients correspond
to b coefficients in the logistic regression equation, the standardized logit
coefficients correspond to beta weights, and a pseudo R square (R2) statistic is
available to summarize the strength of the relationship. Unlike OLS regression,
however, logistic regression does not assume linearity of relationship between
the independent variables and the dependent, does not require normally
distributed variables, does not assume homoscedasticity, and in general has less
stringent requirements.It does, however, require that observations be
independent and that the independent variables be linearly related to the logit of
the dependent. The predictive success of the logistic regression can be assessed
by looking at the classification table, showing correct and incorrect classifications
of the dichotomous, ordinal, or polytomous dependent. Goodness-of-fit tests
such as the likelihood ratio test are available as indicators of model
appropriateness, as is the Wald statistic to test the significance of individual
independent variables.

The idea of this study focusing on MLR model, that we believe it is important and
useful for analyzing categorical data. Therefore, the problem is:
By using real data, how can we apply a new statistical method (multinomial
logistic regression model) for analyzing categorical data?

The paper is organized as follows: Section 2 recalls the technical background of

multinomial logistic regression model. Section 3 physical violence data. Section 4
building of multinomial logistic model. Section 5 conclusion.

II. Multinomial logistic regression model

The logit (logistic) regression model

In fact, the multinomial logistic regression (MLR) model is a fairly straightforward
generalization of the binary model, and both models depend mainly on logit
analysis or logistic regression. Logit analysis in many ways is the natural

272 Pak.j.stat.oper.res. Vol.VIII No.2 2012pp271-291

An Application on Multinomial Logistic Regression Model

complement of ordinary linear regression whenever the response is categorical

variable. When such discrete variables occur among the explanatory variables
they are dealt with by the introduction of one or several (0, 1) dummy variables,
but when the response variable belongs to this type, the regression model breaks
down. Logit analysis provides a ready alternative.

For a response variable Y with two measurement levels (dichotomous) and

explanatory variable X, let:  ( x )  p (Y  1| X  x )  1  p (Y  0 | X  x ) , the
logistic regression model has linear form for logit of this probability
  (x )   (x )
Logit[ (x )]  log       x , where the odds  ,
 1   (x )  1   (x )

The odds = exp(   x ) , and the logarithm of the odds is called logit, so
  (x ) 
Logit[ π(x)]  log    log exp(   x ) =    x
 1   (x ) 

The logit has linear approximation relationship, and logit = logarithm of the odds.
The parameter β is determined by the rate of increase or decrease of the S-
shaped curve of π (x). The sign of β indicates whether curve ascends (β > 0) or
descends (β < 0), and the rate of change increases as |β| increases.

Multiple logistic regressions

The logistic regression can be extending to models with multiple explanatory
variables. Let k denotes number of predictors for a binary response Y by
x1,x2,...,xk , the model for log odds is
Logit[P(Y  1)]    1x1  2x 2  ...  k x k

And the alternative formula, directly specifying  ( x ) , is

exp(  1x 1  ...   k x k )
 ( x) 
1  exp(  1x 1  ...   k x k )

The parameter  i refers to the effect of x i on the log odds that Y =1, controlling
other x j , for instance, exp(  i ) is the multiplicative effect on the odds of a one-
unit increase in x i , at fixed levels of other x j .

If we have n independent observations with p-explanatory variables, and the

qualitative response variable has k categories, to construct the logits in the
multinomial case, one of the categories must be considered the base level and
all the logits are constructed relative to it. Any category can be taken as the base
level, so we will take category k as the base level. Since there is no ordering, it
is apparent that any category may be labeled k. Let π j denote the multinomial

Pak.j.stat.oper.res. Vol.VIII No.2 2012 pp271-291 273

Abdalla M. EL-HABIL

probability of an observation falling in the j th category, to find the relationship

between this probability and the p explanatory variables, X1, X2,........., X p , the
multiple logistic regression model then is

j  
 x 
log 
i
     x   x  ...   x ,
 k  
 x 
i 
0i 1 j 1i 2 j 2i pj pi

Where j= 1, 2, … , (k-1), i = 1, 2, … ,n. Since all the  ’s add to unity, this

reduces to

exp  0i  1 j x 1i  2 j x 2i  ...   pj x pi 
   
log  j x i k 1
,

1   exp  0i  1 j x 1i  2 j x 2i  ...   pj x pi
j 1

For j = 1, 2, … , (k-1), the model parameters are estimated by the method of ML.
Practically, we use statistical software to do this fitting, Chatterjee and Hadi
(2006).

Baseline-Category Logit Model

In MLR model, the estimate for the parameter can be identified compared to a
baseline category. We defined bold letter as matrix or vector, let
 j ( x )  p (Y  j | x) at a fixed setting x for explanatory variables, with 
j
j ( x)  1 ,

for observations at that setting, we treat the counts at the J categories of Y as

multinomial with probabilities, 1 (x),...,  J (x) , logit models pair each response
category with a baseline category, often the most common model is:
 j (x)
log   j  βj x,
 J (x)
where j = 1,…, ( J 1), simultaneously describes the effects of x on these (J-1)
logits, the effects vary according to the response paired with the baseline, these
(J-1) equations determine parameters for logits with other pairs of response
 ( x)  ( x)  ( x)
categories. Since log a  log a  log b
 b ( x)  J ( x)  J ( x)
with categorical predictors, Pearson chi-square statistic  2 and the likelihood
ratio chi-square statistic G 2 goodness-of-fit statistics provide a model check
when data are not sparse. When an explanatory variable is continuous or the
data are sparse, such statistics are still valid for comparing nested models
differing by relatively few terms, Agresti (2002).

III. Physical violence data

We used real data of Youth Survey, 2003 which conducting by Palestinian
Center Bureau of Statistics (PCBS) for application of the MLR model. The data

274 Pak.j.stat.oper.res. Vol.VIII No.2 2012pp271-291

An Application on Multinomial Logistic Regression Model

were used for the purposes of scientific research according to a special

agreement between PCBS and Al – Azhar University- Gaza. In this survey, youth
were defined as the young people, there age was (10-24) years, includes
preteen, teen-agers (10-19) and young (20-24). Also in this survey age with
complete years were used, the data referred to calculating the vital rates and
ages as it was on 1/3/2003, the target population of this survey was consist of all
Palestinian households that usually reside in the Palestine Territory.

Sample size and design frame used on the Youth Survey

The sample size had been 5,570 households of whom 4,830 households
responded, 3,256 households residing in the West Bank, 1,574 households in
Gaza Strip. The sample strata had been designed on two levels, first level was
the governorate (16 governorates), and second level was the type of locality
(urban, rural and camps). The survey frame was a list of enumeration areas
peculiar to the 1997 Population, Housing and Establishment Census,
enumeration areas was a residential area containing about 150 households. It
should be noted that thedata had been collected according to the procedures,
rules andmethodologyestablished by PCBS to achieve the highest data-quality,
PCBS user guide (2003).

Our resource of data was the file "Youthfile.sav", conducting by PCBS of the
Youth Survey, 2003, and the user guide, survey questionnaire, and methodology
book. According to this survey, total number of persons in Palestinian National
Authority on the age group (10-24 Years) in the year 2003 was (1,189,282),
51.0% male, 49.0% female, and 62.3% in West Bank, and 37.7% in Gaza Strip.

Target population group used

The target group used in this study was youth people living in Gaza governorate
in the age group 10-14 years in both sex (male and female). To achieve this goal
we selected population group with the following specifications: first we selected
the youth group in Gaza governorate (160,948), and then, we selected age group
(10-14 Years), (66,935 or 41.588 % of the total).

Selection of the variables

Response variable
Through the review of a questionnaire of the survey, there were two questions
drew our attention directly related to the issue of physical violence, first question
was "have you been subjected to physical violence (beating, burning, biting,
pushing, etc) during the last month? ", with two levels of measures (yes/no),
second was " who did practice physical violence against you?", with 10 levels of
measures, or 10 kinds of people exercised of physical violence against the youth:
father, mother, sibling, wife, other relatives, teacher, employer, peer (schoolmate,
neighbor, etc), Israeli forces, and others.

Pak.j.stat.oper.res. Vol.VIII No.2 2012 pp271-291 275

Abdalla M. EL-HABIL

In fact, the aim ofthis analysis does not focus basically on the phenomenon of
violence, which it is very important, and its need special study, but the aim is to
apply our statistical model, MLR model, on categorical data. For this, we tried to
choose available related data on physical violence from Youth Survey 2003
according to our criteria already mentioned in the sample size and frame. By
merging the two variables we got the response variable with 11 levels of
measures, as the target population was the children 10-14 years, living in Gaza,
wife and Israeli forces were excluded as there is no frequency to these levels.
We tried to focus on the practice of physical violence on children by the family,
father, mother, sibling, and very close environment of the child, schoolmates, the
neighbors, and others. We note that a small number of these levels, in the same
time due to skewness in the response variable we combined these levels: other
relatives, teacher, employer, and other, to be in one category, Takagi et. al.
(2007).

We note that, the skewness before merging was (2.908) and, and after merging
(1.887), and standard error of the skewness was 0.009 for both. The response
variable became as the following "have you been subjected to physical violence
during the last month, and who did practice physical violence against you?", had
four categories, 0-had not been, 1-father/ mother, 2-sibling, 3-peer & other, we
called the response variable as "response physical violence by". The frequencies
of response variable according to these categories are shown in Table 3.1.

Table 3.1: The frequencies of the response variable categories

Physical violence categories Frequency Percent
0 Had not been 52545 78.5
1 Father/Mother 4201 6.3
2 Sibling 7025 10.5
3 Peer & other 3163 4.7
Total 66935 100.0

Baseline category (reference) of the response variable

Any category of response variable can be chosen to be the baseline or reference
category, the model will fit equally well, achieving the same likelihood and
producing the same fitted values, only the values and interpretation of the
parameters will change, Schafer (2006). In our situation we used the category
with highest frequency so we selected category of (0-had not been). This means,
the comparison will be against the children whom did not been suffering of the
physical violence in the last month of survey date.

The explanatory variables

We tried to select a set of explanatory variables, that, we believed it has an effect
in somehow, on the physical violence against children in age 10-14 years in
Gaza in the year 2003. Some of these explanatory variables describe the
environment around the child, some belongs to child himself. We will review
these explanatory variables in some details.

276 Pak.j.stat.oper.res. Vol.VIII No.2 2012pp271-291

An Application on Multinomial Logistic Regression Model

A set of questions talk about "In your opinion, are the following behavior exists
among youth in the locality where lives?":
 Hh04-a"Alcohol consumption" with two categories (1-no/ little, 2-
yes widely).
 Hh04-b" Smoking" with two categories (1-no/ little, 2-yes widely).
 Hh04-c "Reckless driving" with two categories (1-no/ little, 2-yes
widely).
 Hh04-d "Drug abuse" with two categories (1-no/ little, 2-yes
widely).
 Hh04-e "Verbal violence" (e.g., harassment, swearing), with two
categories (1-no/ little, 2-yes widely).
 Hh04-f "Begging" with two categories (1-no/ little, 2-yes widely).
 Hh04-g "Assault on properties",(Stealing, pillage, plundering ), with
two categories (1-no/ little, 2-yes widely).
 Hh04-h "Physical violence", (Beating, rape, etc), with two
categories (1-no/ little, 2-yes widely).
Notes: In this set of variables we considered the answer don't know as missing
system, as this answer does not give an opinion
Another set of questions talked about the child himself:
 Hh01-a "How do you evaluate your physical health status" with two
categories (1-good, 2-moderate/poor).
 Hh01-b"How do you evaluate your mental health status"with two
categories (1-good, 2-moderate/poor).
 Hr04 "Sex" with two categories (1-male, 2-female).
 Hh02 "Do you want your current weight to ", we merged the categories to
three only (1-remain as it is, 2-to decrease, 3-to increase).
 Hr08 "Enrolled in education status", with two categories (1-currently
enrolled, 2-not enrolled now).
 S01 "Free time you have" with three categories (1-little, 2-enough, 3- too
much).
Another set of questions talked about the family circumstances:
 Loctype "locality type" with two categories (1-urban, 2- camps).
 Ir04 "Total number of household members" (numeric variable).
 Hr07 "Refugee status" with two categories (1-refugee, 2-not refugee).
 Ho5 "Current status of parents" the variable was with 7 categories (1-living
together, 2-divorced, 3-father is dead, 4-mother is dead, 5-both are dead,
6-one of them works abroad, 7-others),some of these without frequencies,
so we merged the categories to three only as: 1-living together, 2-one of
them dead, 3-divorced & others.

We used the same variable's name and codes used by PCBS survey. Full
detailed of the explanatory variables was summarized in table 3.2. This table
prepared by using frequency command of (Statistical Package for Social
Sciences) SPSS.

Pak.j.stat.oper.res. Vol.VIII No.2 2012 pp271-291 277

Abdalla M. EL-HABIL

Table 3.2: List of the explanatory variables and their frequencies

Explanatory variables Frequency % Valid % Cumulative %
Ho5 1-living together 62,905 94.0 94.0 94.0
2-one of them dead 2,615 3.9 3.9 97.9
3-divorced 7 other 1,415 2.1 2.1 100.0
S01 1-little 14,134 21.1 21.1 21.1
2-enough 32,903 49.2 49.2 70.3
3-too much 19,898 29.7 29.7 100.0
Hh01-a 1-good 61,143 91.3 91.3 91.3
2-moderate/poor 5,792 8.7 8.7 100.0
Hh01-b 1-good 53,884 80.4 80.4 80.4
2-moderate/poor 13,091 19.6 19.6 100.0
Hh02 1-remain as it is 29,382 43.9 48.9 48.9
2-to decrease 12,292 18.4 20.4 69.3
3-to increase 18,446 27.6 30.7 100.0
4-don't know (missing) 6,815 10.2
Hh04-a 1-n0/little 44,472 66.4 91.3 91.3
2-yes widely 4,250 6.3 8.7 100.0
3-don't know (missing) 18,212 27.2
Hh04-b 1-n0/little 10,058 15.0 15.2 15.2
2-yes widely 55,982 83.6 84.8 100.0
3-don't know (missing) 895 1.3
Hh04-c 1-n0/little 43,809 65.5 68.1 68.1
2-yes widely 20,560 30.7 31.9 100.0
3-don't know (missing) 2,565 3.8
Hh04-d 1-n0/little 41,617 62.2 91.4 91.4
2-yes widely 3,920 5.8 8.6 100.0
3-don't know (missing) 21,398 32.0
Hh04-e 1-n0/little 34,610 51.7 52.6 52.6
2-yes widely 31,169 46.6 47.4 100.0
3-don't know (missing) 1,156 1.7
Hh04-f 1-n0/little 50,580 75.6 79.6 79.6
2-yes widely 12,952 19.4 20.4 100.0
3-don't know (missing) 3,403 5.1
Hh04-g 1-n0/little 52,709 78.7 85.9 85.9
2-yes widely 8,669 13.0 14.1 100.0
3-don't know (missing) 5,557 8.3
Hh04-h 1-n0/little 48,037 71.8 77.7 77.7
2-yes widely 13,813 20.6 22.3 100.0
3-don't know (missing) 5,084 7.6
Hr04 1-male 32,387 48.4 48.4 48.4
2-female 34,548 51.6 51.6 100.0
Hr07 1-refugee 39,840 59.5 59.5 59.5
2-not refugee 27,095 40.5 40.5 100.0
Hr08 Currently enrolled 66,374 99.2 99.2 99.2
Not enrolled now 560 0.8 0.8 100.0
Loctype 1-urban 53,791 80.4 80.4 80.4
2-camps 13,144 19.6 19.6 100.0
Ir04 Mean= 9.28 Std deviation 2.529 Median= 9

278 Pak.j.stat.oper.res. Vol.VIII No.2 2012pp271-291

An Application on Multinomial Logistic Regression Model

IV. Building of MLR model

We chose 18 explanatory variables which we believed it had an influence on
physical violence issues against children see table 4.2. We tried to explore the
effects of these variables by building MLR model and then examined of the
results. To achieve this goal, we used SPSS software version 13, and used
NOMERG command to calculate the MLR model with response variable and all
explanatory variables to make the primary model.

Checking of response variable

From the table 4.1 of case processing summary we can check some points:

Table 4.1: Case processing summary by using 18 explanatory variables

Response variable categories N Marginal percentage
0 Had not been 25317.33 75.7
1 Father/Mother 2858.29 8.6
2 Sibling 3235.41 9.7
3 Peer & other 2011.97 6.0
Valid 33423.00 100.0
Missing 33511.57
Total 66934.56
Subpopulation 126
The dependent variable has only one value observed in 125(99.2%) subpopulation

Table 4.1 is a portion of large table contains all variables, response variable and
explanatory variables. We are focusing on the response variable, as we see in
Table 4.3; the number of the valid observations used in our model is 33,423
distributed among the four categories. The marginal percentage column lists the
proportion of valid observations found in each of the response variable' groups,
75.7% of the valid case (had not been) subjected to physical violence, 8.6% had
been subjected by (father/mother), 9.7% by (sibling), and 6.0% by (peer and
other).

Subpopulation
Subpopulation indicates the number of subpopulations contained in the data. A
subpopulation of the data consists of one combination of the explanatory
variables specified for the model. The SPSS footnote for table 4.1 provides how
many of these combinations of the explanatory variables consist of records that
all have the same value inthe response variable. In our model there are 126
combinations that appear in the data and 125 of these combinations are
composed of records with the same response variable categories

Pak.j.stat.oper.res. Vol.VIII No.2 2012 pp271-291 279

Abdalla M. EL-HABIL

Missing
Missing indicates the number of cases in the dataset where data are missing of
the response variable or any of explanatory variables. In primary model we found
the missing almost 50%. Brannon et al (2007) suggests that we can calculate
scales with missing items if at least two thirds of the items were completed and
others were dropped. Anyhow, this model still under checking, but we refer to
some explanatory variables like alcohol consumption, smoking, etc, a category of
"I don't know" is not a valid decision in this situation. It was considered as
missing system as this procedure will not affect the final result, Moorman and
Carr (2008).

Computing by chance accuracy

The proportional by chance accuracy rate can be computed by calculating the
proportion of cases for each group based on the number of cases in each group
of the response variable. The squaring and summing the proportion of cases in
each group are (0.757 2  0.086 2  0.097 2  0.06 2 )  0.593454= 59.34% (rounding
error taken into account). The benchmark that used to characterize MLR model
as useful is a 25% improvement over the rate of accuracy achievable by chance
alone, so the proportional by chance accuracy criteria are: 1.25 * 0.593454 =
74.18%. This proportion will be compared with the overall percentage of the final
model.

Checking of explanatory variables

The ML method used to calculate MLR by using an iterative fitting process that
attempts to cycle through repetitions to find an answer. Sometimes the method
will break down and not be able to converge or find an answer, or sometimes will
produce widely improbable results. Reporting that one- unit change in an
explanatory variable increases the odds of the model by unreasonable results.
These implausible results can be produced by multicollinearity, or categories of
explanatory variables having no cases or zero cells. If we faced a situation
comparable to that, this means we have a numerical problem and should not
interpret the results. The practical solution of this problem is checking the
standard errors for the parameter's explanatory variables that are larger than 2,
Schwab (2007). This information is available in the parameter estimates table of
the MLR results. This table 4.4 is consisting of three parts as the response
variable has four categories. After checking this table we found there were five of
the explanatory variables had standard error more than 2, we summarized the
results of these variables in the Table 4.2.

280 Pak.j.stat.oper.res. Vol.VIII No.2 2012pp271-291

An Application on Multinomial Logistic Regression Model

Table 4.2: The parameter estimates with more than 2 units of standard
error
Response physical violence by B Std .Error Wald df sig Exp(B)
1-father/mother
Ho5=1 14.510 160.461 0.008 1 .928 2001804.550
H05=2 -.652 190.610 0.0 1 .997 .521
Hh04-b=1 -14.034 46.734 0.090 1 .764 8.04E-007
2-sibling
H05=1 14.050 152.557 0.008 1 .927 1264463.302
H05=2 13.386 152.557 0.008 1 .930 650756.508
3-peer & other
H05=1 12.302 103.476 0.014 1 .905 220089.142
H05=2 -5.735 135.662 0.002 1 .966 .003
Hh01-a=1 35.119 46.309 0.575 1 .488 1786163….
Hh02=2 -32.769 34.640 0.895 1 .344 5.87E-015
Hr08=1 15.417 319.464 0.002 1 .962 4960497.108

From Table 4.2, there are five explanatory variables causing a numerical
problem, these variables: H05 "current status of parents", hh04-b "smoking
among youth", hh01-a "evaluation of physical health status", hh02 "current
weight", and hr08 "enrolling in education". We note these parameter estimates
gave unreasonable results by one unit change in the explanatory variable.

Selection of the model

In the second phase we re-calculated the model after excluding the five
variables, which had causing a numerical problem, and scanned the result and
re-calculate of the model again after excluding the variables that had parameter
estimates not statistically significant. This procedure had been replicated four
times, and then we stopped as we found all parameter estimates are significant.
The results were summarized for these four models in Table 4.3.

Table 4.3: The specifications of four models

Description Model (1) Model(2) Model(3) Model(4)
Number of explanatory variables 13 10 8 6
Valid cases 37202 37715 37715 38184
Missing cases 29732 29219 29219 28749
Subpopulation 133 117 82 31
Chi-square value (likelihood ratio test) 13423.84 12425.41 9817.82 9073.22
Df 42 33 24 18
R- square Cox and Senell 0.303 0.281 0.229 .211
R-square Nagelkerke 0.377 0.351 0.287 .265
R-square Mc Fadden 0.222 0.205 0.162 .149
Number of correct predicted "had not been 27473 28158 28180 28798
Number of correct predicted "father/mother 571 571 571 571
Number of correct predicted "sibling" 342 1198 513 684
Number of correct predicted "peer & other" 520 520 0 0
Classification Overall percentage 77.7 80.7 77.6 78.7

Pak.j.stat.oper.res. Vol.VIII No.2 2012 pp271-291 281

Abdalla M. EL-HABIL

Information of the four models showed that the Model (2) is the best to be
appropriate to the data comparing with the other models. It has the highest
classification overall percentage, includes 10 independent variables. Also worked
to increase the valid cases and reduce the missing cases. R-square factor which
is usually influenced by a number of variables, had given the values comparable
to the other models. For these reasons we selected model (2) with the following
explanatory variables (10 variables): Hh01-b "Evaluation of mental health status",
Hh04-a "Alcohol consumption", Hh04-d "Drug abuse", Hh04-g "Assault on
properties", Hh04-h "Physical violence", Hr04 "Sex", Hh04-f "Begging", Ir04
"Total number of household members", S01 "Free time you have", Hh04-e
"Verbal violence".

Checking of the selected model

There is no guarantee that the model fits the data well for any particular logistic
regression model, Agresti (2007). Even though, there some ways used to detect
lack of fit like likelihood ratio test.

The sample size requirements

The minimum number of valid cases for each of explanatory variable according
to a guideline provided by Hosmer and Lemeshow is 10, and preferring case to
variable ratio 20 to 1 explanatory variable, Schwab (2007). In selected model the
ratio is 377 cases to 1 explanatory variable. Table 4.4 presents the response
variable and their categories frequencies used in selected model.

Table 4.4: Case processing summary

Response variable Categories N Marginal %
0- Had not been 28767.61 76.3
1-Father/Mother 3091.03 8.2
2-Sibling 3577.71 9.5
3-Peer & Other 2279.11 6.0
Valid 37715.46
Missing 29219.10
Total 66934.56
Subpopulation 117

Pseudo R-square
There are three pseudo R-square values can be calculated by SPSS for logistic
regression table 4.5. Pseudo R –square does not have an equivalent to R2 in
OLS regression (the coefficient of determination). R2 summarizes the proportion
of variance in the response variable associated with explanatory variables, but
pseudo R-square does not means what R2 means in OLS regression but we can
use it as indicator for different areas of application. The model with the largest
pseudo R-square statistic is best according to the measures; however,

282 Pak.j.stat.oper.res. Vol.VIII No.2 2012pp271-291

An Application on Multinomial Logistic Regression Model

classification coefficients as overall affect size measures are preferred over

pseudo R-square measures as they have some severe limitations for this
purpose, Garson (2009).

Table 4.5: Pseudo R-Square

Cox and Snell 0.281
Nagelkerke 0.351
McFadden 0.205

Comparing accuracy rates

The classification overall percentage computed by SPSS was 80.7% while was
greater than the proportional by chance accuracy criterion of 74.1% or
(1.25*59.34) =74.1.

The criterion for classification accuracy is satisfied. See Table 4.6.

Table 4.6: The classification table of the selected model

Predicted
Observed Had not been Father/Mother Sibling Peer & other Percent correct
Had not been 28158.16 171.15 438.30 0 97.9%
Father/Mother 2220.95 571.87 0 298.21 18.5%
Sibling 2246.08 133.57 1198.05 0 33.5%
Peer & other 1758.86 0 0 520.25 22.8%
Overall percentage 91.2 2.3 4.3 2.2 80.7%

Absence of Multicollinearity
Multicollinearity can be occurred in logistic regression, as the correlation
increases among the independent variables, the standard errors of the logit
parameters will become inflated. Multicollinearity does not change the estimates
of the parameters, only their reliability, Garson (2009). We check first the
standard error; variables with more than 2 units were ignored. We checked the
asymptotic correlation matrix which is a matrix of parameter estimate correlation.
In this matrix we found that the majority of correlation coefficients were less than
0.10, another 4 were between (0.20 and 0.27), only one coefficient was 0.54, this
means we do not have serious problem with multicollinearity among the
explanatory variables that used in the model. Correlation between total number
of household members (ir04) and evaluation of mental health status (hh01-b) is
0.221. Correlation between alcohol consumption among youth in locality (hh04-a)
and drug abuse among youth in locality (hh04-d) is 0.208, also between drug
abuse and assault on properties (hh04-g) is 0.54, between assault on properties
and physical violence (hh04-h) is 0.276, and between the physical violence and
begging (hh04-f) is 0.265.

Pak.j.stat.oper.res. Vol.VIII No.2 2012 pp271-291 283

Abdalla M. EL-HABIL

Goodness- of-fit measures

The likelihood ratio test is based on deviance [-2 Log Likelihood (LL)], the
significance of the difference between the (-2LL) for our selected model minus
likelihood ratio for a reduced model (intercept only) as in Table 4.9. A common
use of the likelihood ratio test is to test this difference (it called chi-square model)
dropping an interaction effect.If the chi-square model is significant, the interaction
effect is contributing significantly to the full model and should be retained. The
presence of a relationship between the response variable and combination of
explanatory variables is based on the statistical significance of the final model
chi-square. In our model, the p-value of the model chi-square (12425.41) was
0.000, less than the level of significance 0.05. We reject the null hypothesis
which states that there was no difference between the model without explanatory
variables and the model with explanatory variables. The existence of a
relationship between the explanatory variables and the response variable was
supported.

Table 4.7: Model fitting information

Model Fitting Criteria Likelihood Ratio Tests
Model AIC BIC -2log Chi-Square df Sig
Likelihood
Intercept only 53577.558 53603.171 53571.558
Final 41218.146 41525.508 41146.146 12425.411 33 0.000

AIC (Akaike Information Criterion) and BIC (Bayesian Information Criterion) judge
a model by how close its fitted values tend to be to the true expected values, as
summarized by a certain expected distance between the two, the optimal model
is the one that tends to have its fitted values closest to the true outcome
probabilities. In our model AIC and BIC and -2log likelihood are very close.
Table (4.8): Likelihood ratio tests of the selected model

Model Fitting Criteria Likelihood Ratio Tests

-2 Log
AIC of Reduced BIC of Reduced Likelihood of
Effect Model Model Reduced Model Chi-Square df Sig.
Intercept a
41218.146 41525.508 41146.146 .000 0 .
ir04 41434.453 41716.202 41368.453 222.307 3 .000
hh01_b 42633.441 42915.189 42567.441 1421.294 3 .000
hh04_a 42357.903 42639.651 42291.903 1145.756 3 .000
hh04_d 42920.768 43202.517 42854.768 1708.622 3 .000
hh04_g 42830.898 43112.646 42764.898 1618.752 3 .000
hh04_h 43045.503 43327.251 42979.503 1833.357 3 .000
hr04 44271.829 44553.577 44205.829 3059.683 3 .000
hh04_f 41683.817 41965.565 41617.817 471.670 3 .000
s01 42895.663 43151.798 42835.663 1689.517 6 .000
hh04_e 42475.932 42757.680 42409.932 1263.785 3 .000
The chi-square statistic is the difference in -2 log-likelihoods between the final model and a reduced model.
The reduced model is formed by omitting an effect from the final model. The null hypothesis is that all
parameters of that effect are 0.
a. This reduced model is equivalent to the final model because omitting the effect does not increase
the degrees of freedom.

284 Pak.j.stat.oper.res. Vol.VIII No.2 2012pp271-291

An Application on Multinomial Logistic Regression Model

In Table 4.8, we checked the same point with all explanatory variables used to
build model separately. The result was referred that the existence of a
relationship between each of the explanatory variables and the response variable
was supported.

The criteria of Odds ratio explanation

The major difference from the binomial situation is explanatory variable in
selected model has three (b) parameters with three odds ratios (exp (b)), one for
each level of the response physical violence except its reference category
(0=had not been), (Table 4.9), which is assumed but does not show in the tables.
The "exp(b)'' column in SPSS's label for odds ratio of the explanatory variables
with the response variable, it is predicted change in odds for a unit increase in
the corresponding explanatory variable. Odds ratios less than 1 correspond to
decreases and odds ratio more than 1.0 correspond to increases. Odds ratios
close to 1.0 indicates that unit changes in that explanatory variable does not
affect the response variable.

Estimating response probabilities

The MLR model has an alternative expression in terms of the responses
j jx
e
probabilities, that is  j  , j=1,…,J. In our model, we will denote the
 e
h
h
 h x

probability of the child had not been faced physical violence (baseline category)
by  0 and the estimate by ˆ0 . The physical violence by father/ mother by  1 and
the estimate by ˆ1 . The physical violence by sibling by  2 and the estimate by
ˆ2 , and the physical violence by peer and other by  3 , and the estimate by ˆ3 ,
3
the response probability satisfying 
j 0
j
 1 , our baseline category is (had not

been=0), from table (4.9) of parameter estimates we can calculate these

probabilities by two steps:

 ˆ   ˆ   ˆ 
First, we can calculate log  1  , log  2  , and log  3  , as the response
 ˆ   ˆ   ˆ 
 0  0  0
variable has four categories (J=4), which means that there are 3 equations as
following:
 ˆ   ˆ   ˆ 
let y1 = log  1  , and y2 = log  2  , and y3 = log  3  , so
 ˆ   ˆ   ˆ 
 0  0  0

Pak.j.stat.oper.res. Vol.VIII No.2 2012 pp271-291 285

Abdalla M. EL-HABIL

y1 = 1.381 - 0.116(ir04) - 0.162(hh01-b=1) - 0.244(hh04-a=1) - 1.895(hh04-d=1)

+ 0.465(hh04-g=1) - 0.445(hh04-h=1) - 1.163(hr04=1) - 0.599(hh04-f=1)
+ 0.677(s01=1) + 1.291(s01=2) - 1.501(hh04-e=1) (4.1)

y2 = - 1.954 - 0.044(ir04) +1.069(hh01-b=1) + 0.940(hh04-a=1) + 1.908(hh04-d=1)

- 2.326(hh04-g=1) - 1.620(hh04-h=1) - 1.964(hr04=1) + 0.843(hh04-f=1)
- 0.539(s01=2) (4.2)

y3 = 0.082(ir04) - 1.753(hh01-b=1) - 2.417(hh04-a=1) - 1.192(hh04-d=1)

- 0.862(hh04-g=1) + 1.983(hh04-h=1) + 1.195(hr04=1) - 0.536(hh04-f=1)
- 0.886(s01=1) - 1.853(s01=2) +0.839(hh04-e=1) (4.3)

We cannot make corresponding statement about variables "s01=1" and "hh04-

e=1" in equation (2), and intercept in equation (3) as that odds ratios are non-
significant. As Agresti (2007) says "Statistical significance should not be the sole
criterion for whether to include a term in a model. It is sensible to include a
variable that is important for the purposes of the study and report its estimated
effect even if it is not statistically significant. Keeping it in the model may help
reduce bias in estimating effects of the other predictors and may make it possible
to compare results with other studies where the effect is significant"

Second we calculate ˆ1 , ˆ2 , ˆ3 , ˆ0 , as following, where exp or e = 2.71828 is the
base of the system of natural logarithms:
exp(y1 )
ˆ1  (4.4)
1  exp(y1 )  exp(y2 )  exp(y 3 )

exp(y2 )
ˆ2  (4.5)
1  exp(y1 )  exp(y2 )  exp(y 3 )

exp(y 3 )
ˆ3  (4.6)
1  exp(y1 )  exp(y2 )  exp(y 3 )

1
ˆ0  (4.7)
1  exp(y1 )  exp(y2 )  exp(y 3 )

Where the (1) term in each denominator and in the numerator of ˆ0 represents
exp(ˆ  ˆ x ) , for ˆ  ˆ  0 , Agresti (2007).
0 0 0 0

286 Pak.j.stat.oper.res. Vol.VIII No.2 2012pp271-291

An Application on Multinomial Logistic Regression Model

Table (4.9): The parameter estimates of the selected model

response Physical
a
violence by B Std. Error Wald df Sig. Exp(B)
1 Father/Mother Intercept 1.381 .133 108.121 1 .000
ir04 -.116 .010 123.252 1 .000 .891
[hh01_b=1] -.162 .049 10.789 1 .001 .850
[hh01_b=2] b
0 . . 0 . .
[hh04_a=1] -.244 .064 14.439 1 .000 .784
[hh04_a=2] b
0 . . 0 . .
[hh04_d=1] -1.895 .068 780.216 1 .000 .150
[hh04_d=2] b
0 . . 0 . .
[hh04_g=1] .465 .077 36.714 1 .000 1.592
[hh04_g=2] b
0 . . 0 . .
[hh04_h=1] -.445 .054 67.278 1 .000 .641
[hh04_h=2] b
0 . . 0 . .
[hr04=1] -1.163 .046 645.567 1 .000 .312
b
[hr04=2] 0 . . 0 . .
[hh04_f=1] -.599 .053 129.126 1 .000 .549
[hh04_f=2] b
0 . . 0 . .
[s01=1] .677 .066 106.005 1 .000 1.967
[s01=2] 1.291 .060 455.276 1 .000 3.635
[s01=3] b
0 . . 0 . .
[hh04_e=1] -1.501 .050 888.151 1 .000 .223
[hh04_e=2] b
0 . . 0 . .
2 Sibling Intercept -1.954 .174 126.245 1 .000
ir04 -.044 .010 17.415 1 .000 .957
[hh01_b=1] 1.069 .060 314.681 1 .000 2.913
[hh01_b=2] b
0 . . 0 . .
[hh04_a=1] .940 .090 110.122 1 .000 2.561
[hh04_a=2] b
0 . . 0 . .
[hh04_d=1] 1.908 .096 396.509 1 .000 6.737
[hh04_d=2] b
0 . . 0 . .
[hh04_g=1] -2.326 .061 1468.420 1 .000 .098
[hh04_g=2] b
0 . . 0 . .
[hh04_h=1] -1.620 .046 1215.457 1 .000 .198
[hh04_h=2] b
0 . . 0 . .
[hr04=1] -1.964 .050 1556.703 1 .000 .140
[hr04=2] b
0 . . 0 . .
[hh04_f=1] .843 .057 218.119 1 .000 2.323
b
[hh04_f=2] 0 . . 0 . .
[s01=1] -.020 .049 .169 1 .681 .980
[s01=2] -.539 .052 108.999 1 .000 .583
[s01=3] b
0 . . 0 . .
[hh04_e=1] -.069 .045 2.300 1 .129 .933
[hh04_e=2] b
0 . . 0 . .
3 Peer& other Intercept .145 .146 .980 1 .322
ir04 .082 .011 59.833 1 .000 1.085
[hh01_b=1] -1.753 .058 899.007 1 .000 .173
[hh01_b=2] b
0 . . 0 . .
[hh04_a=1] -2.417 .080 923.704 1 .000 .089
[hh04_a=2] b
0 . . 0 . .
[hh04_d=1] -1.192 .078 232.146 1 .000 .304
[hh04_d=2] b
0 . . 0 . .
[hh04_g=1] -.862 .077 126.518 1 .000 .422
b
[hh04_g=2] 0 . . 0 . .
[hh04_h=1] 1.983 .100 395.992 1 .000 7.262
[hh04_h=2] b
0 . . 0 . .
[hr04=1] 1.195 .064 347.519 1 .000 3.305
[hr04=2] b
0 . . 0 . .
[hh04_f=1] -.536 .064 70.418 1 .000 .585
[hh04_f=2] b
0 . . 0 . .
[s01=1] -.886 .063 199.821 1 .000 .412
[s01=2] -1.853 .068 733.513 1 .000 .157
[s01=3] b
0 . . 0 . .
[hh04_e=1] .839 .061 192.266 1 .000 2.315
b
[hh04_e=2] 0 . . 0 . .
a. The reference category is: 0 Had not been.
b. This parameter is set to zero because it is redundant.

Pak.j.stat.oper.res. Vol.VIII No.2 2012 pp271-291 287

Abdalla M. EL-HABIL

Predications by using MLR model

Each case consists of a combination of explanatory variables. Prediction is
based on classifying this combination in one of the four groups of the response
variable. The model estimates the probabilities of this combination of the four
groups of the response variable and then according to the largest probability will
classify the case (we have 117 subpopulation group in the model). For the
application of the model we had selected randomly of two cases of the data and
we used the model to predict from which groups will classify it, the model
consists of three equations, to estimate the four response probabilities
( 0, 1, 2,3 ) by using equations (4.1), (4.2), (4.3).

Table 4.10: Selected cases information

Explanatory variables Case 12049 Case 13102
Total number of household members (ir 04) 8 4
Evaluation of the mental health status (hh01-b) Moderate/ poor Moderate/ poor
Alcohol consumption among youth in locality (hh04-a) No/little No/little
Drug abuse among youth in locality (hh04-d) No/little Yes widely
Assault on properties among youth in locality(hh04-g) Yes widely Yes widely
Physical violence among youth in locality (hh04-h) Yes widely Yes widely
Sex (hr04) female Female
Begging among youth in locality (hh04-f) Yes widely Yes widely
Free time child has (s01) enough Little
Verbal violence among youth in locality (hh04-e) Yes widely Yes widely

For case number 12049,

By using information from table (4.9)
 ˆ 
y1 = log  1  = 1.381 - 0.116(8) - 0.162(0) - 0.244(1) - 1.895(1) + 0.465(0) -
 ˆ 
 0
0.445(0) - .163(0) - 0.599(0) + 0.677(0) + 1.291(1) - 1.501(0) = -0.395
 ˆ 
y2 = log  2  = - 1.954 - 0.044(8) +1.069(0) + 0.940(1) + 1.908(1) - 2.326(0) -
 ˆ 
 0
1.620(0) - 1.964(0) + 0.843(0) - 0.539(0) = 0.542
 ˆ 
y3 log  3 
=  ˆ0  = + 0.082(8) - 1.753(0) - 2.417(1) - 1.192(1) - 0.862( 0) + 1.983(0) +
1.195(0) - 0.536(0) - 0.886(0) - 1.853(1) +0.839(0) = -4.806

By using equations (4.4, 4.5, 4.6, and 4.7), we can calculate the estimated
probability to occur in each category as the following:
exp(0.395)
ˆ1  = 0.1980
1  exp(0.395)  exp(0.542)  exp(4.806)

288 Pak.j.stat.oper.res. Vol.VIII No.2 2012pp271-291

An Application on Multinomial Logistic Regression Model

exp(0.542)
ˆ2  = 0.5055
1  exp(0.395)  exp(0.542)  exp(4.806)

exp(4.806)
ˆ3  = 0.0024
1  exp(0.395)  exp(0.542)  exp(4.806)

1
ˆ4  = 0.2940
1  exp(0.395)  exp(0.542)  exp(4.806)

These probabilities appeared that the case number 12049 has probability of
0.198 to occur in category that the child had facing physical violence by his
father/mother, and probability of 0.5055 by sibling, and 0.0024 by peer and other
and finally had not been facing physical violence with probability of 0.2940. So
the conclusion here is that the child was facing physical violence by sibling has
the largest probability comparing with other groups or categories.

For case number 13102,

 ˆ 
y1 = log  1  = 1.381 - 0.116(4) - 0.162(0) - 0.244(1) -1.895(0) + 0.465(0) -
 ˆ 
 0
0.445(0) - 1.163(0) - 0.599(0) + 0.677(1) + 1.291(0) - 1.501(0) = 1.35
 
y2 log  ˆ2 
=  ˆ  = - 1.954 - 0.044(4) +1.069(0) + 0.940(1) + 1.908(0) - 2.326(0) -
 0
1.620(0) - 1.964(0) + 0.843(0) - 0.539(1) = -1.729

 ˆ3 
y3
= log  ˆ  = + 0.082(4) - 1.753(0) - 2.417(1) - 1.192(0) - 0.862(0) + 1.983(0) +
 0
1.195(0) - 0.536(0) - 0.886(1) - 1.853(0) +0.839(0) = -2.975

By using equations (4.4, 4.5, 4.6, and 4.7), we can calculate the estimated
probability to occur in each category as the following:
exp(1.35)
ˆ1  = 0.7584
1  exp(1.35)  exp(1.729)  exp(2.975)
exp(1.729)
ˆ2  = 0.0349
1  exp(1.35)  exp(1.729)  exp(2.975)
exp(2.975)
ˆ3  = 0.0100
1  exp(1.35)  exp(1.729)  exp(2.975)
1
ˆ4  = 0.1966
1  exp(1.35)  exp(1.729)  exp(2.975)

Pak.j.stat.oper.res. Vol.VIII No.2 2012 pp271-291 289

Abdalla M. EL-HABIL

These probabilities appeared that the case number 13102 has probability of
0.7584 to occur in category that the child had facing physical violence by his
father/mother, and probability of 0.0349 by sibling, and 0.0100 by peer and other
and finally had not been facing physical violence with probability of 0.1966. So
the conclusion here is that the child was facing physical violence by
father/mother has the largest probability comparing with other groups or
categories.

V. Conclusion
We have reviewed the results of the model and carried out some tests to make
sure that the model is fit of the data according to statistical terms. Also we have
reviewed the estimates of parameters and interpreted these estimates focusing
on odds ratio scale. Likelihood ratio tests showed all explanatory variables were
significance but the effects and contribution of each variable were not the same,
so it were sorted according to their effects on the model. "Sex" variable was the
most significant, followed by "Spread of physical violence among young people",
"Spread of drug abuse", "Free time the children had", "Assault on properties
among the youth", "Mental status", "Spread of verbal violence", "Spread of
alcohol among youth", "Spread of begging", and finally "Total number of
household members". The model ability of prediction had been checked by
choosing two cases of the data randomly and applying the model to predict in
any of the response variable's group can be classified of these cases. The model
has been successful in one classification.
The crucial conclusion can be presented by several important points:
1. The usage of the MLR model gives us the opportunity to deal with a
response categorical variable with more than two levels and variety of
explanatory variables.
2. MLR indicates the effect of each of explanatory variables as well as its
additive effect by used in the analysis simultaneously which we are aiming
of the study of this model.
3. MLR enables building a statistical model showing those complex and
interrelated relationships, particularly as we are dealing with a qualitative
response variable has more than two categories. These equations could
measure accurately the effect of each of explanatory variables and
excludedthose variables which did not have statistical significant.
4. MLR model, also has proved its ability to predict, and has reached the
precision with which exhibited 80.7% in our model.
5. The model will help researcher who will try to study the subject of physical
violence by gave him an idea about variables importance and effects, of
course it can be made comparisons between the effects that are
calculated from models if used the similar variables.
6. The logistic regression model is a suitable model to many types of data
when the response variable with more than two categories. MLR has no
any restrictions about the explanatory variables; this model is most
common in the categorical data analysis. MLR can be used in many areas
of social, educational, health, behavioral and even scientific experiments.

290 Pak.j.stat.oper.res. Vol.VIII No.2 2012pp271-291

 An Application on Multinomial Logistic Regression Model

References
1. Agresti, A. (2007). An Introduction to Categorical Data Analysis. John
Wiley & Sons, Inc.
2. Agresti, A. (2002). Categorical Data Analysis. John Wiley & Sons, Inc.
3. Brannon, D., Barry, T., Kemper, P., Schreiner, A., and Vasey, J.(2007).
Job Perception and Intent to Leave Among Direct Care Workers: Evidence
from the Better Jobs Better Care Demonstrations. The Gerontologist
47:820-829, The Geronotological Society of America.
4. Chatterjee, S., and Hadi, A. (2006). Regression Analysis by Example.
John Wiley & Sons.
5. Garson, D. (2009). Logistic Regression with SPSS. North Carolina State
University, Public administration Program.
6. Moorman, S.M., and Carr, D. (2008).Spouses Effectiveness as End-of-Life
Health Care Surrogates: Accuracy, Uncertainty, and Errors of
Overtreatment or Undertreatment. The Gerontologist 48:811, The
Gerontological Society of America.
7. PCBS (2003). User guide, Youth Survey 2003. Palestinian Central Bureau
of Statistics.
8. Schafer J.L. (2006). Multinomial logistic regression models. STAT 544-
Lecture 19.
9. Schwab J., (2007). Multinomial Logistic Regression Basic
Relationships.www.utexas.edu/courses/schwab/sw388r7/SolvingProblems
/Analyzi.
10. Takagi, E., Silverstein, M., and Crimmins, E. (2007). Intergenerational
Coresidence of Older Adults in Japan: Conditions for Cultural Plasticity.
The Journals of Gerontology Series B: Psychological Sciences and Social
Sciences 62:S330-S339, The Gerontological Society of America.

Pak.j.stat.oper.res. Vol.VIII No.2 2012 pp271-291 291

View publication stats