1

CHAPTER ONE
1.0 INTRODUCTION
1. 1 BACKGROUND
Health issues have been one of the primary and indeed a paramount concern in mans
quest for knowledge since time immemorial. Good health is basic to our survival. Adequate
food and sound nutrition are essential to good health; they are crucial not only for
human survival, but also for prevention of and recovery from illness (Ghana
Demographic and Health Survey, GDHS, 2003). This is especially so for newborns. A total
of 19 million newborns are born every year in the developing world with low birth
weight, which is related to maternal malnutrition. Worldwide, more than 9 million
children under 5 years of age die each year. Malnutrition underlies a majority of these
under 5 deaths, 70% of which occur in the first year of life. Infant and young child
feeding practices directly impact the nutritional status and, ultimately, the child survival
of children under 2 years of age. Therefore, improving infant and young child feeding is
critical to ensuring their optimal health, nutrition and development (Unicef.org, 2008).
Appropriate feeding practices are of fundamental importance for the survival, growth,
development, health and nutrition of infants and children. Feeding practices are one of
the underlying determinants of childs nutritional status which in turn influence the risk
of illness and ultimately death (GDHS,2003).
Breastfeeding, according to the Oxford Advanced Learners Dictionary, means giving the
baby breast milk soon after birth. The baby is put to breast and begins to suck the
mothers nipple to stimulate the breast to release milk. Also en.wikipedia.org defines
2

breastfeeding as the feeding of an infant or young child with breast milk directly from
female human breast ( i.e. via lactation) rather than from a baby bottle or other containers.
Breastfeeding is sufficient and beneficial for an infants nutrition in the first six months
of life. Exclusive breastfeeding implies feeding the infant only with breast milk without
any additional food or drink, not even water (WHO, 2003). According to
UNICEF.org.media (2008), high average with optimal breastfeeding practices, especially
exclusive breastfeeding for the first six months of life, could have the single largest
impact on child survival, with the potential to prevent 1.4 million under five deaths. Yet
in 2008 rates of exclusive breastfeeding were only about 37% in developing countries.
The World Health Assembly (WHA), in May 2001, recommended exclusive breastfeeding
for the first six months of a childs life. It emphasized the media role in promoting and
protecting exclusive breastfeeding to improve child health. Since Ghana adopted this
policy (Exclusive breastfeeding for the first six months), various seminars and programs are
being held and information being spread through National and Private Television stations, radios
and news papers to promote and encourage the practice. All women are encouraged to
breastfeed their infants exclusively for the first six months and then complement the
breastfeeding with nutritious food for at least two years. Pre-lacteal feeding (given
something other than breast milk in the first three days of life) is generally discouraged
since it inhibits breastfeeding and expose the new born infant to illness (GDHS,2003).
Despite the recommendation by the WHA, some infants are fed with complementary foods in
addition to breast milk. There are other infants, especially those in orphanage homes in Ghana,
who do not receive any breast milk.
3

One way of measuring a healthy baby is to look at the weight gain. If the baby gains
enough weight during its first year then everything is fine. This study is intended to
compare the growth of babies under four forms of feeding practices namely exclusive
breastfeeding, breastfeeding with water, breastfeeding with complimentary foods and
feeding without breast milk. This would be done by comparing their weights for the
first nine months of their life.
Children gain weight and grow more rapidly during infancy and childhood than at any other time
in life. However, some children fail to gain weight at a normal rate, either because of expected
variations related to genes, being born prematurely, or because of under nutrition, which may
occur for a variety of reasons (Rebecca, T. K., 2006). Under nutrition is sometimes called a
growth deficit or failure to thrive. It is important to recognize and treat children who are not
gaining weight normally, because it may be a sign of under nutrition or an underlying medical
problem that requires treatment. This, according to Rebecca, T. K. (2006), is because, each year,
under nutrition contributes to the deaths of about 5.6 million children younger than 5 years in the
developing world. Another 146 million children within this age group are underweight and are at
an increased risk of early deaths, illness, disability and underachievement. In least developed
countries, 42 percent of children are stunted and 36 percent are underweight. Under nutrition can
have complications such as a weakened immune system, shorter than normal height, or
difficulties with learning apart from deaths in children. These complications are more common in
children who are undernourished for a long period of time (Rebecca, T. K., 2006).
Poor weight gain is defined as gaining weight at a slower rate than other children who are the
same age and sex. "Normal" ranges for weight and height are based upon the weight and height
4

of thousands of children. In the United States, standard growth charts are published by the
Centers for Disease Control and Prevention; these charts are available for boys and girls and are
appropriate for all races and nationalities. Weight gain normally follows a predictable course
from infancy through adolescence. However, some children do not gain weight normally from
birth, while other children gain weight normally for a while, then slow or stop gaining weight.
Weight gain usually slows down before the child slows or stops growing in length. A child is
said to have poor weight gain if he or she does not grow at the expected rate for their age and
sex. Poor weight gain is not a disease, but rather a symptom, which has many possible causes
(Rebecca, T. K., 2006).
Three factors account for poor weight gain in children. The first factor includes not consuming
adequate amount of calories or the right combination of protein, fat, and carbohydrates. The
second factor is when the child is not absorbing adequate amount of nutrients, which leads to
poor weight gain. The third factor has to do with requiring a higher than normal amount of
calories (Rebecca, T. K., 2006). Poor weight gain can occur as a result of a medical problem, a
developmental or behavioral problem, and lack of adequate food, a social problem at home, or
most frequently, a combination of these problems (Rebecca, T. K., 2006).
If an infant or a child slows or stops gaining weight, it is important to try to determine and treat
the underlying cause. The first step is a complete medical history and physical examination.
Most children will not require blood testing or imaging tests, although testing may be
recommended in certain situations. Parents should also mention if they have eliminated foods
from the child's diet due to concern about the effects of these foods (e.g. abdominal pain,
diarrhea, "hyperactivity"). The provider may also ask about the child's household, including who
5

lives in the child's house, if there have been recent changes or stresses (e.g. divorce, illness,
death, new sibling), or if anyone in the house has a medical or psychiatric illness. The provider
may also ask about the food supply (e.g. if there have been days when anyone in the family went
hungry because there was not enough money for food). Although these questions can be difficult
to answer, it is important to be honest (Rebecca, T. K., 2006).
1.1.1 Growth Monitoring In Infants and Children
For a relatively small expenditure per child, growth monitoring can greatly strengthen preventive
health programmes. Growth is the best general index of the health of an individual child, and
regular measurements of growth permit the early detection of malnutrition, frequently
association with diarrhea, and other illnesses, when remedial action is relatively easy. Although
acute signs of malnutrition are easily noted by health workers, it is often too late, and always
more expensive, to help the severely malnourished children (UNICEF, 2006).
For early detection of children with growth retardation and high risk of malnutrition and
mortality, health workers need special tools and training in growth monitoring. The growth status
of children is a measure of the health and well-being of the whole community. Birth weight is of
a particular significance in determining the nutritional status of a community, as low birth weight
is a good indicator of subsequent illness and death in children (UNICEF, 2006).
Various body measurements are used to access growth. Some are easier to use, more accurate
and more useful than others. Monitoring the growth of a child usually requires taking the same
measurements at regular intervals and seeing how they change. A single measurement only
indicates the childs size or weight is increasing, staying the same, or declining. Careful repeated
measurements and comparisons with previous measurements are necessary because most
6

children will continue to grow a little, unless they are very ill, and it is easy to mistake some
growth for adequate growth. Growth measures are usually compared to a reference population.
Gathering data to establish a local reference population is a major undertaking. Therefore,
western standards are usually used for comparison, such as Tanner and Boston, or more recently,
those of the National Centre for Health Statistics (UNICEF, 2006).
1.1.2 Weight-For-Age of Children
To obtain weight-for-age, the weight of the child (in kilograms) is compared with that of an
ideally healthy child of the same age from a reference population. This is the basis of the weight-
for-age, or Gomez classification of nutritional status. A child weighing less than 60 per cent of
the reference weight-for-age is considered to be severely malnourished. For these reasons,
countries have different ways of assessing the growth of children through weight taking and at
different places. For instant, in Indonesia 2.5 million infants and young children are being
weighed regularly at the traditional monthly meetings of village women. The results are entered
on growth charts kept by the mothers themselves. In Thailand, a programme based on the home
use of growth charts by parents in several villages, helped to eliminate completely third degree
malnutrition, and reduced second degree malnutrition by 44 per cent during 1981 1982, even
though no additional food was provided. In Colombia, improvements in weight gain for a
majority of children suffering from mild, moderate and severe malnutrition have been achieved
in poor communities by nutrition programmes incorporating the Carnet de Salud or health card
kept at home by mothers. In Jamaica, a systematic programme to improve the health and growth
of over 6,000 young children using growth charts, immunization and nutrition education and
7

milk supplements, has resulted in a 40% decline in the prevalence of malnutrition and a 60
percent fall in infant mortality (UNICEF, 1985).
In Ghana, mothers and family members have traditionally used variety of indicators and
milestones when assessing their childrens growth. The signs used and local standards children
are compared to vary culturally, but are important factors affecting child care practice. These
were taken into consideration and used by the community health workers to understand the
communitys perceptions, so that they could be related to the use of growth charts.
One quite productive community survey in central Ghana did examine closely the growth related
beliefs and practices of local mothers and families in order to suggest ways of applying the
information to improve rate monitoring at the local level (Love et al 1985). One key question
concerned how mothers know when their children are growing or not growing well. The
responses were rich and comprehensive in scope, covering indicators such as appetite, stages of
physical and mental development, general health appearance and the feel and look of the skin.
Height and weight were also mentioned, as well as variation in mood and activity level
(Brownlee, 1990).
A wide range of physical changes both for periods of good and poor growth were mentioned.
Children growing well cry and move normally. When growth is poor the abdomen becomes
distended, bigger than normal, the hair turns brown or grayish, the face puffy or pale, the
fontanel sunken or large. Observations were both sensitive and detailed, and often included
accurate descriptions of the signs of anemia and dehydration, two major causes of child mortality
in Ghana. Size was emphasized, with clothing often mentioned as a useful indicator: in poor
8

growth, dresses become loose around the body and clothes bought for a young child are still a
good fit many months later (Brownlee, 1990).
Traditionally anthropometric measures were some of the most interesting. It is customary in
central Ghana, as in many other cultures, to make strings or beads for a newborn and put them
around the waist, wrist and legs. They are intended as decoration but used by many parents to
assess growth. One mother explained that by the time the child had reached the age of 5
months, the bead-strings around the waist should have been changed or adjusted five times.
Other items mentioned included metal bracelets, necklaces and finger rings (Brownlee, 1990).
1.1.3 Interventions to Improve Weight Gain of Infants and Children
In 2000, the Center for Diseases Control and prevention (CDC) established the International
Micronutrient Malnutrition Prevention and Control Program (IMMPaCt) to support the global
effort to eliminate vitamin and mineral deficiencies, or hidden hunger in both developed and
developing nations. Through the IMMPaCt program, CDC provides funding and/or technical
assistance directly to countries through cooperative and interagency agreements with UNICEF,
the World Health Organization (WHO), the U.S. Agency of International Development
(USAID), and the Global Alliance for Improved Nutrition (GAIN), and the Micronutrient
Initiative (MI). With these partners, CDC has assisted countries in assessing the burden of hidden
hunger through national surveys and surveillance systems that allow countries to monitor the
coverage and impact of their food fortification and micronutrient supplementation programs. In
addition, computer and web-based training tools and regional and national training workshops
developed by CDC have strengthened the capacity of countries to assess the burden of poor
weight gain through malnutrition, track the effectiveness of interventions strategies through
9

surveillance systems, and plan social marketing and health communication strategies to promote
the consumption of vitamin- and mineral-fortified foods.
In 2002, in collaboration with the WHOs Eastern Mediterranean Regional Office (EMRO),
CDC provided funding support and consultation toward a national micronutrient survey to
generate baseline data on iron status of adult women and preschool children in order to monitor
the impact of the recently initiated national flour fortification program in Jordan.
To help improve nutrition worldwide, the CDC IMMPaCt Program helped launch the Flour
Fortification Initiative (FFI) in 2002. The Initiative was formalized in 2005. The FFI Leaders
Group, a network of government and international agencies, wheat and flour industries,
academia, and consumer and civic organizations, was established to promote flour fortification.
FFI supports fortification of flour with essential vitamins and minerals, especially folic acid and
iron, as one important way to help improve the nutritional status of populations, especially
women and children, around the world.
In many Asian and African countries commercially produced infant foods are either not
commonly used or readily accessible through markets in remote areas. Through the IMMPaCt
program, CDC is actively planning pilot interventions in Kenya and Tajikistan to assess the
feasibility of alternative approaches to sustainable distribution, through small local markets and
house-to-house sales, of easy-to-use, in-home fortificants to enrich baby foods. These efforts
will require public-private-civic sector partnerships to be nurtured and strengthened over time.

10

1.2 STATEMENT OF THE PROBLEM
Weight gain is one of the observable signs of growth especially at the early stages of life.
Regular assessment of weight gain among babies has been a way of assessing their health status
and growth. Too much of it leads to child obesity while too little is a sign of ill-health. Ghana is
much concerned about the health and growth of her children and consequently instituted a
monitoring mechanism for babies (both prenatal and postnatal) up to the age of sixty months.
This mechanism involves monthly assessment of weight right after birth and compulsory
immunization against some diseases (six childhood killer diseases) such as Tuberculosis,
Poliomyelitis, Diphtheria, Yellow fever etc at every public or private health center. One could
access these weight records when a visit is made to these centers.
Weight change, especially among individual children, could be biological (variation related to
genes), due to socio-economic status of the mother or due to childs nutritional status. Children
who show poor weight gain could be a sign of malnutrition. Appropriate feeding practices are of
fundamental importance for the growth, survival, development, health and nutrition of infants
and children (GDHS, 2003).
Parents do not see the need to practice a particular feeding type and feel reluctant to participate
in the monthly weighing exercise as they do not know the importance of them. It is in the light of
this that the study seeks to examine the relationship between infant feeding practice and weight
gain to enable the parents appreciate the feeding practice that is beneficial and results in good
health of their babies.
1.3 THE OBJECTIVE(S) OF THE STUDY
The main objective is to find out if different forms of feeding practice have effect on
11

weights of babies (growth of babies). To do this the study seeks to:
i. Compare the weights of babies under four different forms of feeding practice namely,
Exclusive breastfeeding, breastfeeding with water only, breastfeeding with any other food
(water inclusive) and feeding without breast milk (non-breastfeeding) .
ii. Investigate how weights vary by months for the four modes of feeding practices.
iii. Determine which mode(s) of feeding practice differ(s) from others.
iv. Find out what happens to the weights for the first three months after introduction of
complimentary (other) foods in case of exclusive breastfeeding.
v. To compare the rate of weight gain for the age periods 1-3 months, 4-6 months and 7-9
months.
1.3.1 Hypotheses
In order to achieve the set objective, the following hypotheses are outlined
1. H: There is no difference in weight gains under the different feeding practices.
(H: = = = )
H ; There is at least one feeding practice that differ in weight gain from the rest.
(H:

= for i and j )
2. H: Feeding practices have no effect on weight of babies
H: Feeding practices affect weight of babies
12

1.4 RESEARCH QUESTIONS
Some concerns which this study is attempting to address are
(a) Is the monthly increase in weight influenced by the birth weight of a baby?
(b) Does the type of feeding practice influence the weight of a baby?
(c) Which age period do we have high rate of weight gain?
1.5 SIGNIFICANCE OF THE STUDY
A baby must have proper nutrition to grow and thrive. According to Ghana Health Service
(GHS) (2003), the most cost effective way to address the pressing public health challenge of
malnutrition is to prevent it. That means ensuring that all of the children who are normal weight
at birth continue within the normal range, and those who are low birth weight at birth are brought
swiftly into a healthy growth range. If the baby weighs too little, it is important to ensure that the
nutrition it needs to grow is given. When the babys weight is too much, and the baby is
overweight or obese, one may need to consult a doctor for a more balanced and healthy diet.
Many researchers including Ghanaian researchers have shown that food insecurity is not the
most important factor that contributes to a high prevalence of under nutrition. Instead
inappropriate feeding and caring practices of children and mothers, poor environment (e.g. poor
sanitation and hygiene), and limited utilization of basic health care services are among the most
important determining factors (GHS, 2003).
From economic point of view, poor weight gain causes ill-health of babies which is a potential
source of drain of parents income. Appropriate feeding method ensures the optimum weight
gain for a baby.
13

The research is to encourage parents, especially lactating mothers, to practice appropriate
feeding method for their babies. This will help reduce instance of diseases associated
with improper feeding practices and ultimately help reduce infant mortality.
1.6 SCOPE OF THE STUDY
The study was carried out at three centers in Mampong in the Ashanti Region of Ghana. The
centers were Mampong Babies Home, Mampong Technical College of Education Primary
School New Darman, and Akyeremade - near Birth and Death Registry. Basically, the data
involving feeding without breast milk was collected from the Babies Home while the data on the
other three feeding practices were obtained from the two centers mentioned. The respondents in
the research were babies of not less than nine months as the data involves the birth weight and
weights for the first nine months after birth.
1.7 THE OUTLINE OF THE STUDY
The study is divided into five chapters. Chapter one is the Introduction and it focuses on the
background of the study, the statement of the problem, the objectives of the study, the research
questions, the significant of the study, the scope of the study and the outline of the study. Chapter
two presents the review of literature related and relevant to the study. Chapter three deals with
the Methodology which involves the sampling procedure, data sources and collection, and the
overview of the analysis and modeling techniques. It also entails the underlying basic theories
and concepts and how they are applied. Chapter four presents the Data analysis and the results of
the analysis. The last chapter, Chapter five, discusses the key findings of the study which are
followed by conclusions and recommendations.

14

CHAPTER TWO
2.0 LITERATURE REVIEW
Parents are primarily responsible for their babys health. It is extremely important for them to
understand the unique nutritional requirement and feeding practices of infants. A lot of studies
have been carried out concerning nutritional requirements and feeding practices of infants.
2.1 FEEDING PRACTICES OF INFANTS
The two main ways of feeding infants, which are orthodox, are Breastfeeding and Non-
breastfeeding. Breastfeeding has been the basic norm for humans from the genesis of creation
while Non-Breastfeeding is artificial to human race. Non-Breastfeeding takes place in most cases
when Breastfeeding is medically contraindicated such as health of mother, the baby being unable
to breastfeed, absence of mother, personal preferences, beliefs and experiences.
Breastfeeding is practiced in three main ways. It can be done exclusively or with water only or
with water and other foods. The third form, breastfeeding with water and other foods may be due
to a baby being considered at risk for malnutrition, lactation insufficiency, decision or preference
of parents etc. Both Non-Breastfeeding and breastfeeding with water and other foods are
considered automatic, for example Bottle-feeding, and have the benefit of allowing freedom to
leave the baby with others (Wambach and Koehn, 2004).
2.1.1 Breastfeeding
Breastfeeding was the rule in ancient times up to recent human history and babies were
carried with the mothers and fed as required. With 18
th
and 19
th
century industrialization
in the Western world, mothers in many urban centers began dispensing with
15

breastfeeding due to work requirement in urban Europe. Breastfeeding declined
significantly from 1900 to 1960, due to improved sanitation, nutritional technologies, and
increasingly negative social attitude towards the practice (Riordan, 1980). By the 1960s the
predominant attitude to breastfeeding was that it was something practiced by the
uneducated and those lacking temperament of lower classes. The practice was considered
old fashioned and a little disgusting, left for those who could not afford infant
formula and discouraged by medical practitioners and media of times. In fact, letters and
editorials to the womens magazine Chatelaine from 1945 to as late as 1995, regarded
breastfeeding with a predominately negative attitude. However, since the middle 1960s
there has been a steady resurgence in the practice of breastfeeding in Canada and the
US, especially among more educated affluent women (Nathoo, 2009).
After the addition of solid food, mothers are advised to continue breastfeeding for at
least a year. The World Health Organisation recommends nursing for at least two years
or more. The World Health Organisation (WHO) and the American Academy of
Pediatrics (AAP) emphasize the value of breastfeeding for mothers as well as children.
Both recommend exclusive breastfeeding for the first six months of life. The AAP
recommends that this be followed by supplemented breastfeeding for at least one year,
while WHO recommends that supplemented breastfeeding continues two years or more.
While recognizing the superiority of breastfeeding , regulating authorities also work to
minimize the risk of artificial feeding (Nathoo, 2009).
According to Ip S, Chung M, Raman G, et al (2007) , scientific research such as the studies
summarized in a 2007 review for the US Agency for Healthcare Research and Quality
(AHRQ) and a 2007 review for the WHO (Horta BL, 2007) have found numerous benefits
16

of breastfeeding for the infant. According to AAP, research shows that breastfeeding
provides advantages with regard to general health, growth and development. Not
breastfeeding significantly increases risks for a large number of acute and chronic
diseases including lower respiratory infection, ear infections, bacteremia, bacterial
meningitis, botulism, urinary tract infections, and necrotizing enterocolitis (Lucas, A; Cole, T.J,
1990). There are a number of studies that show a possible protective effect of breast milk
feeding against sudden infant death syndrome, insulin dependent diabetes mellitus,
Crohns disease, ulcerative colitis, lymphoma, allergic diseases, digestive diseases and a
possible enhancement of cognitive development (PEDIATRICS, Feb., 2005).
Under modern health care, human breast milk is considered the healthiest form of milk
for babies (Picciano, M., 2001). Breastfeeding promotes health for both mother and infant and
helps to prevent diseases (Riordan, JM., 1997). Longer breastfeeding has also been associated
with better mental health through childhood and into adolescence (Ody, Wendy H.; Kendall GE.;
et al, 2010). Exclusive breastfeeding for the first six months of life has the single largest
impact on child survival (Unicef.org. media ,2008). Supplementing breast milk before six
months is unnecessary and is strongly discouraged because of the likelihood of
contamination, the unaffordability of breast milk substitutes, and the resulting increased
risk of diarrheal diseases. The early introduction of liquids and solids reduces milk output
because the production and release of milk is influenced by the frequency and intensity of
sucking (GDHS, 2003).
Despite the recommendations that babies be exclusively breastfed for the first 6 months,
less than 40% of infants below this age are exclusively breastfed worldwide. The
overwhelming majority of American babies are not exclusively breastfed for this period
17

in 2005 under 12% of babies were breastfed exclusively for the first 6 months, with 60%
of babies of 2 months of age being fed formula and approximately one in four breastfed
infants having infant formula feeding within two days of birth (WHO, 2008).
Despite the high breastfeeding prevalence (97%) in Ghana, the majority of infants are
not fed in compliance with WHO / UNICEF recommendations (World Health Assembly,
2001). The recommendations call for a period of exclusive breastfeeding for six months
and the introduction of complementary foods after the age of six months. Fifty-three
percent (53%) of children under six months of age are exclusively breastfed in Ghana as
at 2003, (GDHS, 2003), which is a slight increase over the proportion reported in the
1998 GDHS (Ghana Statistical Service and MI, 1999). Exclusive breastfeeding drops
sharply from 65% at age 2 3 months to 39% at age 4 5 months. Further, 6% of
children aged 2 3 months and 32% of children aged 4 5 months are receiving
complementary foods in addition to breast milk. This indicates that there are many
infants who are at risk of being exposed to bacterial contamination and poor quality
foods, even if they started out well with early initiation of breastfeeding (GDHS, 2003).
2.1.2 Artificial Feeding
Artificial Feeding involves not breastfeeding and supplementing breastfeeding with other foods.
Experts agree that breastfeeding is beneficial and have concerns about the effects of
artificial formulas. Artificial feeding is associated with more deaths from diarrhea in
infants in both developing and developed countries. There are few exceptions, such as
when the mother is taking certain drugs or is infected with hum T lymphotropic virus,
or has active untreated tuberculosis. In developed countries with access to infant formula
18

and clean drinking water, maternal HIV infection is an absolute contraindication to
breastfeeding (regardless of maternal HIV viral load or antiretroviral treatment) due to the
s risk for mother to child HIV transmission (Horton, 1996).
Infant formula is a manufactured food designed and marketed for feeding to babies and
infants under 12 months of age, usually prepared for bottle feeding or cup feeding
from powder (mixed with water) or liquid (with or without additional water). The US
Federal Food, Drug and Cosmetic Act (FFDCA) defines infant formula as a food which
purports to be or is represented for special dietary use solely as a food for infants by
reason of its simulation of human milk or its suitability as a complete or partial
substitute for human milk (Wells, J.C.K., 1996).
A 2001 WHO report found that infant formula prepared in accordance with applicable
Codex Alimentarius Standards was a safe complementary food and a suitable breast milk
substitute. In 2003, the WHO and UNICEF published their Global Strategy for infant
and young child feeding which restated that processed food products for infants and
young children should when sold or otherwise distributed, meet applicable standards
recommended by the Codex Alimentarius Commission, and also warned that lack of
breastfeeding and especially lack of exclusive breastfeeding during the first half year
of life are important risk factors for infant and childhood morbidity and mortality. In
particular, the use of infant formula in less economically developed countries is linked to
poorer health outcomes because of the prevalence of unsanitary preparation conditions,
including lack of clean water and lack of sanitizing equipment (WHO, 2003).
UNICEF(2007), estimates that a formula fed child living in unhygienic conditions is
19

between 6 and 25 times more likely to die of diarrhea and four times more likely to die
of pneumonia than a breastfed child.
Throughout history, Schuman (2003) stated, mothers who could not breastfeed their babies
either employed a wet nurse or less frequently, prepared food for their babies, a process
known as dry nursing. Baby food composition varied according to region and economic
status (Oliver, 2004). As early as 1846, scientists and nutritionists noted that an increase in
medical problems and infant mortality was associated with dry nursing. Improvement of
quality of manufactured baby foods gave birth to Raw milk formulas (mixture of cow
milk, water, cream, and sugar or honey in specific ratios) to achieve the nutritional
balance believed to approximate human milk reformulated in such a way as to
accommodate the believed digestive capability of the infant (Fomon, 2001).
In 1920s and 1930s, evaporated milk began to be widely commercially available at low
prices, and several clinical studies suggested that babies fed with evaporated milk
formula thrive as well as breastfed babies. Then came commercial Formulas followed by
Generic brand formulas and later Follow on toddler formulas (Fomon, 2001).
The global infant formula market has been estimated at $7.9 billion, with North America
and Western Europe accounting for 33% of the market and considered largely saturated,
and Asia representing 53% of the market. Infant formula is the largest segment of the
baby food market, with the fraction given as between 40% and 70% (Markos, 2005).
Leading health organizations (e.g. WHO, US CDC and Department of Health and Human
Services) are attempting to reduce the use of infant formula and increase the prevalence of
breastfeeding from birth through 12 to 24 months of age through public health awareness
20

campaigns (CDC, 2007). The specific goals and approaches of these breastfeeding promotion
programs, and the policy environment surrounding their implementation, vary by country. As a
policy basic framework, the International Code of Marketing of Breast-Milk Substitutes, adopted
by the WHOs World Health Assembly in 1981, requires infant formula companies to preface
their product information with statements that breastfeeding is the best way of feeding babies and
that a substitute should only be used after consultation with health professionals. The Baby
Friendly Hospital Initiative also restricts use by hospitals of free formula or other infant care aids
provided by formula companies (WHO, 1981).
2.2 NUTRITIONAL STATUS AND PREVALENCE OF UNDERWEIGHT CHILDREN
IN AFRICA
On nutritional status of children in Ghana, according to the 2003 GDHS, 30% of
children under five are stunted and 11% severely stunted. Weight for age results show
that 22% of children under five are underweight, with 5% severely underweight.
Dinesh et al (2006) studied the nutritional status of under-five children and whether infant
feeding practices are associated with the under nutrition in anganwari (AW) areas of urban
Allahabad. Under-five-years children and their mothers in selected four anganwari areas of urban
Allahabad (UP) participated in the study. Nutritional assessment by WHO criterion (SD-
classification) using summary indices of nutritional status: weight-for-age, height-for-age and
weight-for-height were done. Normal test of proportions, Chi-square test for testing association
of nutritional status with different characteristics and risk analysis using odds ratios with 95%
confidence intervals was also done. Among all under five children surveyed, 36.4% underweight
(<2SD weight- for -age), 51.6% stunted (<2SD height- for- age), and 10.6% wasted (<2SD
21

weight- for- height). Proportions of underweight (45.5%) and stunting (81.8%) were found
maximum among children aged 13-24 months. Wasting was most prevalent (18.2%) among
children aged 37-48 months. Initiations of breast-feeding after six hours of birth, deprivation
from colostrums and improper complementary feeding were found to be significant (P<0.05) risk
factors for underweight. Wasting was not significantly associated (P>0.10) with any infant
feeding practice studied. ICDS benefits received by children failed to improve the nutritional
status of children. They concluded that delayed initiation of breast-feeding, deprivation from
colostrums, and improper weaning are significant risk factors for under nutrition among under-
fives. There is need for promotion and protection of optimal infant feeding practices for
improving nutrition.
According to Dramane and Carolyn (2006), malnutrition of children (0-59 months) is a public
health concern in Africa, particularly in the Sahelian countries. In their study to evaluate the
trends of the malnutrition status of children under five, from the project monitoring /impact
indicators, they determine that, nutritional status of children under-five continues to deteriorate
in spite of better agro climatic conditions and agricultural production in many of these countries.
Using data collected from the GFSI project, they conducted their analysis base on the following
indicators such as Underweight (percent of children of a given age range with weight-for-age z
score less than -2 or less than -3 standard deviation) and Stunting (percent of children of a given
age range with height-for-age z score less than -2 standard deviation).
In the first exploratory phase of the analysis, descriptive statistics and factor analyses (Principal
Component Analysis - PCA, Multiple Correspondences Analysis - MCA) were used. The PCA
was used for the analysis of underweight children data and for the analysis of the stunted
children under-five data, the use of the MCA was chosen.
22

The analysis of anthropometric data on the prevalence of underweight children (under five) over
the period from December 2000 to September 2005 shows that this form of malnutrition
continues to be a challenge (with rates > 20%) in the study zone. It reaches its peak in June with
an average rate varying from 43% to 44% of weighed children. The rate of severely
malnourished children (Weight/Age < -3 SD) follows the same trends as that of the rate of global
malnutrition. It is largely influenced by the food availability/access at the household level (see
figure2.1). The minimum rate of severe malnutrition is observed in September. The maximum
recorded rate is 11% in June 2005. That means that in June 2005, 11 children out of 100 in the
project zone should have been referred to hospitals as emergency cases due to severe
malnutrition problems.

Figure 2.1: Evolution Child Malnutrition and Food Availability
The rate of underweight children under-five is more important among the sedentary population
than the pastoralists. The average rate of global (severe + moderate) underweight children under-
0
2
4
6
8
10
12
D
e
c
e
m
b
e
r
D
e
c
e
m
b
e
r
M
a
r
c
h
J
u
n
e
S
e
p
t
e
m
b
e
r
D
e
c
e
m
b
e
r
M
a
r
c
h
J
u
n
e
S
e
p
t
e
m
b
e
r
D
e
c
e
m
b
e
r
2000 2003 2004 2005
%
0
1
2
3
4
5
6
7
8
N
b

o
f

m
o
n
t
h
% Severe Malnutris Children Nb months of food availability
23

five in the pastoral zone is 37.5% (CV=38%) against 36%, 40 %( CV=26%) and 41% (CV=37%)
for the children under-five in the zones of river, Irrigated Village Perimeters (IVP), and of the
Lakes respectively. However the average rate of severely underweight children is higher among
the children of pastoralists (8.84%) than those of sedentary IVP cereal producers (7.60%), and
flooded rice growers (6.50%). The rate of severely underweight children in the livestock zone
remains lower than the rate of malnourished children living in the villages located around the
lakes (9.91%).
Fawzi et al (1998) studied the relationship between prolonged breastfeeding and child growth by
examining prospectively among 28,753 Sudanese children less than 36 months of age enrolled in
a broader cohort study of child health and nutrition. 81% of children were breast-fed at 12
months, but this prevalence declined to 62% at 18 months and 27% at 24 months. At baseline
and at each of three 6-monthly follow-up visits breastfeeding status was assessed and all subjects
were weighed and measured. Their results show that undernourished children were more likely
to be breastfed for a longer period of time compared with normal children. They found a small
difference between breastfed and fully weaned children in the gain in height over the following
6-month period. However, breastfed children were likely to gain significantly less weight,
particularly among children who were aged 6-12 months. Similar findings were noted when
these associations were examined among children who were normally nourished at the time of
breastfeeding assessment. The inverse association between breastfeeding status and weight gain
was significantly larger among children of poor or illiterate mothers compared with children of
relatively more affluent or literate mothers, respectively.
24

In conclusion, their findings suggest that the inverse association is not causal, and may be
explained by poorer complementary feeding among breastfed compared with weaned children.
Children from poorer households and whose parents are illiterate are more likely to have less
than adequate complementary feeding. The importance of adequate complementary feeding in
the second half of infancy needs to be stressed in nutrition education programmes.
2.3 FACTORS AFFECTING THE PREVALENCE OF MALNUTRITION
Salah E. et al (2006) states that, malnutrition affects growth, morbidity, mortality, cognitive
development, reproduction, and physical work capacity, and it consequently impacts on human
performance, health and survival. It is an underlying factor in many diseases for both children
and adults, and is particularly prevalent in developing countries, where it affects one out of every
3 preschool-age children. A well-nourished child is one whose weight and height measurements
compare very well with the standard normal distribution of heights and weights of healthy
children of the same age and sex. Factors that contribute to malnutrition are many and varied.
Their study was to evaluate the level of malnutrition and the impact of some socio-economic and
demographic factors of households on the nutritional status of children under 3 years of age in
Botswana. Factors included: the number of children under 3 years of age in the family,
occupation of the parents, marital status, family income, parental education, maternal nutritional
knowledge, residence location (urban or rural), gender, and breastfeeding practices. The study
was a cross-sectional descriptive survey using a structured questionnaire and measurements of
25

weight and height. Four hundred households and mothers of children under three, representing
the 23 Health Regions of Botswana, participated in the study. Reference standards used were
those of the National Center for Health Statistics (NCHS). EPI Info software (version 5) was
used for data entry and analysis. The results show that the level of wasting, stunting, and
underweight in children under three years of age was 5.5 %, 38.7 %, and 15.6 % respectively.
Malnutrition was significantly (p < 0.01) higher among boys than among girls. Underweight was
less prevalent among children whose parents worked in the agricultural sector than among
children whose parents were involved in informal business. Children brought up by single
parents suffered from underweight to a significantly (p < 0.01) higher level than children living
with both parents. The prevalence of underweight decreased significantly (p<0.01) as family
income increased. The higher the level of the mothers education, the lower the level of child
underweight observed. Breastfeeding was found to reduce the occurrence of underweight among
children. The study findings imply that efforts for redressing child undernutrition issues in
Botswana should focus on factors associated with development outcomes such as maternal
income, maternal education, and the creation of employment or economic engagements that do
not compromise important child care practices such as breastfeeding.
Martens (2007) studied into the Impact of School Feeding Programme in Ghana with the aim of
determining the nutrient intake from school meals and the out-school food consumption of the
children and the impact on demand for locally produced foods. Data were collected in 4 primary
26

schools in 4 different districts in Central Region in Ghana from February to April 2007 with the
study population of 129 3
rd
grade children aged 7 to 16 years.
Anthropometric measurements were taken to determine nutritional status. Data collection on
nutrient intake from school meals was done using 1-day weighed dietary records and weighing
the portion sizes of the selected 3rd grade children. The primary caretakers of the 3rd grade
children were interviewed by trained translators using 24hr recalls to determine the out-school
food consumption. The demand for locally produced foods per district was determined via the
production figures of staple foods gathered from the district Agricultural Extension Service and
the information from the weighed dietary records. The food composition data were derived from
3 tables, of which the leading table was: Table de composition daliments du Mali (Barikmo,
2004) (supplements: Foods commonly used in Ghana (Eyeson, 1975) and the NEVO table
(NEVO, 2006)). The statistical analysis comprised of descriptive analyses on all data, ANOVA
followed by a Tukey test for normally distributed variables, Kruskal-Wallis followed by a
Dunnett test for skewed variables, a Spearman correlation test, an independent sample T-test and
a paired sample T-test.The total food consumption had a DDS of 6.4 (1.0), which is on average
1 (0.8) food group higher than the home food consumption (sig. 0.001). Only 22.6% of the
children still received a home lunch. The nutrient intake recommendations for energy (30-45% of
RDA) and protein (60-70% of RDA), formulated by the Ghana SFP National Secretariat have
been met (31.9% and 67.6% respectively). The iron intake (6.15.8) is low compared to the
weighted DRI of this age group of 22.9 mg/day (bioavailability of 5%). The demand for locally
produced staples is on average 0.07% of the total production of staples in the district. This may
potentially increase to 1.11% in the year 2010.
27

The Ghana SFP succeeded in increasing the dietary diversity of the diet of the school children in
the selected schools. This may reflect in the nutritional adequacy and nutritional status of the
primary school children, bet no internationally agreed upon cut-off points are available to use as
a reference. The Ghana SFP meets their recommendations for energy intake and protein intake.
Vitamin A intake is probably sufficient, but the iron intake remains low, which raises concern.
The impact of the Ghana SFP on the local demand for staple foods at district level seems limited.
Maseta et al (2008) conducted a comparative cross-sectional study to compare childcare
practices and nutritional status of children aged 636 months in Mwembesongo and Mjimpya
wards that had long and short experiences respectively with the Child Survival, Protection and
Development (CSPD) programme. The purpose of the study was to establish whether the long-
term implementation of the CSPD programme had an impact compared to that of a short-term
programme. The findings showed that the children from Mwembesongo were exclusively breast-
fed for a significantly longer period (50 days) than those in the Mjimpya ward (32 days) and that
significantly more mothers (95.7%) in Mwembesongo than in Mjimpya (84.5%) attended growth
monitoring programmes. On the other hand, significantly more mothers in Mjimpya (71.5%)
compared to those in Mwembesongo (51.8%) breast-fed immediately (less than one hour) after
birth. The study revealed that there was no significant difference in childrens nutritional status
(wasting and underweight) between the two wards, except for stunting. More children in
Mwembesongo (39.7%) than in Mjimpya (27.5%) were stunted. The findings have demonstrated
that financial capacity to support childrens food and care requirements forms a springboard
from which to launch additional efforts for improved nutritional status.
Janet and Mabel (1992) used Demographic Health Survey data I Malawi to assess the association
between breast-feeding practices, socio-economic and morbidity variables, and the nutritional
28

status of children under the age of five years using multilevel models. About 27% of under-five
children in Malawi are underweight, and nearly 50% are stunted. The results of this study
suggest that socio-economic factors, morbidity, and inappropriate feeding practices are some of
the factors associated with malnutrition in Malawi. High socio-economic status, as measured by
urban residence, the presence of modern amenities, and some maternal education, is associated
with better nutritional status, whereas morbidity within two weeks before the survey is associated
with low weight-for-age Z scores. Breast-feeding is almost universal and is carried on for about
21 months, but the introduction of complementary food starts much too early; only 3% of
Malawian children under the age of 4 months are exclusively breastfed. Children aged 12 months
or older who were still breastfeeding at the time of the survey were of lower nutritional status
than those who had stopped breastfeeding. The analysis also showed a significant intra-family
correlation of weight-for-age Z scores of children of the same family of about 39%.
In his study to the causes of morbidity and mortality in developing world, Adam (2005)
identified under-nutrition and infection as the major causes. These two problems are interrelated.
Under-nutrition compromises barrier function, allowing easier access by pathogens, and
compromises immune function. Thus malnutrition predisposes to infection. Studies done
worldwide by UNICEF indicate that Africa has the worst nutritional status with high levels of
underweight, stunting and wasting countrywide (UDHS, 2000 - 2001 and 1995). UNICEF also
estimated that 40% of all deaths among children under five years of age were related to
malnutrition, with high levels of underweight up to 12%. The general objective of this study was
to assess the nutritional status of children under five years of age in Rweibaare Parish, Bushenyi
District. A descriptive cross-sectional study using quantitative data collection methods was used
to collect data from mothers or caretakers of 215 children under five years. A structured
29

questionnaire was used to collect the data and anthropometric measurements (weight and height)
were taken and recorded in a table. The measurements were transformed into the standard
anthropometrical measurements and standard indices of physical growth were considered. The
data collected was analyzed manually using frequency tables. The results revealed that the rate of
malnutrition was high, of the 21 5 children assessed 32% of the children were stunted, 13%
underweight and 2.8% wasted. The factors contributing to this were likely to be related with
socio-economic status of the mothers which included age of the mother, education level and
main occupation. Lack of information about nutrition was also a contributing factor as evidenced
by inadequate breastfeeding with children that stopped breast feeding at four months and below
having a high likelihood of malnutrition. Chronic malnutrition was a predominant form of
malnutrition in this study as seen by malnourished children being mostly stunted (32%) and
underweight (13%). Nutrients requirements are high and the diets given are often inadequate
because of lack of suitable food, parental poverty, ignorance and the social customs to which
they are subjected. One of the most striking problems facing the World Health Organisation is
the high morbidity and mortality rates in infants and under fives, (WHO, 2002). Under nutrition
is described as the most pervasive human problem especially in less developed countries. It has
an adverse effect on the quality of life and social and economic development (Gabir 1995). The
problem of under nutrition is most prevalent among vulnerable groups especially in developing
countries notably among children, adolescents, pregnant and lactating women and people living
in difficult situations, such as the landless, the displaced and the poor (MOH, 2002). In Uganda,
as in other developing countries children below five years of age have been the target groups for
nutrition programmes due to the high rates of mortality and morbidity in this age group.
30

UNICEF (1998), reported that the infant mortality rate was 97/1,000 live births and the under
five mortality rate was 141/1,000 live births in Uganda. Thirty to forty percent of the children
under five years of age are malnourished, 38% in the same age group are stunted 5% wasted, and
61% of the population living below the poverty line, (UDHS, 1995). Despite the apparent
abundance of food in Uganda, subsistence agriculture, the favorable climate, and enormous
natural resources, malnutrition still exists. This is due to regional food imbalances due to
ecological differences between the various regions of the country, the under-developed inter-
regional food marketing system, poor road network and inadequate food processing and storage
facilities. The nutritional status of children is influenced by both diet and frequency of infections.
Under-five children have high mortality rates; they suffer from frequent infections such as
malaria, measles, intestinal worms and skin diseases, (Jelliffe, 1996). It was therefore, concluded
that malnutrition is a significant problem in Rweibaale Parish Bushenyi district. It can also be
concluded that malnutrition is caused by many factors hence a need for a multi-sectoral approach
if the problem is to be solved. The results could also be an indicator of a need for reproductive
health services especially in planning for their families, caring for their children and have a
source of income. It is therefore, recommended that another study be done to cover whole
District to give a broader picture of nutritional status of children since this study covered only
one Parish. Other studies that would use other indicators in addition to what was used (height
and weight) such as upper arm circumference and head circumference is necessary. Health
workers should be trained and re-oriented to do growth monitoring and promotion of good
nutrition as part of integrated management of childhood illnesses.
According to Khaled (2009), Major progress has been made over the last 30 years in reducing
the prevalence of malnutrition amongst children less than 5 years of age in developing countries.
31

However, approximately 27% of children under the age of 5 in these countries are still
malnourished. His work focuses on the childhood malnutrition in one of the biggest developing
countries, Egypt. The study examined the association between bio-demographic and
socioeconomic determinants and the malnutrition problem in children less than 5 years of age
using the 2003 Demographic and Health survey data for Egypt. In the first step, separate
geoadditive Gaussian models with the continuous response variables stunting (height-for-age),
underweight (weight-for-age) were used, and wasting (weight-for-height) as indicators of
nutritional status in the case study. In a second step, based on the results of the first step, we
apply the geoadditive Gaussian latent variable model for continuous indicators in which the 3
measurements of the malnutrition status of children are assumed as indicators for the latent
variable nutritional status.









32

CHAPTER THREE
3.0 METHODOLOGY
3.1 INTRODUCTION
This chapter reviews the statistical methods and tools used for the study and their underlying
concepts and theories. The rational for adopting the methods and the procedures involved in the
analyses would be determined. Paramount among these methods in this regard are univariate and
multivariate repeated measures analyses of variance, general linear model ANOVA and general
linear mixed model ANOVA. These methods would be considered under longitudinal data
analysis. Thorough exploratory data analysis for some descriptive statistics, preliminary and
further inferential analyses would be conducted.
3.2 SAMPLING PROCEDURE
All nursing mothers presented at the three Weighing Centers at Mampong Metropolis for
scheduled clinic visits from Monday 18
th
March to Friday 22
nd
March were invited to
participate in the study. The qualitative strand of the data which involves the feeding practice
of the babies was collected using a brief questionnaire.
3.3 SOURCES AND COLLECTION OF DATA
Data were collected at three Weighing Centers at Mampong namely, Mampong Babies Home,
Mampong Technical College of Education Primary School and Akyeremade (near Birth and
Death Registry). These Centers have a monthly weighing programme for children of age
one month after birth to sixty months (five years) after birth. Every month children within
this age group are brought to the Weighing Centers for weighing. It is from this exercise
33

that the data (birth weights and monthly weights of babies for the first nine months ) was
extracted and the form of breastfeeding the baby was undergoing during that period was
solicited from the mother. The childs weighing card was picked to record the weights and
the mother supplied a kind of food the baby was given within the first six months after
birth. Also the weighing cards of children within the age group in the orphanage who
were supposedly not undergoing the breastfeeding were requested and the weights
extracted.
A sample of 109 babies altogether was obtained from all the centers. Thus the weight of total
of 109 babies were used for the study.
3.4 DATA ANALYSIS
The following exploratory data analyses were carried out:
1. Tabulated statistics for sex and feeding practice
2. Tally for discrete variables involving sex and feeding practice
3. Descriptive statistics involving maximum, minimum and mean monthly weights of
babies under each form of feeding practice
4. A component line graph of the mean weights of babies under each feeding practice and a
plot of weight by time with respect to gender to observe the trend of weight gain
5. Tabulated correlation coefficients of monthly weights of babies
At the preliminary inferential analysis stage repeated measures analysis of variance for both
univariate (ANOVA) and multivariate (MANOVA) cases were used. Furthermore, General
Linear Model was employed for some inferential analysis. Linear mixed models and General
linear mixed models were applied for further inferential analysis.
34

Computer software packages, Excel, Minitab and SAS were used to analyze the data.
3.5 MODEL FORMULATION
The response variable was the weight of the child with the explanatory variables as the
four types of feeding practices of the child. Other variables that might affect the weight
increase such as the childs birth weight and the gender of the baby were considered under the
explanatory variables. Each child was assessed for the first nine months after birth.
Since the outcomes were repeated measurements over time on the same subjects (weight
of the same babies when they were 1 month, 2 months, . , 9 months) the data is a
longitudinal data. This called for analyzing the data based on repeated measures analysis and
longitudinal data analysis
3.5.1 Repeated Measures Analysis
Repeated measurement analysis sometimes known as longitudinal studies are used in many fields
of study and the need to analyze this unique data is growing increasingly. Sometimes a
distinction is drawn between longitudinal designs (where subjects are followed for extended
periods of time) and repeated measures designs (where the measurements are collected over a
relatively short period) (Ware, 1985). Repeated measures designs come in several forms such as
split-plot, change over, sources of variability, and longitudinal studies. Split plot designs are
mostly used in agricultural experiments where the field is split into multiple plots. Such plots are
treated with different fertilizer, and crops are randomly assigned to the subplots within each
fertilized plot (Kotz & Johnson, 1988). Each fertilized plot will produce several types of crops
and a measure of each will be collected, yielding repeated measures for each plot. Change over
designs may be used when testing two types of drugs. With this type of design, selected people
35

are usually put into two groups with half given drug A and the others given drug B. Then drug A
and drug B are switched and the experiment is rerun. This design is a repeated measures
experiment because every person will have two measures, one for drug A and one for drug B
(Kotz& Johnson, 1988). Source of variability studies may include taking several randomly
selected items from a manufacturing process and allowing several people to test each one,
possibly over several days (Kotz& Johnson, 1988). Longitudinal studies address situations such
as the growth of chicks, where weights of each chick may be measured every few days. Usually,
in longitudinal designs, the observational units are sampled over an extended period of time.
However, there exists a subset of longitudinal designs where time is not an independent variable
(Ware, 1985). In this case, the interest is studies that are not broken into intervals of time but
rather are categorized according to variables composed of concepts, items, or locations in space
(Kotz& Johnson, 1988).
3.5.2 Model 1: (Where Every Subject Is Tested At Every Level)
In the need to test the efficacy of a drug in a clinical experiment, two methods of experimentation are
available. First, an experimenter may collect observations on N I J = people, randomly assign
th
i
of them to one of
th
j treatment groups, and give each treatment group a different dose of the drug.
Under this design, exactly one reading per person is taken. Alternately, an experimenter could collect
observations on
th
i person and test it at each of the
th
j levels of the factor. This second method is
categorized as repeated measures because the same set of people is used in multiple tests. The
repeated measures experimental design is beneficial as it eliminates the variation due to subjects,
which can significantly reduce the mean square error, ultimately making detecting real differences
between the treatment doses easier (Montgomery, 1984).

36

In this Model, we will simulate an experiment with J treatments and I people. Each person is tested
once per treatment, which helps minimize the amount of variation due to the subjects. This sort of
study is used in many fields; one such example is in the field of psychology where several subjects
were asked to read a set of sentences (Lorch& Myers, 1990). Each sentence was timed, so each
subject provided multiple readings. In this particular model, we have I subjects with indices
1, 2, 3, , i I = in our experiment with
th
j repeated measures per person with indices 1, 2, 3, , j J =
person is timed as they read J sentences, and a time is recorded for each sentence. Let N I J =
denote the total number of observations.

In Model 1,
ij
y is the measurement of the time it took the
th
i person to read the
th
j sentence. In this
example,
1
X will be a dummy coded variable for which sentence is being read. But in other
experiments,
1
X could just as easily be a continuous variable, such as the dose of a drug taken for a
study. Other independent variables may exist as well, such as the age of the person, gender, or IQ
score, denoted

, however, only one independent variable is used in these simulations.

is a vector of the coefficients, and is the vector of errors. Let be the covariance matrix of the
errors.
The model will be of the form:
y X| c = + . 3.1
where X is a matrix of the independent variables, | is the vector of the coefficients, and c is the vector
of errors
In matrix notation we have:
37

11
12
1
21
j
ij
y
y
y y
y
y
(
(
(
(
(
= (
(
(
(
(
(

,
11 1
12 2
1
11 1
1
1
1
1
1
1
k
k
j kj
k
j kj
x x
x x
X x x
x x
x x
(
(
(
(
(
= (
(
(
(
(
(

,
0
1
k
|
|
|
|
(
(
(
=
(
(

,
11
12
1
21
j
ij
c
c
c c
c
c
(
(
(
(
(
= (
(
(
(
(
(


11 11 12 11 1 11 21 11 11
11 12 12 12 1 12 21 12 12
11 1 12 1 1 1 21 1 1
1
cov( , ) cov( , ) cov( , ) cov( , ) cov( , )
cov( , ) cov( , ) cov( , ) cov( , ) cov( , )
cov( , ) cov( , ) cov( , ) cov( , ) cov( , )
cov(
j ij
j ij
j j j j j ij j
c c c c c c c c c c
c c c c c c c c c c
c c c c c c c c c c
c
E =
1 21 12 21 1 21 21 21 21
11 12 1 21
, ) cov( , ) cov( , ) cov( , ) cov( , )
cov( , ) cov( , ) cov( , ) cov( , ) cov( , )
j ij
ij ij j ij ij ij ij
c c c c c c c c c
c c c c c c c c c c
(
(
(
(
(
(
(
(
(
(
(


It is also noted that these matrices were created for a regression analysis. If RM ANOVA is utilized,
which requires categorical variables, the X matrix must be adjusted.
In this model, the X matrix used for RM ANOVA would have two categorical variables, which
would be dummy coded with J - 1 levels for the treatments and I - 1 dummy coded variables for the
subjects and result in rank(X) = (J - 1) + (I - 1) + 1. Figure 3.1 shows how data from such a model
would look if graphed, where each observational unit is given its own symbol. For example, if there
was a study on a blood pressure drug the graph could look as following, where a patient (denoted
+,,, or ) has three blood pressure readings taken, one at each dosage of a drug (maybe 1, 2, and
3mgs).
38


Figure 3.1 Graphical Representation of Model 1
Crowder and Hand (1990) describe many inefficient methods of analyzing this experiment type.
These methods include performing multiple t-tests on different groupings of the factors and
observational units. This analysis tends to be invalid because the involvement of multiple hypothesis
tests increase the chance of a Type I error (Bergh, 1995). Others suggest performing multiple
regressions on these crossover design experiments, allotting one regression for each observational
unit. Some suggest performing an area under the curve analysis, but this examination only inspects
one very specific feature of the regression curve. Similarly, other aspects like time to the peak, half-
life of the curve, or distance back to the baseline are not accounted for (Crowder & Hand, 1990).
39

Crowder and Hand (1990) suggest the best way to analyze data is to have one all-inclusive ANOVA
or regression model, so as not to have multiple statistical tests.
3.5.3 Repeated Measures Analysis of Variance (RM ANOVA)
In Model 1, each measurement on a subject is taken at each of the levels of treatment. Therefore, if,
as before, an experimenter wanted to test a drugs effectiveness he or she could simply test each
person at, for example, J = 3 dosage levels. All subjects must be tested at all treatment levels for
repeated measures ANOVA (RM ANOVA) to be a valid test (Montgomery, 1985). The null
hypothesis is that effectiveness is statistically the same at each treatment level (Misangyi et al.,
2006). If the F value is large enough, then the null hypothesis will be rejected, and the treatments will
be considered statistically distinct. The calculations for the F values as shown in Table 3.1 is given as
MSTR
F
MSE
= . 3.2
MSBS
F
MSE
= .3.3

MSTR is mean square treatment, MSBS is mean square between subject and MSE is mean
square error.
The interest is in the first F value in the above equations because that is used to test the null
hypothesis that the treatments levels are the same. This analysis is identical to randomized complete
block design with subjects as the blocks (Montgomery, 1985).





40

Table 3.1 ANOVA with One Within-Subject Factor

Sources df SS MS F
Between Subject i-1 SSBS MSBS MSBE/MSE
Within subject i(j-1) SSWS MSWS
Treatment J-1 SSTR MSTR MSTR/MSE
Error (j-1)(i-1) SSE MSE
Total IJ-1 SST

In a standard ANOVA setting, without repeated measures, all of the error terms are assumed
independent and normally distributed, with equal variance. Also, in ANOVA, the independent
variable is considered as a factor variable that contains levels. In RM ANOVA, since the
variation broken into two parts, one due to the observational units and one due to the factor
levels, the assumptions must be modified. The assumptions of independence and constant
variance no longer necessarily hold, and this assumption is change to one of sphericity, or
circularity, which restricts the variances and correlations of the measurements (Bergh, 1995).
Compound symmetry, a condition where all variances of the factor levels are equal and all
covariances between each pair of factor levels are equal (OBrien & Kaiser, 1985), can be
examined in place of the less restrictive sphericity assumption because, if it does not hold, the
sphericity assumption usually will not hold. First, the variance of each level of the factor must be
identical so as not to violate the assumptions. For example, we may test children at ages 2, 4, 6,
and 8; with sphericity it is assumed that the correlation between observations taken at ages 2 and
4, and the correlation between observations taken at ages 4 and 6 would be the same. Usually, it
41

would not be the case that the correlation between observations taken at ages 2 and 8 is the same
as the correlation between observations taken at ages 2 and 4, but compound symmetry requires
this supposition (OBrien & Kaiser, 1985). The sphericity assumption generally will be violated
when individuals are being tested over a time range or more than two measurements are taken:
two characteristics of most longitudinal designs (OBrien & Kaiser, 1985). In this way, the
sphericity assumption is almost always violated in the repeated measures situation making RM
ANOVA (OBrien & Kaiser, 1985) and may lead to an increase in Type I errors if the degrees of
freedom are not adjusted (Misangyi et al., 2006).
3.5.4 Univariate Repeated Measure Model
In the univariate repeated measures ANOVA the correlation is assumed to come from the
individual specific random effects; this is due to the fact that each subject is assumed to have an
underlying level of response that persists over time and influences all measurements on that
subject. The times of measurement are treated as a within-subject factor and the effect of time is
assumed to be the same for all subjects. The response for the i
th
subject is assumed to be related
to discrete covariates and is assumed to be different from the population mean .
Repeated measures ANOVA can be expressed
ij j i ij
e V y + + + = t ,

=
=
t j
s i

, 2 , 1
. 2 , 1
3.4

where ( )
j ij ij
V X Y E + = =
'

The parameter
j
v is the effect of time. The parameter ( )
2
, 0
t
o t N
i
~ is the random subject effect
that gives the between-subject variance, and ( )
2
, 0
e ij
N e o ~ is a within-subject measurement error
and it gives the within-subjects variance. The covariance matrix of the ANOVA has a compound
42

symmetry structure, where the variance and covariance are homogeneous across time and equal
to
2 2
2
e
o o
o
t
t
+

3.5.5 Multivariate Analysis of Variance (MANOVA)
Multivariate analysis of variate (MANOVA) is an extension of the univariate analysis of variate
(ANOVA) to accommodate two or more dependent variables. MANOVA is a technique that
measures the differences for two or more metric dependent variables base on the set of
categorical variables acting as independent variables.
Like ANOVA, MANOVA is concerned with differences between groups or experimental
treatments. MANOVA is termed multivariate procedure because it is used to assess group
difference across multiple metric dependent variables simultaneously. In MANOVA, each
treatment is observed on two or more dependent variables.
The concept of multivariate analysis of variance was introduced more than seventy (70) years
ago by Wilks (1932). However, it was not until the development of appropriate test statistics
with table distributions and more recent wide spread availability of computer programs to
compute the statistics that it became a practical tool for researchers.
3.5.6 Multivariate Procedure for Assessing Group Difference
As statistical inference procedure, both the univariate techniques (t test and ANOVA) and their
multivariate extensions (Hotelling
2
T and MANOVA) are use to assess the statistical
significance of difference between groups. In the t test ANOVA the null hypothesis tested is the
equality of single dependent variable means across groups. In multivariate techniques, the null
43

hypothesis tested is the equality of vectors of means on multiple dependent variables across
groups. The differences between the hypotheses tested in both cases are shown below:
n o
H = = = =
3 2 1
:
(3.5)

Null hypothesis
o
H = all the group means are equal, that is they come from the same population.
This is for the univariate analysis of variance case.
For the multivariate analysis of variance (MANOVA), the null hypothesis
o
H is stated as
follows:
(
(
(
(
(
(

=
(
(
(
(
(
(

=
(
(
(
(
(
(

pk
k
k
p p
o
H

2
1
2
22
12
1
21
11
........ :
. (3.6)

Unlike the ANOVA, a variate is tested for equality in MANOVA and the researchers actually
have two variates, one for the dependent variables and another for the independent variables. The
dependent variable variate is of more interest because the metric-dependent measures can be
combined in a linear combination as in multiple regression and discriminant analysis. The
uniqueness of MANOVA is that the variate optimally combines the multiple dependent measures
into a single value that maximizes the differences across groups.


44

3.5.7 The Two-group MANOVA
Assume that we have two groups of multivariate observations, as in the example above, and that
there are
1
n observations in group 1 and
2
n observations in group 2. Hence we observe the
vectors p
ij
y ,
1 2
1, 2, , , 1, 2, , j n i n = = , or equivalently expressed, we observe the
( ) p n n +
2 1
matrix . Formally, the statistical model that underlies the MANOVA is "just" an
ANOVA model for each variable together with an error structure that allows for covariance
between response variables:
ij i ij
y e = +
. (3.7)

For
1
1, 2, i n = and
2
1, 2, j n =
where, if the number of variables equals three: 3 = p
|
|
|
.
|

\
|
=
31
21
11
1

and
|
|
|
.
|

\
|
=
32
22
12
2

(3.8)

are the two expected group means. As in univariate ANOVA, we interested in an overall
statistical significance test for the hypothesis of no group difference at all
0 1 2
1 1 2
:
:
H
H


=
=

Note, that this hypothesis expresses that the groups are equal in all the response variables, and
that the alternative expresses that in at least one variable there is a group difference. So this is a
composite hypothesis jointly in all variables. Computationally, the analysis is handled by
computing the group and overall mean:
45

1 2
1 1 2 2
1 1
1 1
1 1
,
n n
j j
j j
y y y y
n n
= =
= =

.. (3.9)

=
+
=
2
1 2 1
1
i
i i
y n
n n
y . (3.10)
the (weighted) between group variance-covariance structure of these averages:
( )( )

=
=
2
1 i
T
i i i Between
y y y y n SS
and the within-group covariances of the group-corrected data:
( )( )

= =
=
2
1 1 i
n
j
T
i ij i ij Within
i
y y y y SS
From a statistical perspective, the two group centroids are estimates of the unknown group
centroids
1
and
2
and the within group error variance-covariance matrix Eis estimated as
1 2
1
2
w within
S SS
n n
E = =
+
. (3.11)
in analogy with the univariate case. From the example above it is clear that the key information
about the group differences is given by the distance between the two centroids. This difference
can be measured statistically by the Mahalanobis distance:
2 1
1 2 1 2
( ) ( )
T
w
MD y y S y y

= . (3.12)

Recall, how this distance measure down weighs directions in which there is a high residual
variability and up weighs directions of low residual variability (allowing for covariances between
46

each variable). In the case of two groups the
2
MD related to the multivariate generalization of the
student's t-test, Hotelling's
2
T as:
2 2 1 2
1 2
n n
T MD
n n
=
+
..(3.13)

and the
2
T can be transformed into an exact F-ratio as follows:
( )
2 1 2
1 2
1
2
n n p
F T
n n p
+
=
+
..(3.14)


which has an F-distribution with p and 1
2 1
+ p n n degrees of freedom (under the null
hypothesis). Multivariate confidence ellipsoids for group mean and/or mean difference can be
found from the F-distribution and the inverse within group covariance matrix
1
w
S .
3.5.8 The Multiple group one-way MANOVA
Assume that we have I groups of multivariate observations with
i
n observations in group i,
1, 2. , i I = . Hence we observe the vectors p ,
ij
y 1, 2, ,
i
i n = , 1, 2, , j I = . The model is
expressed exactly as for the two-group case:
( )
, var
ij i ij ij
y e e = + =
. (3.15)

For 1, 2. , i I = and 1, 2, ,
i
j n =
The hypothesis of no group difference at all is given as
0 1
:
I
H = =
(3.16)

47

with the alternative that for at least on variable there is at least two groups that differ. Again, this
is a composite hypothesis jointly in all variables, even to a higher degree than for the two-group
case, since it is composite in as well the group structure as in the variables.
Computationally, the analysis is handled exactly as above. First the averages are computed as;
Treatment mean
1
1
1 1
1
n
ij
j
y y
n
=
=


And the total mean as
1
1
I
i i
i
y n y
n
=
=

(3.17)
where
I
n n n n ,
2 1
+ + = .
Then the between-group information is found:
1
( )( )
I
T
Between i i i
i
SS n y y y y
=
=

(3.18)
and the within-group covariance of the group-corrected data:
( )( )
1 1
i
n I
T
Within ij i ij i
i j
SS y y y y
= =
=

(3.19)
And as before the group means are estimates of the unknown group mean
i
and the within
group error variance-covariance matrix Eis estimated as
48

1
W Within
S SS
n I
E = =

(3.20)
In the univariate hierarchy of methodology the F-test is introduced as a necessary tool when
going from two to more than two groups. This is because there is no direct generalization of the
t-statistic (a difference between two groups) to more than two groups, since e.g. three groups
require two distances to summarize the group difference information. So instead of comparing
the (two) groups directly they are all compared with the overall mean/centroids of the data, that
is, the between group variation. In the univariate case, the F-statistic has the form (assuming for
a moment that the SS's are univariate)
/
/
Between Between
Within Within
SS df
F
SS df
=
(3.21)

In the multivariate case the measure of group differences in the light of the within-group
variability is given by the matrix
1
Within Between
SS SS

(3.22)
In fact the unique decomposition of the total variability often expressed in the univariate case
also holds for multivariate:
Total Between Within
SS SS SS = +

(3.23)
Where
Total
SS
( )( )
1 1
i
n I
T
Total ij ij
i j
SS y y y y
= =
=

(3.24)
49

The exercise of doing a statistical hypothesis test for the hypothesis of no group differences is to
evaluate the degree of extremeness of this statistic compared to what would be expected under
the null hypothesis. Due to the multivariate nature, extremeness can be defined in various ways.
Four of these are usually standard output from statistical software:
1
,
1
K
k
k
k
Pillai sTrace

=
=
+

1
,
K
k
k
Hotelling sTrace
=
=




1
1
'
1
K
k k
Wilk s

=
A =
+
[



max
max
' arg
1
Roy s L est Root

=
+





Where
K
, ,
2 1
are the eigenvalues of the
Between within
SS SS
1
matrix. The number of eigen
values K is equal to the minimum of the dimension of the variables space and the number of
groups minus 1, ( ) 1 , min = I p K . Being familiar with PCA a way to express this would be:
Perform a PCA on the
2 / 1
w
S pre-processed group average data and compare the explained
variances of each component with what comes out of data where the groups are truly equal. Note
that eigenvalues are the variations of the principal directions of the variables that are
corresponding to the covariance matrix
Between
SS . In other words the first (largest) eigenvalue of
this matrix express the largest possible variation (between group means) of a new combined
variable (the first principal component). The largest eigenvalue of
Between within
SS SS
1
expressed
50

the largest possible variation between group means taking the within-covariance (point scatter
shape) into account.
The theoretical null hypothesis distribution of these statistics is in principle known and described
in the literature. However, often they are substituted by approximate F-ratios. We give here no
formulas for these approximations. However in the two group case all of these statistics are
equivalent and can be exactly equivalently expressed by the F-ratio and/or Hotelling's
2
T .
In Sharma (1995), Olson (1974) is referred to as showing that the Pillai's Trace was the most
robust and had adequate power to detect true differences under different conditions, and hence
Pillai's Trace is recommended to test multivariate significance.
3.6 LONGITUDINAL DATA ANALYSIS CONCEPTS
The objectives of longitudinal data analysis are to examine and compare responses over time.
The defining feature of a longitudinal data model is its ability to study changes over time within
subjects and changes over time between groups. For example, longitudinal models can estimate
individual-level (subject-specific) regression parameters and population-level regression
parameters (Patetta et. al, 2002).
Longitudinal data sets differ from time series data sets because longitudinal data usually consists
of a large number of a short series of time points. However, time series data sets usually consist
of a single, long series of time points (Diggle, Liang, and Zeger 1994). For example, the monthly
average of the Dow Jones Industrials Index for several years is a time series data set, and the
efficacy of a drug treatment over time for several patients is a longitudinal data set.
To illustrate the ability of longitudinal models to study changes over time, consider cross-
sectional studies in which a single outcome is measured for each subject. In the figure 3.2 below
51

where each point represents one subject, blood pressure appears to be positively related to age.
However, you can reach no conclusions regarding blood pressure changes over time within
subjects.
Blood
Pressure





Age

Figure 3.2 Cross-Sectional Analysis

Now expand the cross-sectional study of baseline data to a longitudinal study with repeated
measurements over time. The baseline data still shows a positive relationship between blood
pressure and age. However, now you can distinguish changes over time within subjects from
differences among subjects at their baseline or initial starting values. Cross-sectional models
cannot make this distinction (Diggle, Liang, and Zeger 1994).


Blood 3
Pressure
1 4
2
Age
Figure 3.3 Longitudinal analysis

52

An example of a longitudinal study is the Baltimore Longitudinal Study of Aging (Shock, et al.
1984). This is a multidisciplinary observational study where participants return approximately
every two years for three days of biomedical and psychological examinations. One objective of
the study is to look for markers, which can detect prostate cancer at an early stage. One marker
with this potential is the prostate-specific antigen (PSA), which is an enzyme produced by both
normal and cancerous prostate cells and its level is related to the volume of prostate tissue.
However, an elevated PSA level is not necessarily an indicator of prostate cancer because
patients with benign prostatic hyperplasia can also have increased PSA levels. Therefore,
researchers have hypothesized that the rate of change in the PSA level might be a more accurate
method of detecting prostate cancer in the early stages of the disease. A longitudinal model can
address this hypothesis (Patetta et. al, 2002).
Another example of a longitudinal study is the Indonesian childrens health study (Sommer
1982). In this study over 3000 children had quarterly medically exams for up to six visits to
assess whether they suffered from respiratory or diarrheal infection and xerophthalmia. One
objective of the study is to determine whether children who have a vitamin A deficiency are at
increased risk of respiratory infection.
3.6.1 Notation
Let
it
Y be the response for the
th
i subject ( 1, 2,..., ) i N = at the
th
t occasion where ( 1, 2,..., )
i
t n = .
The total number of subjects is equal to
N
i
i
n

, 1
i i
y n = is the vector of responses, and 1
i
X p =
is the covariate vector for subject i at time t. The matrix of covariates is
i i
X n p = for subject i
and will usually include an intercept. For example in a 14 repeated measures data with four
factors, if i = subject, 1, ,14 t = (number of repeated measures), 14 1
i
y = , 4 1
it
x = and
53

14 4
i
X = . If the fixed effects are age and gender because they do not change throughout the
duration of the study and the within individual effects are time and body weight, the data set will
be a balanced with time. This meaning that all subjects will be measured at a common set of
occasions and there will be no missing data.
3.6.2 Advantages and Disadvantages
There are several advantages and disadvantages to using longitudinal studies.
Some of the advantages are: subjects serving as their own controls which mean the direct study
of change can be measured; fewer subjects are required because the measurements are being
repeated, between-subject variation is excluded from the error, and longitudinal data can separate
aging effects from cohort effects. Some of the disadvantages are: the dependence of the
measurements which must be accounted for in the analysis, models are not as well developed;
the risk of attrition, carry-over effects, and the improvement or the decline could be caused by
treatment or fatigue.
3.6.3 Problems with Model Assuming Independence
Special methods of statistical analysis are needed for longitudinal data because the set of
measurements on one subject tends to be correlated, measurements on the same subject close in
time tend to be more highly correlated than measurements far apart in time, and the variances of
longitudinal data often change with time. These potential patterns of correlation and variation
may combine to produce a complicated covariance structure. This covariance structure must be
taken into account to draw valid statistical inferences. Therefore, standard regression and
ANOVA models may produce invalid results because two of the parametric assumptions
(independent observations and equal variances) may not be valid (Patetta et. al, 2002).

54

3.6.4 Variance Covariance Matrix for continuous outcome
To illustrate the differences between longitudinal data models and other types of models,
consider the variance-covariance matrix of the response variable below.
Subject X Y
1 4 10
2 2 7
3 6 12
4 8 11

If you have four observations, you would have a 4-by-4 variance-covariance matrix with the
variances on the main diagonal and the covariances (a measure of how two observations vary
together) on the off diagonals. In linear regression with continuous responses, the assumptions
are that the responses have equal variances and are independent. Therefore, the variances along
the diagonal are equal and the covariances along the off-diagonals are 0 (Patetta et. al, 2002).

Longitudinal data
Subject X

1 4 10 6 6
2 2 7 5 3
3 6 12 9 8
4 8 11 14 16
With longitudinal data (as above), there are now multiple measurements taken on each subject.
You not only can examine the differences between subjects, but you can also examine the change

0 0 0
0

0 0
0 0

0
0 0 0

55

within subjects across time. There are still only four subjects and the response is continuous.
How does this change your variance-covariance matrix?
3.6.5 Variance-Covariance Matrix for Longitudinal Data
Subject Time X Y
1 1 4 10
1 2 4 6
1 3 4 6
2 1 2 7
2 2 2 5
2 3 2 3
3 1 6 12
3 2 6 9
3 3 6 8
4 1 8 11
4 2 8 14
4 3 8 16

For longitudinal data models fit in the MIXED procedure, the number of observations is not the
number of subjects but rather the number of measurements taken on all the subjects. Because
there are three repeated measurements on each subject, you now have 12 observations and a 12-
by-12 variance-covariance matrix (as shown above). For a simple longitudinal model, the matrix
is now a block-diagonal matrix in which the observations within each block (the block
corresponds to a subject) are assumed to be correlated and the observations outside of the blocks

0 0 0


0

0 0


0 0

0


0 0 0

56

are assumed to be independent. In other words, the subjects are still assumed to be independent
of each other and the measurements within each subject are assumed to be correlated.
a). Effect on Time-Independent Predictor Variables
If the observations are positively correlated, which often occurs with longitudinal data, then the
variances of the time-independent predictor variables (variables that estimate the group effect (or
between-subject effect) such as gender, race, treatment, and so on) are underestimated if the
data is analyzed as though the observations are independent. In other words, the Type I error rate
(rejecting the null hypothesis when it is true, in other words, a false positive) is inflated for these
variables (Dunlop 1994).
b). Effect on Time-Dependent Predictor Variables
For time-dependent predictor variables (variables that measure the time effect (or within-subject
effect) such as how the measurements change over time), ignoring positive correlation leads to a
variance estimate that is too large. In other words, the Type II error rate (failing to reject the null
hypothesis when it is false, in other words, a false negative) is inflated for these variables
(Dunlop 1994). Because the variances of the group effects will be underestimated and the
variance of the time effects will be overestimated if positive correlation is ignored, it is evident
that correlated outcomes must be addressed to obtain valid analyses.
3.7 STATISTICAL METHODS FOR CONTINUOUS RESPONSES
The following are statistical methods for continuous responses
-Univariate ANOVA in PROC GLM
-Multivariate ANOVA in PROC GLM
-Linear mixed model in PROC MIXED
57

The GLM procedure enables users to perform univariate and multivariate ANOVA of repeated
measures data and the MIXED procedure enables users to fit linear mixed models.
3.7.1 Univariate ANOVA using PROC GLM
Univariate ANOVA for repeated measures data, also called split-plot in time ANOVA, has a data
structure where the number of observations equals the number of measurements taken on all
subjects. The assumptions are that the measurements have equal variances at all time points and
that the correlation between measurements on the same subject is the same regardless of the time
lag between measurements. If these assumptions are valid, then univariate ANOVA for repeated
measures data is an optimal method of analysis (Littell, Henry, and Ammerman 1998).
However, these assumptions are frequently unrealistic with longitudinal data. Measurements
close in time are often more highly correlated than measures far apart in time. Therefore, the
inferences from univariate ANOVA for repeated measures data may not be valid
The condition required for the validity of the univariate ANOVA tests is called the Huynh-Feldt
(H-F) condition, which is equivalent to differences between each pair of measurements having
equal variances. While the H-F condition is mathematically more general than equal variances
and correlations, the differences are minor from a practical point of view (Patetta et. al, 2002).
3.7.2 Multivariate ANOVA using PROC GLM
Multivariate ANOVA has a data structure where the rows are subjects and the columns are time
points in the study. For example, if there were five time points in the study, then there would be
five response variables. For balanced longitudinal data where all subjects have the same number
of repeated measurements taken at time points that are the same for all subjects, multivariate
ANOVA has some useful features. However, for observational longitudinal studies, this data
structure has major limitations. First, it is unreasonable to assume that observational data will be
58

balanced. Second, because PROC GLM uses complete case analysis, if a subject has one missing
measurement, then PROC GLM deletes the entire subject (Patetta et. al, 2002).
Multivariate ANOVA uses a covariance structure that assumes that the correlation between every
pair of measurements within a subject is unique. For example, if there were five time points, then
there would be 15 (((1+5)*5)/2) unique parameters that are estimated. For observational data
with many time points, a simpler covariance structure would probably be sufficient to
characterize the data. This is a limitation for multivariate ANOVA because if the covariance
structure (and thus your model) is more complex than the within-subject correlation structure,
then you sacrifice power and efficiency in your statistical tests. In other words, if a simpler
covariance structure can be used to model the within-subject correlations, then the inferences in
the longitudinal study can be severely compromised if multivariate ANOVA is used (Littell,
Freund, and Spector 2001).
3.7.3 Complete Case Analysis
In complete case analysis, only those observations without any missing values are used in the
analysis. Because each time point represents a separate column in multivariate ANOVA, if one
measurement is missing then PROC GLM deletes the entire subject. Therefore, a great deal of
valid data may be discarded.
Another drawback of the complete case analysis method is that severe bias in the inference can
result if the missing data is not missing completely at random (MCAR). A simple check of
MCAR is to divide the subjects into 2 groups: those with a complete set of measurements and
those with missing measurements. If the MCAR assumption holds, then both groups (with their
measurements) should be random samples of the same population of measurements. In other
59

words, the probability of missing is independent of the observed measurements and the
measurements that would have been available had they not been missing (Patetta et. al, 2002).
The MCAR assumption is much more restrictive than the missing at random assumption (MAR).
If the measurements are MAR, then the subjects with missing measurements are a random
sample of the population of measurements that would have been available if they were not
missing, but they do not have to have the same distributional characteristics as the observations
with complete measurements. In other words, the probability of missing is independent of the
measurements that would have been available had they not been missing but does not have to be
independent of the observed measurements (Little and Rubin 1987).
3.7.4 Linear Mixed Model using PROC MIXED
The linear mixed model allows a very flexible approach to modeling longitudinal data. The data
structure is similar to univariate ANOVA where the number of observations equals the number
of measurements for all the subjects. This means the data does not have to be balanced.
An advantage of fitting linear mixed models is that PROC MIXED uses all the available data in
the analysis. This method differs from complete case analysis in which any observation with a
missing value is dropped from the analysis. The method PROC MIXED uses, called a likelihood-
based ignorable analysis, leads to a valid analysis when the missing data is MAR (which is a less
restrictive assumption than MCAR). When the probability of missing is informative (probability
of missing is dependent on the measurements that would have been available if they were not
missing), then the ignorable analysis is not valid and more complex modeling is required
(Verbeke and Molenberghs 1997).
PROC MIXED offers a wide variety of covariance structures. This enables the user to directly
address the within-subject correlation structure and incorporate it into a statistical model. By
60

selecting a parsimonious covariance model that adequately accounts for within-subject
correlations, the user can avoid the problems associated with univariate and multivariate
ANOVA using PROC GLM (Littell, Freund, and Spector 2001).
3.8 MODEL-BUILDING STRATEGIES IN LONGITUDINAL DATA ANALYSIS
The first step in any model building process is to conduct a thorough exploratory data analysis.
For longitudinal data this involves plotting the individual measurements over time and fitting a
smoothing spline over time. Plotting different groups over time and illustrating cross-sectional
and longitudinal relationships are also important steps in exploratory data analysis.
The second step is to fit a complex mean model in PROC MIXED and output the ordinary least
squares residuals. These residuals can be used to create a sample variogram, and the pattern in
the sample variogram can be helpful in selecting a covariance structure (Patetta et. al, 2002).
The third step is to fit the linear mixed model in PROC MIXED using the selected covariance
structure. Eliminating unnecessary terms and fitting a parsimonious model are important steps in
the model building process. After a candidate model is selected, the final steps of the model
building process are to evaluate model assumptions and to identify potential outliers.
3.8.1 Exploratory data analysis.
The first step in any model-building process is exploratory data analysis. In this step you create
graphs that expose the patterns relevant to the scientific question. The recommendations below,
given by Diggle, Liang, and Zeger (1994), are used to produce the graphs.
Recommendations
- Graph as much of the relevant raw data as possible.
- Highlight aggregate patterns of potential scientific interest.
- Identify both cross-sectional and longitudinal patterns.
- Identify unusual individuals or observations.
61

3.8.2 Individual Profiles
A scatter plot of the response with time is a useful graph. Connecting the repeated measurements
for each subject over time shows you whether there is a discernible pattern common to most
subjects. These individual profiles can also provide some information on between-subject
variability. For example, the figure 3.4 below is a graph of weight over time for several subjects.
These individual profiles illustrate several important patterns (Diggle, Liang, and Zeger 1994).
1. All of the subjects are gaining weight.
2. The subjects that are the heaviest at the beginning of the study tend to be the heaviest
throughout the study.
3. The variability of the measurements is smaller at the beginning of the study compared to the
end of the study.
Weight





Time
Figure 3.4 Individual Profiles

3.8.3 Group Profiles
Besides plotting the response over time, it is also useful to include different subgroups on the
same graph to illustrate the relationship between the response and an explanatory variable over
62

time. For example, in the slide below it appears that both males and females have decreasing
blood pressures over time. However, the slope for the males seems to be more pronounced than
the slope for the females, which may indicate an interaction between gender and time.

Blood
Pressure
Males

Females
Time
Figure 3.5 Group Profile

3.9 GENERAL LINEAR MODEL
The General Linear Model (GLM) is written as
y = +
Where y is the matrix of observed responses; is the design matrix of predictor variables;
is the matrix of regression parameters; is the vector of random errors.

The GLM incorporates a number of different statistical models; ANOVA, ANCOVA,
MANOVA, MANCOVA, ordinary linear regression, t-test and F-test. The general linear model
is a generalization of multiple linear regression model to the case of more than one dependent
variable. If y, , and were column vectors, the matrix equation above would represent multiple
linear regression.
63

Hypothesis test with the GLM can be made in two ways: multivariate or as several independent
univariate tests. In multivariate test, the columns of Y are tested together whereas in univariate
tests the columns of Y are tested independently as multiple univariate tests with the same design
matrix. The standard linear regression model, which is used in the GLM procedure, models the
mean of the response variable by using the regression parameters. The random errors are
assumed to be independent and normally distributed with a common variance. If these parametric
assumptions are valid, then the estimated regression parameters are the best linear unbiased
estimates (Patetta et. al, 2002).
The general linear mixed model is an extension of the general linear model. The standard linear
regression model, which is used in the GLM procedure, models the mean of the response
variable by using the regression parameters. The random errors are assumed to be independent
and normally distributed with a common variance. If these parametric assumptions are valid,
then the estimated regression parameters are the best linear unbiased estimates (BLUE).
3.10 GENERAL LINEAR MIXED MODEL
General Linear Mixed Model is written as
y = + +
Where is the design matrix of random variables; is the vector of random-effect parameters;
is no longer required to be independent and homogeneous.
The general linear mixed model extends the general linear model by the addition of random
effect parameters and by allowing a more flexible specification of the covariance matrix of the
random errors. For example, general linear mixed models allow for both correlated error terms
and error terms with heterogeneous variances. The matrix Z can contain continuous or dummy
64

predictor variables, just like the matrix X. The name mixed model indicates that the model
contains both fixed-effect parameters and random-effect parameters (Patetta et. al, 2002).

3.10.1 Fixed and Random Effects

A represents all possible
levels of the variable
Drug B
represents all levels in which
C inferences are to be made

Figure 3.6 Fixed Effects
A
Drug B -Fixed Effects
C
7 - Levels represent only a random sample
Clinic 18 of a larger set of potential levels
23 -Interest is in drawing inferences that are valid
41 for the complete population of levels

Figure 3.7 Fixed and Random Effects

65

Variable effects are either fixed or random depending on how the levels of the variables that
appear in the study are selected. For example, the figure 3.6 above represents a clinical trial
analyzing the effectiveness of three drugs. If the three drugs are the only candidates for the
clinical trial and the conclusions of the clinical trial are restricted to just those three drugs, then
the effect of the variable drug is a fixed effect.
However, suppose the clinical trial was performed in four clinics and the four clinics are a
sample from a larger population of clinics. The conclusions of the clinical trials are not only
restricted to the four clinics but rather to the population of clinics. The appropriate model in this
study is a general linear mixed model with drug as a fixed-effect variable and clinic as a
random-effect variable (figure 3.7) (Patetta et. al, 2002).
3.11 MODEL ASSUMPTIONS IN PROC MIXED
i. Random effects and error terms are normally distributed with means of 0.
ii. Random effects and error terms are independent of each other.
iii. The relationship between the response variable and predictor variables is linear
iv. Variance-covariance matrices for random effects and error terms exhibit structures available
in PROC MIXED.
Even with these assumptions, the general linear mixed model is an extremely flexible model. By
appropriately defining the model matrices for fixed and random effects, and the covariance
structures for the random effects and the error terms, you can perform numerous mixed model
analyses.
3.12 LINEAR MIXED MODELS FOR LONGITUDINAL DATA
y = + +
where represents
-parameters that are allowed to vary over subjects
66

- subject-specific regression coefficients that reflect the natural heterogeneity in the population.
represents
- parameters that are assumed to be the same for all subjects
- factors that affect subject-specific regression coefficients
and
- represents within-subject variation
- has a covariance matrix that is block diagonal with each block corresponding to a subject.
The general linear mixed model can easily be fitted to longitudinal data. The model assumes that
the vector of repeated measurements on each subject follows a linear regression model where
some of the regression parameters are population-specific (fixed-effects), but other parameters
are subject-specific (random-effects). The subject-specific regression coefficients reflect how the
response evolves over time for each subject. These subject-specific models can be very flexible,
but in practice polynomials involving time will often suffice. However, extensions of this
flexibility, such as fractional polynomial models or extended spline functions, can be considered
as well (Verbeke and Molenberghs 2000).
How the subject-specific regression coefficients relate to known covariates is reflected in the
fixed-effect parameter estimates. These parameter estimates represent the population average. A
model with the assumption that all variability in subject specific intercepts and slopes is
attributed to the fixed-effect variables can be obtained by omitting the random-effect variables.
Omitting only the random variables reflecting subject-specific slopes generates a random
intercepts model (Patetta et. al, 2002).
In the model proposed by Diggle, Liang, and Zeger (1994), it is assumed that the error terms
have a constant variance and can be decomposed as
67

=
()
+
()

where

()
is the measurement error reflecting the variation added by the measurement process

()
is the error associated with the serial correlation in which times closer together are
more correlated than times further apart
i denotes the subject.
If you assume that the measurement errors have an independent covariance structure (

), then
you should only concern yourself with covariance structures that reflect the serial correlation.
3.13 SELECTING THE APPROPRIATE COVARIANCE STRUCTURE
When finding reasonable estimates for covariance structure of the random errors (denoted by R),
if you choose a structure that is too
- simple, you risk increasing the Type I error rate
- complex, you sacrifice power and efficiency.
The validity of the statistical inference of the general linear mixed model depends upon the
covariance structure you select for R. Therefore, a large amount of time spent on building the
model is spent on choosing a reasonable covariance structure for R.
Because the covariance structure models the variability not explained by the fixed effects,
selecting the appropriate mean model is critical. For models dealing with data collected in an
experiment, a saturated model is usually recommended. However, for models dealing with
observational data, saturated models are not feasible. Therefore, it is important to include all the
important main effects and interactions (Patetta et. al, 2002).
After a candidate mean model is selected, fitting the model using ordinary least squares
regression and examining the residuals may help determine the appropriate covariance structure.
68

For example, plotting the residuals by subject over time may indicate whether random intercepts
(between-subject variability at baseline) and random slopes (between-subject variability over
time) are needed.
You can also use the information criteria (such as the AIC and BIC) produced by PROC MIXED
as a tool to help you select the most appropriate covariance structure. The smaller the
information criteria value, the better the model.
Examining a plot of the autocorrelation function of the residuals may be useful for equally
spaced data that is roughly stationary. (The residuals have constant mean and variance and the
correlations depend only on the length of the time interval.) An alternative function that
describes the association among repeated measurements and is easily estimated with irregular
observation times is the variogram (Diggle 1990).
3.13.1 Information Criteria
In a simulation study conducted by Guerin and Stroup (2000), several information criteria were
compared in terms of their ability to choose the right covariance structure. In terms of Type I
error control, assuming the Kenward-Roger (KR) adjustment is used, Guerin and Stroup showed
that it is better to err in the direction of a more complex covariance structure. More complex
covariance structures tend to have inflated Type I error rates only if you fail to use the KR
adjustment, while excessively simple covariance structures have inflated Type I error rates that
the degrees of freedom adjustment cannot correct. However, because complex covariance
structures reduce power, erring too far in the direction of complexity is also not recommended.
Guerin and Stroup believe the AIC is the most desirable compromise in practice. However, if the
sample size is relatively small, the finite-sample corrected version of AIC, called AICC, may be
the most desirable.
69

3.13.2 Variance Component
The simplest covariance structure is the independent or variance component model, where the
within-subject error correlation is zero. This is the default structure for both the RANDOM and
REPEATED statements. For the between-subject errors, the simple covariance structure may be
a reasonable assumption. However, for the within-subject errors, the simple covariance structure
may be a reasonable choice if the repeated measurements occurred at long enough intervals so
that the correlation is effectively zero relative to other variation.

0 0 0
0

0 0
0 0

0
0 0 0

3.13.3 Compound Symmetry
The covariance structure with the simplest correlation model is the compound symmetry
structure. It assumes that the correlation () is constant regardless of the distance between the
time points. This is the assumption that univariate ANOVA makes, but it is usually not a
reasonable choice in longitudinal data analysis. However, this covariance structure may be
reasonable when the repeated measurements are not obtained over time.
1
1
1
1

70

3.13.4 Unstructured Covariance
The unstructured covariance structure is parameterized directly in terms of variances and
covariances where the observations for each pair of times have their own unique correlations.
The variances are constrained to be nonnegative and the covariances are unconstrained. This is
the covariance structure used in multivariate ANOVA.
The correlation coefficient for row 1 column 2 is

)

There are two potential problems with using the unstructured covariance. First, it requires the
estimation of a large number of variance and covariance parameters. This can lead to severe
computational problems, especially with unbalanced data. Second, it does not exploit the
existence of trends in variances and covariances over time, and this can result in erratic patterns
of standard error estimates (Littell, et. al. 1998). If a simpler covariance structure is a reasonable
alternative, then the unstructured covariance structure wastes a great deal of information, which
would adversely affect efficiency and power.

71

3.13.5 First-Order Autoregressive (AR1) Covariance
The first-order autoregressive covariance structure takes into account a common trend in
longitudinal data; the correlation between observations is a function of the number of time points
apart. In this structure, the correlation between adjacent observations is , regardless of whether
the pair of observations is the 1st and 2
nd
pair, the 2nd and 3rd pair, and so on. The correlation is

for any pair of observations 2 units apart, and

for any pair of observations d units apart.

Notice that the AR (1) model only requires estimates for just two parameters,

and , whereas
the unstructured models require estimates for( (1+T )*T)/2 parameters (where T is the number of
time points).
The assumption in the AR (1) model is that the longitudinal data is equally spaced (Littell, et. al.
1996). This means that the distance between time 1 and 2 is the same as time 2 and 3, time 3 and
4, and so on. The AR (1) structure also assumes that the correlation structure does not change
appreciably over time (Littell, et. al. 2002).

1

1

3.13.6 Toeplitz Covariance
The Toeplitz covariance structure is similar to the AR (1) covariance structure in that the pairs of
observations separated by a common distance share the same correlation. However, observations
d units apart have correlation

instead of

. The Toeplitz structure requires the estimation of

72

T parameters instead of just two parameters. You can also specify a banded Toeplitz structure in
which you specify the number of time points apart the measurements are still correlated. For
example, a TOEP (3) (Toeplitz with 3 bands) structure would indicate that measurements are
correlated if they are three or fewer time points apart. If they are four or more time points apart,
the correlation is zero. As with the AR (1) structure, the Toeplitz structure assumes that the
observations are equally spaced and the correlation structure does not change appreciably over
time (Littell, et. al. 2002).

1

1
3.13.7 Spatial Power
Covariance structures that allow for unequal spacing are the spatial covariance structures. These
structures are mainly used in geostatistical models, but they are very useful for unequally spaced
longitudinal measurements where the correlations decline as a function of time. The connection
between geostatistics and longitudinal data is that the unequally spaced data can be viewed as a
spatial process in one dimension (Littell, et. al. 1996).
The spatial power structure provides a direct generalization of the AR (1) structure for equally
spaced data. In fact, the AR (1) structure is more efficient when you have equal spacing. Only
two parameters are estimated ( 2 and ).


73

1.0

1.0

1.0

1.0

3.14 OUTPUT FROM PROC MIXED USING SAS
1. The Model Information table shows the name of the data set, the dependent variable, the
covariance structure used in the model, the subject effect, the estimation method to compute the
parameters for the covariance structure, and the method to compute the degrees of freedom. The
default estimation method is REML. The METHOD= option can be used in the PROC MIXED
statement to specify other estimation methods.
There are four methods for handling the residual variance in the model. The profile method
concentrates the residual variance out of the optimization problem, whereas the fit method retains
the residual variance as a parameter in the optimization. The factor method keeps the residual
fixed, and none is displayed when a residual variance is not a part of the model. The
NOPROFILE option in the PROC MIXED statement changes the method based on the chosen
covariance structure.
The fixed effects standard error method describes the method used to compute the approximate
standard errors for the fixed-effects parameter estimates and related functions of them. The
default method can be changed using the EMPIRICAL option in the PROC MIXED statement.
This option requests robust standard errors obtained from using the sandwich estimator, which
has shown to be consistent as long as the mean model is correctly specified. However, if there
are any missing observations, the EMPIRICAL option only provides valid inferences for the
74

fixed effects under the MCAR assumption. The EMPIRICAL option is not used here because it
cannot be used with the Kenward-Roger degrees of freedom calculation.
2. The Dimensions table lists the sizes of the relevant matrices. This table can be useful in
determining CPU time and memory requirements.
3. The Iteration History table describes the optimization of the residual log likelihood. The
minimization is performed using a ridge-stabilized Newton-Raphson algorithm, and the rows of
the table describe the iterations that this algorithm takes in order to minimize the objective
function.
4. The Covariance Parameter Estimates table shows the parameter estimates for the compound
symmetry covariance structure.
5. The Fit Statistics table provides information you can use to select the most appropriate
covariance structure. Akaikes Information Criterion (AIC) (Akaike 1974) penalizes the 2
residual log likelihood by twice the number of covariance parameters in the model. The smaller
the value, the better the model. The finite sample version of the AIC (AICC) is also included.
The Schwarzs Bayesian Information Criterion (BIC) (Schwarz 1978) also penalizes the 2
residual log likelihood, but the penalty is more severe. Therefore, BIC tends to choose less
complex models than AIC or AICC.
6. The Null Model Likelihood Ratio Test table shows a test that determines whether it is
necessary to model the covariance structure of the data at all. The test statistic is 2 times the log
likelihood from the null model (model with an independent covariance structure) minus 2 times
the log likelihood from the fitted model. The p-value can be used to assess the significance of the
model fit.
75

7. The Type 3 Tests of Fixed Effects table shows the hypothesis tests for the significance of each
of the fixed effects. A p-value is computed from an F distribution with the numerator and
denominator degrees of freedom. You can use the HTYPE= option in the MODEL statement to
obtain tables of Type I (sequential) tests and TYPE II (adjusted) tests in addition to or instead of
the table of TYPE III (partial) tests. You can also use the CHISQ option to obtain Wald chi-
square tests of the fixed effects (Patetta et. al, 2002).
3.15 EVALUATING COVARIANCE STRUCTURE
3.15.1 Importance of covariance structure
Covariance structures
-model all the variability in the data, which cannot be explained by the fixed effects
- represent the background variability that the fixed effects are tested against
- must be carefully selected to obtain valid inferences for the parameters of the fixed effects.
Obtaining valid inferences in a mixed model is much more complex than in a general linear
model. For example, inferences are obtained in the GLM procedure by testing the fixed effects
against the error variance (residual variance). However, in PROC MIXED the inferences are
obtained by testing the fixed effects against the appropriate background variability, which is
modeled by the covariance structure. This background variability may consist of several sources
of error, so selecting the appropriate covariance structure is not a trivial task.
3.15.2 Sources of Error
Longitudinal models usually have three sources of random variation. The between subject
variability is represented by the random effects. The within-subject variability is represented by
the serial correlation. The correlation between the measurements within subject usually depends
on the time interval between the measurements and decreases as the length of the interval
76

increases. A common assumption is that the serial effect is a population phenomenon
independent of the subject. Finally, there is potentially also measurement error in the
measurement process.
The covariance structure that is appropriate for your model is directly related to which
component of variability is the dominant component. For example, if the serial correlation
among the measurements is minimal, then the random effects will probably account for most of
the variability in the data and the remaining error components will have a very simple covariance
structure. Diggle, Liang, and Zeger (1994) believe that in most applications, the serial correlation
is very often dominated by the combination of random effects and measurement error.
Furthermore, Chi and Reinsel (1989) found that models with random effects and serial
correlation might sometimes over parameterize the covariance structure because the random
effects are often able to represent the serial correlations among the measurements. They conclude
that methods for determining the best combination of serial correlation components and random
effects are an important topic that deserves further consideration.
However, suppose the autocorrelation among the measurements is relatively large, and the
between-subject variability not explained by the fixed effects is relatively small. Then choosing
the appropriate serial correlation function in the covariance structure becomes important.





77

CHAPTER FOUR
DATA ANALYSIS AND RESULTS
4.1 INTRODUCTION
This chapter describes the data obtained in easily understandable terms. The chapter examines
the data collected, the detailed analysis of the data and the presentation of results. Table 4.1
presents the snapshot of the data used for the analysis. It contains data on one hundred and nine
(109) babies of age nine months old. The data was a primary data from three weighing centers in
the Mampong metropolis.
Table 4.1 Snapshot of the Data Collected for Analysis

Respondent

Sex
Feeding
Practice

BW

W1

W2

W3

W4

W5

W6

W7

W8

W9
1 1 1 3.2 5.2 6.5 7.3 7.9 7.8 8.2 8.5 9.1 9.1
2 2 2 2.6 3.9 4.8 5.5 6.2 6.9 7.6 7.6 7.8 8.0
3 2 2 3.0 4.9 5.6 6.8 7.7 8.5 9.6 10.2 10.5 10.9
4 2 1 3.3 5.0 6.4 7.2 8.0 8.8 9.7 10.5 10.5 10.4
5 2 2 3.0 5.0 6.5 6.9 7.5 8.3 9.3 10.0 10.0 11.0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
108 1 4 2.4 3.1 3.3 3.6 3.8 4.3 5.0 5.6 5.6 6.1
109 2 4 2.7 3.5 3.8 4.2 4.2 4.6 5.8 7.0 6.9 7.0
78

The table consists of 13 columns. Column 1 represents the respondents (babys) serial number,
column 2 the sex of the baby (1 for female and 2 for male), column 3 the feeding practice of the
baby, column 4 the birth weight of the baby and the successive 9 columns (i.e. column 5 to 13)
represents the weight of the baby for 1-9 months (denoted W1, W2, . . ., W9) respectively. The
feeding practices in column 3 are in four forms namely exclusive breastfeeding coded 1,
breastfeeding with water only coded 2, breastfeeding with water and any other food coded 3 and
feeding without breast milk coded 4. (These are shown in Appendix A).
4.2 EXPLORATORY DATA ANALYSIS
4.2.1 Descriptive Statistics
The birth weight of each of the babies was recorded and their monthly weights were also
recorded repeatedly over nine months. The sex and feeding practice of the baby were recorded to
be used to determine if sex, birth weight and feeding practices were the factors that affect the
weight gain of the babies.
For the random sample of 109 babies obtained for the study, 52 representing 47.71% were
females while 57 representing 52.29% were females. Thirty two (32) of the babies, representing
29.36%, were under exclusive breastfeeding, 30 representing 27.52% were under breastfeeding
with water only, 26 representing 23.85% were under breastfeeding with water and any other food
while 21 representing 19.27% were feeding without breast milk.
Statistics on sex and feeding practice indicates that, for the females, 16 were under exclusive
breastfeeding, 16 were under breastfeeding with water only, 13 were under breastfeeding with
water and any other food and 7 were under feeding without breast milk. For the males, 16 were
79

under exclusive breastfeeding, 14 were under breastfeeding with water only, 13 were under
breastfeeding with water and any other food and 14 were under feeding without breast milk.
Table 4.2 presents the mean weights of the babies under the four different forms of feeding
practice as well as their maximum and minimum weights.
Table 4.2 Summary statistics for feeding practices


Month


Measure
FEEDING PRACTICE
Exclusive
Breastfeeding
Breastfeeding
with water
Breastfeeding
with water
and any other
food
Feeding
without
breast
Milk

Mean
Maximum
Minimum
4.688
6.5
3.2
4.63
6.1
3.1
4.777
5.9
3.7
3.662
5.0
3.0

Mean
Maximum
Minimum
5.672
7.5
4.0
5.477
8.0
3.7
5.665
7.4
4.2
4.405
6.0
2.9

Mean
Maximum
Minimum
6.409
8.1
4.4
6.323
9.7
4.9
6.177
7.5
5.1
4.929
6.6
3.3

Mean
Maximum
Minimum
6.969
9.5
4.4
6.83
10.0
5.2
6.719
8.4
5.3
5.548
7.8
3.7
80

Mean
Maximum
Minimum
7.369
10.2
4.3
7.307
10.8
5.3
7.146
9.2
5.5
6.143
8.8
4.1

Mean
Maximum
Minimum
7.763
10.5
4.6
7.700
11.4
5.4
7.554
10.1
5.6
6.724
8.8
4.1

Mean
Maximum
Minimum
8.081
10.8
4.7
8.117
11.9
5.9
7.742
10.0
5.6
7.267
9.8
5.1

Mean
Maximum
Minimum
8.397
11.5
5.0
8.443
12.2
6.2
8.054
9.8
5.8
7.648
9.9
5.6

Mean
Maximum
Minimum
8.603
11.7
5.0
8.760
11.8
6.0
8.408
10.3
6.0
8.024
9.7
6.1

The pattern in the table shows a general increasing trend of weights for all the forms of feeding
practice. From the table, the mean weights of babies are higher for the exclusive breastfeeding
from the second month to the sixth month than the other three modes of feeding. Breastfeeding
with water only became higher than all the other three modes of feeding from the 7
th
to the 9
th

month. Weight for non-breastfeeding babies was lowest throughout the period.

81

4.2.2 Mean Weight of Babies under each feeding practice
A composite line graph for the four modes of feeding practices for the babies using mean
Fig 4.1 Mean weights of Babies by modes of feeding

weights of the babies is plotted in figure 4.1. From the graph it is observed that weights increase
by months for all modes of feeding. Weights of exclusive breastfeeding are higher after one
month and remain consistently higher than others up to about the seventh month. Feeding
without breast milk is the lowest for the whole study period. Breastfeeding with water only is the
WEIGHT
2
3
4
5
6
7
8
9
10
11
12
13
MONTH
1 2 3 4 5 6 7 8 9
PRACTICES Breastfeeding+water only
Exclusive Feeding without breastmilk
breastfeeding+water+other foods
82

second highest after the exclusive breastfeeding followed by breastfeeding with any other food
throughout the period.
4.2.3 Weight Gain
Table 4.3 represents the mean weights for all the babies for each month for the nine months.
Table 4.3 Mean weight of babies
Month Mean weight
1
2
3
4
5
6
7
8
9
4.495
5.372
6.045
6.597
7.062
7.495
7.853
8.183
8.488

The mean weight gain for the first three months is 1.55kg (6.045-4.495), fourth to sixth month is
0.898kg (7.495-6.597) and for seventh to ninth month is 0.635kg (8.488-7.853). This suggests
that the weight gain is higher in the first three months of life.
For various groups of feeding practices the mean weight gains were respectively 1.721kgs
(6.409-4.688), 1.639kg (6.323-4.630), 1.4kg (6.177-4.777) and 1.267kg (4.929-3.662) for
83

exclusive breastfeeding, breastfeeding with water only, breastfeeding with water and any other
food and feeding without breast milk for the first three months. The mean weight gains for the
next three months, from the 4
th
to the 6
th
months, were respectively, as in the same order,
0.794kg (7.763-6.969), 0.87kg (7.7-6.83), 0.835kg (7.554-6.719) and 1.176kg (6.724-5.548).
After the introduction of complimentary foods in the case of exclusive breastfeeding (from
seventh to ninth month) the weight of babies under breastfeeding with water only became highest
of all the different modes of feeding. This is shown in column four of Table 4.2 as 8.117, 8.443
and 8.760 for the 7
th
, 8
th
and 9
th
months respectively. The mean weight gain, however, for this
period were 0.522kg (8.603-8.081) for exclusive breastfeeding, 0.643kg (8.760-8.117) for
breastfeeding with water only, 0.666kg (8.408-7.742) for breastfeeding with water and any other
food, and 0.757kg (8.024-7.267) for feeding without breast milk. The result is the reverse order
of the mean weight gain by various feeding practices as compared to the first three months. This
information suggests that after the introduction of complimentary foods in the case of exclusive
breastfeeding, the babies rate of gaining weight slows down as compared to the other forms of
feeding practices. For the first six months, exclusive breastfeeding babies gain 3.08kgs (7.763-
4.688), breastfeeding with water only gain 3.07kgs (7.7-4.63), breastfeeding with water and other
foods gain 2.78kgs (7.554-4.777) and non-breastfeeding babies gain 3.06kgs (6.724-3.662).
A plot of weight by time with respect to gender indicates that males in general have higher rate
of weight gain than females. This is especially so for babies of age one to nine months. Figure
4.2 indicates this result. The plot shows the graph of males being consistently higher than that of
84

females, suggesting that males are generally heavier than females at the infant stage of life.

Figure 4.2 Plot of Weight by Month with respect to Gender.
4.2.4 Correlation between weights of Babies
General Linear models are employed in situations where the responses are correlated rather than
independent random variables. This occur, for instance, if they are repeated measurements on the
same subject or measurement on a group of related subjects obtained, for example, from
clustered sampling (Dobson, 2002). Table 4.4 below is drawn to show the correlation between
the monthly weights.

WEIGHT
2
3
4
5
6
7
8
9
10
11
12
13
MONTH
1 2 3 4 5 6 7 8 9
GENDER Female Male
85

Table 4.4 Correlation Coefficient for the weights of the babies

Weight for month
Weight
for
month

1

2

3

4

5

6

7

8

9
1
2
3
4
5
6
7
8
9

0.845
0.783
0.739
0.679
0.610
0.556
0.539
0.482


0.892
0.861
0.814
0.772
0.717
0.677
0.631



0.922
0.869
0.801
0.748
0.700
0.648




0.953
0.884
0.843
0.805
0.759





0.941
0.900
0.865
0.819






0.963
0.917
0.894







0.956
0.932








0.964


Generally, there are strong positive correlations between the weights for any two months. The
strength of the correlation, however, declines with increasing gap between the months. For
instance correlation coefficient between month 1 and month 2 is 0.845; month 1 and month 3 is
0.783 and continues to decline to 0.482 for month 1 and month 9. The P-value for each of the
correlation is 0.00 which signifies that there is significant correlation between any two monthly
weights.
86

4.3 INFERENTIAL STATISTICS
4.3.1 Introduction
Taken cognizance of the seeming disparities in the weights of the babies, this section seeks to
delve into what accounts for the phenomenon and to clarify any perception.
The tests were done with the level of significance of 5% (CI = 95%). The results from the
various analyses performed were obtained using the Minitab and SAS software.
4.3.2 One-way ANOVA for mean weights of three monthly periods
The nine months were divided into three age periods, each consisting of three months, which are
the first three months, 4-6 months and 7-9 months. Table 4.5 shows results of a one-way analysis
of variance of General Linear Model where age group (1-3 months, 4-6 months, 7-9 months) is
the factor with mean weight as the response variable.
Table 4.5 One-way ANOVA for mean weights of three monthly periods.
Source of
variation
DF SS MS F P
Age 2 393.17 196.58 96.49 0.000
Error 93 189.47 2.04
Total 95 582.64
In the table 4.5, the results show the age as been significant with the P-value of 0.000 at = 0.05
level of significant. This means that the assertion that the weight gains by age groups are the
same over the time period would be rejected and confirms the previous suggestion that in the
87

first three month the weight gain is higher. There are differences in weight gains by babies for
different months.
4.3.3 One-way ANOVA for Weights
Using the means of the birth weights and each of the nine months, a one-way analysis of
variance was conducted to find out if there exist differences. Table 4.6 shows the results from the
output of the analysis as produced by Minitab.
Table 4.6 One-way ANOVA: Birth weight and nine monthly weights


The P-value of 0.00 is less than =0.05 which implies that differences of the mean weights were
statistically significant. Thus there exists difference in mean weights at 5% significant level.
4.3.4 ANOVA for Birth weight
Considering feeding practice and sex as factors and Birth weight as response variable, an
analysis of variance conducted using Minitab gave results as shown in Table 4.7 below.



Source DF SS MS F P
Factor
Error
9
1080
2964.68
1347.84
329.41
1.25
263.96 0.00
Total 1089 4312.52
88

Table 4.7 ANOVA for Birth weight
Source DF SS MS F P
Feeding
Sex
Error
3
1
104
0.5809
0.0014
22.6917
0.1936
0.0014
0.89
0.01
0.45
0.936
Total 108

Since the P-value = 0.45 = 0.05, the test is not significant. Also P-value = 0.936 = 0.05,
the test is not significant. This indicates that birth weights are the same for the different feeding
practices and sexes. Naturally, birth weights can never be influenced by feeding practice.
4.3.5 ANOVA for each of the nine months using Adjusted SS for Tests.
Using weight as response variable, feeding practice and sex as factors, an analysis of variance
was conducted for each of the nine months, using Adjusted Sum of Squares for tests. The results
of this from Minitab output are shown in Table 4.8. There is indication from the output that there
were significant differences in weights of the babies under the four forms of feeding practices for
the first six months of life. For each of the six months the P-value is less than an alpha level of
0.05. However, for the 7
th
, 8
th
and 9
th
months, the weights were found to be statistically not
significantly different for the various feeding practices since the P-value were either greater or
equal to 0.05 (0.05, 0.08, 0. 17 respectively).


89

Table 4.8 ANOVA for each of the nine months weight
MONTH SOURCE DF Seq SS Adj SS Adj MS F P
1 FEEDING 3 18.3740 19.5497 6.5166 14.46 0.000
SEX 1 1.6422 1.6422 1.6422 3.64 0.059
ERROR 104 46.8715 46.8715 0.4507
TOTAL 108 66.8877
2 FEEDING 3 25.0907 26.1206 8.7069 12.13 0.000
SEX 1 1.2028 1.2028 1.2028 1.68 0.198
ERROR 104 74.6639 74.6639 0.7179
TOTAL 108 100.9574
3 FEEDING 3 33.1999 35.2999 11.7666 13.27 0.000
SEX 1 2.8572 2.8572 2.8572 3.22 0.076
ERROR 104 92.2327 92.2327 0.8869
TOTAL 108 128.2897
4 FEEDING 3 29.565 30.952 10.317 9.44 0.000
. SEX 1 1.667 1.667 1.667 1.53 0.220
ERROR 104 113.657 113.657 1.093
TOTAL 108 144.889





90

5 FEEDING 3 22.732 24.514 8.171 6.20 0.001
SEX 1 2.717 2.717 2.717 2.06 0.154
ERROR 104 137.066 137.066 1.318
TOTAL 108 162.516
6 FEEDING 3 16.130 18.003 6.001 4.31 0.007
SEX 1 3.655 3.655 3.655 2.63 0.108
ERROR 104 144.642 144.642 1.391
TOTAL 108 164.428
7 FEEDING 3 11.291 13.208 4.403 2.77 0.045
SEX 1 5.213 5.213 5.213 3.29 0.073
ERROR 104 165.047 165.047 1.587
TOTAL 108 181.551
8 FEEDING 3 9.950 11.582 3.861 2.35 0.077
SEX 1 4.442 4.442 4.442 2.70 0.103
ERROR 104 170.818 170.818 1.642
TOTAL 108 185.210
9 FEEDING 3 7.336 8.719 2.906 1.69 0.173
SEX 1 3.965 3.965 3.965 2.31 0.132
ERROR 104 178.533 178.533 1.717
TOTAL 108 189.834

Considering the sex, there were no significant differences in weights for all the nine months.
Thus, differences in weights of babies are not dependent on sex of the baby.
91

4.3.6 Repeated Measures Analysis of Variance for Gender
As Tests for Hypotheses for Between Subjects Effects, the comparison test between male and
female weight gain at the first nine months of life was performed using Univariate Repeated
Measures Analysis of Variance. Table 4.9a below shows the results as for the dependent variable
(Weight) versus the independent variable (gender) using SAS.
Table 4.9a Analysis of Variance for Gender (Between Subject Effects)
(Test of Hypotheses for Between Subjects Effects)
Source DF Type 111 SS Mean Square F-Value Pr F
Gender 1 10.616209 10.616209 1.07 0.3038
Error 1.07 1063.992438 9.943855
The results show that the between subject effect is not significant. Thus gender is not statistically
significant in terms of weight gain. This is indicated by P-value of 0.3038 being greater than =
0.05 level of significance. This means that we fail to reject the hypothesis that the mean weight
gain for the nine months is the same. Sex is therefore not an influential factor on growth at the
infant level as indicated earlier using the Minitab.
Univariate tests of Hypotheses for Within Subjects effects produced the results given in Table
4.9b using SAS.


92

Table 4.9b Univariate Analysis of Variance for within Subject effects
(Univariate Tests of Hypotheses for Within Subjects Effects)
Adj. Pr > F
Source DF Type111 SS Mean Square F-Value Pr F G-G H-F
Sex 7 884.6088 126.3727 522.75 0.0001 0.0001 0.0001
Sex*Sex 7 1.9994 0.2856 1.18 0.3108 0.3119 0.3126
Error (sex) 749 181.0679 0.2417
The results in table 4.9b present insignificant test for the sex by sex interaction with P value of
0.3108 at 5% level of significance. It is concluded therefore that the effect of interaction of sexes
is not significant on weight gain by the children. It must be noted here that since sex is composed
of two mutually exclusive sets (male and female), there is no possibility of interaction of sexes to
produce any effect.
4.3.7 Univariate Analysis of Variance by Feeding Practice
Table 4.10a and 4.10b shows the Repeated Measures ANOVA test of Hypotheses for Between
Subjects Effects and for within subject effects respectively.


93

Table 4.10a Tests of Hypotheses for Between Subjects Effect. (ANOVA for Feeding
Practice)
Source DF Type 111 SS Mean Square F-value Pr F
Practice 3 140.0670 46.6890 5.25 0.0021
Error 105 934.5417 8.9004
The Between Subjects effects test in table 4.9a produces a P-value of 0.0021. This does not
support the hypothesis of equal mean effects. Hence the test for feeding practice is significant
indicating that the weight of babies is influenced by the feeding practice.
Table 4.10b Univariate Analysis of Variance for within Subject Effects
(Univariate Tests of Hypotheses for Within Subjects Effects)
Adj. Pr > F
Source DF Type111 SS Mean Square F-Value Pr F G-G H-F
Practice 7 888.4970 126.9281 555.84 0.0001 0.0001 0.0001
Prac.*Prac. 21 15.2276 0.7251 3.18 0.0001 0.0030 0.0025
Error (Prac.) 735 167.8398 0.2284
In the within subject effect test above (Table 4.8b), the interaction of feeding practices is shown
to be significant at = 0.05 significance level with P-value of 0.0001.

94

4.3.8 Repeated Measures Multivariate Analysis of Variance (MANOVA)
The factors, feeding practices and sex are considered for multivariate analysis using repeated
measures analysis. Tables 4.11 (a, b and c) show the output for extreme measures for the factors.
There are seven columns in each of the tables. The effects of the factors are in the first column,
the extreme measures for each factor (statistic) in column two, the values of the extreme
measures in column three and F values in column four. Numerator degrees of freedom,
Denominator degrees of freedom and the P-value are in columns five, six and seven respectively.
Table 4.11 Multivariate Repeated Measures Analysis on Weight Gain
(a) MANOVA Test Criteria and Exact F Statistics for the Hypothesis of no SEX Effect
Effect Statistic Value F Num DF Den DF Pr F
Sex Wilks Lambda 0.0930 140.74 7 101 0.0001
Pillais Trace 0.9070 140.74 7 101 0.0001
Hotelling-Lawley Trace 9.7545 140.74 7 101 0.0001
Roys Greatest Root 9.7545 140.74 7 101 0.0001




95

(b) MANOVA Test Criteria and Exact F Statistics for the Hypothesis of no PRACTICES
Effect
Effect Statistic Value F Num DF Den DF Pr > F
Practice Wilks Lambda 0.0880 146.70 7 99 0.0001
Pillais Trace 0.9121 146.70 7 99 0.0001
Hotelling-Lawley Trace 10.3729 146.70 7 99 0.0001
Roys Greatest Root 10.3729 146.70 7 99 0.0001
(c) MANOVA Test Criteria and F Approximations for the Hypothesis of no
PRACTICES*PRACTICES Effects.
Effect Statistic Value F Num DF Den DF Pr > F
Prac*Prac Wilks Lambda 0.7320 1.56 21 284.82 0.0591
Pillais Trace 0.2908 1.55 21 303 0.0606
Hotelling-Lawley Trace 0.3355 1.57 21 200.45 0.0607
Roys Greatest Root 0.1983 2.86 7 101 0.0091
The results in the tables show that both sex and feeding practice are influential factors in
determining the weight gain of children of age 1 9 months. This is as a result of p-value for all
the extreme measures being 0.0001 at 5% significance level.

96

4.4 FURTHER INFERENTIAL ANALYSIS
Table 4.12 represents covariance structures investigated in selecting an appropriate covariance
structure of the random errors.
Table 4.12 Covariance Structures Investigated
COVARIANCE STRUCTURE AIC AICC BIC
COMPOUND SYMMETRY 2062.8 2062.8 2068.2
UNSTRUCTURED 1271.3 1275.7 1392.4
FIRST ORDER AUTOREGRESSIVE 1407.4 1401.4 1406.7
FIRST ORDER ANTE-DEPENDENCE 1257.6 1258.2 1303.3
HETEROGENEOUS FIRST ORDER
AUTOREGRESSIVE
1350.0 1353.2 1379.7
HETEROGENEOUS COMPOUND
SYMMETRY
1963.8 1964.1 1990.7
HETEROGENEOUS TOEPLITZ 1333.7 1334.3 1379.4
FIRST ORDER AUTOREGRESSIVE
MOVING AVERAGE
1394.8 1394.9 1402.9
HUYNH-FELDT 1974.5 1974.7 2001.4
TOEPLITZ 1373.5 1373.7 1397.7
VARIANCE COMPONENT 2876.3 2876.3 2879.0
From the table the smallest value is 1257.6 which is in the row of the First Order Ante-
Dependence corresponding to the AIC column. Thus the best covariance structure to be used is
the Ante-Dependence.
97

4.4.0 Longitudinal Model with Ante-dependence Covariance Structure (Mixed Procedure)
4.4.1 Model Information
The Model Information table, Table 4.13, shows the name of the data set, the dependent variable,
the covariance structure used in the model, the subject effect, the estimation method to compute
the parameters for the covariance structures (the default estimation method is REML), the
residual variance method, the fixed effects standard error method and the method to compute the
degrees of freedom.
Table 4.13 Model information
Data Set SASUSER.SEK
Dependent Variable WEIGHT
Covariance Structure Ante-dependence
Subject Effect RESPONDENTS
Estimation Method REML
Residual Variance Method None
Fixed Effects SE Method Model-Based
Degrees of Freedom Method Satterthwaite

4.4.2 Dimensions
The Dimensions table below, Table 4.14, lists the sizes of the relevant matrices. The table is
useful in determining CPU time and memory requirements.

98

Table 4.14 Dimensions
Covariance Parameters 17
Columns in X 9
Columns in Z 0
Subjects 109
Max Obs Per Subject 9
Number of Observations Read 986
Number of Observations Used 981
Number of Observations Not Used 5

4.4.3 Iteration History
The Iteration History table (Table 4.15) describes the optimization of the residual log likelihood.
The minimization is performed using a ridge-stabilized Newton-Raphson algorithm, and the
rows of the table describe the iterations that this algorithm takes in order to minimize the
objective function.






99

Table 4.15 Iteration History
Iteration Evaluations -2 Res Log Like Criterion
0 1 2874.32193062
1 2 1238.12934187 0.04301972
2 1 1230.70751929 0.02156881
3 1 1224.01108546 0.00134184
4 1 1223.58269601 0.00009122
5 1 1223.55544908 0.00000076
6 1 1223.55523153 0.00000000
*Convergence criteria met
4.4.4 Covariance Parameter Estimates
The covariance Parameter Estimates table (Table 4.16 in Appendix C) shows the parameter
estimates for the Ante-dependence covariance structure.
4.4.5 Fit Statistics
The Fit Statistics table provides information to be used to select the most appropriate covariance
structure. Akaikes Information Criterion (AIC)(Akaike,1974) penalizes -2 residual log
likelihood by twice the number of covariance parameters in the model. The smaller the value, the
better the model. The finite-sample version of the AIC (AICC) is also included. The Schwarzs
Bayesian Information Criterion (BIC)(Schwarz,1978) also penalizes the -2 residual log
likelihood. In The Fit Statistics table below (Table 4.17) AIC is the best option.

100

Table 4.17 Fit Statistics
-2 Res Log Likelihood 1223.6
AIC (smaller is better) 1257.6
AICC (smaller is better) 1258.2
BIC (smaller is better) 1303.3
4.4.6 Null Model Likelihood Ratio Test
The Null Model Likelihood Ratio Test shows a test that determines whether it is necessary to
model the covariance structure of the data at all. The test statistic is -2 times the log likelihood
from the null model (model with an independent covariance structure) minus -2 times the log
likelihood from the fitted model. The p-value is used to assess the significance of the model fit.
Therefore for the p-value of 0.0001, the model is considered significant at 5% level of
significant. This is shown in table 4.18.
Table 4.18 Null Model Likelihood Ratio Test
DF Chi-Square Pr ChiSq
16 1650.77 0.0001
4.4.7 Solution for Fixed Effects
Table 4.19 contains the solution for Fixed Effects. It contains the estimates, standard errors, the
degrees of freedom, the t-value and the p-values for each of the fixed effects listed in column
one. From the table, the initial weight (birth weight), monthly weight gains, and the gender are
shown to be significant. This is so as the p-value for each of the effects mentioned is 0.0001.
101

Feeding without breast milk has p-value 0.023 which shows that the practice is significant.
Exclusive breastfeeding and Breastfeeding with water only are found not to be significant.
Table 4.19 Solution for Fixed Effects
Effect Gender Practices Estimate Std. Error DF t-value Pr |t|
Intercept 1.8824 0.3830 108 4.91 0.0001
Birth weight 0.7969 0.1151 108 6.92 0.0001
Month 0.4275 0.01241 157 34.43 0.0001
Gender Female -0.2450 0.1063 108 -2.31 0.0230
Gender Male 0 . . . .
Practice BFW -0.03157 0.1483 108 -0.21 0.8318
Practice Exclusive -0.09133 0.1449 108 -0.63 0.5298
Practice NBF -1.0019 0.1630 108 -6.15 0.0001
Practice BWFO 0 . . . .
4.4.8 Type 3 Tests of Fixed Effects
The Type 3 Test of Fixed Effects table (Table 4.20) shows the hypothesis test for the
significance of each of the fixed effects. A p-value is computed from an F distribution with the
numerator and denominator degrees of freedom. The p-values are given as 0.0001 for each of the
following fixed effects: Birth weight, monthly weight gain and Feeding Practices. The p-value
for Gender is 0.0230. Thus these effects are considered significant under this test.
102

Table 4.20 Type 3 Test of Fixed Effects
Effect Num DF Den DF F-Value Pr F
Birth weight 1 108 47.92 0.0001
Month 1 157 1185.55 0.0001
Gender 1 108 5.32 0.0230
Practices 3 108 16.95 0.0001












103

CHAPTER FIVE
5.0 DISCUSSIONS, CONCLUSIONS AND RECOMMENDATIONS
5.1 DISCUSSIONS OF RESULTS
The study used a random sample of 109 babies. The weights of these babies were recorded
repeatedly for nine months. The birth weight and the feeding practice of each of the babies were
also recorded. The study was intended to find out whether the monthly weight of babies recorded
at the weighing centers were significantly different and if birth weight, age, sex and feeding
practice could contribute to weight changes.
At the exploratory analysis stage it was found out that there was generally an increasing trend of
weights for all the babies. The descriptive statistics of the mean monthly weights of the babies
suggested that those under exclusive breastfeeding had higher weights than the others for the first
six months after birth. The babies under breastfeeding with water only and breastfeeding with
water and any other food were second and third respectively in terms of higher weights. The
non-breastfed babies had the lowest mean monthly weights throughout the study period.
Considering the mean monthly weight gain for the three age groups, the babies within ages 0-3
months had 2.977kg while those within 4-6 and 7-9 had 1.45kg and 0.993kg respectively. This
suggested that the weight gain was higher in the first three months of life and decreased through
the next age groups. For the four groups of feeding practices the weight gains were respectively
3.312kg, 3.34kg, 3.015kg and 1.948kg for exclusive breastfeeding, breastfeeding with water
only, breastfeeding with water and any other food and feeding without breast milk for the first
three months of life. The mean weight gains for the 4
th
, 5
th
, and the 6
th
months were 1.354kg for
exclusive breastfeeding, 1.377kg for breastfeeding with water only, 1.377kg for breastfeeding
104

with water and any other food and 1.795kg for feeding without breast milk. The weight gains for
the first six months after birth confirmed the assertion that exclusive breastfeeding is the best
form of feeding for babies. The weight gains were 4.7kg for exclusive breastfeeding, 4.7kg for
breastfeeding with water only, 4.4kg for breastfeeding with water and any other food and 3.7kg
for non-breastfeeding.
There was seemingly a drastic decrease in weight gain for babies under exclusive breastfeeding
from the 7
th
to the 9
th
month which was just after introduction of complimentary food to the
babies although there was general decreasing trend of weight gain. The weight gains as in the
same sequence were 0.84kg, 1.06kg, 0.85kg and 1.3kg. A composite line graph of mean weights
against age in months in figure 4.1 suggested such a shocking revelation of non-breastfed babies
having high rate of weight gain at the latter stages of the study period. A Plot of weight by time
with respect to gender indicated that males have high rate than females (fig 4.2) as the graph for
male was consistently higher.
A table of correlation coefficient gave strong positive correlations between any two different
months. The strength of the correlation, however, declines from the closer months to the months
wider apart. This prompted the use of general linear model for further analysis.
An analysis of variance conducted using General linear model method revealed that age is
significant in determining the weight of babies under nine months old. Thus the older the baby
the heavier the baby as exposed by the increasing trend in weight by month is confirmed.
At 5% significant level the one-way ANOVA revealed that there exist differences in the mean
weights of babies. There was also an indication that there were significant differences in weights
of the babies under the four forms of feeding practice for the first six months. For the 7
th
, 8
th
and
105

9
th
months, the analysis indicated that there were not significant differences in weights under the
feeding practices.
Contrary to the suggested assertion that male babies weigh more than female as revealed by the
plot in fig 4.2, the analysis by the general linear model showed that sex does not influence
weight gain. For all of the nine months, sex was found to be not statistically significant at 5%
level. This called for further investigation.
Using Repeated Measures Multivariate Analysis of variance (MANOVA), sex was found to be
significant for all the statistics; Wilks Lambda, Pillais Trace, Hotelling-Lawley Trace and
Roys Greatest Root at 5% level. Similarly, the MANOVA test criteria and exact F statistics for
the hypothesis of no feeding practices effect confirmed again that there are differences in weights
under the feeding practices. Thus, feeding practice affects the weight gain. There were no
interaction effects of sexes and feeding practices on the weights.
Since the measurement were repeated measurements over time on the same subject the data is a
longitudinal data requiring longitudinal data analysis. Furthermore, the plot of mean monthly
weight of the feeding practices against the time (age in months) illustrated longitudinal
relationship in the exploratory analysis. The observations were positively correlated, which often
occurs with longitudinal data, and meant that the variances of the time-independent predictor
variables are underestimated if the data is analyzed as though the observations are independent.
In other words the Type I error rate is inflated for these variables. For time-dependent predictor
variables also, ignoring positive correlation leads to a variable estimate that is too large. In other
words, the Type II error rate is inflated for these variables.
106

Special methods of statistical analysis are needed for longitudinal data such as this because the
set of measurements on one subject tends to be correlated; measurements on the same subject
close in time tend to be more highly correlated than measurements far apart in time and the
variances of longitudinal data often change with time. These potential patterns of correlation and
variation may combine to produce a complicated covariance structure. This covariance structure
must be taken into account to draw valid statistical inferences. Therefore, standard regression
and ANOVA models may produce invalid results because two of the parametric assumptions
(independent observation and equal variances) may not be valid as stated and documented in
chapter 3 (methodology).
The linear mixed model allows a very flexible approach to modeling longitudinal data. The data
structure is similar to univariate ANOVA where the number of observations equals the number
of measurements for all the subjects. This means the data does not have to be balanced. An
advantage of fitting linear mixed model is that PROC MIXED uses all the available data in the
analysis.
PROC MIXED offers a wide variety of covariance structures. This enables the user to directly
address the within-subject correlation structure and incorporate it into a statistical model. By
selecting a parsimonious covariance model that adequately accounts for within-subject
correlations, the user can avoid the problems associated with univariate and multivariate
ANOVA using PROC GLM (Little, Freund and Spector, 2001).
Using the information criteria in selection, it was realized that 1
st
order Ante-Dependence was
the most appropriate covariance structure for the data for the study. In Model Information
SASUSER SEX was the data set, Weight was the dependent variable, RESPONDENTS were the
107

Subject Effect, Estimation Method employed was REML, Residual variance method was none,
the Fixed Effects Standard Error Method was Model-Based and the Degrees of Freedom Method
was Satterthwaite.
The Dimensions for the matrices were 17 covariance parameters, 9 columns in X matrix with 0
columns in Z matrix and maximum observations per subject were 9. There were 109 subjects,
986 Number of observations read, 981 Number of observations used and 5 Number of
observations not used.
The iteration history indicated that convergence criteria were met, Covariance Parameter
estimates were given as; respondents are the subjects; variables (1), variables (2), ,variable (9);
Rho (1), Rho (2), , Rho (9) are the covariance parameters with their corresponding estimates
listed respectively.
AIC was the appropriate Fit statistics used to select the most appropriate covariance structure
which was the Ante-dependence.
The model validated the graph of the mean monthly weights in the exploratory analysis section
and other assertion such as relationship between feeding practice and weight gain, initial weight
(or birth weight) and weight gain and monthly weight gains. Thus, it confirmed that feeding
practices affect weight of babies, birth weight affect weight of babies and monthly weight gains
are different. Sex was found to be influential in weight of babies which is the biggest change.
This illustrates how changing the covariance structure can change the inferences made from
model.
Using an output generated by General Linear Model, and judging from the contributions of each
of the feeding practices to the model, the exclusive breastfeeding is found to be very significant,
108

followed by breastfeeding with water only with non-breastfeeding and breastfeeding with water
and any other food being not significant for the first six months. Furthermore, considering the
solution for fixed effects, using BFWO as the base, both EBF and BFW are found not to be
significant while NBF is found to be significant.
5.2 CONCLUSIONS
From the review of the related research works, proper nutrition helps give every child the best
start in life. Malnutrition of children (0-59 months) is a public health concern in Africa,
particularly in the Sahelian countries. Breastfeeding has been found to reduce the occurrence of
the pressing public health challenge of malnutrition.
From the analyses, it is concluded that the type of feeding practice of a baby affect weight gain
significantly.
It is also concluded that the mean weight gain is higher in the first three months of life compare
to that of 4
th
6
th
and 7
th
9
th
months.
The birth weight of a baby has a significant influence on weight gain at the early stage of growth.
The Exclusive Breastfeeding babies rate of weight gain is highest for the first six months but
slows down drastically and becomes the second lowest after the introduction of complimentary
foods. The non-breastfed babies have rate of weight gain to be the lowest throughout the study
period. The study has therefore confirmed that Exclusive Breastfeeding is the best form of
feeding practice for a baby for the first six months after birth.
Comparatively, there is no clear difference in weight gain between female and male babies as
there was inconsistency in the results provided by various methods of analysis.
109

5.3 RECOMMENDATIONS
1. There is need for promotion and protection of optimal infant feeding practice for improving
nutrition since infant feeding practices constitute a major component of child caring practices.
2. Nursing mothers should be encouraged to frequently report at the weighing centers to
facilitate the early detection of ill-health in babies and other early childhood diseases.
3. The ministry of health should open more weighing centers and put up mobile teams to go
round the communities to solve the unfriendly overcrowding situations in the existing ones that
deter mothers from participating in the exercise.
4. Educational programs on Exclusive Breastfeeding for the first six months of life should be
intensified, especially by radios in order to reach the remote areas of the country.
5. Parents or guardians who have babies who do not have the benefit of breastfeeding should
consider timely admission to Babies Homes as urgent and important.
6. The importance of adequate complementary feeding in the second half of infancy needs to be
stressed in nutrition education programmes.





110

REFRENCES
1. Bergh, D. D. (1985) Problems with Repeated Measures Analysis: Demonstration with a
study of Diversification and Performance Relationship. The Academy of Management
Journal, 38(6), 1692 1708.
2. Center for Disease Control and Prevention (May 2007), Promotion and support of
Breastfeeding and obesity Prevention.
3. Chi, E.M., and Reinsel, G.C. (1989), Models for longitudinal data with random effects
and AR(1) errors, Journal of the American StatisticalAssociation, 84:452-459.
4. Crowder, M. J. and Hand, D. J. (1990) Analysis of Repeated Measures. New York:
Chapman and Hall.
5. Diggle, P.J. (1990), Time Series: a Biostatistical Introduction, Oxford: Oxford University
Press.
6. Diggle, P.J., Liang, K., and Zeger, S.L. (1994), Analysis of Longitudinal Data, New
York: Oxford University Press.
7. Dinesh K., Goel N.K. and Poonam C. (2006). Influence of Infant-Feeding Practice on
Nutritional Status of Under-Five Children. Department of Community Medicine, GMCH,
Sector 32a, Chandigarh, India.
8. Dramane M. and Carolyn H (2006). An Exploratory Analysis of child Nutritional Status
in the Sahel, the Goundam Circle Case Study-Timbuktu Region-Mali.
9. Dunlop, D.D. (1994), Regression for Longitudinal Data: A Bridge from Least Squares
Regression, The American Statistician, 48:299-303.
111

10. Fawzi, W. Herrera, MG. Nestel, P. Amin, A. and Mohammed, A. (1998). Longitudinal
study of prolonged breastfeeding in relation to child undernutrition. Int. J.Epidemiol.
27(2): 255-60.
11. Fomon, Samuel J. (2001). Infant Feeding in the 20
th
century: Formula and Beikost. San
Diego, CA: Department of Pediatrics, College of Medicine, University of Iowa.
12. Guerin, L., and Stroup, W.W. (2000), A Simulation Study to Evaluate PROC MIXED
Analysis of Repeated Measures Data, Proceedings of the Twelfth Annual Kansas State
University Conference on Applied Statistics in Agriculture, April 30 May 2, 2000
13. Ghana Statistical Service (GSS) and Macro International Inc. (MI)(1999). Ghana
Demographic Survey 1998. Calverton, Maryland, U.S.A.
14. Ghana Statistical Service (GSS) and Macro International Inc. (MI)(2003). Ghana
Demographic Survey 1998. Calverton, Maryland, U.S.A.
15. Horta, BL. Bahl, R. Martins, JC. Victoria. CG (2007) Evidence on the long-term effects
of Breastfeeding: Systematic reviews and meta-analyses Geneva, Switzerland: WHO
ISBN 978-92-4-159523-0
16. Horton, S. Sanghvi, T. Phillips, M. et al (1996) Breastfeeding promotion and priority
setting in health. Health Policy Plan 11(2) : 156- 68 dio:10.1093/heapol/11.2 156
PMID 10158457
17. Ip, S. Chung, M. Raman, G. et al (2007). Breastfeeding and maternal and infant health
outcomes in developed countries. Evid Rep Technol Assess (Full Rep)(153): 1-186PMID
17764214.
112

18. Jelliffe, Derrick B. and E. F. Patrice, Jelliffe, (1990) Growth Monitoring and
Development of Young Children: Selection of Methods and Techniques of Training. New
York; Oxford University Press.
19. Kapur, D. Shama, S. and Agarwal, KN (2005). Dietary Intake and Growth Pattern of
Children 9-36 months of Age in an Urban Slum in Delhi. India pediatrics 42: 351-356
20. Kotz, S. and Johnson, N.L. (1998). Encyclopedia of Statistical Sciences (Vol.8). New
York: John Wiley and Sons. 125.
21. Lucas, A. Cole, T.J. (1990) Breastmilk and neonatal necrotizing enterocolitis Lancet
336 (8730): 1519-152 doi: 10.1016/0140-6736(90)93304-8 PMID 1979363.
22. Littell, R.C. Freund, R.J. Spector, P.C. (2002) SAS System for Linear Models, Fourth
Edition, Cary, NC: SAS Institute Inc.
23. Littell, R.C. Henry, P.R. and Ammerman, C.B. (1998) Statistical Analysis of Repeated
Measures Data Using SAS Procedures, Animal Journal of Animal Science, 76:1216-
1231.
24. Littell, R.C. Milliken, G.A. Stroup, W.W. and Wolfinger, R.D. (1996) SAS System for
Mixed Models, Cary, NC: SAS Institute Inc.
25. Little, R.J.A. and Rubin, D.B. (1987) Statistical Analysis with Missing Data, New York:
John Wiley & Sons.
26. Love, H., J. De Graat and G. Gordon, (1985) How Mothers Measure Growth: Results of
Survey in Ghana. Mothers and Children Bulletin on Infant Feeding and Maternal
Nutrition 5(1): 1-4.
27. Markos, Kaminis. (Jan 11, 2005) Business Week. A Growing Boost for Baby Formula.
113

28. Maseta, E. Kogi-Makau, W.and Omwega, A.M. (2008) Children Practices and
Nutritional Status of Children Aged 6-36 months. University of Nairobi, Kenya.
29. Montgomery, Douglas C. Peck, Elizabeth A. and Vining, G. (2001) Introduction to
Linear Regression Analysis. New York: John Wiley and sons, Inc.
30. Mustila, T. Raitanen, J. Keskinen P. (2011). Lifestyle counseling targeting infants
mother during the childs first year and offspring weight development until 4 years of
age: follow-up study of a cluster RCT. BMJ Open 2012; 2: e000624.
doi: 101136/bmjopen-2011-000624
31. Nathoo,Tasnim. Ostry, Aleck (2009) The one best way? : Breastfeeding History, Politics
and Policy in Canada. Wilfrid Laurier University Press.
32. National Family Health Survey (1998-1999). International Institute for Population
Sciences, Mumbai, India.
33. Ody, Wendy H. Kendall, GE. et al (2010). The long-term effects of breastfeeding on
child and adolescent mental health: a pregnancy cohort study followed for 14 years. The
journal of Pedriatrics 156(4): 568-74 doi: 10-1016/j.j peds. 2009. 10.020 PMID
20004910
34. Oliver, Lynne-(2004), Food Timeline-history notes: baby food.
35. Patetta, M. et al, (2002) Longitudinal Data Analysis with Discrete and Continuous
Responses Course Notes. SAS Institute Inc., Cary, NC 27513, USA.
36. PEDIATRICS. Volume 115 No. 2 Feb.2005 pp 496-506 doi: 10.1542/peds. 2004-2491
Policy statement: Breastfeeding and the use of Human milk, section on Breastfeeding(1)
37. Picciano, M.2001. Nutrient composition of Human milk. Pediatric in North Americal
48(1): 53-67 doi: 10.1016/50031-3955(05)70285-6 PMID 11236733
114

38. Riordan, BA. Countryman, BA. (Jul-Aug 1980) Basics of breastfeeding. Part 1:Infant
feeding patterns past and present. JOGN Nurs 9(4):207-210 doi: 10.1111/j 1552-
6909.1980.tb 02778 PMID 7001126
39. Riordan, JM. (1997) The cost of not breastfeeding: a commentary. J Hum Lact 13(2):
93-97 doi: 10.1177/089033449701300202 PMID 9233193
40. Sala E.O. (2006) Factors Affecting Prevalence of Malnutrition among Children under
Three years of Age in Botswana. Rural Outreach Program, Nairobi-Kenya.
41. Schuman, A. (2003) A concise history of Infant formula (twists and turns included).
Contemporary Pediatric.
42. Shock, N.W. et al. (1984) Normal Human Aging: The Baltimore Longitudinal Study of
Aging, National Institutes of Health Publication 84- 2450. Washington, DC: National
Institutes of Health.
43. Sommer, A. Katz, J. and Tarwotjo, I. (1984) Increased risk of respiratory infection and
diarrhea in children with pre-existing mild vitamin A deficiency, American Journal of
Clinical Nutrition, 40:1090-1095. A.3 References A-19
44. Tharakan, C.T. and Suchindran, C.M. (1999) Determinants of child malnutrition- an
intervention model for Botswana. Nutr. Res. 19(6): 843-860.
45. UNICEF. (1998) A UNICEF Nutrition Information strategy improving decision-making
at household, community and national levels by U. Jonsson Urban, D. Pelletier and R.
Shrimpton. New York.
46. UNICEF. (2007) Infant and Young Child Feeding and Care..
47. UNICEF.org.media, (2008): Key Facts http://www.unicef.org.media/media
30011.html(accesed: 23/09/12)
115

48. United Nations Childrens Fund, (1996), United Nations, T-Plaza 4-1234C, New York.
NY10017, USA.
49. Verbeke, G. and Lesaffre, E. (1997) The effect of misspecifying the random effects
distribution in linear mixed models for longitudinal data, Computational Statistics and
Data Analysis, 23:541-556.
50. Verbeke, G. and Molenberghs, G. (1997) Linear Mixed Models in Practice: A SAS-
Oriented Approach, New York: Springer-Verlag, Inc.
51. Verbeke, G. and Molenberghs, G. (2000) Linear Mixed Models for Longitudinal Data,
New York: Springer-Verlag, Inc.
52. Wells, J.C.K. (1996) Nutritional Considerations in infants formula design. Seminars in
Fetal and Neonatal Medicine, 1(1): 19-26
53. World Health Assembly, (2001) Global strategy on infant and young child feeding.
Provisional agenda item 13.1. Geneva: WHO.
54. WHO, (1981) International Code of Marketing of Breast-milk Substitutes.
ISBN 978-92-4-154160-2
55. World Health Organization (WHO), (1998) Complementary feeding of young children in
developing countries. A review of current scientific knowledge. Geneva: WHO
56. WHO/UNICEF, (2003) Exclusive Breastfeeding
http:/www.who:int/nutrition/exclusive breastfeeding/(accessed: 12/01/13)
57. WHO, (2003) Global strategy for infant and young child feeding-Geneva.
58. Zeger, S.L. and Liang, K. (1986) Longitudinal Data Analysis for Discrete and
Continuous Outcomes, Biometrics, 42:121-130.
116

59. Zeger, S.L. and Liang, K (1992) An Overview of Methods for the Analysis of
Longitudinal Data, Statistics In Medicine, 11:1825-1839.