Chapter Two

2. Determinants of Total Fertility

Introduction
As the previous chapter notes, the demographic transition in Venezuela during the
1970s and 1980s presents certain deviations from the norm. Its decline stalls and
reverses twice, first from 1973 to 1978 and then again from 1987 to 1991. Since these
stalls and reversals seem to defy the natural decline in fertility expected from a typical
demographic transition, it is hypothesized that the anomalies in the period fertility rates
are not associated with changes in total fertility, but with an acceleration of childbearing.
Certain circumstances, such as extreme or unusual changes in economic conditions, can
cause an acceleration in childbearing that distorts the period fertility rate in Venezuela,
but leaves the total fertility rate almost unchanged or at a lower level. If this were the
case, the period fertility rates could have increased considerably and later decreased,
leaving the total fertility rate at the same or a lower level. To mechanize this argument, a
simple two period model of fertility is developed in chapter three. It illustrates how
females may change their fertility timing preferences by responding to temporary
changes in income, caused by either changes in wages or changes in the amount of
transfers, thereby leaving their total fertility after the two periods at a lower level or
unchanged.
Before the two period model is developed, two aspects of the mid 1970s and early
1980s need to be explored empirically to determine the true nature of the increases in the
period fertility rates presented in Chapter one. To answer the question of whether the
stall and increase in the period fertility rates during the 1970s are associated with changes
in total fertility or whether they are associated with changes in the timing of births, a look
at total fertility and the timing of fertility is needed. This chapter, therefore, concentrates
on exploring total fertility and its determinants during the period in question. By
estimating a model of total fertility by union cohort, the hypothesis that the total fertility
did not increase (as suggested by the increases in period fertility rates), but instead has
decreased for every subsequent cohort since 1967, is tested. In other words, is it possible
to associate the increase in the period fertility rate during the mid 1970s, presented in

Page 19
Chapter one, with changes in total fertility for females, even after controlling for
individual characteristics of females? To do this, a model of total fertility is estimated
and union cohort dummies are used to detect changes in the total fertility for different
union cohorts.

2.1. The Theory of Demand for Children

According to the theory of demand for children, people who decide to have a
child weight the benefits of an additional child against the additional cost. The theory’s
framework is the maximization of utility; if greater utility can be obtained from an
alternative to an additional child, then that alternative will be chosen. Decline in fertility
thus implies that the relative price of a child has increased, couple incomes have fallen or
there has been a change in the shape of the couple’s utility function for children versus
other goods (Becker 1981).
Most theoretical explanations for changes in fertility based on the demand theory
focuses on a shift in the price (or cost) of a child. In 19th century Europe, costs of
children rose with the advent of compulsory education (Caldwell 1981), which places a
new emphasis upon the quality of children rather than their quantity and also leads to
increases in the costs of children. The indirect or opportunity costs of having children
(lost market earnings due to having children) are contingent upon the feasibility of
combining market work with children. Where mothers are unable to work if they have
children, the indirect costs of having children are high. However, when the labor force
participation rates of mothers increase, as they have done in many places in recent
decades, the indirect costs of children fall (Chapman et al. 1999). Consecuently, the
indirect costs rise as potential market incomes rise. When the birth of the first child is
delayed, women are able to increase their potential market income through education and
the accumulation of on-the-job skills and experience, thereby increasing the indirect costs
of having a child.
In the context of contemporary low fertility, Coleman states that, while the cost of
children can be figured in dollar terms, there are no dollar benefits. Instead, the benefits
or utility of children consists of dimensions of a psychological nature that are not readily
quantifiable. He refers to these benefits as ‘immanent values’ (Coleman 1998: 20). One

Page 20
way to think about the utility of a child in this circumstance is in terms of net benefit
thresholds (the psychological benefits less the psychological costs). That is, people make
some calcullation of the psychological gain (utility) to them of having the next child. If
the costs of children rise or if the shape of the utility function shifts in favor of other
goods, some individual benefit thresholds will be crossed and decisions will be made not
to have the next child. There is also an argument that delaying the first child enhances
tastes for goods other than children, thereby shifting the utility function towards goods
other than children.
Surprisingly little in known about the determinants of the utility of children
despite its crucial status in the determination of present-day fertility levels. However,
utility does vary from individual to individual. There are some who are willing to subject
themselves to the most torturous privations in order to have a child while others are
unwilling to have a child or another child under any circumstances.
It also needs to be remembered that children come in highly discontinuous lumps
and that the utility of having a child will vary according to the birth order of the child.
Having the first child provides the benefits of the status of being a parent, of ‘being a
family’, of having offspring who will carry on the family name, of meeting expectations
of others, of having a baby who will be fun and will grow loving his her parents, of
fulfilling childhood dreams, of providing vicarious pleasure from the child’s success, and
so forth. The decision to have a second child can result form the notion that each child
should have at least one sibling or to the desire of having a child of the opposite sex. It is
possible that those who have a third child believe it takes at least three children to make a
‘real’ family, or they may be still trying for the other sex that they do not have yet. Those
who have a fourth child may simply love children, or the child may have been a mistake.
It is likely that the level of the net threshold of psychological benefits falls as the birth
order rises. That is, the highest psychological threshold relates to the first child. Also, it
is very likely that the level of the threshold falls as people grow older. In other words, it
is possible that a woman at age 29 years feels more inclined to have a second child than a
woman at age 39 years. Psychological costs probably rise with age or, perhaps, as argued
above, increased age leads to shifts in the utility function towards other goods.

Page 21
Accordingly, as ages at childbearing increase, people will be less likely to have additional
children.
The demand theory implies that, if there is to be a positive impact on fertility
decision-making, it is necessary to try to reduce the economic costs of children, increase
parental incomes or shift the utility function towards children and away from other goods.
Therefore, increases in market incomes would increase the opportunity costs of children
unless there are provisions in place that allow parents, both mothers and fathers, to
combine market work with children. Also, raising incomes of parents with government
transfer payments to those who have children or subsidies of child costs, particularly of
services such as health and education, would increase fertility.

2.2. A Poisson Model of Total Fertility

Individual household fertility decisions are modeled in various ways in the
literature. Barmby and Cigno (1990) estimate fertility patterns using a sequential
probability model. Sobel and Arminger (1992) use a non-linear simultaneous probit
model. In recent years, the modeling of household fertility decisions has utilized Poisson
type models. Caudill and Mixon (1995) have developed censored regressions for fertility
data. Winkelmann and Zimmermann (1994) have developed the generalized event count
model, which subsumes the Poisson, the negative binomial, and the binomial models.
In many empirical studies of fertility, the number of children in a household is
modeled as a function of other social and economic variables, such as a female’s
education level and family income. The commonly used models include the standard
Poisson and negative binomial regression models. These models account for the fact that
the number of children in a family is non-negative. Under the Poisson regression model,
the conditional mean and variance of the dependent variable are constrained to be equal
for each observation. This is sometimes referred to as equi-dispersion. In practice, this
assumption is not satisfied, because the variance can be either larger or smaller than the
mean. If the variance is not equal to the mean, the estimates in Poisson regression
models are still consistent but inefficient. Therefore, an inference based on the estimated
standard errors is no longer valid. The negative binomial regression model is more
flexible than the standard Poisson model and is frequently used to study count data with

Page 22
over-dispersion. For example, Cameron and Trivedi (1986) have analyzed factors
affecting how often a person visits the doctor. Other examples included Goodwin and
Sauer (1995), Englin and Shonkwiler (1995), and Winkelmann and Zimmermann (1995).
As Winkelmann and Zimmermann (1994) have noted, the number of children in a
household often exhibits under-dispersion when the mode is two. Therefore, the standard
Poisson regression model, which assumes equi-dispersion, and the negative binomial
regression, which accommodates over-dispersion, do not account for cases of under-
dispersion. However, more recently Wang and Famoye (1997) have estimated a
generalized Poisson regression model of household fertility decisions to accommodate
under-dispersion as well as over-dispersion. Therefore, to model fertility behavior using
the standard Poisson regression model seems a good starting point. However, assessing
the validity of equi-dispersion is necessary to account for either over or under-dispersion.

2.3. Data and Variable Definitions

Becker (1960) has stimulated most of the research on household fertility decisions
by developing an economic theory of the family. In his seminal work, he suggests that
children can be viewed as durable goods, yielding primary psychic income to parents, in
the neoclassical economic framework. Household fertility decisions are determined by
female wage and family income, which are supposed to measure the time cost of raising
children and earnings potential (Becker 1960; Becker and Lewis 1973). Several
subsequent empirical analyses (Leibowitz 1974; Heckman and Walker 1990; Caudill and
Mixon 1995; Wang and Famoye 1997) support the Becker-Lewis hypothesis.
With this neoclassical framework in mind, the 1998 National Survey of
Population and Family collected in Venezuela by The Central Office of Statistics and
Information is used to model household fertility decisions by estimating a data count
model. The present study, however, deviates somewhat from the standard modeling of
household fertility, because its objective is to model the household fertility decision for a
group of females that entered a first union between 1967 and 1982. The 1998 National
Survey of Population and Family contains complete birth histories that allow computing
the number of children born to each female during the first 14 years after entering the
first union. It is hypothesized that entering a first union signals the desire to have

Page 23
children in a planned context. Since all females are observed over the same amount of
time, comparisons of total fertility between union cohorts are appropriate.
As is true of most surveys, different households have different probabilities of
being selected for the sample either by design or by accident. In the 1998 National
Survey of Population and Family, some females are overrepresented relative to others as
part of the design. This creates a bias problem, because if females in the sample have
different probabilities of being selected for the sample, sample means will be biased
estimators of population means. To undo this bias, the sample data need to be re-
weighted to make them representative of the sample. When estimating total fertility,
observations are weighted according to the survey weights coming from its particular
design.
Also the 1998 National Survey of Population and Family was collected in two
stages, first with the selection of sampling clusters and then with the selection of
households from within each cluster. This is a very common two stage stratification
design. It is well known that this design affects the sampling variability of estimates,
because households within clusters are often similar to one another in their relevant
characteristics. It is, therefore, frequently the case that clustering can increase variability
when compared to simple random sampling. It can be a serious mistake to treat two-
stage sample as if it were a simple random sample. Therefore, when estimating the
parameter the two-stage design is also taken into account.
The purpose of this chapter is to compare all first union cohorts between 1967 and
1982 to ascertain if these cohorts followed a lower fertility pattern as the demographic
transition suggests or if any subsequent cohort ended with a higher fertility level than a
previous one during the first 14 years after entering the first union. The former would
suggest that the stalls and reversals presented in the first chapter are caused by changes in
the timing of births and not by changes in total fertility levels.
On the one hand, females entering a first union during the period between 1967
and 1975 experienced the sharp increase in economic activity fueled by sharp increases in
the price of oil. On the other hand, females entering a first union after 1975 experienced
most of the sharp decline or bust that followed after 1978. All females are observed for
14 years and the total number of children born in the 14 years after a woman enters the

Page 24
first union is recorded and viewed as a proxy for total fertility. To assess if increases in
period fertility rates documented in the first chapter are associated with total fertility
increases, the total fertility for females that entered first union during these years is
explored. The sample consists of 2,038 females whose ages ranges from 26 to 49 years at
the time of the survey. The dependent variable, the total number of children born during
the first 14 years after a woman enters her first union, is a non-negative integer ranging
from 0 to 10 in the sample. In line with the neoclassical theory of fertility developed by
Becker (1960) and Becker and Lewis (1973), explanatory variables common to fertility
studies are included. An introduction of the explanatory variables used in this analysis
follows in the next five subsections.

2.3.1. Place of Residence and Migratory Experience

The place of residence (dominio) generally tends to have a notable impact on
fertility, as it reflects major differences in a series of economic conditions faced by
families. For instance, urban families will have fewer children than rural families,
because it costs less to raise children on a farm. Also, rural communities lag behind
urban centers in the distribution of contraceptive knowledge. Table 2.1, which presents a
distribution of the variables of place of residency, length of residency (lenres), and place
of socialization (chood), shows that of the total of females in the sample, 25.07% live in
the Caracas Metropolitan Area, a little more than one-half (54.76%) reside in a city with
25,000 or more people, while 20.17% reside in urban centers of less than 25,000 people
or in rural areas.
Like the place of residence, a migratory experience can influence the change of
fertility to the extent that migrants can broaden the channels of dissemination by being
carriers of new ideas and modern lifestyles, which they can transfer to their counterparts
in the places where they are received or of origin, as the case may be. Although, there
have been periods of extensive spatial mobility in the Venezuelan population, it is not
known how this process bears on reproductive behavior changes. Information about the
length of residence and the place of socialization allow a limited examination of how the
migratory experience of females affects household fertility decisions. Also, Table 2.1
shows that of the total number of females in the sample, 58.39% stated they had had

Page 25
some migratory experience, while 41.61% said they had always lived in the current place
of residence. An examination of this behavior by participant age reveals, as expected,
that the percentage of migrants is greater among the older women, with almost more than
two out of every three (63.63%) having changed residence at least once.

Table 2.1 Distribution of place of residency, length of residency, and place of

socialization
Variable Frequency Percent

Place of residence (dominio )

Caracas Metropolitan Area 511 25.07
Cities with 25,000 people or more 1,116 54.76
Urban centers with less than 25,000 people or rural areas 411 20.17
Total 2,038 100.00

Length of residence (lenres)

Always 848 41.61
Less than 1 year 51 2.50
1 to 4 years 100 4.91
5 to 9 years 147 7.21
10 years or more 892 43.77
Total 2,038 100.00

Place of socialization (chood )

Caracas 360 17.66
Another city 963 47.25
A town 360 17.66
Farm 183 8.98
Abroad 172 8.44
Total 2,038 100.00

Place of socialization, which has been defined as the place where the interviewed
females spent the first 12 years of their lives, is important because it reflects the
differences in the behavior of females, as it is related to the transmission of norms and
values. For instance, modern guidelines for family size, reproductive behavior, and
contact with other reproductive alternatives would have the greatest influence on those
who spent their youths in urban areas and, as a result, affect fertility negatively. Among
the group of females interviewed, nearly two out of every three (64.91%) spent most of
their early years in Caracas or another urban area, while the cultural characteristics of a
town or farm influenced the early socialization experience of 26.64% of the sample.

Page 26
Finally, 8.44% of the females spent the first 12 years of their lives abroad. These women
took part in the major migratory current that occurred in the country during the 1970s.
Since this is a retrospective study of fertility during the 1970s and 1980s, it is important
to include the effect that these migratory currents can have on fertility.

2.3.2. Family Income

According to neoclassical economic theory, the effect of family income on
fertility is ambiguous. Since children are treated as durable goods, an increase in the
family income will have a positive effect on fertility (income effect). However, family
income can also have a negative substitution effect (price effect). Becker and Lewis’
(1973) quantity-quality approach predicts that a substitution effect is likely from quantity
to quality of children with a rising family income. An increase in quality per child
implies an increase in the cost of raising a child, which, in turn, decreases fertility. The
net effect of income on fertility depends on the relative strength of the income effect
when compared to the substitution effect. Becker (1960) argues that the substitution
effect can be large compared to the income effect. Unfortunately, the survey used in this
study lacks a direct measure of family income or information on wages of females or
heads of households. Nevertheless, an attempt is made to determine the family’s level of
income using characteristics regarding the house and households where females reside,
because these normally reflect the level of income of the family. These characteristics
are the type of house (thouse) and profession of the head of the household (profe).

Table 2.2 Distribution of type of house and profession of head of household

Variable Frequency Percent
House type
House 470 23.06
Apartment 414 20.31
Slum, rural house, urban/rural shack 1,154 56.62
Total 2,038 100.00

Profession of head of household

University educated professional 206 10.11
Technician 159 7.80
Employee, no university education 491 24.09
Specialized laborer 540 26.50
Specialized worker 642 31.50
Total 2,038 100.00

Page 27
 Table 2.2, which shows the distribution of these variables, reveals that about a
quarter of the sample (23.06%) live in a house, 20.31% live in an apartment and 56.62%
live in a slum, a rural house or an urban/rural shack. Also, Table 2.2 shows that 42.00%
of the sample reside in a household where the head is either a university educated
professional, technician or employee with no degree. However, the rest (58.00%) reside
in a house where the head is laborer or worker. A natural way to proceed is to explore
these variables to find out how they reflect the level of income which affects, in theory,
household fertility decisions.

2.3.3. Work
In the neoclassical framework, labor activities of the interviewed females (work)
are expected to be related to fertility decisions. At the time of this survey, 56.38% of
females in the sample were working. A female’s participation in the labor force is
expected to be negatively related to fertility.
According to neoclassical theory (Becker 1981), working wives have a higher
opportunity cost of time than non-working wives. Therefore, households with working
females are expected to have fewer children than those with non-working females. Table
2.3, which presents the distribution of this variable, shows that 56.38% of the sample
worked at the time of the survey.
When work status is looked at by area of residency, some differences in the levels
emerge. In the Caracas Metropolitan Area, the percentage of females who work totals
63.80%. In cities with 25,000 or more people, it totals 58.15%, while in the rest of the
country it drops to 42.34%. These figures show that the percentage of females who work
is distinctly higher the more urbanized the area of residence is.
When the level of education completed is considered, major differences arise in
the levels of work status. Females who have reached a higher level of education can
develop greater skills and abilities that make the incursion into the labor market easier.
Females who have no education have a participation rate of 39.32%. However,
participation increases to 48.51% for females who have a middle school education and
65.56% for those who have attended high school. Participation reaches 77.70% when the

Page 28
education level is an university undergraduate degree and 95.34% when the education
level is equal to a graduate degree.

Table 2.3 Distribution of work status, work status by area of residency, and by level
of education
Variable Frequency Percent

Work status (work )

Working 1,149 56.38
Not working 889 43.62
Total 2,038 100.00

Work Status by Area of Residency

Caracas Metropolitan Area
Working 326 63.80
Not working 185 36.20
Total 511 100.00
Cities with 25,000 people or more
Working 649 58.15
Not working 467 41.85
Total 1,116 100.00
Urban centers with less than 25,000 people or rural areas
Working 174 42.34
Not working 237 57.66
Total 411 100.00

Work Status by Female's Level of Education

None
Working 46 39.32
Not working 71 60.68
Total 117 100.00
Middle School
Working 587 48.51
Not working 623 51.49
Total 1,210 100.00
High School
Working 238 65.56
Not working 125 34.44
Total 363 100.00
University (undergraduate)
Working 237 77.70
Not working 68 22.30
Total 305 100.00
University (graduate)
Working 41 95.35
Not working 2 4.65
Total 43 100.00

Page 29
 2.3.4. Education
The educational attainment of the female is predicted to be directly related to her
opportunity cost of time and inversely related to her fertility decision. The inverse
relationship is expected to be stronger, the higher the female’s educational attainment.
Therefore, illiteracy, the level of formal education reached (ledu) and the number of years
of schooling completed (yedu) by females are important variables to consider in the
analysis of fertility.

Table 2.4 Literacy rates and distribution of literacy by area of residence

Variable Frequency Percent

Literate
Yes 1,891 92.79
No 147 7.21
Total 2,038 100.00

Caracas Metropolitan Area

Literate
Yes 490 95.89
No 21 4.11
Total 511 100.00
Cities with 25,000 people or more
Literate
Yes 1058 94.80
No 58 5.20
Total 1116 100.00
Urban centers with less than 25,000 people or rural areas
Literate
Yes 343 83.45
No 68 16.55
Total 411 100.00

Table 2.4, which tabulates literacy rates and literacy by area of residence, shows
that a total of 147 females, representing 7.21% of the total females in the sample, said
they did not know how to read or write. Most often the illiterate women are older, and
their illiteracy reflects greater restrictions of access to education in the past and is closely
related to the place of residence, with a 4.11% rate of illiteracy in the Caracas
Metropolitan Area, 5.20% in cities of 25,000 people or more and 15.55% in the rest of
the country. In the less urbanized parts of the country, the percentage of illiterate females
is almost four times (3.78) greater than in the Caracas Metropolitan Area. By the same

Page 30
token, the percentage of illiterate females is three times (2.99) greater in the cities of
25,000 people or more than in the Caracas Metropolitan Area.

Table 2.5 Distribution of level of education

Variable Frequency Percent

Level of formal education reached (ledu )

None 117 5.74
Middle School 1,210 59.37
High School 363 17.81
University (undergraduate) 305 14.97
University (graduate) 43 2.11
Total 2,038 100.00

Level of education reached by area of residency

Caracas Metropolitan Area
None 10 1.96
Middle School 258 50.49
High School 95 18.59
University (undergraduate) 126 24.66
University (graduate) 22 4.31
Total 511 100.00
Cities with 25,000 people or more
None 51 4.57
Middle School 667 59.77
High School 226 20.25
University (undergraduate) 152 13.62
University (graduate) 20 1.79
Total 1,116 100.00
Urban centers with less than 25,000 people or rural areas
None 56 13.63
Middle School 285 69.34
High School 42 10.22
University (undergraduate) 27 6.57
University (graduate) 1 0.24
Total 411 100.00

In order to explore the level of education reached by a Venezuelan female, a brief

overview of the Venezuela formal educational system follows to allow a better
understanding of what the levels mean. The Venezuelan formal educational system is
divided into five levels, kindergarten, basic, diversified and professional, technical
superior, and university. Kindergarten is required for children under the age of six and
usually lasts about two years. Basic, which is also obligatory, is for children over six

Page 31
years old. It is divided into three stages of three years each and lasts nine years.
Diversified and professional, the third level of education, prepares students for higher
education or for the labor market. It lasts two years and offers two options. The
diversified option offers concentrations in sciences, humanities, and technologies such as
agriculture, industry, commerce, health, administration, and art and is intended for
students who plan to continue their formal education. The professional option offers
concentrations in agriculture, industry, commerce, health, and administration and is
intended for students who plan to join the labor force after completion of this third level.
The technical superior level, which can be compared to a two-year college, is usually
offered by polytechnic institutes. Finally, there is the university level.
For the purpose of the following table and the rest of the analysis, kindergarten is
merged with the no education category. Basic is called middle school; diversified and
professional are grouped and called high school; technical superior and undergraduate
university are grouped and called university (undergraduate); and the rest are called
university (graduate). Table 2.5, which tabulates the distribution of level of education
according to this new grouping, shows that 18.65%, or approximately one out of every
four females in the sample, stated that she completed some year of higher education,
18.47% reached high school, and 57.58% completed middle school. Only 5.31% of those
interviewed stated they had no access to school. When grouped by age, the results are
similar to those for literacy. Older females show lower achievement levels that are
associated with the greater restrictions of access to education that existed in the past.
It is possible that the more or less urban nature of the place of residence
influences the number of years a female spends in school, because there are more
opportunities to study in the more urbanized areas. Consequently, the educational profile
of females residing in the Caracas Metropolitan Area is higher, with 28.97% having
attended universities as undergraduate or graduate students, compared to 17.08% for the
entire sample. Conversely, the educational profile of females residing in small urban
centers or rural areas is much lower, with 6.81% having attended universities as
undergraduates or graduates, compared to the 17.08% for the entire sample.

Page 32
 2.3.5. Age
Since older women have had a longer time to reproduce than younger women, it
is reasonable to expect younger females to have a lower fertility than older females.
Therefore, a female’s age is included as an explanatory variable, because age (Age)
reflects characteristics of the life cycle and can, therefore, condition some of the
explanatory variables mentioned above. Table 2.6, which tabulates the distribution of
age, shows that 99.17% of the sample is at least 30 years old, while 87.34% of the sample
is at least 35 years old.
The mean age at first union is 18.82 years and the mean age at the fourth child is
28.17 years. These numbers reveal that most of the fertility in this sample takes place
early in the life cycle of females. In general, females enter their first union young and
have most of their children in their twenties. This fact means that using the number of
children born during the first 14 years after entering the first union as a proxy for total
fertility is appropriate.

Table 2.6 Distribution of age

Age group Frequency Percent
25-29 17 0.83
30-34 241 11.83
35-39 618 30.32
40-44 690 33.86
45-49 472 23.16
Total 2038 100.00

Table 2.7 Frequency distribution of the number of children born during the first 14
years after entering the first union
Number of children 0 1 2 3 4 5 6 7 8 9 10
Frequency 59 226 514 561 336 186 101 35 16 2 2

Finally, it is important to consider the distribution of the dependent variable, the

number of children born during the first 14 years after entering the first union. Table 2.7,
which tabulates the distribution of the children born during this period, shows that most
of the females have three children after the first union. It also shows that the frequency
of having 7, 8, 9 or 10 children is very small by comparison.

Page 33
 2.4. The Poisson Regression Model and its Estimation
The dependent variable is the number of children a female gives birth to over the
course of the first 14 years after entering her first union. Ordinary Least Squares (OLS)
has been frequently used to estimate the parameters in models of this type, where the
dependent variable is an integer describing a count. Despite its wide use in this situation,
the OLS estimation has several weaknesses. When the dependent variable is a count, the
OLS estimation is inefficient, and estimates of standard errors are inconsistent. It is also
possible that the estimation by OLS could lead to negative count predictions. These
problems have pushed researchers to seek alternatives.
A popular solution when the dependent variable, is a count is found in the
estimation of either a Poisson regression model or a negative binomial regression model.
Coleman (1964) outlines the basic idea of a Poisson regression. Goodwin and Sauer
(1995) have recently employed these models to estimate publication by an economist
over the life cycle; Ozuna and Gomez (1994) have used them to estimate a model of
recreational demand; Michener and Tighe (1992) have utilized them to estimate highway
fatalities; and Mayer and Chappell (1992) have employed them to estimate net entry in
several industries.
Although the estimation of Poisson regression models and negative binomial
models is common, the development of an interest in the negative binomial regression
model arises from dissatisfaction with the extreme assumptions of the Poisson regression
model. The major difficulty with the estimation of Poisson regression models is that the
mean and variance are constrained to be equal for each observation. In practice, this
assumption is frequently false, because usually the variance exceeds the mean. This
situation is known as over-dispersion, and its consequence for parameter estimates in
Poisson regression models is like the problem of heteroscedasticity in linear models. The
regression parameters are consistently estimated, but the standard errors are biased
downward leading to the rejection of too many false null hypotheses. Therefore, several
tests for over-dispersion have been developed. And if there is evidence of over-
dispersion, the less restrictive negative binomial regression model provides an alternative
to the Poisson model.

Page 34
 The dependent variable Yi , which is a count variable, is defined as the number of

children born to a female during the first 14 years after entering the first union. Although
this period is smaller than the entire number of childbearing years, it can be viewed as a
proxy for total fertility, because in Venezuela during the 1970s and 1980s, most of the
pregnancies took place in the very early years of the female’s childbearing years.
Following Famoye (1993), the probability function of Yi is given by:

e − µi ( µi ) yi
fi ( yi ; µi ) = where yi = 0,1, 2,...,14
yi !
ln ( ti ) + β 0 + β1 x1,i +...+ β k xk ,i
and µi = ti µi ( xi ) = e = eln( ti )+ xi β , where xi is a (k − 1) dimensional vector
of explanatory variables including personal characteristics of both female and husband or
partner as well as some demographic attributes of the family, and β is a k dimensional
vector of regression parameters. The mean and the variance of yi are given by

E (Yi | xi ) = V (Yi | xi ) = µi
The likelihood function of the Poisson regression model is
N N
e − µi ( µi ) yi
L( β | y, X ) = ∏ fi ( yi ; µi ) = ∏
i =1 i =1 yi !
After taking the natural log, numerical maximization can be used to find a
maximum and estimate the coefficients. Since the likelihood function is globally
concave (Maddala, 1983), if a maximum is found, it will be unique.

2.5. Results
The model is estimated using a Poisson regression. The estimation results are
presented in Table 2.8, which shows that place of residency, place of socialization, type
of house, conjugal status, level of education, age, and union cohort seem to be significant
determinants of fertility. In particular, the result most relevant to this work is that
compared to the 1967-1968 union cohort, all subsequent union cohorts have a lower total
fertility after 14 years of entering the first union. However, this result is only statistically
significant for the 1973-1974 through the 1981-1982 union cohorts.
This raises the possibility that the period fertility rate stalls and reversals
presented in the first chapter are not be associated with rises in total fertility for females

Page 35
entering a union during the period considered, because they did not end with more
children when compared with previous union cohorts. The first stall and reversal takes
place in the mid 1970s, but females entering their first union during this period have
fewer children than previous cohorts 14 years after the first union.

Table 2.8 Determinants of fertility

Variable IRR
Place of Residency Coef. IRR Std. Err. z P>z
Caracas Metropolitan Area (omitted)
Cities with 25,000 people or more 0.1252 1.1333 0.0414 3.42 0.001
Urban with < 25K people or rural areas 0.1876 1.2063 0.0550 4.11 0.000
Migratory Experience
Never moved -0.0228 0.9775 0.0266 -0.84 0.402
Place of Socialization
A city (omitted)
A town 0.0183 1.0185 0.0358 0.52 0.602
Farm 0.0877 1.0917 0.0501 1.91 0.056
Abroad -0.1341 0.8745 0.0462 -2.54 0.011
House Type
House (omitted)
Apartment -0.0883 0.9154 0.0427 -1.89 0.058
Slum, rural house, urban/rural shack 0.0750 1.0779 0.0360 2.24 0.025
Conjugal Status
Widowed/divorced/separated -0.1060 0.8995 0.0296 -3.22 0.001
Work
Working at time of survey -0.0405 0.9603 0.0263 -1.48 0.138
Level of Education
None (omitted)
Middle School -0.1070 0.8986 0.0446 -2.15 0.031
High School -0.2350 0.7905 0.0476 -3.90 0.000
University (undergraduate) -0.3220 0.7247 0.0483 -4.83 0.000
University (graduate) -0.4273 0.6522 0.0812 -3.43 0.001
Demographic Control
Age -0.0202 0.9800 0.0039 -5.07 0.000
Union Cohort
67-68 (omitted)
69-70 -0.0726 0.9300 0.0613 -1.10 0.271
71-72 -0.1022 0.9029 0.0600 -1.54 0.124
73-74 -0.2287 0.7956 0.0533 -3.41 0.001
75-76 -0.2517 0.7775 0.0523 -3.74 0.000
77-78 -0.3552 0.7011 0.0484 -5.15 0.000
79-80 -0.3608 0.6971 0.0497 -5.06 0.000
81-82 -0.4547 0.6346 0.0483 -5.97 0.000
Constant 2.2716

Page 36
 2.6. Assessing the Fit
For the Poisson distribution, the mean and the variance are equal, which implies
that the deviance and the Pearson statistic divided by the degrees of freedom should be
approximately one. Values greater than one indicate over-dispersion, meaning that the
true variance is bigger than the mean; values smaller than one indicate under-dispersion,
meaning that the true variance is smaller than the mean. Evidence of under-dispersion or
over-dispersion indicates the inadequate fit of the Poisson model. Corrective measures
include using the deviance or Pearson chi-square divided by degrees of freedom as an
estimate of the dispersion parameter instead of setting it to one, or, in the case of over-

Observed Predicted

0.3

0.25

0.2
Probabilities

0.15

0.1

0.05

0
0 1 2 3 4 5 6 7 8 9 10

Year

Figure 2.1 Observed and Predicted Probabilities

dispersion, running the negative binomial regression instead of the Poisson regression.
Figure 2.1, which graphs the observed and predicted probabilities for this Poisson model,
seems to suggest an adequate fit.

Table 2.9 Assessing goodness of fit

Criterion DF Value Value/DF
Deviance 1,695 1180.50 0.6965
Pearson Chi-Square 1,695 1073.02 0.6330
Log Likelihood -2706.83

Page 37
 Table 2.9 shows that the deviance and the Pearson chi-square statistic in the
Poisson estimation are 1,180.501 and 1,073.018, respectively, and the degrees of freedom
are 1,695. Dividing the deviance and the Pearson Chi-Square statistic by the degrees of
freedom obtains 0.7 and 0.6, respectively. These results indicate a somewhat adequate fit
and, therefore, offer some evidence that the fit of the Poisson model is somewhat
adequate.

2.7. Interpreting the Results

The factor or percentage change in the expected count E ( y | x) can be computed
simply from the parameters of the model so that
E ( y | x, xk ) = e β0 + β1x1 +...+ βk xk +...+ β K xK = e β0 e β1x1 ...e β k xk ...e β K xK

where E ( y | x, xk ) makes explicit the value of xk . If xk changes by δ , then

E ( y | x, xk + δ ) = e β0 + β1x1 +...+ βk xk +...+ β K xK = e β0 e β1x1 ...e β k xk e β kδ ...e β K xK

The factor change in the expected count for a change of δ in xk equals

E ( y | x, xk + δ ) e β0 e β1x1 ...e βk xk e βk δ ...e β K xK

= β0 β1 x1 β k xk β K xK
= e βkδ
E ( y | x, xk ) e e ...e ...e

Therefore, for a change of δ in xk , the expected count changes by a factor of e βk , where

all other variables remain constant. This is the so-called incidence rate. Alternatively,
the percentage change in the expected count for a δ unit change in xk , when all other
variables remain constant, can be computed as
 E ( y | x, xk + δ ) − E ( y | x, xk )  βkδ
100   = 100(e − 1)
 E ( y | x, xk ) 
The factor changes or incidence rate for each variable is presented in Table 2.8.
However, since most of the variables in the model are indicator variables, the discrete
change in the expected value of y for a change in xk starting at xS and ending at xE
seems a better way to interpret the effect of these variables. This discrete change equals
∆E ( y | x)
= E ( y | x, xk = xE ) − E ( y | x, xk = xS )
∆xk

Page 38
Therefore, for a change in xk from xS to xE , the expected count changes by

∆E ( y | x) / ∆xk , when all other variables remain constant. Then, the effect of a binary

variable is obtained by letting xk change from 0 to 1. Table 2.10, which presents these
descrete changes, confirms that the place of residence generally tends to have a notable
impact on fertility, as it reflects major differences in a series of economic conditions
faced by families. Urban families will have fewer children than rural families, because it
costs less to raise children on a farm. Also, rural communities lag behind urban centers
in the distribution of contraceptive knowledge. Residing in cities with 25,000 people or
more increases the expected count by 0.3691 children compared to the more urban
metropolitan area of Caracas. Further, residing in urban centers of less than 25,000
people or rural areas also increases the expected count by 0.5890 children when
compared to the metropolitan area of Caracas.
The discrete changes presented in Table 2.10 also confirm that place of
socialization reflects differences in the behavior of females, as it is related to the
transmission of norms and values. Modern guidelines for family size, reproductive
behavior, and contact with other reproductive alternatives have the greatest influence on
those who spent their youths in urban areas and, as a result, affect fertility negatively. As
Table 2.10 shows, having socialized in a town increases the expected count by 0.0546
children compared to having socialized in a city. Further, having socialized on a farm
increases the expected count by 0.2697 children compared to having socialized in a city.
A very interesting result is that having socialized abroad decreases the expected count by
0.3764. This finding suggests that females that immigrate into Venezuela have different
guidelines for family size and reproductive behavior.
The results for the type of house, also depicted in Table 2.10 and intended to
capture family income, show that living in an apartment decreases the expected count by
0.2556; however, living in a slum, rural house or urban/rural shack increases the expected
count by 0.1626. The results of educational attainment confirm the hypothesis that the
educational attainment of the female is directly related to her opportunity cost of time and
inversely related to her fertility decision, and the inverse relationship is expected to be
stronger, the higher the female’s educational attainment. When compared to no level of
education, middle school, high school, university (undergraduate), and university

Page 39
(graduate), decrease the expected count by 0.3205, 0.6475, 0.8564, and 1.0403 children,
respectively.

Table 2.10 Discrete change in expected values

Variable Disc.
Place of Residency Change
Caracas Metropolitan Area (omitted)
Cities with 25,000 people or more 0.3691
Urban with < 25K people or rural areas 0.5890
Migratory Experience
Never moved -0.0675
Place of Socialization
A city (omitted)
A town 0.0546
Farm 0.2697
Abroad -0.3764
House Type
House (omitted)
Apartment -0.2552
Slum, rural house, urban/rural shack 0.2212
Conjugal Status
Widower/divorced/separated -0.3052
Work
Working at time of survey -0.1204
Level of Education
None (omitted)
Middle School -0.3205
High School -0.6475
University (undergraduate) -0.8564
University (graduate) -1.0403
Demographic Control
Age -1.4754
Union Cohort
67-68 (omitted)
69-70 -0.2089
71-72 -0.2906
73-74 -0.6219
75-76 -0.6833
77-78 -0.9392
79-80 -0.9561
81-82 -1.1756
Constant

Finally, every subsequent cohort reduces the expected count, without exception,
when compared to the 1967-1968 union cohort. This last is a key result in the context of
this study, because if the number of children born during the first 14 years of entering the

Page 40
first union is used as a proxy for the total fertility of a female, the result means that every
cohort that entered an union had fewer children than the previous one in relation to the
1967-1968 cohort. Therefore, increases in the period fertility rate in the mid 1970s do
not seem to be associated with increases in total fertility. Hence, the fertility decline was
never really interrupted as the stalls and reversals of the period fertility rates examined in
the first chapter suggest. This last hints that the stall and reversal of the period fertility
rates during the 1970s is in fact due to changes in the timing of childbearing.

2.8. Summary
Poisson regression results appear to indicate that the independent variables are
useful in explaining differences in the number of children a female has. The results show
that place of residence, place of socialization, type of house, conjugal status, level of
education, age, and union cohort are significant at the 0.10 level or better. The
coefficient of each of the highly significant variables has the predicted sign, including
place of residency, place of socialization, work, level of education, and union cohort.
Place of residence generally tends to have a notable impact on fertility, as it
reflects major differences in a series of economic conditions faced by families. Urban
families will have fewer children than rural families, because it costs less to raise children
on a farm. Also, rural communities lag behind urban centers in the distribution of
contraceptive knowledge. Place of socialization reflects differences in the behavior of
females, as it is related to the transmission of norms and values. Modern guidelines for
family size, reproductive behavior, and contact with other reproductive alternatives have
the greatest influence on those who spent their youths in urban areas and, as a result,
affect fertility negatively. The educational attainment of a female is directly related to
her opportunity cost of time and inversely related to her fertility decision, and the inverse
relationship is stronger as the female’s educational attainment increases. Conjugal status
reduces the expected number of children, because it reflects the lower risk that widowed,
divorced or separated females have of childbearing. Also, age is an important
determinant of fertility, because it reflects differences in life cycles among females.
Finally, every subsequent cohort reduces the expected number of children,
without exception. This last is a key result in the context of this study, because if the

Page 41
number of children born during the first 14 years after entering the first union is used as a
proxy for the total fertility of a female, this result means that every cohort that entered an
union had fewer children than the previous one in relation to the 1967-1968 cohort.
Therefore, increases in the period fertility rate in the mid 1970s do not seem to be
associated with increases in total fertility. This finding is consistent with the hypothesis
that changes in the timing of births could have taken place as suggested by the period
fertility rates presented in the first chapter. However, since increases in the period
fertility rate in the mid 1970s did not change the total fertility, the fertility decline was
never really interrupted, as the period fertility rates suggest. This is further evidence that
the distortions in the period fertility rates introduced in the first chapter are the result of
changes in fertility timing and not changes in total fertility, and there is circumstantial
evidence to suggest that the oil based boom and bust may have induced households to
change the timing of births.

Page 42