Professional Documents
Culture Documents
The Growth of Community Colleges in The American
The Growth of Community Colleges in The American
The Growth of Community Colleges in The American
Abstract
What accounts for the variation in the number of community colleges in the Amer-
ican states? Existing studies point to a number of factors, but there have been few
studies that seek to explain the differential rate of growth of these institutions across
states. We test theories from the economic, sociological, and political science literature
using a multi-level Poisson regression model, utilizing Bayesian methods for estimation
and inference. Our results indicate support for the idea that community colleges grew
in response to changes in state populations, and that states with large number of other
types of institutions of higher education saw slower growth. Little support is found for
theories regarding community colleges as engines of stratification.
Introduction
The establishment of community colleges in the American states stands as one of the
constitute 35% of the enrollment in American higher education, while the colleges them-
selves represent around 25% of all institutions. The size and scope of the community
1
college system in the United States reflects both a desire for educational opportunity
the American system of higher education, one major question about these institutions
remains unanswered. Why did certain states create large systems of community col-
The simplest answer is that large states created more community colleges. How-
ever, size alone does not predict the number of community colleges across states. Some
large states, such as New York have relatively few community colleges (population
19 million—35 community colleges), while some smaller states, such as North Car-
olina have larger numbers of these institutions (population 8.8 million—59 community
colleges) .
This study attempts to determine the characteristics of states that are associated
within the states has been characterized as an effort to both respond to increasing
levels of student demand and also to “coordinate higher education along the lines of a
hierarchically segmented system” (Brint and Karabel, 1989, p. 72). The simultaneous
pressures to both expand access and create more differentiation in systems of higher
education remain today. To better understand how states might respond in the future
to these two pressures, this study attempts to refine our understanding of what may
This study is also part of a broader effort to understand the antecedents of state
policy for higher education generally. We still know relatively little about why certain
states are more or less likely to adopt certain policies for higher education. Establishing
the characteristics of states that are associated with one of the most important policy
innovations in the history of higher education will further our knowledge in this area.
2
Literature Review
The literature review looks to past research to understand the following questions:
What explanations have been offered for the emergence of two-year colleges? Further,
setting aside the question of initial establishment, what explanations have been offered
for the differential growth rates of community colleges in the American states?
This literature review focuses on four possible explanations for the development of
community colleges. First, researchers operating from within the economic literature
have argued that the development of community colleges came about as a result of
increased demand for higher education, coupled with business leaders’ interests in
developing more young people with specific vocational skills that did not necessarily
require a bachelor’s degree. Second, many researchers from within the sociological
literature have argued that these institutions were created as a result of broader social
forces, although they disagree about both the nature of those social forces and the
that community college development had more to do with broader societal forces that
functionalists within the sociological framework argue that community colleges were
describes the development of these institutions in explicitly political terms, as the result
constituents and important lobbies. Last, the organizational ecology literature speaks
founders’ ability to fill a particular “niche”, where they would not compete for resources
We describe each of the four strands below, with a particular emphasis on those
studies which seek to provide explanations for the development of these institutions, as
the four approaches suggests specific, testable propositions. We develop these propo-
3
sitions in our conceptual framework.
Economic Perspectives
lege emergence that deal primarily with economic considerations. According to this
literature, community colleges were developed as a result of either (1) labor market
conditions the need of technology in trained workforce, the demand for new careers
and retraining, especially during economic crises (Betts and Farland, 1995; Griffith
and Connor, 1994) and (2) consumer demands (Blau et al., 1994). The common view
is that vocational education and training are instrumental to making a nation more
competitive; thus, community colleges among other institutions are directed to serve
the needs of capital through provision of trained workforce (Levin, 2001). Their crit-
ical role in workforce preparation, pre-service training, in-service training and assis-
tance for small business development is greatly influenced by purposive state policies
(Dougherty, 2001). These government policies resulted from the economic concerns:
governments, and responsive to business needs (Levin, 2001). Brint and Karabel (1989)
attribute the community college vocationalization to the fiscal crises of the 1970s, in
tion strategies. Many authors mention the demographic factor as one of the leading
causes of the two-year college emergence. The demographic pressures have been ac-
companied by the growing demand for educational opportunity and additional years of
schooling (Witt et al., 1994; Cohen and Brawer, 2003; Blau et al., 2000; Baker, 1994)
At the same time, some believe that recent increases in technology and changing
idea. Most of these demands are now met by community colleges, and, as a result, the
latter have lost their leadership role and identity, become less rigorous, served diverse
4
student bodies while trying to recruit new populations (Cain, 1999). Most important is
the observation that, in the pursuit of economic goals, community colleges altered their
missions and moved away from social needs of local communities toward local market
needs, abandoning their earlier goals of accessibility, personal and social development,
Several studies in this strand provide reliable and extensive data based on the use
college foundings: state affluence, as measured by per capita income, the existence of
both large manufacturing sectors and diverse industrial sectors, and, partly, compe-
tition (Blau et al., 1994, 2000). These researchers find that community colleges are
responsive, among other factors, to the labor force needs of a diversified economy. Ex-
amining the impact of business cycles on community colleges, Betts and Farland (1995)
find that full-time and part-time enrollments show similar anti-cyclical patterns, in con-
trast to pro-cyclical state and local per-student appropriations. In spite of this failure
to align education policy with labor markets, the recessions drive people into two-year
founding back to a set of specific demands from the workplace. These institutions arose
in the marketplace to meet that demand, ensuring that citizens within state would have
access to the appropriate level of training required for the changing economies of states.
In contrast social theorists to be discussed next see these institutions as being part of
broader social forces, and not necessarily as a necessary outgrowth of particular market
needs.
Social Theories
The next strand of literature describes the development of community colleges as re-
flecting broader societal forces, rather than the specific supply-demand relationships
posited by economists. Within this literature, there are two opposite ideological camps.
The first is the class-reproduction school, which according to Dougherty (1988b) the-
5
orize that “the community college’s fundamental social role is to reproduce the class
The elitists of the barrier function promote one overarching idea: two-year institu-
tions have emerged as a means of social stratification and served to keep the undesirable
out” function (Clark, 1960), community colleges help softly deny students of certain
backgrounds and aspirations the entry into the higher levels of educational hierarchy.
attending college, there is a wide discrepancy between the hopes of high school gradu-
ates and the means of realizing them in higher education. This provides for inevitable
failure within the college. Junior colleges become a “soft response” alternative to early
dismissal of those with little academic abilities so that to minimize stress both for the
Taking Clark’s argument one step further, the proponents of the social stratification
idea posit that community colleges perform a contradictory task: they simultaneously
provide educational opportunities and serve to divert numerous students from four-
year institutions. They believe that community college vocationalization poses a grave
threat to equal educational opportunity and the dream of social upward mobility (Pin-
cus, 1980). Meeting the need to select and sort students destined to occupy different
positions in the job structure of a capitalist economy, junior colleges were designed as
a buffer between a high school and a four-year college and were never meant to provide
more than a terminal education for majority of its students. In many states they were
sities and providing for exclusion (Cain, 1999; Griffith and Connor, 1994). Brint and
Karabel (1989) posit that it was the junior college expansion of the 1950s and 1960s
6
that made it possible for state higher education institutions to become more exclusive.
idea—believe that the major role of two-year institutions is providing broader access
to, and less expensive entry into, higher education (Witt et al., 1994; Pedersen, 1997).
They claim that the primary purpose of two-year institutions is greater inclusion and
trace back the junior college emergence to the local citizens demands (Griffith and
Connor, 1994). In the study of the educational opportunity context, Keith (1996)
the transfer opportunities depend on the way the community colleges are integrated
Social theorists generally agree that community colleges arose as part of broader
trends at play in society, while disagreeing about what those trends might be and
how they specifically resulted in the kinds of institutions we see today. The political
theorists to be described next emphasize the role of individuals’ self-interest and desire
Political Theories
The development of community colleges in the states can also be viewed as a result
of policymakers and citizens acting in their own interests to create these institutions.
Political theorists have identified several possible rationales for the creation of commu-
nity colleges: state governments seeking to expand their educational systems; (Cohen,
2001; Brint and Karabel, 1989) ; elected officials responding to demands and pursu-
from vocational and liberal arts education (Griffith and Connor, 1994); higher educa-
tion authorities seeking to expand their own responsibilities; (Kane and Rouse, 1999);
business sector and manufacturing elites hoping to have a supply of workers trained
with public money (Blau et al., 1994; Cohen and Brawer, 2003); local communities
and their leaders looking for ways to enhance community prestige (Griffith and Con-
7
nor, 1994; Dougherty, 1994; Pedersen, 2005); superintendents and teachers seeking for
college-level jobs (Cohen and Brawer, 2003), and high school authorities seeking to
many involved had less to do with specific educational or demographic pressures, and
more to do with specific groups seeking to maximize their influence and power through
governmental actions. While they may differ on which groups were most important
and how these groups interacted to create institutions, they share a common interest in
development of political power through the creation of these new institutions of higher
education.
Cohen and Brawer (2003) note that the overarching reason for the community
college growth could have simply been an increasing number of demands placed on
schools at every level. Many believe that the original proponents of the idea were
research universities seeking to move away from the liberal arts education towards
more research specialization. These lobbying efforts were supported by local citizens
motivated by a number of desires (Griffith and Connor, 1994; Dougherty, 1988b). There
is evidence that, in comparison with the federal government and local authorities,
state governments have been the most influential constraining force for community
colleges: they accredited colleges, supplied most of their resources, prevented them
(Cohen, 2001; Brint and Karabel, 1989). In face of these views, Pedersen (2005; 1997)
posits that junior colleges were often their communities last choice (after failing to
secure traditional institutions or trying to sustain the community status after a private
college closure) and not the outcome of a university lobbying, advocacy of a school
students.
Political theorists suggest that these institutions were not the result of specific la-
bor market demands, nor the part of broader social forces, but instead the result of
efforts by individuals to use public policy for their own ends. The last perspective, de-
8
scribed next, suggests that these institutions may have been created to fill a particular
organizational “niche”, which existing institutions could not or would not fill.
Organizational Ecology
focus on the ways in which overall organizational competition for resources and func-
tional differentiation (and subsequent organizational legitimation) affect both the birth
and death of specific organizations. Clearly, this perspective may have much to offer
in terms of understanding the founding of community colleges, yet few studies have
utilized this perspective in order to understand why and under what circumstances
the concept of density, the number of organizational units at any one time within a
defined type of organizational population (Baum and Oliver, 1996). Density has impli-
cations for two key aspects of organizational founding: legitimation and competition.
increasing from very few, but then decreasing after competition becomes too fierce.
Another key insight from this field is the concept of resource differentiation. This
concept explains the co-existence of related organizations without competition. The key
differentiate, they can develop into either complementary or even symbiotic partners
which will of course face different pressures for survival and different modes of com-
petition than non-profit entities such as community colleges. However, as Baum and
Oliver (1996) argue in their study of the founding of child-care centers, non-profit or-
ganizations face more challenges in the area of legitimacy—that without strong social
legitimacy, these organizations are much less likely to form and persist, but with social
The literature has several implications for the founding of the community colleges.
First, the concept of density is key–the size and scope of the system of higher education
in states should play a role in the number of community colleges formed in any year.
Second, the competing pressures of legitimacy and competition will also most likely play
a role. This implies that other similar organizations and previously existing community
colleges may play a role in the founding of additional institutions. Last, the high levels
of social legitimacy that this literature say are required for non-profit firms may serve to
dampen competition between them and instead encourage some form of complementary
relationship.
Two studies on the founding of community colleges merit special attention, both of
which were cited above because of their overlapping approaches to the study of com-
munity college development. Dougherty (1994; 1988b; 1988a) finds that the business
and students demand was too weak to account for the extent of community college
expansion and the degree of responsiveness to their interests. His explanation is that
the community college policy emerged out of combination of demands from the interest
groups and relatively autonomous actions of government officials. The officials could
act on their own in pursuit of their self-interests; however, their actions favored student
10
and business interests. Thus, the main initiative of establishing community colleges
lay with the government officials. Although limiting his analysis to only five states
the states.
foundings in 28 states in the period of their most rapid growth (1942-1979) to test
model, the researchers conclude that the community college expansion was not directly
related to demand; a diverse economy is important in their expansion, with the large
manufacturing sectors and diverse industrial sectors translating their need for trained
labor into educational policies; manufacturing elites play an important role if there is
a surplus of high school graduates; and the states with Democratic traditions tend to
have higher community college founding rates than states with Republican traditions.
Based on the two-stage contextual analysis and reliable estimation techniques, these
The Blau study, while an important contribution, has several shortcomings. First,
its panel is limited to four ten year segments, with just 28 states included as part of
the sample. Second, while their model does incorporate a multilevel approach, they
do not employ the types of estimation techniques that allow them to simultaneously
estimate first and second level parameters. Last, their model does not incorporate
state-specific time effects, which, as will be shown later, constitute an important facet
From the literature, we find the following propositions regarding the expansion of
community colleges. The economic literature suggests that community colleges formed
as a result of an increased demand for a more educated workforce that did not nec-
essarily require a bachelor’s degree. Social theorists posit that the development of
community colleges served primarily as a tool to further stratify higher education. The
political perspective suggests that ideology and competition for government resources
11
may have determined the development of community colleges within states. Finally,
the organizational ecology literature emphasizes the importance of the existing com-
petition for scarce resources in the area of higher education in a given state.
Conceptual Framework
Our conceptual framework develops on the literature reviewed to posit hypotheses
directly related to each of the theoretical areas outlined previously. We first hypothesize
a simple supply and demand relationship, which would suggest that overall population
size as well as the size of the young population should account for the expansion of
community colleges. We then explore the social stratification literature, which would
suggest that in states where pressures for social stratification are large, we would
expect to see the creation of more community colleges. The political explanation for the
expansion of community colleges would suggest that states with a more liberal ideology
would be more likely to create more community colleges, as these institutions are the
types of solutions that favor equity over an approach that emphasizes funding for more
of the expansion of community colleges: states with high levels of population growth
ties for more of their citizens. This includes the young population—it is reasonable
to expect that states with a higher proportion of young people would have been more
The economic lens would suggest, in parallel with Betts and Farland (1995), that
community colleges themselves would have been developed in concert with the expan-
sion of the young population, allowing more young people to attain higher levels of
respond to the perceived need on the part of young people for more education, which
can not be entirely accommodated by the existing system of four-year colleges and
universities.
For the purposes of this study, we examine the extent to which growth in the young
other populations might also be part of the demand for more community colleges,
we posit that this group should be particularly important for state policymakers and
Social theorists generally posit that community colleges are part of broader social
the existing social order. We focus in this conceptual framework on the set of hypothe-
ses generated from those who emphasize this social stratification aspect of community
college systems.
Beginning with the “cooling out” theory described by Clark(Clark, 1960), many
have suggested that community colleges were created in order to divert students from
their educational goals. According to this theory, the function of community colleges in
the educational system is to realign individuals’ expectations with their actual status in
a gentle way, cooling them out so that their lack of access to postsecondary education
Similarly, Brint and Karabel (1989) suggest that instead of becoming engines of
opportunity, community colleges instead diverted students away from fulfilling their
promise in the system of higher education. We suggest that these and other class-
reproduction theories can be broadly grouped under the concept of stratification: the
idea that community colleges, instead of being created to enhance educational oppor-
tunity, were instead designed to increasingly stratify the system, diverting some parts
In general, the social stratification theory would suggest that states with more
13
inequality between classes would have more of a stake in reproducing the existing
order. This could also extend to states with more heterogeneous populations, in which
those populations that traditionally held advantages sought to maintain their status
Based on this conception of the social stratification literature, we posit two possible
relationships based on the stratification theory. First as more of a state’s young people
are non-white, this theory would suggest that the state would be more likely to expand
its community college system. this follows from the above theory in that privileged
groups in the state would attempt to divert young people from other racial or ethnic
groups from four year colleges by creating a system of community colleges. In addition,
We expect to see more community colleges in states with higher levels of class
income inequality should therefore be more likely to put community college systems in
colleges has been tied to the general policy liberalism that was prevalent in the early
1970’s (Klingman and Lammers, 1984). Like other state and federal projects, commu-
among all citizens (Dougherty, 1988b). It is possible, too that conservative politicians
may have favored the development of community colleges, as they typically are a more
The organizational ecology literature also has several important implications for
this study . The first is the recognition of the influence of existing institutions on
the creation of new institutions. It seems likely that existing institutions might view
Again, following Dougherty, we hypothesize that states with a more competitive post-
community colleges. Thus, four-year institutions in these states would seek to deprive
these institutions of the legitimacy that they would need in order to be founded in
number of community colleges in any year. Many states have an explicit policy that
all citizens should be within a certain range of a community college. If this policy were
followed explicitly, we should find that states with lower population densities created
more community colleges, after controlling for population levels. This also follows
from the organizational ecology literature, which would suggest that in areas where
organizations are less likely to face heavy competition, foundings are more likely.
The organizational ecology literature also emphasizes the importance of the richness
will serve as a constraint on the types of actions that can be undertaken. We include
income in the state as an important predictor of this state policy, as legislators can not
develop any policy if they can not draw from a resource base to fund it.
a long period of development of community colleges in the state. Second, this section
describes the models to be estimated, and the Bayesian framework for estimation and
inference utilized for this study. Both the data and the methods employed have utility
15
Data
The data for this study come from a unique dataset, compiled from a variety of sources
of state-level data. The data cover all 50 states, for the time period 1969-2002 2 . Data
The dependent variable in this study is the number of community colleges in each
state in each year. This variable was collected from the Office of Education and Welfare
and later Department of Education statistical compendia, which eventually became the
Digest of Education statistics (U.S. Department of Education, 2001). The same source
is also used for the number of public and private four year institutions in each state.
Population data by age and race are available from the National Cancer Insti-
tute Surveillance, Epidemiology and End Results program (National Cancer Institute,
2007). Young population is defined as the proportion of state population aged 20-
24, while the proportion of the young population that is non white is defined as all
Income inequality comes from the University of Texas’ income inequality project
(Galbraith and Hale, 2006). The researchers compute this measure by predicting the
Gini coefficient for family income inequality from state-level between-industry wage in-
equality data from the Bureau of Economic Analysis. Between-industry pay inequality
is computed using Theil’s T statistic. The resulting measure correlates highly with
state are more or less liberal. This measure, as developed by Berry, is based on the
voting record of the congressional delegation from the state (Berry et al., 1998). The
ideological position of state legislators and governors from each party are assumed to be
the same as the members from their party that are sent to the U.S. Congress. Higher
levels on the ideology index indicate a more liberal outlook among those elected to
state government.
2
The first year of this study was chosen as it is the first year that reliable data across the dependent
variable and all of the covariates are available. The last year was chosen as more current information was
not available at the time of writing
16
Income and population data are drawn from the Bureau of Economic Analysis,
while state land area (the denominator for population density) is available from the
Census Bureau.
Methods
The dependent variable in this study takes on a limited number of positive integer
values, suggesting that the Poisson distribution best describes the underlying data
generating process. The probability density function for the Poisson model in a panel
exp(−λist )λyist
P r(yist |λist ) = (1)
yist !
is also known as the “rate”. The Poisson distribution has only the single parameter λ,
which is equal to both the expectation (E(yi st)) and the variance. Failure to meet this
exp(−λist )λyist
P r(β|y, X) = (2)
yist !
Where λ = Xβ.
The plan of analysis for this paper is to estimate the underlying model using four
tercept for all states, a no-pooling model, which estimates a separate intercept for all
50 states, a partial pooling model, which pools information to estimate separate inter-
cepts, and a partial pooling model with a state-specific time effect 3 . Each model is
evaluated relative to its overall fit with the data. As we will show, the partial pooling
3
We follow Gelman and Hill (2006) in running models in this order
17
model with a state specific time effect provides the best fit to the data.
This study utilizes a Bayesian framework for inference and estimation. The critical
difference between this approach and a standard (frequentist) approach is that Bayesian
statistics assume that the population parameter under study is a random variable with
some distribution, rather than a fixed point. While a full description of the Bayesian
approach is beyond the scope of this paper, the results are based on posterior density
are therefore described in distributional terms, not as point estimates with standard
errors. Most of the inferences are drawn using the median, central 50% and central
95% of posterior distributions, which refer to quantiles drawn from the thinned Markov
chains for the parameter or prediction of interest. Given the state of knowledge in this
We utilize a Bayesian framework for estimation and inference for two reasons. First,
we are concerned with estimating models and coefficients for population as opposed
to sample data. As Gill (2001) explains, “in population models variance exists around
coefficients, but this variance is not an indication of estimator reliability. Rather it mea-
sures the variability of the observed effect size (β) subject to models misspecification .
. .” (Gill, 2001, p.340). Given that the coefficients to be reported can not be subjected
to traditional null hypothesis significance testing, reporting their point estimates and
standard errors does not provide enough information to judge the reliability of the es-
timates. Instead, taking the Bayesian perspective that the population parameters are
A second motivation for our use of Bayesian methods comes from Gelman and Hill
(2006), who state that the Bayesian approach for estimation and inference “averages
over the uncertainty in all the parameters in the model” (Gelman and Hill, 2006, p.
345). This is particularly helpful, they note, when the number of groups in a multilevel
18
model is relatively small, and the model itself is relatively complex. Given some of the
state-specific effects we wish to estimate, we believe that this approach will provide
The complete pooling model is the simplest of the models to be considered. Given the
α ∼ N (0, 100)
βk ∼ N (0, 10)
This model in effect assumes that most states’ processes for creating community
colleges can be reliably estimated using a common intercept, along with covariates.
No-Pooling Model
The no-pooling model involves estimating a separate intercept αs for all 50 states,
In the no-pooling (fixed effects) model, the rate λ is estimated using the following
specification:
βk ∼ N (0, 10)
19
As with fixed effects, the no-pooling model does not allow for cross-state inferences,
as all of the unit-specific heterogeneity has been absorbed by the αi term. This type of
specification is common in studies where states are the unit of analysis, as a correlation
between state characteristics and predictors can bias results (Greene, 2003; Hausman,
1978).
The no-pooling model, while providing a reasonable fit to the data, has several draw-
backs. First, it is inefficient, which is not a major concern for this study. Second, it
prevents the analyst from making any cross-unit inferences, since any systematic dif-
a partially pooled model, so called because cross-state differences are partially pooled
when estimating separate intercepts for each state. This is commonly referred to as a
“random effects” model in the econometric literature (Greene, 2003; Hausman et al.,
1984). This model has been shown to have superior efficiency in the estimation of
To overcome the problem of group effects that correlate with predictors, we use the
within-state means of all predictors to estimate the partially pooled intercept terms.
This is the approach described by Bafumi and Gelman (2006). This approach removes
the state-level heterogeneity that is of concern by controlling for different within state
The partial pooling model, like the no-pooling model, estimates a separate intercept
for each state. However, unlike the no-pooling model which has an uninformative
prior, the partial pooling model estimates the intercept from the data itself, with a
The pooled intercept term αi is modeled using state means of all covariates x¯s .
20
αs ∼ N (µ, σ 2 )
(6)
µ = bxs
βk ∼ N (0, 10)
bk ∼ N (0, 10)
σ ∼ U (0, 5)
The partial pooling model provides a compromise between the two extremes of the
important unit-specific heterogeneity in the data, the no-pooling model absorbs all
of the heterogeneity in an intercept term, preventing the analyst from making any
cross-state inferences. Traditionally no-pooling models are used when group effects
and predictors are correlated (Bafumi and Gelman, 2006). Controlling for within-
state means obviates the need for the no-pooling model, which allows for cross-state
inferences.
The emphasis in this study is on the characteristics of states that made them more
likely than other states to develop a system of community colleges. This implies that we
must use cross-state, as opposed to within-state inferences, which are only possible in
the partial pooling models. For this reason, we employ this model—to make cross-state
inferences possible.
The partial pooling with time effects model is identical to the partial pooling model,
21
The state specific time coefficient has a non informative prior of the form:
γs ∼ N (0, 100)
The effect of adding a state specific covariate on time is to allow each state to have
its own linear growth pattern, which can be modeled directly from the data 4 . The
interpretation of β changes to allow for the impact of covariates beyond what would
Estimation
All of the above models were estimated using Markov Chain Monte Carlo techniques,
specifically a Gibbs sampler. For each model, three separate chains of 100,000 iterations
each was run. The first 50,000 iterations were discarded, and the second 50,000 were
thinned by 50, resulting in 1000 draws from the estimated posterior density from each
chain for each unknown parameter. Missing data were multiply imputed using the
same techniques 5 .
Results
Results are presented for each of the models. We investigate first whether the functional
form for the model adequately fits the data, before proceeding to inferences regarding
Pooled Model
The pooled model estimates a standard Poisson regression, with a common intercept
term for all states. This model provides a poor fit to the data. Figure 1 shows the 95%
central interval for the value exp(λ̂) for all states in all years plotted against the actual
4
This follows the approach suggested by Singer and Willett (2003) in their work on longitudinal data
analysis
5
Data for this analysis, along with Bugs code for all models, and R code for all pre- and post-estimation
transformations of data and results can be obtained from the authors upon request.
22
number of institutions in that year. As the figure shows, the model is relatively good
at predicting values for states in the middle of the distribution, but provides a poor fit
for very large states such as California or very small states such as Delaware.
This lack of fit can be summarized in several ways. One relatively straightforward
method is to compare the posterior estimates of exp(λ̂) with the number of community
colleges in the state, and observe the number of times the 50 and 95% intervals do not
For the pooled model, the 50% interval includes the true value of community colleges
17% of the time, while the 95% interval includes the true value of community colleges
19% of the time. This indicates relatively poor model fit for the complete-pooling
model.
No-Pooling Model
The no-pooling model, as described above, estimates a separate intercept for every
state, thus removing any time-constant differences among states from the model. In-
For the no-pooling model, the true value of y lies within the central 50% prediction
interval 30% of the time, while the 95% central predication interval captures the true
value 54% of the time. While an improvement on the pooled model, this method of
estimation also does not fit the data well. A plot of the predicted ranges from the no
pooling model and the actual numbers of community colleges in the states can be seen
in figure 2.
Given the above considerations, these results do allow us to make the following
inferences:
The conceptual framework for this study suggested that states with more young peo-
ple would be more likely to create community colleges to respond to demand. Within
any state, an increase in the young population is associated with an increase in the
number of community colleges (see figure 5). Also, within-state growth in the pro-
portion of the young population that is non-white is also associate with an increasing
23
number of community colleges.
The conceptual framework also suggests that competition from existing institutions
of higher education could slow the development of community colleges within states,
do find that increases in population density also show a positive, although highly un-
the number of both public and private four year institutions within any given state are
Dougherty (1994, 1988b) who suggests that the competitive actions of within-state four
year institutions may have slowed the growth of community colleges. This also provides
initial support for the theories derived from the organizational ecology literature.
The partial pooling model, as described before, estimates separate intercepts for each
state, but generates these estimates from a common distribution. Following Gelman
and Hill (2006), the pooling model understates the variation across units, as figure 1,
shows, while the no-pooling model may overstate the variation, as figure 2 shows. The
The partial pooling model does not fit the data any better than the no-pooling
model. In this model, the true value of the number of community colleges lies within
the 50% interval 29% of the time, and within the 95% interval 54% of the time. As
figure 5 shows, this model does provide more efficiency, in that the posterior density
time trends. Take the example of Louisiana in figure 3. The rapid increase in the
number of community colleges in the state in the last period to be considered is not
modeled well, leading to poor predictive power. For this reason, we consider state-
24
Partial Pooling with Time Effects
The last model to be considered is essentially the same as the partial pooling model,
with the important addition of a state-specific time effect. This allows the estimates in
The partial pooling model with time effects provides a superior fit to the data. The
true value lies within the 50% predictive interval 43% of the time. The 95% interval
contains the true value 74% of the time. This model works better in capturing the data
than the other models considered in this paper 7 . Figure 4 shows that this model is
better at handling the kind of shocks that occurred in specific states, such as Louisiana,
Because it provides a better fit to the data, I will consider the results of the partial
pooling with time effects model in the most detail. As mentioned before, posterior
density estimates for the coefficients in the form of medians and 50 and 95% intervals
As figure 5 shows, the posterior density of the coefficient for young population in the
state is positive in all model specifications. As the proportion of the population aged
20-24 increased, in most states so did the number of community colleges. Figure 6
shows the predicted impact of an increase in the proportion of population aged 20-24
This figure plots the probability of any number of community colleges from 0-100
for each of four states: Illinois, New York, North Carolina, and Washington 8 . In each
6
This model shares much in common with the approach described by Singer and Willett (Singer and
Willett, 2003).
7
Posterior estimates of the deviance information criterion (DIC) were also calculated from all samples.
The DIC for the partial pooling with time effects has the lowest DIC, indicating the best out-of-sample
predictive accuracy of all models (Gelman et al., 2004)
8
These states were chosen because, as figure 1 they have either much higher or lower numbers of commu-
nity colleges than one might expect from a simpler model
25
states, the 25th, 50th and 75th percentile of the posterior predictive density is plotted
for each of the two hypothetical scenarios: under the lower value, the proportion of
young people is set to one standard deviation below the mean, while under the higher
value, the proportion of young people is set to one standard deviation above the mean.
exp(−µ)µm
P r(y = m) =
m!
Where µ = exp(λ̂). This prediction is carried out for values of λ̂ at each of the
The first panel in figure 6 shows the impact of increasing the proportion of young
people on the expected number of community colleges in Illinois. As the figure shows,
the median number of colleges centers around 40 in the lower hypothetical scenario
and around 60 in the higher hypothetical scenario. While the size of the difference
varies, the impact of an increasing proportion of young people can be observed in the
This provides support for the demand-side hypotheses from the conceptual frame-
to see that in states with more heterogeneous young populations, more community
colleges would be created. We found mixed evidence on this question. In the no-pooling
and partial pooling models, the coefficient for this variable is positive. However, once
we include a parameter for time, the 95% central density estimate for this coefficient
overlaps with 0.
As figure 5 shows, the predicted impact of the proportion of the young population
that is non-white is positive for most draws from the posterior distribution, but is
highly variable. Figure 8 plots the impact of an increase in the size of the proportion
26
of the population that is non-white on the predicted number of community colleges.
As the figure shows, in all three states the impact is substantively large, but highly
uncertain. The dotted gray lines indicate the 25th and 75th percentiles of the posterior
predictive distribution for λ̂. In each state, these range considerably from the median
estimate. This suggests that although the impact of this variable has been estimated
We did not find support for the class-reproduction theory in terms of income in-
equality in the states. In the no-pooling, partial-pooling and time effects model, the
central density estimate for the coefficient on the non-white young population variable
is centered on 0.
of community colleges. In our conceptual framework, we posit that states with more
liberal policymakers should be more likely to create more community colleges. Figure
5 shows that liberal ideology has a small but negative relationship with the number of
community colleges across states. Figure 7 shows the predicted impact of an increase in
liberal ideology, using the same method as previously. A the figure shows, the impact
of even fairly large changes in ideology (from one standard deviation above the mean to
one standard deviation below) have a very small impact on the number of community
colleges. While the relationship between liberalism and the building of community
to suspect that competition from other institutions would slow the growth of commu-
nity colleges, as would higher levels of population density. State resources, on the other
hand should increase the rate of founding of community colleges. In the no-pooling
and partial pooling model, the coefficient for the number of both private and public
posterior density estimates for the private institutions center on 0, while the 50% cen-
tral density estimate for public institutions is negative. This provides partial support
for the idea that competition from public institutions may have slowed the growth of
27
community colleges within states. Given the inclusion of the state-specific time param-
eter in this model, this means that a higher number of public four-year institutions is
Conclusion
This paper has attempted to model in a comprehensive fashion the development of
at the outset that the enormous complexity of such a phenomenon cannot be entirely
modeled through the use of only a handful of variables. Yet we also find that using
a relatively parsimonious but flexible model we can accurately predict the range of
In terms of our significant findings, we find first of all that community colleges did
indeed appear to grow as a direct result of the rapidly increasing proportion of young
people in states that occurred, particularly during the early part of our sample. This
provides support for the economic perspective examined in the literature review, which
suggested that these institutions came into existence in response to increased demand
for their services. Second, we find some evidence that the proportion of the population
that is non-white is also associated with the growth of these institutions, but we cannot
significant and negative, but substantively small, relationship between the level of state
liberalism and growth in community colleges. Last, we also find evidence that states
with more four year colleges may have seen slower growth in their community college
We derive several implications for the community college literature. First, there is
evidence that states were being directly responsive to their own needs, and less evi-
dence that factors like income inequality or political factors such as ideology impacted
the formation of these institutions. However, politics surely could be at work in the
lobbying efforts of public four year colleges. Much of the qualitative literature points
28
to resistance from these institutions in many states to the formation of community
colleges. The result of this could have been the slower growth of community colleges
Based on these specific findings, we suggest the following possible implications for
the theoretical frameworks described in our literature review and conceptual frame-
work. The results as stated do provide some evidence for the propositions derived
from the economic framework, but clearly more work needs to be done in this area.
In particular the specific demands of local labor markets have not been tied to the
development of these institutions. Future research could examine the link between
increased labor market demands for students with some postsecondary education and
The evidence from this study does not provide conclusive proof regarding the social
theories we examined earlier. Certainly, we did not find any evidence that income
inequality can be tied to the development of these institutions. The findings on racial
heterogeneity are not as clear, and suggest the need for a finer-grained approach to
understanding how forces for social stratification may have played a role in the devel-
Our evidence regarding the role of political ideology goes against what we expected,
based on extant theories. Theories regarding the politics of higher education are still
very much under development (McLendon, 2003). It could be that the role of polit-
ical liberalism did not extend to the development of these institutions, although the
Last, our findings regarding the impact of four-year institutions on the develop-
ment of community colleges are intriguing. One of the key questions to be answered
about more recent developments in higher education is the role of existing institutions
in the development of for profit institutions of higher education. Given the role of
existing institutions during the time period in this study, we would expect to see that
in many states these organizations would seek to limit the development of alternatives,
that are as-yet unknown but may be as large in impact as the community college could
develop. This study, by providing some insight into the development of one of the
largest higher education policy innovations in the 20th century, could provide clues as
to the conditions which may favor the development of the new new thing in higher
education.
References
Bafumi, J. and Gelman, A. (2006). Fitting multilevel models when predictors and
group effects correlate. Technical report, Columbia University, New York. 20, 21
Berry, W. D., Ringquist, E. J., Fording, R. C., and Hanson, R. L. (1998). Measuring
citizen and government ideology in the American states, 1960-93. American Journal
30(4):741–765. 4, 5, 12
Blau, J. R., McVeigh, R., and Land, K. C. (1994). The expansion of two-year col-
Blau, J. R., McVeigh, R., and Land, K. C. (2000). The expansion of two-year colleges:a
dynamic, multi level model. Community College Journal of Research and Practice,
24:127–143. 4, 5, 9, 11
30
Brint, S. and Karabel, J. (1989). The Diverted Dream:Community Colleges and the
leges:Policy in the Future Context, pages 3–22. Ablex Publishing, Westport, CT. 7,
Forum, 3(3):400432. 10
and Futures of the Community Colleges. State University of New York Press, Albany,
Dougherty, K. J. (2001). State policies and the community college’s role in workforce
leges:Policy in the Future Context, pages 129–147. Ablex Publishing, Westport, CT.
4
31
Galbraith, J. K. and Hale, T. (2006). State income inequality and presidential election
turnout and outcomes. Technical Report 33, The University of Texas Inequality
Gelman, A., Carlin, J. B., Stern, H. S., and Rubin, D. B. (2004). Bayesian Data
Gelman, A. and Hill, J. (2006). Data Analysis Using Regression and Multilevel/ Hier-
19, 20
Hannan, M. T., Carroll, G. R., Dundon, E. A., and Torres, J. C. (1995). Organizational
gium, britai, france, germany and italy. American Sociological Review, 60:509–528.
1271. 20
Hausman, J., Hall, B. H., and Griliches, Z. (1984). Econometric models for count data
20
7, 14
32
Keith, B. (1996). The context of educational opportunity:states and the legisla-
105(1):67101. 7
Levin, J. S. (2001). Public policy, community colleges, and the path to globalization.
National Cancer Institute (2007). Surveillance, epidemiology, and end results (seer)
P., B. W. and Carroll, G. R. (1987). Competition and mutualism among early telephone
of its historical origins. In Goodchild, L. F. and Wechsler, H. S., editors, The History
of Higher Education, pages 499–509. Pearson Custom Publishing, Boston, MA, 2nd
edition. 6, 7, 8
college. In Katsinas, S. G. and Palmer, J. C., editors, New Directions for Community
Periodicals. 8
Pincus, F. L. (1980). The false promises of community colleges: Class conflict and
33
Singer, J. D. and Willett, J. B., editors (2003). Applied Longitudinal Data Analysis:
Modeling Change and Event Occurrence. Ocford University Press, New York. 22, 25
ington, D.C. 4, 6, 7
34
Figure 1: Posterior Predictive Densities and Actual Values for All States, Pooled Model
1970 1990 1970 1990 1970 1990 1970 1990 1970 1990
AK AL AR AZ CA CO CT DE FL GA
● ●●●
●●●●●●●
●● ● ●●●● ●
●● ●
100 ●●
●
●●
●●
80
60 ●●●
●●
●
●
●
40 ●●●
●
● ●●
●●●●
●●
●
● ●
20 ●●●●●
●●●●●●
●●
●●
●
● ●●●
●●●●●
●●●●●
●
●
●●●● ●
●●●●●
●●● ●●●
●●●●●●
●
●●●●● ●● ●●●● ● ●●●● ●●● ●●●
●●
●●
●●●●●●
● ●● ●●●● ● ●●
0 ●●●●●● ● ●●●●●●●●
●●● ●●●●●● ●●●●●●
HI IA ID IL IN KS KY LA MA MD
100
80
60
●●● ●● ● ●●●●●● ●●
●● ●●● ● ● ● ●●●
●●●
40
●
●●●●
●●
●●
●●
20
●●● ●
● ● ●●●●●●●●●●● ●● ●●●● ● ●●●●●
●●●● ●● ● ●●●●●●●● ● ●●●●●● ●●● ●● ●●●● ● ●●●●●●●●
●● ● ●● ●●●● ● ●●●●●●●● ●●● ●●
●●●● ● ●●● ●● ●●●● ● ●●●●●●●● ●●● ●●●● ●●●● ●●●
●● ● ●● ●● ● ●●●●●●●● ●●
●●●●
●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●● ●●●●● ●● ●●●● ● ●●●●●
●●
●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●● ●●●●●● ●●●●●●●●●●● ●● ●●●●●●
0
ME MI MN MO MS MT NC ND NE NH
Community Colleges
100
80
60 ●
● ●
●●●
●●●● ●● ●●●● ● ●●●●●●●●
●●●
● ●●●
40
●
●●●
●
●●
20
●● ●●●●●●
●●●●● ● ●● ● ● ● ●●
●●●●●● ● ● ●●● ●●●●●●●●●●● ●● ●●●
● ●●●● ●●●
●●●●● ●● ●●●● ● ● ●●●● ● ● ●●
●●●●● ●●● ● ●●
● ● ●
●● ●● ● ●●●●●●●● ●●● ●●●● ●● ● ● ● ●●●
●●●●●●● ●● ●●●● ● ●●●●●●
●●●●●●●● ●●● ● ● ●● ●● ● ●● ●●●
● ●● ●●●● ● ●●●●●●●●●●● ●● ●●●
●●●●● ●● ●● ●● ●●●
0 ●●●●●●● ●●●●●●
●●● ●●●●
NJ NM NV NY OH OK OR PA RI SC
100
80
60
● ● ●●●
●● ●● ●●●●●●● ● ● ●
●●●●●● ● ● ●
●● ●●●
●●
●
●●●
● ●
● ●●●●● ●● ●
●●
●●●●●
●● ●●● 40
20
●●
●●● ●●● ●●
● ● ●●●●●●●● ● ●●●●● ●● ●●●● ● ●●●●●●●● ●●●
● ● ●●●●●●●● ●●● ●● ●
●●●●●●● ●● ●●● ● ● ●●● ● ●●●●● ●● ●●●● ● ●●●●●●●● ●●●
●●● ●●●
●●●● ● ●●●● ●●●●●●●●●●● ●● ●●●● ● ●●●●
●
●●●
●●● ●
●●●● ●
●●● ●● ●● ●●
●● ●
●
●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●●
●●●●●● ●● ●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●●
0
SD TN TX UT VA VT WA WI WV WY
100
80
●● ●●●
● ● ●●●●●●
60 ●●
●●● ●● ●●
●
●●
●
●●
40 ●
●●●
●●●
●
●●●●● ●● ●●●● ● ●●●●
●●●●
●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●●
20 ●● ●●
● ● ●●● ●●●
● ● ●●●●●●● ●● ●●●● ● ●●●●
● ●
●●● ●● ●●●● ● ●●● ●●●● ●●● ●
●●●●●●
●●
●●● ●●● ●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●● ●●●●● ●● ●●●● ● ● ●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●●
●●●● ●●●●● ●● ●●●
0 ●●●●●●●●●●● ●
●
●●●● ● ●●●●● ●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●● ●●
1970 1990 1970 1990 1970 1990 1970 1990 1970 1990
Year
Points indicate actual values, lines indicate 50 and 95% posterior prediction intervals
35
Figure 2: Posterior Predictive Densities and Actual Values for All States, No-Pooling (Fixed
Effects) Model
1970 1990 1970 1990 1970 1990 1970 1990 1970 1990
AK AL AR AZ CA CO CT DE FL GA
● ●●●
●●●●●●●
●● ● ●●●● ●
●● ●
100 ●●
●
●●
●●
80
60 ●●●
●●
●
●
●
40 ●●●
●
● ●●
●●●●
●●
●
● ●
20 ●●●●●
●●●●●●
●●
●●
●
● ●●●
●●●●●
●●●●●
●
●
●●●● ●
●●●●●
●●● ●●●
●●●●●●
●
●●●●● ●● ●●●● ● ●●●● ●●● ●●●
●●
●●
●●●●●●
● ●● ●●●● ● ●●
0 ●●●●●● ● ●●●●●●●●
●●● ●●●●●● ●●●●●●
HI IA ID IL IN KS KY LA MA MD
100
80
60
●●● ●● ● ●●●●●● ●●
●● ●●● ● ● ● ●●●
●●●
40
●
●●●●
●●
●●
●●
20
●●● ●
● ● ●●●●●●●●●●● ●● ●●●● ● ●●●●●
●●●● ●● ● ●●●●●●●● ● ●●●●●● ●●● ●● ●●●● ● ●●●●●●●●
●● ● ●● ●●●● ● ●●●●●●●● ●●● ●●
●●●● ● ●●● ●● ●●●● ● ●●●●●●●● ●●● ●●●● ●●●● ●●●
●● ● ●● ●● ● ●●●●●●●● ●●
●●●●
●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●● ●●●●● ●● ●●●● ● ●●●●●
●●
●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●● ●●●●●● ●●●●●●●●●●● ●● ●●●●●●
0
ME MI MN MO MS MT NC ND NE NH
Community Colleges
100
80
60 ●
● ●
●●●
●●●● ●● ●●●● ● ●●●●●●●●
●●●
● ●●●
40
●
●●●
●
●●
20
●● ●●●●●●
●●●●● ● ●● ● ● ● ●●
●●●●●● ● ● ●●● ●●●●●●●●●●● ●● ●●●
● ●●●● ●●●
●●●●● ●● ●●●● ● ● ●●●● ● ● ●●
●●●●● ●●● ● ●●
● ● ●
●● ●● ● ●●●●●●●● ●●● ●●●● ●● ● ● ● ●●●
●●●●●●● ●● ●●●● ● ●●●●●●
●●●●●●●● ●●● ● ● ●● ●● ● ●● ●●●
● ●● ●●●● ● ●●●●●●●●●●● ●● ●●●
●●●●● ●● ●● ●● ●●●
0 ●●●●●●● ●●●●●●
●●● ●●●●
NJ NM NV NY OH OK OR PA RI SC
100
80
60
● ● ●●●
●● ●● ●●●●●●● ● ● ●
●●●●●● ● ● ●
●● ●●●
●●
●
●●●
● ●
● ●●●●● ●● ●
●●
●●●●●
●● ●●● 40
20
●●
●●● ●●● ●●
● ● ●●●●●●●● ● ●●●●● ●● ●●●● ● ●●●●●●●● ●●●
● ● ●●●●●●●● ●●● ●● ●
●●●●●●● ●● ●●● ● ● ●●● ● ●●●●● ●● ●●●● ● ●●●●●●●● ●●●
●●● ●●●
●●●● ● ●●●● ●●●●●●●●●●● ●● ●●●● ● ●●●●
●
●●●
●●● ●
●●●● ●
●●● ●● ●● ●●
●● ●
●
●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●●
●●●●●● ●● ●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●●
0
SD TN TX UT VA VT WA WI WV WY
100
80
●● ●●●
● ● ●●●●●●
60 ●●
●●● ●● ●●
●
●●
●
●●
40 ●
●●●
●●●
●
●●●●● ●● ●●●● ● ●●●●
●●●●
●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●●
20 ●● ●●
● ● ●●● ●●●
● ● ●●●●●●● ●● ●●●● ● ●●●●
● ●
●●● ●● ●●●● ● ●●● ●●●● ●●● ●
●●●●●●
●●
●●● ●●● ●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●● ●●●●● ●● ●●●● ● ● ●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●●
●●●● ●●●●● ●● ●●●
0 ●●●●●●●●●●● ●
●
●●●● ● ●●●●● ●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●● ●●
1970 1990 1970 1990 1970 1990 1970 1990 1970 1990
Year
Points indicate actual values, lines indicate 50 and 95% posterior prediction intervals
36
Figure 3: Posterior Predictive Densities and Actual Values for All States, Partial Pooling
Model
1970 1990 1970 1990 1970 1990 1970 1990 1970 1990
AK AL AR AZ CA CO CT DE FL GA
● ●●●
●●●●●●●
●● ● ●●●● ●
●● ●
100 ●●
●
●●
●●
80
60 ●●●
●●
●
●
●
40 ●●●
●
● ●●
●●●●
●●
●
● ●
20 ●●●●●
●●●●●●
●●
●●
●
● ●●●
●●●●●
●●●●●
●
●
●●●● ●
●●●●●
●●● ●●●
●●●●●●
●
●●●●● ●● ●●●● ● ●●●● ●●● ●●●
●●
●●
●●●●●●
● ●● ●●●● ● ●●
0 ●●●●●● ● ●●●●●●●●
●●● ●●●●●● ●●●●●●
HI IA ID IL IN KS KY LA MA MD
100
80
60
●●● ●● ● ●●●●●● ●●
●● ●●● ● ● ● ●●●
●●●
40
●
●●●●
●●
●●
●●
20
●●● ●
● ● ●●●●●●●●●●● ●● ●●●● ● ●●●●●
●●●● ●● ● ●●●●●●●● ● ●●●●●● ●●● ●● ●●●● ● ●●●●●●●●
●● ● ●● ●●●● ● ●●●●●●●● ●●● ●●
●●●● ● ●●● ●● ●●●● ● ●●●●●●●● ●●● ●●●● ●●●● ●●●
●● ● ●● ●● ● ●●●●●●●● ●●
●●●●
●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●● ●●●●● ●● ●●●● ● ●●●●●
●●
●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●● ●●●●●● ●●●●●●●●●●● ●● ●●●●●●
0
ME MI MN MO MS MT NC ND NE NH
Community Colleges
100
80
60 ●
● ●
●●●
●●●● ●● ●●●● ● ●●●●●●●●
●●●
● ●●●
40
●
●●●
●
●●
20
●● ●●●●●●
●●●●● ● ●● ● ● ● ●●
●●●●●● ● ● ●●● ●●●●●●●●●●● ●● ●●●
● ●●●● ●●●
●●●●● ●● ●●●● ● ● ●●●● ● ● ●●
●●●●● ●●● ● ●●
● ● ●
●● ●● ● ●●●●●●●● ●●● ●●●● ●● ● ● ● ●●●
●●●●●●● ●● ●●●● ● ●●●●●●
●●●●●●●● ●●● ● ● ●● ●● ● ●● ●●●
● ●● ●●●● ● ●●●●●●●●●●● ●● ●●●
●●●●● ●● ●● ●● ●●●
0 ●●●●●●● ●●●●●●
●●● ●●●●
NJ NM NV NY OH OK OR PA RI SC
100
80
60
● ● ●●●
●● ●● ●●●●●●● ● ● ●
●●●●●● ● ● ●
●● ●●●
●●
●
●●●
● ●
● ●●●●● ●● ●
●●
●●●●●
●● ●●● 40
20
●●
●●● ●●● ●●
● ● ●●●●●●●● ● ●●●●● ●● ●●●● ● ●●●●●●●● ●●●
● ● ●●●●●●●● ●●● ●● ●
●●●●●●● ●● ●●● ● ● ●●● ● ●●●●● ●● ●●●● ● ●●●●●●●● ●●●
●●● ●●●
●●●● ● ●●●● ●●●●●●●●●●● ●● ●●●● ● ●●●●
●
●●●
●●● ●
●●●● ●
●●● ●● ●● ●●
●● ●
●
●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●●
●●●●●● ●● ●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●●
0
SD TN TX UT VA VT WA WI WV WY
100
80
●● ●●●
● ● ●●●●●●
60 ●●
●●● ●● ●●
●
●●
●
●●
40 ●
●●●
●●●
●
●●●●● ●● ●●●● ● ●●●●
●●●●
●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●●
20 ●● ●●
● ● ●●● ●●●
● ● ●●●●●●● ●● ●●●● ● ●●●●
● ●
●●● ●● ●●●● ● ●●● ●●●● ●●● ●
●●●●●●
●●
●●● ●●● ●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●● ●●●●● ●● ●●●● ● ● ●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●●
●●●● ●●●●● ●● ●●●
0 ●●●●●●●●●●● ●
●
●●●● ● ●●●●● ●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●● ●●
1970 1990 1970 1990 1970 1990 1970 1990 1970 1990
Year
Points indicate actual values, lines indicate 50 and 95% posterior prediction intervals
37
Figure 4: Posterior Predictive Densities and Actual Values for All States, Partial Pooling
With Time Effects Model
1970 1990 1970 1990 1970 1990 1970 1990 1970 1990
AK AL AR AZ CA CO CT DE FL GA
● ●●●
●●●●●●●
●● ● ●●●● ●
●● ●
100 ●●
●
●●
●●
80
60 ●●●
●●
●
●
●
40 ●●●
●
● ●●
●●●●
●●
●
● ●
20 ●●●●●
●●●●●●
●●
●●
●
● ●●●
●●●●●
●●●●●
●
●
●●●● ●
●●●●●
●●● ●●●
●●●●●●
●
●●●●● ●● ●●●● ● ●●●● ●●● ●●●
●●
●●
●●●●●●
● ●● ●●●● ● ●●
0 ●●●●●● ● ●●●●●●●●
●●● ●●●●●● ●●●●●●
HI IA ID IL IN KS KY LA MA MD
100
80
60
●●● ●● ● ●●●●●● ●●
●● ●●● ● ● ● ●●●
●●●
40
●
●●●●
●●
●●
●●
20
●●● ●
● ● ●●●●●●●●●●● ●● ●●●● ● ●●●●●
●●●● ●● ● ●●●●●●●● ● ●●●●●● ●●● ●● ●●●● ● ●●●●●●●●
●● ● ●● ●●●● ● ●●●●●●●● ●●● ●●
●●●● ● ●●● ●● ●●●● ● ●●●●●●●● ●●● ●●●● ●●●● ●●●
●● ● ●● ●● ● ●●●●●●●● ●●
●●●●
●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●● ●●●●● ●● ●●●● ● ●●●●●
●●
●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●● ●●●●●● ●●●●●●●●●●● ●● ●●●●●●
0
ME MI MN MO MS MT NC ND NE NH
Community Colleges
100
80
60 ●
● ●
●●●
●●●● ●● ●●●● ● ●●●●●●●●
●●●
● ●●●
40
●
●●●
●
●●
20
●● ●●●●●●
●●●●● ● ●● ● ● ● ●●
●●●●●● ● ● ●●● ●●●●●●●●●●● ●● ●●●
● ●●●● ●●●
●●●●● ●● ●●●● ● ● ●●●● ● ● ●●
●●●●● ●●● ● ●●
● ● ●
●● ●● ● ●●●●●●●● ●●● ●●●● ●● ● ● ● ●●●
●●●●●●● ●● ●●●● ● ●●●●●●
●●●●●●●● ●●● ● ● ●● ●● ● ●● ●●●
● ●● ●●●● ● ●●●●●●●●●●● ●● ●●●
●●●●● ●● ●● ●● ●●●
0 ●●●●●●● ●●●●●●
●●● ●●●●
NJ NM NV NY OH OK OR PA RI SC
100
80
60
● ● ●●●
●● ●● ●●●●●●● ● ● ●
●●●●●● ● ● ●
●● ●●●
●●
●
●●●
● ●
● ●●●●● ●● ●
●●
●●●●●
●● ●●● 40
20
●●
●●● ●●● ●●
● ● ●●●●●●●● ● ●●●●● ●● ●●●● ● ●●●●●●●● ●●●
● ● ●●●●●●●● ●●● ●● ●
●●●●●●● ●● ●●● ● ● ●●● ● ●●●●● ●● ●●●● ● ●●●●●●●● ●●●
●●● ●●●
●●●● ● ●●●● ●●●●●●●●●●● ●● ●●●● ● ●●●●
●
●●●
●●● ●
●●●● ●
●●● ●● ●● ●●
●● ●
●
●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●●
●●●●●● ●● ●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●●
0
SD TN TX UT VA VT WA WI WV WY
100
80
●● ●●●
● ● ●●●●●●
60 ●●
●●● ●● ●●
●
●●
●
●●
40 ●
●●●
●●●
●
●●●●● ●● ●●●● ● ●●●●
●●●●
●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●●
20 ●● ●●
● ● ●●● ●●●
● ● ●●●●●●● ●● ●●●● ● ●●●●
● ●
●●● ●● ●●●● ● ●●● ●●●● ●●● ●
●●●●●●
●●
●●● ●●● ●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●● ●●●●● ●● ●●●● ● ● ●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●●
●●●● ●●●●● ●● ●●●
0 ●●●●●●●●●●● ●
●
●●●● ● ●●●●● ●●●●●●●●●●● ●● ●●●● ● ●●●●●●●● ●●● ●●
1970 1990 1970 1990 1970 1990 1970 1990 1970 1990
States
Points indicate actual values, lines indicate 50 and 95% posterior prediction intervals
38
Figure 5: Posterior Estimates of Coefficients, All Models
Complete Pooling Model No Pooling Model
Instititutions Institutions
Institutions Institutions
Log of Population ●
Log of Population ●
Density Density
Legislative Ideology ●
Legislative Ideology ●
Income Inequality ●
Income Inequality ●
Non White ●
Non White ●
Young Population ●
Young population ●
−0.2
0.0
0.2
0.4
0.6
0.8
1.0
−0.2 0.0 0.2 0.4 0.6 0.8 1.0
Estimate Estimate
Institutions Institutions
Institutions Institutions
Log of Population ●
Log of Population ●
Density Density
Legislative Ideology ●
Legislative Ideology ●
Income Inequality ●
Income Inequality ●
Non White ●
Non White ●
Young population ●
Young population ●
−0.2
0.0
0.2
0.4
0.6
0.8
1.0
−0.2
0.0
0.2
0.4
0.6
0.8
1.0
Estimate Estimate
Points represent medians, black lines indicate central 50%, gray lines indicate central 95%
of posterior estimate
39
Figure 6: Predicted Impact of Young Population on Number of Community Colleges
Illinois New York
0.06
0.06
0.05
0.05
Predicted Probability
Predicted Probability
0.04
0.04
0.03
0.03
0.02
0.02
0.01
0 20 40 60 80 100 0 20 40 60 80 100
0.08
0.07
0.05
0.06
Predicted Probability
Predicted Probability
0.04
0.05
0.03
0.04
0.02
0.03
0.01
0.02
0 20 40 60 80 100 0 20 40 60 80 100
Black lines indicate lower value of young population, gray lines indicate upper. 50th and
75th percentile are represented by dotted lines, medians by solid lines
40
Figure 7: Predicted Impact of Liberal Ideology on Number of Community Colleges
Illinois New York
0.05
0.05
0.04
0.04
Predicted Probability
Predicted Probability
0.03
0.03
0.02
0.02
0.01
0.01
0 20 40 60 80 100 0 20 40 60 80 100
0.07
0.06
0.04
Predicted Probability
Predicted Probability
0.05
0.03
0.04
0.03
0.02
0.02
0.01
0.01
0 20 40 60 80 100 0 20 40 60 80 100
Black lines indicate lower value of liberal ideology, gray lines indicate upper. 50th and 75th
percentiles of prediction are represented by dotted lines, medians by solid lines
41
Figure 8: Predicted Impact of Non-White Young Population on Number of Community
Colleges
Illinois New York
0.06
0.06
0.05
0.05
0.04
0.04
Predicted Probability
Predicted Probability
0.03
0.03
0.02
0.02
0.01
0.01
0.00
0 20 40 60 80 100 0 20 40 60 80 100
0.06
0.04
Predicted Probability
Predicted Probability
0.03
0.04
0.02
0.02
0.01
0.00
0.00
0 20 40 60 80 100 0 20 40 60 80 100
Black lines indicate lower value of non-white young population, gray lines indicate upper.
50th and 75th percentile are represented by dotted lines, medians by solid lines
42
Figure 9: Predicted Impact of Public Four Year Institutions on Number of Community
Colleges
Illinois New York
0.06
0.06
0.05
0.05
0.04
0.04
Predicted Probability
Predicted Probability
0.03
0.03
0.02
0.02
0.01
0.01
0.00
0 20 40 60 80 100 0 20 40 60 80 100
0.06
0.04
Predicted Probability
Predicted Probability
0.03
0.04
0.02
0.02
0.01
0.00
0.00
0 20 40 60 80 100 0 20 40 60 80 100
Black lines indicate lower value of public four year institutions, gray lines indicate upper.
50th and 75th percentile are represented by dotted lines, medians by solid lines
43