You are on page 1of 13

417

Higher Education 7 (1978) 417 -429


9 ElsevierScientificPublishing Company,Amsterdam - Printed in the Netherlands

F O R E C A S T I N G U N I V E R SI T Y E N R O L L M E N T S BY RATIO SMOOTHING

MARILOU T. HEALEY
Division of Undergraduate Education, State University of New York at Buffalo, U.S.A.

and

DANIEL J. BROWN
Centre for the Study o f Administration in Education,
University of British Columbia, B.C., Canada

ABSTRACT

A smoothing model was applied to the transition ratios of the number of university
students in successive time periods. Each application was conditional on the department
and year of student. The model was trained and validated, using a single university for
one, two, and three time periods in advance. Results showed a reasonable error under
conditions of conservative testing. Further, some enrollment trends were made apparent
within the university studied. Since the model was simple in conception, not difficult
to apply, and the results may be applied to any level of aggregation in the institution,
implementation prospects appear to be relatively good.

Introduction

Recent higher education enrollment trends suggest a need for new


strategies and techniques for planning, especially in the area o f quantitative
systems and forecasting methods. Unstable enrollments, curriculum moderni-
zation and scarcity o f funds are all pressures which no longer permit an
institution the luxury o f rule-of-thumb procedures. Rourke and Brooks
(1971) found that public universities generally fare much bet t er in their
interactions with the c o m m u n i t y and the state if administrators utilize
space-utilization formulas, cost analysis and ot her quantitative measuring
devices in justifying their requests for support.
This research has m a ny implications for university administrators
interested in short-range planning relating to enrollment projections. As
a consequence o f the growing scarcity o f resources, decision-makers are
becoming aware o f the need to have flexible plans available. Hence, knowl-
edge o f forecasting tools which enable planners to k n o w where the demand
for services will be in the near future is valuable. Inform at i on regarding
trends in student enrollment permits decision-makers to set priorities and
418

assess the status of a university department and thereby make needed deci-
sions regarding what departments should be expanded or, though more
difficult, what departments should face resource retrenchment.
The purposes of this research were: (1) to test a new method of fore-
casting university enrollments using the ratio smoothing model "and (2) to
uncover any significant enrollment trends which might be representative of
those occurring generally in institutions of higher education.

OVERVIEW OF PROJECTION TECHNIQUES

Several different planning techniques are currently utilized in the


development of enrollment estimates. It seems reasonable to classify them
into six general groups: judgmental methods, Markov methods, trend anal-
ysis, regression analysis, simulation, and ratio methods. An elaboration of
these models follows.
(1) Forecasts are often made strictly on some qualitative or judgmental
basis. This occurs under conditions of the failure of known planning models
to project with sufficient accuracy, or simply a lack of knowledge of any
mathematical models which might be applied to the problem.
(2) A second class of methods has utilized Markov processes to forecast
future enrollments within educational systems. Wasik (1971) referred to
these as quasi-Markov models since information about the probability of
individuals with certain characteristics moving from one academic location
to another can be used to predict future movement of individuals with the
same characteristics. If the distribution of students among specific categories
is known for a particular time interval, then the observed movement may be
used to forecast the distribution of students for the next time period.
Markov processes as presented by Johnstone and Philip (1973) along with
cohort survival (Correa, 1967), provide the advantage of a system-wide view
of student flow. Under relatively stable conditions of movement, such
models have potential. However, both these models demand that students be
tracked individually and consequently the college or university information
system must have on hand the particular history of movement of each pupil.
Such knowledge is normally either not available or inaccessible, probably
because the need for it has not heretofore been apparent.
(3) The use of trend lines and time series applied to actual enrollment
counts is another general method by which forecasts may be made. A
variation on this theme usually referred to as classical decomposition may be
used to account for periodic fluctuations in enrollments that occur with
monthly patterns within any given year, or for semester or quarterly patterns
if applicable. One of the" most common techniques used is trend analysis
which consists of determining the functional relationship that exists between
past enrollments and years. The relationship is then projected to the years
419

for which the enrollment figures are needed. The chief assumption made
with this method is that trends based on historical information will continue
and that influences of the past are indicative of those factors which will
occur in the future (Lins, 1960). Relatively sophisticated time series tech-
niques such as that pro~,ided by Box and Jenkins (1970) appear promising
because of their normally high degree of forecast accuracy. However, the
paucity of relevant observations coupled with a fairly high cost of implemen-
tation make such planning models comparatively unattractive at this time
(Wheelwright and Makridakis, 1973).
(4) The approach using indicator methods and multiple regression is
viable because once the key indicator variables and their lead times are
determined, they may be used to predict enrollment change in an explantory
rather than simply a projective manner. This kind of model is, in a sense,
independent of the past history of the enrollment process. The changes in
enrollments may be found to be related to a single variable or to a combina-
tion of variables (Lins, 1960). Examples of independent variables used in
making projections are economic indices and the population of 18-21 year
olds. The difficulty with multiple regression lies chiefly with the determina-
tion of the indicator variables (they tend to vary from college to college) and
the high costs involved in the acquisition of relevant data.
(5) A simulation approach involving many input variables and probabal-
istic outcomes is provided by Denham (1971). The advantage is that fore-
casts have their errors explicitly attached to them and hence are more valu-
able to the educational decision-maker. Unfortunately, the simulation
model requires considerable data gathering, and the errors involved are
relatively large.
(6) The ratio method is also widely used because of its simplicity.
This technique determines a ratio between numbers of students enrolled at
two different times. The ratio method can be used to forecast freshmen,
sophomores, juniors and seniors separately, but it has rarely been applied
in this way (Lins, 1960). Rather, in most studies, the ratio used is calculated
by the division of the total university enrollment by the total university-age
population defined as all persons in a region who are 18-21 years of age.
But ratios may be used as transition measures between periods, such as
from one year, semester, or quarter to another. They are a quotient of two
groups of people invariably containing many of the same individuals. The
ratio measure is a number near 1.0, closer to 0.9 when there are losses from
the group between the periods, and nearer 1.1 when there are gains.
As demonstrated by Healey (1976), once future ratios are forecasted,
future enrollments may be predicted. The challenge is to select a forecasting
model which makes sense and forecasts accurately.
420

The Model

The underlying assumption of ratio smoothing is based upon. the causal


factors, either unknown or not quantifiable, which influence each transition
ratio. Such factors include the marketability of program skills, the student
spirit of the times (liberal or conservative), administrative decisions to
augment or curtail programs, rates of promotion of students, and the retain-
ing power of universities. If it is assumed that the ratios remain approxi-
mately constant apart from the random variations induced by these causal
factors, then it is possible to derive a formula for projecting a future ratio
on the basis of past observations of the same ratio. The projection works
in this manner: the most recent ratio is given a weight, the next most recent
a lesser weight, and so forth into the past. This outcome also appears plausi-
ble, since intuitively, ratios from the recent past should be more important
to the forecast than those from the distant past. When a trend in the ratios
is apparent, an analogous model called ratio trend smoothing is applicable.
Both these models and their accompanying equations are discussed in the
appendix.
Smoothing has been applied previously to raw enrollment data in man-
agement schools (Planisek et al., 1974) and to public school district data by
Brown (1975) and Watson (1975).

The Modeling Process

Ratio and ratio trend smoothing were used to project ratios for all
grade classifications for three semesters into the future. The forecasted ratios
were translated into actual enrollments by multiplying an enrollment figure
by the ratio to arrive at a new projected enrollment.

TRAINING AND VALIDATION

The models were trained separately for: each student year (freshmen,
sophomores, etc.), each of the two model variations (ratio and ratio trend
smoothing), each transition (fall-to-spring and spring-to-fall), and each
projection (one, two, and three semesters into the future). Projections and
error calculations were made for past data. From the resulting error calcula-
tions, the minimum error was found, its associated weight value noted, and
that value was used for the next "actual" projection. In this way, the model
is an adaptive version of'smoothing, since it presumably finds its optimal
weight value. This adaptive method was used because the nature of the tran-
sition process has changed. That is, student transition behavior has altered
between 1965 and 1975. The actual number of students overpredicted or
421

underpredicted relates directly to the number of professors hired or released


and hence, errors in either direction were viewed as equal in importance.
Results from the validation phase of ratio smoothing were contrasted to
those of ratio trend smoothing for each department. The model selected as
the better, or optimal, for each enrollment category was that model which
exhibited the least average error.
Training, validation, and projection for all student years, transitions,
models, projections, and departments was accomplished by a FORTRAN
program in about five minutes using a Control Data Corporation 6400 com-
puter.

Results

AVERAGEERROR

When all combinations of student year, transitions, models, projections,


and departments were considered, there were 1056 times when a choice was
necessary between ratio and ratio trend smoothing. The latter produced the
least error in 234 cases or only about 22% of the time, suggesting that for
the most part, trends in transition ratios are n o t evident in universities.
As found in Table I, the university-wide average error was 17 students
or 56% of the average group size. These errors were calculated by finding an
average error for each student year, for three projections into the future.
The four student years were subsequently averaged in order to arrive at an
error for each department. The 44 undergraduate departmental errors were
then averaged. As for the three projections, one obvious conclusion is that
the further into the future projections occur, the larger the error. Since
earlier forecasts were used to generate later forecasts, errors necessarily
increased.

TABLE I

Overall Average Error for 3 Projections into the Future

Projection Projection Projection Overall


I II III Average

Average mean
error 10.58 17.62 23.83 17.34
Average mean
percent error 31.40 60.11 76.99 56.17
422

Apart from distance into the future, another factor contributing to


error is size of student year, the average being 31 with considerable disper-
sion. Because of the existence of some departments containing groups of less
than 10, percentage errors were inflated. Due to the presence of departments
with group sizes numbering about 200, the numerical errors were inflated.
Another consideration was the small number o f observations available.
Some departments possessed less data to train and validate the model than did
other departments. Specifically, 29 departments had 8 years o f past data,
while 15 departments had only 6 to 7 years to train and validate the models.
A final problem is one o f data error. A change in the registration system
in the university at about year 5 resulted in data o f questionable validity.
However, this disturbance was n o t corrected since (1) these were the best
data available and (2) college record systems are subjected to disturbance
o f this kind quite frequently.
As a consequence of the inclusion o f b o t h small and large student
groups, the small amount of time for training and validation, and disturbance
in the data, it may be said that the smoothing models were subjected to a
very stringent test.

ERROR BY YEAR OF STUDENT

As indicated in Table II, average numerical error and average percent


error were found to be the least for the senior group for two projections.
Seniors appear to be the most stable because it is very difficult to change
department majors at that juncture. Sophomores represented the year with
the highest average percent error for the second and third projections and
the second highest numerical errors. The years with the most error were the
sophomores and juniors which can be understood in that most changes in
undergraduate majors occur during these two middle years of college life.

TABLE II
Student Year Average Error for 3 Projections into the Future

Projection Projection Projection


I II III
Av. Mean Av. Mean Av. Mean Av. Mean Av. Mean Av. Mean
Err. Pete. Err. Err. Perc. Err. Err. Perc. Err.

Freshmen 10.68 35.29 16.01 55.76 18.05 66.44


Sophomores 12.71 30:84 19.94 71.65 27.20 88.96
Juniors 13.48 31.89 24.50 65.86 37.33 75.51
Seniors 9.77 24.46 15.47 42.55 21.15 68.41
423

ERROR BY TYPE OF STUDENT

Departments were grouped into their associated divisions or schools


because if trends were occurring on a departmental level, the likelihood o f
the trend being apparent on a division-wide level was quite great. As shown in
Table III, for the first semester projection, Health Sciences clearly portrayed
the smallest average percentage error o f 20.39, though it also incurred the
largest average numerical error. This can be explained by the fact that
Health Sciences is composed o f departments which are, in general, larger in
enrollments than the typical department in the other five divisions. On the
other hand, Educational Studies had the largest average percent error o f
48.50 and the smallest associated average numerical error, partly because it is
comprised o f only two very lightly enrolled departments and both o f which
had less than the 16 semesters o f past data. Social Sciences and Natural
Sciences b o t h had very close average percent errors. For the most part, b o t h
of these schools have large departmental enrollments which would account for
their average numerical errors. Also, both divisions were comprised of depart-
ments which had 16 semesters of past enrollment data available, thus account-
ing in some measure for their smaller-than-average percent error. The School of
Engineering had a larger-than-average percent error, 40.47. However, it is
important to note that in no department within this division were there
more than 14 semesters of past data available. The associated average numer-
ical error was 10.76 which is very close to the average for all departments
during the first projection. From a brief scan of the data, it is apparent that
engineering departments approximate the average size o f all university

TABLE III
Faculty-Wide Average Error for 3 Projections into the Future

Projection Projection Projection


I II III
Av. Mean Av. Mean Av. Mean Av. Mean Av. Mean Av. Mean
Err. Perc. Err. Err. Perc. Err. Err. Perc. Err.

Arts and letters 8.95 35.28 16.60 88.34 17.63 81.09


Natural sciences 9.21 27.74 15.81 51.31 20.70 50.13
Engineering 10.76 40.47 17.71 59.21 21.15 67.71
Educational
studies 5.89 48.50 8.79 78.39 11.67 179.67
Health sciences 12.80 20.39 16.70 29.94 26.43 46.47
Social sciences 11.69 27.42 19.86 54.67 32.18 93.13
424

departments. Arts and Letters is very close to the mean for all university
departments for average percent error and average numerical error. Arts and
Letters seems to represent that school which can be most accurately identi-
fied as typifying the "normal" university division. For the second and third
projection into the future most of the patterns identified for the first projec-
tion were again evident.

ENROLLMENT TRENDS

Further analysis shows that the trend in most of the arts and letters
departments is toward gradually decreasing enrollments while the Faculty
of Educational Studies trend is sharply downward. In the more career-
oriented and pre-professional programs, enrollments are definitely increasing
as shown by the upward trend in engineering departments, the school of
management and health sciences related programs. Social science departments
portrayed mixed results, although programs with human service orientations,
such as psychology, indicated rising enrollments. The more humanistically
oriented departments appear to be losing students which was most obviously
seen in the philosophy department. The last school studied, natural sciences
and mathematics, generally indicated increasing enrollments.

Discussion

MODELING

Observation of the model training indicates that ratio smoothing tends


to produce less average error when a small amount of data is available, even
when a trend is evident, while ratio trend smoothing appears to produce less
average error as more data are used during the process, providing a trend is
pronounced and continuing. The evidence suggests that ratio trend
smoothing is the more sensitive planning model though it must have sub-
stantial past information in order to function properly. The conclusion
drawn from the results is that at least ten past data points are necessary for
ratio trend smoothing to begin performing with much accuracy. Ratio trend
smoothing might have been reported as the optimal model more frequently,
and thereby have indicated the occurrence of additional trends, if more data
had been available.
The selection of the optimal weighting value is another issue. The value
did not remain stable, but rather, ranged from 0.05 to 1.00, fluctuating
considerably. Implicit in these results is the notion that departmental enroll-
ment stability is associated with weighting value stability. The implication is
that if the data portray dramatic fluctuations at some point because of a
425

factor not inherent in the actual enrollments, sufficient observations must


be available to give the model time to overcome this problem.
A further modeling consideration basic to the selection of the optimal
weighting value is the weight of the base series versus the weight of the trend
component. In this study, the trend weight was always set equal to the
weight of the base series. However, Planisek et al. (1974) weighted the trend
component with a different value than was used in the base series, resulting
in a matrix of weights from which they chose the ones resulting in minimum
error.
The ratio smoothing technique may be a partial solution to the problem
of forecasting majors which is encountered by the Resource Requirements
Prediction Model developed by the National Center for Higher Education
Management Systems at the Western Interstate Commission for Higher
Education. If headcounts can be converted to full-time equivalent students,
then providing programmatic changes could be ascertained, the RRPM model
could be used in the prediction mode.
Possible alternatives to the ratio smoothing model used in this paper
include moving averages or weighted moving averages. While they appear less
desirable a priori, their relative accuracy for projection is unknown.

UNIVERSITY PLANNING

Forecasting student enrollment is a necessary prerequisite to the


construction of essential and rational plans within institutions of higher
education. Realistic plans and budgets cannot be developed without the
ability to forecast for some period into the future. Therefore, possessing
accurate information about what is likely to happen in the future regarding
student enrollments is essential when decisions are made concerning the
status of a department or program. Consequently, the implications for
planning that accrue as a result of possessing an implementable forecasting
model are tremendously significant. The trends uncovered in this research
have important implications for the future of higher education. Either the
institutions of higher education will have to expand their facilities in the
areas of increasing enrollment or recognize that increased competition for
the limited enrollment spaces will result. Vocationally oriented, career train-
ing, and pre-professional programs are now facing a surge in higher education
enrollments, while the traditional liberal arts departments are facing severe
retrenchment as a result of their decreasing popularity.
A very important level of planning occurs at the departmental level
such that individual department chairpersons could easily use the results of
this model. By having available information regarding future enrollments,
estimates for student year size, needed space and faculty can be accom-
plished with relative ease.
426

However, pertinent to the problem of retrenchment is the necessity to


recognize the political consequences of attempting to deal with severely
entrenched departments within a university. Nowhere within the university
bureaucracy is the political process more evident than at the departmental
level. It is imperative to be cognizant of the fact that no matter how accurate
a forecasting model, the implications are further reaching than merely having
available good quantitative information for sound decision making. The
necessity of having accurate projections is accepted, but the realities of how
decisions are actually arrived at needs to be recognized. Bureaucracies are
typically composed of varying kinds of political processes. While resources
are readily available, these processes are much easier to deal with for admin-
istrators. In contrast, when resources are scarce, as is currently true, the
political aspects of the organization create severe dilemmas for the decision
makers.
However, though the political components of the institution do com-
plicate the decision processes, administrators must always recognize that
accountability is essential. If an enrollment trend clearly indicates the
demise of a department, an administrator must balance the internal political
pressures to further channel resources in that direction with the increasing
external public demand for justifications concerning the expenditure of
funds.
A second point of utilization is at the division level. Deans would be
able to make enlightened judgements concerning where the resources should
be allocated within each of their associated departments. Though an entire
division may be increasing or decreasing in size, particular departments often
exhibit pronounced trends. The institutional level is the third area of plan-
ning that could benefit considerably the implementation of this model.
Those planners who are responsible for making budget recommendations
would find it useful to forecast the enrollment patterns of each division in
order to determine priorities for resource allocation purposes. The head-
count information for the next three semesters provides a more rational basis
for budget allocations. Furthermore, by being aware of student enrollment
trends, the overall plans concerning the institution's role in serving the com-
munity can be better articulated. At the state or province-wide level, the
model could be utilized in the central planning offices of an entire university
system. By implementing the ratio smoothing technique, those administrators
who must recommend courses of action would be aware of the enrollment
trends occurring at each institution.
Enrollment projection information therefore appears to be crucial at
numerous planning stages in higher education administration. The model
applied in this research could consequently by implemented by decision-
makers at various levels in order for optimal planning to result.
427

Appendix
An observed transition ratio is defined as

Y(t,k)
R(t) = (1)
Y(t-1 ,k-l)

where Y represents enrollment, t is a measure of the time period, and k is the grade or
year of student. Given R(t) as a ratio defined above pertaining to a specific university
department, within a specific class and a specific transition without loss of generality, then
its causal factors may be specified as X ( i , t - j ) , where i is the factor number, t is the year
of measurement, and ] is the lead time dependent on i. That is to say,

R(t) = f [.Z X ( i , t - / ) ] (2)


l

If it can be assumed that the summation of those factors represent a stationary random
process, then

R(t) = a + e (t) (3)

where a is the u n k n o w n mean of the time series of ratios, and e(t) represents the random
variations from the various inputs. Given the further assumption that the time series of
ratios has a memory of strength inversely related to a parameter A which ranges between
0 and 1 and is determined empirically, then from Brown (1963) it is possible to forecast
R(t + 1) from observables using

l~(t+ 1 ) = A R ( t ) + A ( 1 - A ) R ( t - 1)+A(1 - A ) 2 R ( t ' - 2 ) + . . . +


A(1 - A ) n R ( t - n ) + . . . + (1 - A ) t R(0), (4)

the formula which clearly indicates how the impact of each observation declines geometri-
cally as one proceeds into the past, or from

l~(t + 1) = AR(t) + (1 - A ) l~(t) (5)

the iterative version which is simpler to use for computational purposes.


In Equations 4 and 5, the sum of the weights is equal to one. A projection of
enrollments is made by multiplying the number of students in one grade at time t by the
projected ratio l~(t + 1) to give the number of students in the next grade for the following
period:

r + 1, k + 1) = l~(t + 1) V(t,k) (6)

The simple ratio smoothing model is most appropriate if the ratios are approxi-
mately constant. However, if a time series of ratios portrays a trend, then a basic trend
assumption, analogous to Equation 3, is more appropriate:

R(t) = a + b t + e (7)

and a double ratio smoothing model results (Brown, 1963). That is, besides smoothing
428

the actual ratios, the slope of the line joining these ratios is also smoothed and incorpo-
rated into the model. Two smoothing operations are therefore taking place simulta-
neously, one on the actual ratios and one on the changes in ratios. The following equation
pertains:

13(t) = A[l~(t) - R(t - 1)tt- (1 - h ) B ( t - 1) (8)

where 13(t) is the trend being estimated, R ( t ) - R ( t - 1) is the apparent trend, 13(t- 1) is
the trend previously estimated, and t is the time in years.
Both of these smoothed values are combined in developing the following model:

V(t) = R(t) + [(1 - A)/A] 13(t) (9)

where ~ ( t ) is the estimated initial ratio. The final prediction is obtained from:

l~(t + m) = ~/(t) + m l~(t) (10)

where m is the projected period 1, 2, 3, and I~ is the ratio forecasted.


All model applications were given the initial condition of A = 0.50 to permit a
beginning projection for which an error calculation could be made. Parameter values
ranged from 0.05 to 1.00 in steps of 0.05.
Average numerical error was defined as

Y~IE-EI
A= n (11)

where E is the observed enrollment, I~ is the enrollment projected from the previous
year, and n is the number of periods for which the error could be calculated. So that
the training time for any given series would be as long as possible, validation was re-
stricted to the last two periods of observation.

References

Box, G. P. E. and Jenkins, G. M. (1970). Time Series Analysis. San Francisco: Holden-
Day.
Brown, Daniel J. (1975). " A smoothing solution to the school district enrollment projec-
tion problem," Educational Planning, 2( 1): 1 3 - 2 6 .
Brown, Robert (1963). Smoothing, Forecasting and Prediction of Discrete Time Series.
Englewood Cliffs, New Jersey: Prentice-Hall.
Correa, Hector, ed. (1967). Mathematical Models in Educational Planning. Paris: Organiza-
tion for Economic Cooperation and Development.
Denham, Carolyn H. (1971). Probabilistic School Enrollment Projections Using Monte
Carlo Computer Simulation. D.H.E.W., O.E.E.R.I.C. ED 0 6 2 - 7 2 9 .
Healey, Marilou T. (1976). "Enrollment projections: application of a mathematical
planning model to forecast the number of undergraduate majors," Unpublished doc-
toral dissertation, State University of New York at Buffalo.
Johnstone, James N. and Philip, Hugh (1973). "The applications of a Markov chain in
educational planning," Socio-Economic Planning Sciences, 7(3): 2 8 3 - 2 9 4 .
429

Lins, L. J. (1960). "Methodology of enrollment projections for college and universities,"


Paper presented at the annual meeting of the American Association of College Regis-
trars and Admissions Officers.
Planisek, R. J., Krampf, Robert F. and Heinlein, Albert C. (1974). An Evaluation o f Col-
lege and University Forecasting Methods. Paper presented to the Decision Science
Institute Program, "New Approaches to Planning in Higher Education," Kent State
University.
Rourke, Francie E. and Brooks, Glenn E. (1971). "The 'Managerial Revolution' in Higher
Education," in J. Victor Baldridge, ed., Academic Governance, pp. 169-192. Berkeley:
McCutcheon.
Wasik, John (1971). The Development o f a Mathematical Model to Project Enrollments
in a Community College System. Technical Report, O E G - 2 - 7 - 0 7 0 3 4 8 - 2 6 9 8 ,
Center for Occupational Educational Education, North Carolina State University at
Raleigh.
Watson, Cicely (1975). "Project notes - 1 ," Educational Planning, 2(1): 36-56.
Wheelwright, Stephen C. and Makridakis, Spyros (1973). Forecasting Methods f o r Man-
agement. New York: Wiley.

You might also like