Sunteți pe pagina 1din 21

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

net/publication/238836457

A Three-Stage Manpower Planning and Scheduling Model--A Service-Sector


Example

Article  in  Operations Research · June 1973


DOI: 10.1287/opre.21.3.693

CITATIONS READS

131 1,798

4 authors, including:

John C Hershey Sten Wandel


University of Pennsylvania Lund University
102 PUBLICATIONS   5,219 CITATIONS    17 PUBLICATIONS   384 CITATIONS   

SEE PROFILE SEE PROFILE

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

Governing transitions towards Low-Carbon Energy and Transport Systems for 2050 (LETS2050) View project

All content following this page was uploaded by John C Hershey on 14 December 2014.

The user has requested enhancement of the downloaded file.


A Three-Stage Manpower Planning and Scheduling Model-A Service-Sector Example
Author(s): William J. Abernathy, Nicholas Baloff, John C. Hershey and Sten Wandel
Reviewed work(s):
Source: Operations Research, Vol. 21, No. 3 (May - Jun., 1973), pp. 693-711
Published by: INFORMS
Stable URL: http://www.jstor.org/stable/169379 .
Accessed: 23/07/2012 17:46

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Operations Research.

http://www.jstor.org
A Three-Stage Manpower Planning and Scheduling
Model-A Service-Sector Example

William J. Abernathy
HarvardUniversity,Cambridge,
Massachusetts
Nicholas Baloff, John C. Hershey, and Sten Wandel
StanfordUniversity,Stanford,California
(ReceivedOctober12, 1972)

This paper presents a staff planning and scheduling model that has specific
applicationin the nurse-staffing process in acute hospitals, and more general
applicationin many other service organizationsin which demand and produc-
tion characteristics are similar. The aggregate planning models that have
been developed for goods-producing organizations are not appropriatefor
these types of service organizations. In this paper the process for staffing
services is divided into three decision levels: (a) policy decisions, including
the operating procedures for service centers and for the staff-control process
itself; (b) staff planning, including hiring, discharge, training, and realloca-
tion decisions; and (c) short-term scheduling of available staff within the
constraints determined by the two previous levels. These three planning
'levels' are used as decomposition stages in developing a general staffing
model. The paper formulates the planning and scheduling stages as a sto-
chastic programmingproblem, suggests an iterative solution procedure using
random loss functions, and develops a noniterative solution procedure for a
chance-constrained formulation that considers alternative operating proce-
dures and service criteria, and permits including statistically dependent
demands. The discussion includes an example applicationof the model and
illustrations of its potential uses in the nurse-staffing process.

THE PURPOSE of this paper is to present a staff planning and scheduling model
that has been specifically formulated for use in the nurse-staffing process in acute
hospitals. However, the characteristics of the model are sufficiently general that
it can be adapted for profitable application in other service organizations. Benefits
from applying the model are likely to be highest for organizations that experience
highly variable demand for specialized services that must be provided in a timely
manner (not 'backordered')-particularly where the service capacity is not readily
expandable in the short run and the services themselves cannot be 'inventoried.'
Examples of such organizations that come readily to mind include police depart-
ments, ambulance services, fire departments, and commercial banks.
In this introduction to the model formulation, the general nature of the problem
in service-sector organizations is described and contrasted with the familiar work-
force and production-planning problem in goods-producing organizations. Pre-
vious literature relating to the nurse staffing process is then discussed and a three-
stage conceptualization of the process is introduced.
As defined here, the staff-organization planning and control problem concerns
693
694 Abernathy,Baloff,Hershey,and Wandel

the efficient matching of manpower resources with stochastic service demands


placed upon multiple-service (staffing) centers. Solution of the problem requires
decisions, over different time horizons, relating to: (a) the basic organization and
design of service centers and interrelations among numbers and types of staff at
each center; (b) the operation and control of the staffing process itself; (c) planning
personnel hiring, discharge, training, and reallocation actions; and (d) short-term
scheduling and assigning available staff to service centers.
In many service organizations, such as hospitals, it is both important and
difficult to allocate available staff to meet stochastic service demands efficiently,
because of these operating characteristics and constraints: (a) the services demanded
are crucially time-dependent, often of an emergency quality, and cannot be de-
ferred or 'backordered' without incurring very high 'costs'; (b) the services can-
not, by their very nature, be stored or 'inventoried' effectively; (c) uncertain
demand variations must therefore be absorbed through manpower allocations and
reallocations if service-quality specifications are to be met; and (d) the consequences
of inefficient staff organization, planning, and control can be very significant in
dollar terms, because manpower represents a large component of total operating
costs. (See Note 1.)
This staffing problem is somewhat analogous to the familiar workforce and pro-
duction-planning problem found in goods-producing organizations. However, the
differences between the two effectively preclude applying the extensively developed
'aggregate planning models' (see, for example, references 4, 8, 10, 17, 21) in solving
the service-sector problem. These models rely importantly upon the inventory
and backorder capabilities normally present in goods production. Furthermore,
they typically exclude other considerations that must be included in an effective
staff planning and scheduling model for many service-sector applications. For
example, they normally do not consider explicitly matching specific labor skills
with job requirements at individual work centers. Yet this careful matching is
important in organizations like hospitals, where the 'quality' of the service is crucial,
personnel are highly specialized, and no post hoc quality adjustment is possible.
Furthermore, any manpower model should also consider: (a) the various staff
allocation actions and alternatives available to an administrator as decision vari-
ables-not merely overall manpower-capacity adjustments; (b) the time lags be-
tween staff-allocation decisions and changes in manpower capacity at each work
center; (c) the training requirements and losses in manpower efficiency that can
arise from reallocations of staff; and (d) such factors as uncontrollable attrition
that affect the staff planning and scheduling process. In essence, the model must
recognize the interdependence of past, present, and future staffing actions among
work centers in determining total labor costs, rather than simply relating staffing
cost in each planning period to the period's manpower capacity level and the change
from the previous period's level. Furthermore, a tradeoff exists between decisions
about early and inexpensive actions that are based on unreliable information versus
late and expensive decisions made when a more reliable forecast is available. The
model must enable joint consideration of these multiple interdependent decision
options with different forecasts, degrees of aggregation, planning periods, and
review intervals [reference 2, pp. 57-69].
The model formulated in this paper is directed specifically at the nurse-staffing
problem in acute hospitals as an important illustration of the more general service-
ManpowerPlanningand Scheduling 695

sector problem. It has been argued that significant overstaffing often occurs in
these organizations as a direct result of both insufficient staff-allocation planning
and the inflexible organization and operation of nurse staffing units.1 138 Indeed,
the planned staff capacity for each ward in some hospitals is held constant for the
entire year at a level sufficient to meet the peak or near-peak demand expected
during the year, with the resulting underutilization of personnel common to this
type of overcapacity solution.111
To date, nurse-staffing research has concentrated primarily on two short-term
components of the nurse-staffing process-scheduling individuals to shifts and
assigning available staff in each shift to wards [see references 5-7, 9, 12, 14, 15, 16,
19, 23, 25-27]. Many of these previous studies have proposed improvements in
the two components, but they have not considered the larger system of which the
components are a part. They all neglect to consider planning to meet seasonal
fluctuations in demand.t"' Yet, it has been shown that such seasonal fluctuations
can be predicted, that differences as high as 10-15 per cent among monthly means
of patient census can occur, and that these fluctuations are synchronous among
groups of hospitals. 11,28'
An effective solution to the 'total' nurse staffing problem requires that short-
term assignment and scheduling decisions have the support of the longer-term,
multiperiod staff-allocation decisions to assure that the capacity to be scheduled in
each period is in fact supplied in a manner that is economic in terms of present
staffing capabilities and future demands. Conversely, the short-term scheduling
procedures affect the level of staffing that should be provided, since these proce-
dures determine the efficiency of overall allocations. Furthermore, the scheduling
and planning processes are not only highly interconnected in this way, but are also
dependent on the policies used in operating and controlling the staffing process
itself.
Because of these considerations, we propose a three-level formulation of the
problem-policy decision, staff allocation, and short-term scheduling. Following a
general formulation of the problem, we show that these three levels are effective
decomposition stages in an iterative solution procedure. By introducing service
criteria in lieu of random loss functions, we obtain a solution procedure that can
be evaluated numerically for large problems for the allocation and scheduling stages.
The procedure is then used in a hypothetical example to illustrate the type of in-
formation the model can generate. Finally, there is a discussion of how the model
can be used to evaluate alternative organizational structures and policies and to
assist in controlling nurse performance.

FORMULATION OF THE STAFF PLANNING AND SCHEDULING MODEL

THE DEMAND FOR full-time-equivalent (FTE) employees at staff location j is assumed


to be generated by a stochastic process that is stationary during each planning
period i. Hence, the demand in any day for period i and location ] can be expressed
as a nonnegative random variable 2iJ, which is an element of the random matrix
Z= (2ij). The joint cumulative probability distribution of demand F(Z) depends
upon estimated short-term fluctuations within planning periods, accuracy of pa-
696 Abernathy,Baloff,Hershey,and Wandel

tient-load forecasts, and the accuracy of the translations of forecasts to manpower


demands. Generally, the demand 2ij is both autocorrelated and dependent on the
demands of other locations.
There are two general classes of staffing decisions to be made-the choice of a
staffing-organization structure and operating policy and, given this choice, the allo-
cation of staff throughout the organization during each period. Decisions of the
first type are called policy decisions and are formalized as the choice of a policy
vector p from among the set of P = {Ip} possible policies that may be chosen, where
1 indicates the number of a particular policy p. Examples of such policies in a hos-
pital are progressive patient care, fixed staffing by wards, and a variable-staffing
policy that incorporates a float-staff pool. The second type of decision concerns
period-by-period allocations of personnel, described as X = (xkj), indicating the
number of persons moved in the beginning of period i from source k to sink j. A
source or sink can be either within the organization, as with a transfer from one
unit to another, or without, as with new hires and discharges.
Let the state of the system during period i be described by the vector si, and
all states during the planning horizon by the matrix S = (si). If U(S) represents
the utility function, and if S is a function of demand Z, policy p, allocations X, and
the given starting state so, then the problem of maximizing expected utility can be
formalized as
maximizesxEf U[S (Z, p, Xlso)] }. (1)
This provides a concise conceptual statement of the problem, but the complex-
ity and size of actual staffing problems confound direct solution of this formulation.
The explication of a utility function presents the first obstacle. Even with a valid
utility function, methods for optimization are not suitably developed for realis-
tically large problems. However, the problem can be decomposed into three stages,
each representing a level of complexity more consistent with existing solution tech-
niques and methods for quantifying utilities. The three stages are (1) choice of a
staffing policy p from the set available in P, (2) allocations of staff X= (xkj) from
period to period to provide the personnel resources required in each period most
efficiently, and (3) short-term scheduling and allocation of employees in order to
match the effective staff level (number of available FTE employees) D = (dij)
with demand outcome Z. The relation of these elements to the overall optimiza-
tion problem is developed in the following paragraphs.
The number of FTE employees made available from all sources through staffing
allocations X that are made before and during period i is represented in dij. The
matrix D is expressed as a function M (Xlp, so) that models time lags between ac-
tions and capacity changes, efficiency losses during training, etc.
On any particular day, the demand for service in a particular location may be
greater or less than the available staff level dij. For a given policy p, let the pe-
nalties and benefits resulting from such overstaffing and understaffing situations be
expressed as a cost function of demand Z and effective staff level D: G (Z, DIX, p ).
The staffing costs associated with the allocations X necessary to provide the
effective staff level D (such as salary, hiring cost, and retraining costs) are evalu-
ated by the function Q(XID, p, so). Note that this function is sensitive to patterns
of change, unlike simpler and more conventional aggregate staffing-cost relations
Q(Dlp, so), which reflect staffing costs only as a function of effective staff level,
independent of past and future staffing actions.
ManpowerPlanningand Scheduling 697

Given the cost functions G and Q, and the assumptions that utility is only de-
pendent on the chosen policy and the expected cost E(Q+G), then the staffing
problem (1) can be expressed as
maximize U[p, min {IQ(XID, p, so)+Ez[G (Z, DIX, p)]j]. (2)
In problem (2), the Q ( ) term represents the cost of providing staff over the
planning horizon-the principal objective function of the staff planning problem-
and the G ( ) term is a random loss function that includes probabilistic outcomes
and represents the penalties and benefits that are central to the short-term sched-
uling problem.

ITERATIVE SOLUTION PROCEDURE USING RANDOM LOSS FUNCTIONS

ONLY IF Q (XID, p, So) is replaced with an aggregate staffing cost Q (Dfp, so) that
is independent of staffing patterns can the cost part of problem (2), under certain
assumptions, be solved without iteration procedures.20J However, in the general
case, staffing problem (2) can be solved iteratively, as shown in Fig. 1. With an
initial estimate Q'(Dlp, so) of the aggregate cost relation Q (X*fD, p, so)-which is
unknown, since the optimal staff allocations X* have not yet been established-
the staff-level estimation problem can be solved. This gives a first estimate of
the required effective staff level D, for each policy p under consideration:
minimizeD fQ' (DIp, so)+Ez[G(Z, DIp)]j}. (3)
Given the solution D' to problem (3), the staff-allocation planning problem
can then be solved using the known disaggregate cost relation
minimize Q (XID, p, so). (4)
The solution to problem (4) yields staffing allocations X' (the first estimate of
optimal staff allocations X*) and more information about the unknown aggregate
cost function Q(X*ID, p, so). A revised estimate of the cost function can now
be made; for example,
Q"(D Ip, so)= Q (X'JD, p, so). (5)
The new estimate provides a basis for recalculating the effective staff level D/"
through the solution of problem (3). This approach can be used to develop itera-
tion algorithms. To obtain a finite and feasible solution, it is necessary that
E(Q+G) be convex in D and X jointly, which is true for most realistic functions
Q, G, and F. WILLIAMS1241 has suggested using Dantzig's general algorithm for
convex programming for solving successive approximations of problems (3) and
(4). He also derives an upper bound for the difference in expected cost between
the optimum and an iteration point, for linear Q and G functions, that can be used
to determine a satisfactory stopping point.
Finally, when the planning problem has been solved in an optimal or satisfac-
tory way, the solution (Xl*, Di*), together with nonquantified effects of the par-
ticular policy pi, can be evaluated. This can be done for all policies P=J{pd
under consideration, and the maximization problem becomes:
maximize, Ulp, Q(X*ID*, p, so)+Ez [G(Z, D*IX*, p)]}. (6)
698 Abernathy,Baloff,Hershey,and Wandel

Alternatively, a satisfactory rather than optimal solution may be sought, or a


search scheme can be followed to obtain the optimal policy p*.

FIG. 1. The iterative solution procedure using a random loss function.

NONITERATIVE SOLUTION PROCEDURE USING SERVICE-LEVEL CONSTRAINTS

THE DECOMPOSITION approach suggested above simplifies the problem, so that it


can be solved directly through iterations of successive approximations. However,
the method is cumbersome, and it also requires knowledge of the cost G over
ManpowerPlanningand Scheduling 699

the entire range of possible demand outcomes. In most cases, the manager is
better able to establish service policies than to quantify a random loss function;
that is, the G(.) function can be deleted from problem (3) and replaced by chance
constraints. The staff-level estimation problem then becomes
minimizeD Q(X*JD,p, so) subject to pr(D<Z)<A. (7)
The chance constraints require an effective staffing level D such that the probabil-
ity of demand overload is less than some probability A = (aij) established as risk
levels for a given service policy.
Under certain assumptions [see Note 2], the cost part of problem (2) has the
same solution as problem (7), if and only if
aij= (-7rtj+gti)/(gt+g77), (8)

where 7r*=aQ(X*ID*, p, so)/Odij, gqt=expected marginal cost for overstaffing,


7 = expected marginal cost for unmet demand. However, 7rij is the dual cost
and g7
to be obtained from the unknown optimal solution. Hence, in the general case,
the choice of risk levels A is dependent on the establishment of staff allocations X.
Fortunately, if the unknown function Q (X*ID, p, so) is monotonically non-
decreasing in each dij, then the staff level D that satisfies the equation
pr(bZ)=A (9)
yields a lower limit [see Note 3] on the optimal solution, since any dij>_di is an
optimal solution to (7) in the interval where 7rtj(dij) = 0.
Hence, using risk levels A, established independently of allocations X, one can
calculate the lower limit D for a given policy. Figure 2 illustrates how an optimal
solution to the scheduling and planning problem can be obtained without the cum-
bersome iterations suggested in Fig. 1.
A more complete statement of problem (4) as a mathematical programming
problem is
minimizexQ(X) subject to B(XIp, so)=0, M(XIp, so)>b, and X>0. (10)
Here B (. ) represents the balance equations requiring inflow of staff to be equal to
uses and outflow for each location and period, and M () describes effective staff
level as a function of allocations.
MV'ore detailed formulations of the staff-level estimation element (9), and the
staff-planning element (10), are given in the following sections; the joint solution
of both elements is then illustrated by example.

THE SHORT-TERM SCHEDULING MODEL

IN THIS SECTION, a M'Aonte Carlo model for determining the minimal staff level
D that meets specified risk levels A is presented, i.e., equation (9). Assume that
two short-term scheduling policies are under consideration, fixed staffing and con-
trolled-variablestaffing. With a fixed-staffing policy, the daily patient-care demand
in each ward is met by nurses who are permanently assigned to specific wards. In
contrast, controlled-variable staffing procedures provide for a separate manpower
700 Abernathy,Baloff,Hershey,and Wandel

pool of cross-trained float nurses scheduled daily between wards to meet ward fluc-
tuations. The number of nurses permanently assigned to each ward is lower than
under fixed staffing.

FIG. 2, The noniterativesolutionprocedureusing service-levelconstraints.

In the fixed-staffing case, a maximum risk level caij(0<?ai 1 ) is defined to be


the maximum fraction of all days in month i when the stochastic load variable zij
exceeds the effective number of permanently assigned nurses dij. This is equivalent
to a minimum required service level of 1 - aij. Assume the risk level is the same
for each month i and each ward j (vij = ha). Let dtaj be the number of nurses per-
ManpowerPlanningand Scheduling 701

manently assigned to ward j in month i to provide the designated maximum risk


level a exactly. Assume the existence of a unique inverse to the marginal distribu-
tion function Fij ( r j == , ??, z, o, *, o). Then FTj'
(1-a).
With controlled-variable staffing, a particular permanent staffing level is asso-
ciated with a ward; fij (O f ij' 1 ) is defined to be the maximum fraction of days
when the nurses permanently assigned to a given ward are less than the number
required to meet the load zij. Assume the permanent fraction is the same for each
month i and each ward I (fij=3). Let d' be the number of nurses permanently
assigned to ward j in month i that exactly provides the designated risk level d.
Then ddij= Fy (1-:3).
Let the demand for the float pool from ward j be defined as the random vari-
able Zijg,where g indicates the float pool. Then Zjj has the distribution function
(0, for zij <O,
Fijg (zijg)= Fij (d'j ) = 1-3, for zijg= 0,
Fij (d'1j+zijg), for zijg> 0.
When n wards are under consideration, let the total float-pool demand be defined
as ig. Then Zi,= Ej=- Zijjgzj_0. The distribution function Fig(zig) can be
formulated as a multiple convolution of Fijg (Zijg)Vj, but cannot readily be evalu-
ated by analytical methods, such as moment-generating functions, owing to multi-
ple discontinuities at zig,= 0. Furthermore, analytical methods cannot handle un-
equally distributed or mutually dependent demands zijg.
Therefore, a Monte Carlo model is used that establishes three minimal-float
staff levels di, one for each of three alternative criteria, to provide equivalent
patient care with a fixed-staffing case using the same risk level a. The model also
considers correlation coefficients pjj between pairs of ward demands, efficiency
(FTE/nurse) ej for float, and minimum floatable fraction mi of the day. A detailed
discussion of the model is found in a separate paper. E" With the criterion used in
the example in the present paper, the float-staff level diais chosen so that the ex-
pected overloads in the variable-staffing case and the fixed-staffing case are equal;
that is, choose dg such that
- dig|zig>dig) = Eji-1 E (zij-dt'Jijzi>
E(Zig dtj)-
With this model, the minimum staff levels Itaj for the fixed-staffing policy, and
dij and dcifor the controlled-variable staffing policy, are determined independently
for each month i to meet the common risk criterion a. The relative benefits of the
two policies cannot, however, be determined until the staff-planning problem has
been solved.

THE STAFF PLANNING MODEL

IN THE staff-planning component, problem (10), a linear programming formula-


tion is used to determine optimum staff allocations X to satisfy minimal staff levels
D for staffing policy p. Service-policy parameters (e.g., risk level a and permanent
702 Abernathy,Baloff,Hershey,and Wandel

fraction d) of the short-term scheduling model are assumed as contained in the


specification of policy p and are reflected in deterministic equivalent minimal staff
levels D determined by the scheduling model. The formulation, including speci-
fications of functions M, B, and Q, is as follows:

Variables
k
xj -Number of persons to be moved from source k to sink j in the beginning
of period iJl; X-= (xj)>O; xkj for i < 1 are parameters contained in
the starting state so.

Coefficients
k
cj -Cost of salary, moving, and training a person, i.e., unit cost of activ-
ity X4; C= (cai).
f -j Fraction left during period i of the entering number of persons after
attrition during t periods of training in location j and coming from
source k; O ff, <: 1; F = (f).
etk -Efficiency (FTE/person) during period i after t periods of training in
location j and coming from source k; E== (ej).

Right-hand side

di' -Required minimum effective staff level in period i and location j as


determined by the short-term scheduling model; b = (dei).

Problem-structureparameters

I -Set of all periods i, i> 1, under consideration.


MjCK -Set of sources k from which persons can be moved to location jEJ in
the beginning of period i and then be trained there h periods; when
i< 1, Kj are used to define the starting state so.
JkCJ -Set of sinks j to which persons in location keK can be moved in the
beginning of period i.
So -Starting state, containing all X it)j such that (1- t) < 1 and Khjt)jOc
for some t, 1 < t < h.
With these definitions and with suppression of the policy index p, the linear
programming formulation of the staff-planning problem is now defined as
minimize Q(C, X), such that B(F, X, K, J, so)=O, M(F, E, X, K, so)
>13 XO.
With all indices explicit, this becomes
minimizersQ= ZiEeI ZEjeJ ZreK Ckj IXij
ManpowerPlanningand Scheduling 703

such that
ZhzefhhlK(ij-h)p!4') Z k(i-h) i it'X(Vh)j-ieJ'Xjq=O, Yinj[BiJ
EhfthK<-t~~b} Et=0 ZkeK i_ 0 fzi*,ejj x (i_ >' dij, Vi, j[Mij]

Xtj> 0, Vi, i, kx
Note that Bt3's are the personnel-balance constraints (the bookkeeping equa-
tions) that (a) adjust for attrition (1 -fJ t), and (b) update the permanently as-
signed staff by the number of persons who have just completed training that started
h periods ago and those who are kept at the same location from the previous period.
The starting state sodefines Xi-t)j for (i-t) < 1, 1??th. All available staff is then
allocatedto the sinks Jik, k=i.
The supply-demand matching constraints Mti include as supply all persons
working at the location as trainees up to and including the (h- 1)st period after
they started their training. The supply also includes the permanently assigned
staff for the period. The attrition in the training pools during the t periods of al-
ready completed training is accounted for by ft. The efficiency coefficient elkad-
justs the number of persons to FTE's.
The planning model can be operated either in an acyclic (transient) or in a
cyclic (steady-state) mode. The acyclic mode is used when the scheduling and
planning stages are used together as a planning tool. Then the right-hand side of
the constraints is modified by the starting conditions so, which makes the model
sensitive to past decisions x4i where i <1. Owing to the time lags in the model,
the cost coefficients in the last periods should be adjusted to compensate for ex-
pected post-horizon effects, e.g., with dual costs from a cyclic model using expected
seasonal variations. On the other hand, when the model is used for policy analysis,
a cyclic mode is obtained by setting so= sr, where r is the number of periods in the
season. Then the solution becomes independent of both the distance to the plan-
ning horizon and the arbitrarily chosen starting states, which should be dependent
on the choice of policy. In this way, an infinite-horizon solution can be obtained
for both the cyclic and the acyclic mode.

AN EXAMPLE

THE FOLLOWING EXAMPLE is presented to illustrate the types of information that


can be generated for policy and organizational analysis and to guide staff planning
and control activities. The data used are hypothetical, and are not intended to sup-
port general conclusions about the relative merits of alternative planning and
scheduling policies.
The problem setting is a group of four wards in a hospital. The planning period
is one month, and seasonal variations in service demands have a periodicity of one
year (I= { 1, 2, ***, 12}). The float pool is designated location 5. Hiring and dis-
charge of experienced nurses, location 6, is permissible in all months, but inex-
perienced nurses, source 7, can only be hired upon graduation at the beginning of
July (i = 7). Nurses cannot be reallocated during a training assignment. Further-
more, the number of nurses in any location can take on noninteger values, since
704 Abernathy,Baloff,Hershey,and Wandel

TABLE I
FORECASTED MONTHLY MEAN DEMAND Aq
Wardj
Monthi 1 2 3 4

1 8.75 7.84 7.15 7.50


2 8.34 8.13 7.24 7.34
3 7.61 7.35 7.15 7.01
4 8.44 7.35 7.24 7.26
5 7.93 7.08 6.84 6.78
6 6.90 6.21 6.51 6.45
7 6.48 5.45 7.10 6.14
8 7.00 6.70 6.43 6.29
9 6.48 5.64 6.18 5.89
10 6.39 5.54 5.36 5.33
11 7.41 6.98 6.91 6.94
12 6.99 6.31 6.66 6.70

fractions can be filled by nurses working parttime, overtime, part of the day in one
place and part in another, or for fractional-period allocations (reallocated some-
time during the month instead of at the beginning of the month, as assumed in
the model).
Actual monthly average workloads, shown in Table I, are used as a forecast of
the mean demands pij. The standard deviation oij of daily workload is assumed
equal to 7 percent of mean demand pjui. Fij(zij) are assumed to be mutually inde-
pendent normal distributions.
Two short-term scheduling policies are under consideration-fixed staffing and
variable staffing. Using the output of the short-term scheduling model for each
policy, the planning model is solved in its cyclic mode. For purposes of comparison,
the planning model is also solved (for each of the two scheduling policies) assum-
ing that only annual, rather than monthly, forecasts are made. In this case, the
monthly mean demand is simply assumed to be 1?2 of the annual demand. The
total annual variance is determined by summing within month and between month
averages [reference 3, pp. 160-165].
Specific parameters of the model are given in Table II. The efficiency and attri-
tion of trainees are taken to be independent of the source wards, which allows the
introduction of the dummy location a, to which all experienced nurses are moved
immediately before they are allocated to any ward or float pool. This permits the
planning model to be decomposed over wards, and simultaneously brings about a

TABLE II
PLANNING AND SCHEDULING PARAMETERS

Planning
Dummy (j = a) K=. = J {1, 2, 3, 4, 5, 6} Vi
Wards (j = 1, 2, 3, 4) Kli = J {j, a} ViJI; K"j = {7}
Float (j = 5) K15 = {5} and J5 = {5, a} VieI
KS%= {a} Vie{i iel, i 7; = {7, a}
k
Scheduling Pij = OViE, VjeJ, VkEK; e? = 1 and mi = OViE1
oety= 0.05 and #ij = 0O30Viel, VjEJ
ManpowerPlanningand Scheduling 705

reduction in the number of variables in the planning model, in this example, from
425 to 209. The cost assumptions are given in the right column of Table IV.
Attrition is assumed to be 2 percent per month for all nurses, yielding
fk= (1-0.02)'. FTE units are chosen so that a permanently assigned nurse (i.e.,
k=j) has efficiency e?- 1; nurses in dummy locations have zero efficiency, e" =0;
and the efficiencies of trainees, including losses due to attrition while in training are
given in Table III.

POTENTIALUSES OF THE MODEL

THIS HYPOTHETICAL EXAMPLE illustrates the type of information that can be


generated by the model and the utility of these model results (a) for policy evalua-

TABLE III
EFFICIENCY COEFFICIENTS INCLUDING ATTRITION (f} en )
FOR NURSES IN TRAINING

Source Location Length of Efficiencyafter t months


training of training

k j h =0 =
t=1 =2

Experienced, in float a 5 3 0.5 0.6 0.8


Experienced, in ward a 1-4 1 0.8 - -
Inexperienced, in float 7 5 3 0.3 0.5 0.8
Inexperienced, in ward 7 1-4 3 0.6 0.7 0.8

tion and decision purposes, and (b) as an informational basis for actual staff plan-
ning and control [see Note 4]. Looking first at policy evaluation, we see that Table
IV presents information that an administrator might prepare to evaluate the eco-
nomic effects of one year of operation under the four alternatives posed in the
previous section: planning, using monthly forecasts, for each of the two short-term
scheduling policies, and planning based on a single annual forecast for the same two
short-term scheduling policies [see Note 5]. (M'iostof the cells in Table IV show
two numbers-the number of nurses involved and the associated cost; the assumed
cost of an 'activity' is given in the last column.)
As shown in row (9), the monthly-forecasting, variable-scheduling alternative
gives the lowest total cost for this hypothetical example. Indeed, the variable-
scheduling alternatives are superior to fixed scheduling under both monthly and
annual forecasting, despite the higher cost for float nurses. The data in rows (1)
and (2) show that the number and cost of permanently assigned nursing months in
all wards and the float pool is greatest in the annual-forecast, fixed-scheduling case,
lowest for the monthly-forecasting, variable-scheduling alternative, and that vari-
able scheduling gives fewer permanently assigned nurses in both forecasting options.
Planning based on monthly forecasting obviously allows anticipatory hiring and
firing to meet seasonal fluctuations in demand. As the data in rows (3), (6), and
(8) indicate, the two monthly forecasting alternatives result in a significantly
higher rate of hiring, transfer, training, and firing than with the annual forecasting
706 Abernathy,Baloff,Hershey,and Wandel

alternatives. In some cases, the real costs of such employment instability can
become very high, though subjective, and dominate the savings of more 'efficient'
staffing. If, for instance, this were true in our hypothetical example, then
choosing the alternative of annual-forecast, variable-scheduling might be appro-
priate, since it gives the best employment stability without much increase in total
staffing cost.
Table IV is merely suggestive of the types of analysis that might be requested

TABLE IV
EXAMPLE ECONOMIC SUMMARY OF FOUR POLICIES

Planningusing Planningusing
monthly forecasts annualforecast Cost of
- ~~~action
Fixed Variable Fixed Variable ($/nurse)
scheduling scheduling scheduling scheduling

(1) Number (cost)ofpermanentlyas- 358.3 333.2 383.5 344.2


signed nursing months in all $186,320 $173,230 $199,420 $178,970 $520
wards
(2) Number (cost) of permanently as- 9.2 15.1
signed nursing months in float - $6,590 - $10,860 $720
pool
(3) Number (cost) of experienced 11.3 10.7 5.9 5.2
nurses hired $2,260 $2,140 $1,180 $1,040 $200
(4) Number (cost) of inexperienced 2.4 2.2 4.0 1.1
nurses hired and trained three $2,520 $2,270 $4,200 $1,180 $1,050
months for wards
(5) Number (cost) of inexperienced 0.2 1.7
nurses hired and trained three _ $180 _ $1,840 $1,100
months for float pool
(6) Number (cost) of experienced 14.1 13.2 5.9 5.9
nurses trained one month for $7,600 $7,100 $3,180 $3,210 $540
ward assignment
(7) Number (cost) of experienced 0.4 0.3
nurses trained three months for - $670 _ $510 $1,820
float pool
(8) Number (cost) of nurses dis- 6.1 5.7 1.9 0.5
charged $1,220 $1,150 $370 $100 $200
(9) Total personnel costs for one year $199,920 $193,330 $208,350 $197,710

in practice. Innovative administrators are likely to interrogate the model in many


different ways to obtain information for policy evaluation. Here are some ex-
amples of policies that might be studied:
1. Patientscheduling. The effectsof controllingdemandforserviceeitherby usingselec-
tive patient-admissionproceduresor by schedulingactivities on patients based upon avail-
able servicecapacitycan be tested by varyingthe ward-loaddistributionF(Z). The policies
used for assigning patients to wards also affect this distribution. For example, a policy of
progressive patient care (where patients with similar care requirements are assigned to the
same ward) will decrease short-term fluctuations in load but will probably increase seasonal
fluctuations in comparison with multipurpose wards.
2. Specificationof servicelevels. Rather than accepting the service level (1 - a) as given,
the administrator might wish to consider this as a policy parameter that demands as care-
ManpowerPlanningand Scheduling 707

ful investigationas the allocationand forecastingpolicies themselves. Using the model,


he canperformsensitivityanalyseson a to study the tradeoffbetweencost andservice. Sim-
ilarly, he can examinethe cost sensitivity of other parametersin the short-termscheduling
model, such as the service level met by permanentlyassigned nurses (1 -f3), the level
of cross trainingof float nursese, and the numberof wardsn in each groupof wardsusing
the same float pool.
3. Employmentpolicies. Employmentpracticesand predeterminedstaffingallocations
will often restrictthe quantity and timing of hiring,firing,transfer,and trainingof person-
nel. The effectsof such policiescan be analyzedby placing restrictionson the set of staff
allocationsallowed. For example,the cost of a no-firepolicycaneasilybe calculated. Other

TABLE V
STAFF PLANNING INFORMATION FOR WARD 2 USING
MONTHLY FORECASTS AND VARIABLE STAFFING

Month Forecasted
wardload ugees staf a11ocations
Suggested Marginal
cost
oaosopportunity

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Standard Mini- Perma Hire Enter Reallo- Minimal Perma-
Staff nent inexper
Mean devia- malstaff level txperi- cate sta nent
(FTE) tion level assigned enced enced experi-level(/ assigned
Ii2 (FTE) (FT E ) (nurses)
nrs (rses)
nre)(uss (nurse) FT)
Oi2 A ai (nurses) nurse)
da~ ~~X2 X'2 Xi2 X2a O~2~ i2

1 7.84 0.548 8.12 8.12 6.41 2.14 1,013 423


2 8.13 0.569 8.42 8.42 8.38 0.05 541 -72
3 7.35 0.515 7.62 7.77 7.77 0.49 0 -95
4 7.35 0.515 7.62 7.62 7.62 1,150 434
5 7.08 0.495 7.33 7.33 7.33 0.13 516 -200
6 6.21 0.435 6.44 6.44 6.44 0.75 406 -200

7 5.45 0.381 5.65 6.70 6.31 0.64 0.01 0 -88


8 6.70 0.469 6.95 6.95 6.20 0.38 1.11 1,157 441
9 5.64 0.395 5.84 5.84 5.33 418 -200
10 5.54 0.388 5.74 5.83 5.83 0 -100
11 6.98 0.489 7.23 7.23 5.71 1.90 1,042 428
12 6.31 0.441 6.54 6.54 6.54 0.91 11 -95

employmentpolicies,such as the intensity of trainingprograms,efforts taken to decrease


attrition and absenteeism,and vacation scheduling,affect the values of attrition F and
efficiencyE.
4. Operation and control of the staff-planning process. In addition to the preceding types
of policies,the operationalpoliciesemployedin the staff-planningprocessitself can be ana-
lyzed. The two forecastingoptionspresentedin the exampleare illustrationsof operational
policiesfor staff planning. Other policiesincludethe frequencyof plan revisions,intervals
betweenchangesin staff assignments,and length of planninghorizon.
Turning to the informational uses for actual staff planning and control, we see
that the model provides a wealth of information that can be used for making specific
allocations and reallocations on a ward-by-ward and month-by-month basis. Table
V is an example of one set of data for a particular ward that can be developed for
this purpose. In this case, it is assumed that monthly forecasts and variable staffing
have been chosen. The service-demand forecast is given in columns (2) and (3),
708 Abernathy,Baloff,Hershey,and Wandel

and the minimal staff level from the solution of the scheduling problem is given in
column (4). Columns (5)-(9) present suggested staff allocations made by the
planning model. Note that overcapacity occurs in months 3, 7, and 10 (d*2> di2).
These allocations are the result of the ability of the model to recognize the inter-
dependence of past, present, and future staffing actions among work centers.
The model also provides some information that can be used for coordinating
staff allocation and scheduling with other parts of the nursing system. For exam-
ple, column (10) in Table V shows the marginal opportunity cost for increased min-
imal staff levels, which can be used to guide admissions from waiting lists, or other
means of affecting A, a, or p, such as controlled variable allocation of patients to
wards or better forecasting. The marginal opportunity cost of one less permanent
assigned nurse is presented in column (11). This can be used for scheduling va-
cation and training as well as for evaluating actions to control attrition and absen-
teeism. The evaluation of the impact of variations of exogenous variables out-
side the direct control of the coordinator is, of course, not restricted to marginal
analysis; all types of sensitivity analysis can be used for enquiry.
It is important to recognize that information such as that given in Table V
should be interpreted as providing guidelines for staff allocation and reallocation
decisions. Actual staffing decisions involve not only average numbers of nurses, as
the model suggests, but also the selection of individual nurses and the exact timing
of reallocations within the month. The fractions suggested by the model do not
necessarily require the use of part-time, overtime, or variable productivity. These
fractions can also be achieved by having nurses change ward assignments during
the month.
When definite assignment plans have been made, individual nurses can be
scheduled to shifts and days for the coming weeks [see Note 6]. However, reliable
forecasts of the workload for a particular shift can only be made a shift or two in
advance. Hence, minor adjustments to the schedule, allocation of float, and other
means of absorbing short-term fluctuations have to be used on a day-to-day basis.
The model can also be used to study a range of possible loads, and thereby sug-
gest standards for such indices as number of nurses per patient day, costs per
patient day, and number of nurses per FTE to improve the quality and flexibility
of budgeting procedures. Furthermore, the model can be used to control nurse
performance and improve staffing effectiveness. For example, suggested decisions
and actual allocations can be compared and differences analyzed.

CONCLUSION

THE THREE-STAGE model developed in this paper was specifically formulated for
application in the nurse-staffing process. The characteristics of the model are suf-
ficiently general, however, to suggest applications in other organizations. We
expect that the benefits from applying the model are likely to be highest in organiza-
tions facing large and uncertain variations in demand for high-quality or specialized
services that must be provided in a timely manner, particularly where service capac-
ity is not easily expanded and the services themselves cannot be inventoried.
Police forces, ambulance services, fire departments, banks, transportation systems,
ManpowerPlanningand Scheduling 709

and power systems are all examples of service organizations that seem to meet these
criteria. In the last two systems, transportation and power, it is physical capacity,
rather than staff level, that is important, but the approach used here can be adapted
for these applications as well. Furthermore, with some natural extensions, it can
be used for workforce planning and control for goods-producing organizations, and
for nationwide manpower planning.
The formulation of the model has involved certain methodological considerations
that are worthy of mention as being of potential interest in other research efforts.
The two solution procedures used in the model formulation may be of particular
interest for the solution of a wide range of previously unsolvable stochastic and
chance-constrained programming problems. It will be recalled that the iterative
solution procedure was used where the scheduling problem assumes a random-loss
cost function, and a noniterative solution procedure was also developed in which
this cost function is replaced with chance constraints. In the latter case, the use
of a Mlonte Carlo model to obtain these constraints permits the inclusion of sta-
tistically dependent demands among work centers, joint chance constraints, and
realistic operational assumptions and service criteria.
Also of interest is the explicit inclusion of time lags, transient efficiencies and
attritions in the model, which enables the use of linear cost relations. This pro-
cedure differs from that used in other aggregate planning models that assume con-
vex cost functions to describe costs of changes in aggregate manpower levels from
the previous time period. In many cases, such convex cost functions require more
complex optimizing techniques and are impossible to estimate owing to the dynamic
nature of actual staffing patterns.

ACKNOWLEDGMENTS

THE AUTHORS ARE grateful to the Booz-Allen and Hamilton Foundation for their
financial support of this research through the Health Systems Research Project at
Stanford University. The authors also wish to thank ROBERT C. CARLSON,JOEL
S. DEMSKI, and THOMAS A. GONDA, M.D., for helpful suggestions.

NOTES

1. An example drawn from hospitals indicates the potential benefits of more efficient
staff scheduling. A 10 percent increase in manpower productivity would yield an annual
saving of $1.6 billion annually. [29] Improved nurse staffing is one important way in which
greater productivity can be realized, since approximately one-half of annual hospital pay-
roll costs are for nursing services.
2. One set of assumptions, developed by SYMONDS,[2oI is as follows: Q(.) is linear with
linear constraints M(.) = D; G(-) is a separable but not necessarily linear loss function;
the optimal solutions exist; and continuous derivatives of the Langrangian function Q(X) +
Ez[G(2, D)] + 7r[D - M(X)], where iX = (7rij), exist.
3. In the case of a joint risk constraint, dij is a supremum to the lower limit.
4. A discussion of other experiments for policy evaluation with this hypothetical example
can be found in reference 22.
710 Abernathy,Baloiff,Hershey,and Wandel

5. The readeris cautionednot to make any interpretationof the relativeeconomicad-


vantagesof one alternativeover the other. As mentionedearlier,the parametervaluesin the
model are not intendedto be representativeof an actualoperatingsituation. Indeed,values
that do not give extremedifferencesin economicresults are used to avoid any suggestion
that a specificplanning-schedulingalternativeis being recommendedfor generaladoption.
6. Methods that addressthis assignmentproblemare suggestedby references12, 15,
23, 25, and 26.

REFERENCES

1. W. J. ABERNATHY, N. BALOFF, AND J. C. HERSHEY, "The Nurse Staffing Problem:


Issues and Prospects,"Sloan Management Rev.12, 87-99 (1971).
2. J. ANDERSSON, S. LJUNGFELDT,AND S. WANDEL, Produktionsstyrning, Studentlitteratur,
Lund, Sweden, 1968.
3. E. C. BRYANT, StatisticalAnalysis, McGraw-Hill,New York, New York, 1960.
4. E. S. BUFFA AND W. H. TAUBERT, Production-Inventory Systems:Planningand Control,
Irwin, Homewood,Ill., 1972.
5. R. J. CONNOR, "A Hospital Inpatient ClassificationSystem," UnpublishedPh.D. dis-
sertation,The Johns HopkinsUniversity, 1960.
6. C. D. FLAGLE, "The Problemof Organizationfor HospitalInpatient Care,"in Manage-
mentSciences:Modelsand Techniques, Proceedingsof the Sixth InternationalMeeting
of the Institute of ManagementSciences,Paris, 1959,Vol. 2, pp. 275-287, Pergamon
Press, Paris, France,1960.
7. -- ANDJ. P. YOUNG, "Applicationof OperationsResearchand IndustrialEngin-
eeringto Problemsof Health Services,Hospitalsand Public Health,"J. Indust.Eng.
17, 609-616 (1966).
8. F. HANSSMANN AND S. W. HESS, "A Linear-Programming Approachto Productionand
EmploymentScheduling,"ManagementTech.1, 46-52 (1960).
9. J. C. HERSHEY, W. J. ABERNATHY AND N. BALAFF, "Comparisonof Nurse Allocation
Policies-A Monte CarloModel," ResearchPaper No. 98, GraduateSchool of Busi-
ness, StanfordUniversity,August, 1972.
10. C. C. HOLT, F. MODIGLIANI, AND H. A. SIMON,"A LinearDecisionRule for Production
and EmploymentScheduling,"Management Sci. 2, 1-30 (1955).
11. W. J. HORVATH, "OperationsResearchin Medical and Hospital Practice," pp. 127-
157, OperationsResearchfor PublicSystems,P. M. Morse (ed.), The MIT Press,Cam-
bridge,Mass., 1967.
12. R. C. JELINEK,T. K. ZINN, G. L. DELON, AND R. A. ALEMAN, "A Nurse Scheduling
and AllocationSystem in Operation,"TechnicalPapers, 23rd Annual AmericanIn-
stitute of IndustrialEngineersConference,Anaheim,Calif., 1972,pp. 273-281.
13. E. LEVINE, "NurseStaffingin Hospitals,"AmericanJ. of Nursing61, 6568 (1961).
14. J. S. LIEBMAN, "The Developmentand Applicationof a MathematicalProgramming
Modelof PersonnelAllocationin an ExtendedCareNursingUnit,"unpublishedPh.D.
Dissertation,The Johns HopkinsUniversity, 1971.
15. C. MAIER-ROTHEAND H. B. WOLFE,"Cycle Schedulingand Allocation of Nursing
Staff,"ArthurD. Little, Inc., 1972.
16. F. L. OFFENSEND, "Design, Control,and Evaluation of Inpatient Nursing Systems,"
unpublishedPh.D. Dissertation,StanfordUniversity, 1971.
17. E. A. SILVER, "Medium Range Aggregate ProductionPlanning: State of the Art,"
Productionand InventoryManagement 13, 15-41 (1972).
18. H. M. SOMERSANDA. R. SOMERS,Medicare and the Hospitals, The Brookings Institu-
tion, Washington, D. C., 1967.
ManpowerPlanningand Scheduling 711

19. D. H. STIMSON AND R. H. STIMSON, "Operations Research and the Nurse Staffing Prob-
lem," Hospital Administration 17, 61-69 (1972).
20. G. H. SYMONDS, "Chance-Constrained Equivalents of Some Stochastic Programming
Problems," Opns. Res. 16, 1152-1160 (1968).
21. W. H. TAUBERT,"A Search Decision Rule for the Aggregate Scheduling Problem,"
Management Sci. 14, B343-359 (1968).
22. S. WANDEL,"A Combined Linear Programming and Simulation Model for Aggregate
Planning and Scheduling of Nurses in Hospitals," Paper presented at the 41st Na-
tional Meeting of the OPERATIONS RESEARCHSOCIETYOF AMERICA,New Orleans,
Louisiana, April, 1972.
23. D. M. WARNER, "A Two-Phase Model for Scheduling Nursing Personnel in a Hospital,"
unpublished Ph.D. Dissertation, Tulane University, 1971.
24. A. C. WILLIAMS,"Approximation Formulas for Stochastic Linear Programming."
SIAM J. Apple. Math. 14, 668-677(1966).
25. H. WOLFEANDJ. P. YOUNG,"Staffing the Nursing Unit, Part I: Controlled Variable
Staffing," Nursing Res. 14, 236-243 (1965).
26. AND , "Staffing the Nursing Unit, Part II: The Multiple Assignment
Technique," Nursing Res. 14, 299-303 (1965).
27. J. P. YOUNG,"A Conceptual Framework for Hospital Administrative Decision Systems,"
Health Service Res. 3, 79-95 (1968).
28. "Hospital Indicators," Hospitals, Journal of the AmericanHospital Association45,
35-37 (April 16, 1971).
29. Hospitals(Part2), Journalof theAmericanHospitalAssociation45 (August 1, 1971).

View publication stats

S-ar putea să vă placă și