Professional Documents
Culture Documents
Naimark Et Al 2008 The Half Cycle Correction Explained Two Alternative Pedagogical Approaches 1
Naimark Et Al 2008 The Half Cycle Correction Explained Two Alternative Pedagogical Approaches 1
Students of Markov decision models are often taught to add cycle to reduce the latter discrepancy to be intuitive. How-
a half-cycle’s worth of incremental utility to the cumulative ever, students often find the approximate equivalence of
total for each health state. The reason for this ‘‘half-cycle shifting the ideal state membership curve and adding a
correction’’ is often illustrated by a graph of the proportion half-cycle’s worth of incremental utility to the total for the
of the hypothetical Markov cohort remaining in a given state at the beginning of a discrete Markov process to be a
state. The ideal graph is shown as a smooth, declining, difficult cognitive leap. This article describes 2 pedagogical
curve that represents the transition of patients randomly devices, algebraic and intuitive/visual approaches, that
throughout each cycle. On the same graph, the effect of the may assist the instructor of Markov theory to convey the lat-
accounting of state membership at the end of each cycle in ter concept. Elements of adult learning theory are dis-
discrete, computer-based approximations of the ideal Mar- cussed, which may help the instructor to choose which
kov process is shown. Students are able to clearly see that approach to employ. Implementation of the half-cycle cor-
the cumulative incremental utility in the discrete case rection in commonly used decision-analytic software is also
underestimates the desired quantity. Likewise, they find the discussed. Key words: Markov chains; models; educational;
concept of shifting the ideal curve to the right by one-half statistical. (Med Decis Making 2008;28:706–712)
have found that a perennial problem for our stu- The usual pedagogical approach7, 8 is to present
dents and ourselves is the explanation of the half- students with a figure displaying a stepwise, com-
cycle correction employed in Markov models. The puted, state membership graph representing the out-
half-cycle correction is very often used to compen- put from a discrete Markov process in relation to an
sate for the fact that in discrete, computer-based ideal, continuous membership function. The latter is
Markov models, transitions between states are represented as a smoothly declining, monotonic curve
usually modeled to occur at the beginning or end of (Figure 1). The effect of accounting for state member-
a cycle, whereas, in reality, transitions occur in the ship at the beginning or end of each cycle is shown
middle of each cycle on average. with the consequent over- and underestimation of the
true state membership, respectively. The effect of
shifting the computed, stepwise graph one half-cycle
to the right is then shown (Figure 1C) with the desired
Received 18 September 2007 from The Division of Nephrology, Sunny-
effect of reducing the deviation of the computed
brook & Women’s College Health Sciences Centre, Toronto, Canada cumulative state membership from the true value. By
(DMJN); The Division of General Internal Medicine, University Health visual inspection, students can easily appreciate that
Network, Toronto, Canada (MK); The Departments of Medicine and shifting the graph is equivalent to counting state mem-
Health Policy Management and Evaluation, Faculty of Medicine, Univer- bership in the middle of each cycle.
sity of Toronto, Toronto, Canada (DMJN, MK); The Department of Pro- When it comes time for students to build Markov
fessional Practice, Guelph General Hospital, Guelph, Ontario (MB).
models using commonly available decision analytic
Revision accepted for publication 6 January 2008.
software, they find that state membership must be
Address correspondence to David Naimark, MD, Rm A139, 2075 counted either at the beginning or the end of each
Bayview Ave, Toronto, ON, Canada M4N 3M5. cycle rather than in the middle. Students are taught
DOI: 10.1177/0272989X08315241 that, if they use a software application where state
METHODOLOGY 707
NAIMARK ET AL.
The area enclosed by this rectangle will underesti- WHEN THE MARKOV PROCESS
mate the true membership in the state by an amount TERMINATES EARLY
that is equal to the area in the region bounded by the
true state membership curve and the top of the rectan- The algebraic approach may be useful in demon-
gle. Although the region above the rectangle is not strating to students why the half-cycle correction
strictly triangular (due to the curve of the true member- must be adjusted if the Markov process is termi-
ship graph), one can approximate this area with a tri- nated early—that is, when a substantial fraction of
angle provided that the cycle-length is not too long in the cohort remains in the state. If the process is ter-
relation to the termination time of the Markov process. minated after 8 cycles, then the half-cycle correction
Close inspection of these triangular areas will reveal would be too large by an amount approximately
that they sum to an amount equivalent to a half-cycle. equal to the area of the sum of the ninth to the last
However, this fact can be formally proven to students. triangles. This latter area is
The base of each triangle is equal to the cycle-
length ‘‘B,’’ and the height of the ith triangle is equal A9 → n ≈ 1=2B½ðH8 − H9 Þ + ðH9 − H10 Þ
to the difference in height between the (i 1)th + ðH10 − H11 Þ + . . . : + ðHn − 1 − Hn Þ:
rectangle and the ith rectangle (Figure 2):
And canceling terms yields
Ai ≈ 1=2BðHi − 1 − Hi Þ:
A9 → n ≈ 1=2B½H8 − Hn :
Thus, the total area of the triangles is
Now, again, since Hn ∼ 0, one finds that
A = SAi ≈ 1=2Bð1 − H1 Þ + 1=2BðH1 − H2 Þ
+ 1=2BðH2 − H3 Þ + . . . : + 1=2BðHn − 1 − Hn Þ: A9 → n ≈ ½1=2B · ½H8 :
Figure 3 The effect of early termination of a Markov process Determining whether to employ an algebraic or an
when state membership is counted at the end of each cycle is that intuitive/visual explanation of the half-cycle correc-
an amount of incremental utility equivalent to the area of a tion concept may depend on the learning preference
rectangle—with a base equal to half of a cycle and a height equal of the individual student. A large theoretical and
to the proportion of the cohort remaining in the state during the empiric literature exists that explores the issue of
last cycle—must be subtracted from the cumulative utility for the
learning preferences. For example, the Dunn and
state.
Dunn learning model postulates that the educator
should match learning preferences to maximize the
participant’s learning.9 Educational researchers have
underestimates the true number of cycles accrued demonstrated that matching perceptual preferences
by an amount equal to the areas between the tops of improves immediate recall (short-term memory),
the rectangles and the smooth curve (Figure 4a). We delayed recall (long-term memory), and attitudes in
will refer to the latter area as the area of interest. We adult learners.1013 Researchers have also demon-
would like to add the area of interest to the summed strated improved attitudes toward the subject matter
areas of the rectangles to obtain a more accurate when perceptual preferences were matched. These
value for the total number of cycles experienced by findings suggested that matching perceptual prefer-
the cohort in the health state. ences translated into a more satisfactory educational
To improve our estimate, we first create addi- experience.
tional rectangular areas indicated by the dotted lines The terms analytic and global have often been
in Figure 4b. We then disregard the original rectan- used to describe cognitive or information processing
gles for the time being and focus only on the new ones styles. Based on information processing theory,
(Figure 4c). For simplicity, we replace the smooth analytic students learn more easily through a step-
curve with straight line segments, which divides the by-step sequential process, which builds toward
new rectangles diagonally in half (Figure 4d). The area conceptual understanding.11 This learning process
of interest is now represented by the lower left triangle mirrors traditional teaching methods. Conversely,
within each rectangle, It is easy to appreciate that students with a global learning style are able to com-
those lower left triangular regions represent half of the prehend more easily when they understand the
area of the new rectangles. overall concept first, and then concentrate on the
In the next step, we imagine sliding the new rectan- details later. For global students, information is best
gles to the left so that they line up against the y-axis of presented with many examples, which are based on
the graph. It is easy to see, by visual inspection, that actual practice (i.e., case-based).11
the rectangles stack neatly between 0 and 1 and that When confronted with a mathematical construct,
they are all 1 cycle wide (Figure 4e). Since the areas of for example, the analytic student would learn best
interest encompass half of each rectangle, and the rec- via a stepwise algebraic proof using symbolic nota-
tangles stack neatly between 0 and 1, one can see tion. However, the global student would achieve
visually that the total area of interest is equivalent to a understanding most easily if the overall purpose of
rectangle one-half cycle wide and whose height the construct were explained first and then a set of
extends from 0 to 1 (Figure 4f). examples using actual numbers were presented. If
METHODOLOGY 709
NAIMARK ET AL.
Figure 4 An intuitive/visual explanation of the half-cycle correction. A, The underestimation of a discrete Markov process with state
membership counted at the end of each cycle compared to a true continuous process is represented by the area between the tops of the
rectangles and the smooth curve. B, Create new rectangular areas. C, Temporarily disregard the original discrete process rectangles. D,
Replace the smooth curve with straight line segments. E, Slide the new rectangular areas so that they line up against the y-axis. F, The
triangular areas can be seen to be equal to a rectangle one-half cycle wide extending from 0 to 1. G, Add back the original rectangles; the
approximation is better.
the Dunn and Dunn conjecture and if information for the subsequent cycles. If the process terminates
processing theory are both valid, then the algebraic early, then a half-cycle needs to be added back for the
explanation of the half-cycle correction should be proportion of the cohort remaining in that state. Tree-
employed for analytic learners whereas the intuitive Age allows the users to specify the initial utility for
approach should be used for global students. cycle 0, the incremental utilities for all subsequent
cycles, and the terminal utility in case the Markov pro-
cess stops early for each state. One might think that the
HALF-CYCLE CORRECTION ISSUES initial and terminal utilities should have opposite
ASSOCIATED WITH COMMON signs since one is a half-cycle subtraction and the other
DECISION-ANALYTIC SOFTWARE is an addition, respectively. However, this is not the
case. The terminal utility should have a positive sign
Students enrolled in our decision analysis courses because it is a half-cycle addition, whereas, for the
most often use either the TreeAge1 or Decision Maker initial utility, the subtraction is implied: the initial uti-
programs for constructing Markov-based decision lity is only half of the value that would have been
models. It is important that they understand the differ- accrued if the half-cycle correction had not been
ences between these 2 programs with respect to the assessed.
half-cycle correction. Decision Maker accounts for TreeAge refers to incremental utility as incremental
state membership at the end of each cycle so that rewards of which there are 3 for each Markov state:
neither of the pedagogical devices mentioned above ‘‘Init’’ refers to the reward for cycle 0; ‘‘Incr’’ is the
needs to be altered.14 Each Markov state has 3 special reward for all subsequent cycles; and ‘‘Final’’ is the
variables associated with it: ‘‘m.uINIT’’ represents a reward to be assigned if the process terminates early.
quantity of incremental utility to be added at the If ‘‘X’’ represents the incremental reward associated
beginning of the discrete Markov process; ‘‘m.uINCR’’ with the state, then the half-cycle correction in TreeAge
represents the incremental utility of subsequent is achieved by setting Init, Incr, and Final to 0.5∗ X, X,
cycles; and ‘‘m.uTAIL’’ represents the incremental and 0.5∗ X, respectively.
utility to be added when the Markov process termi- Students who use other software for conducting
nates for the portion of the cohort remaining in the Markov analyses should read the accompanying
state. If ‘‘X’’ represents the incremental utility asso- documentation carefully and pay close attention to
ciated with the state, then the half-cycle correction in the order in which calculations are performed dur-
Decision Maker is achieved by setting m.uINIT, ing each cycle. If state transitions occur before state
m.uINCR, and m.uTAIL to 0.5*X, X, and −0.5∗ X, membership is counted, then half-cycle corrections
respectively. should be implemented as it is for Decision Maker.
In contrast, TreeAge accounts for state membership Otherwise, a TreeAge-like approach is required and
at the beginning of each cycle rather than at the end they should determine how their chosen program
and therefore overestimates true state membership.15 handles the initial cycle.
Thus, a half-cycle must be subtracted from the cumu-
lative total for a given health state. If the process termi- CONCLUSION
nates early, then a half-cycle must be added back for
the proportion of the cohort remaining in the state. A We hope that these pedagogical devices and
potentially confusing aspect of TreeAge is that the approaches will prove useful to instructors of Markov
users’ manual indicates that transitions occur between theory as they explain the rough equivalence between
states at the end of each cycle. However, students the half-cycle correction and the state membership
should understand that the important criterion with curve shift to their students. Empirical studies of the
respect to the half-cycle correction is not when the cognitive learning processes and preferences of stu-
transitions occur but, rather, when state membership dents of Markov-based decision analyses is an area for
is counted. future research.
The very first cycle in a Markov process constructed
with the TreeAge program is counted as cycle 0.
Within that initial cycle, the half-cycle penalty is REFERENCES
assessed by adding only half of the incremental utility
to the cumulative total that would normally accrue in 1. Detsky AS, Krahn M, Naglie G, Naimark D, Redelmeier D. Primer
a typical cycle. If no half-cycle correction is to be on medical decision analysis: part 1. Introduction. Med Decis Mak-
assessed, then the value for cycle zero is the same as ing. 1997;17:123–5.
METHODOLOGY 711
NAIMARK ET AL.
2. Detsky AS, Krahn M, Naglie G, Naimark D, Redelmeier D. Primer training achievement, and attitudes of corporate employees. Anony-
on medical decision analysis: Part 2. Tree structure. Med Decis mous. St. John’s University, Queens, NY; 1989. 04-12-1 A.D. Doctor
Making. 1997;17:126–35. of Education. Available at: http://wwwlib.umi.com/dissertations.
3. Naglie G, Krahn M, Naimark D, Redelmeier D, Detsky AS. Primer 11. Tendy SM. Effects of matching and mis-matching perceptual
on medical decision analysis: part 3. Utility measures medical deci- and sociological learning-style preferences on achievement and
sion making. Med Decis Making. 1997;17:136–41. attitude of individuals in a group exercise leadership instruc-
4. Krahn M, Naglie G, Naimark D, Redelmeier D, Detsky AS. Primer tional training program. Anonymous. St. John’s University,
on medical decision analysis: part 4. Sensitivity analysis. Med Decis Queens, NY; 1998. 04-11-1 A.D. Doctor of Education. Available
Making. 1997;17:142–51. at: http://wwwlib.umi.com/dissertations.
5. Naimark D, Krahn M, Naglie G, Redelmeier D, Detsky AS. Primer 12. Raupers P. Effects of accommodating perceptual learning-
on medical decision analysis: part 5. Working with Markov pro- style preferences on long-term retention and attitudes toward
cesses. Med Decis Making. 1997;17:142–51. technology of elementary and secondary teachers in professional
6. Redelmeier DA, Detsky AS, Krahn M, Naimark D, Naglie G. development training. Anonymous. St. John’s University,
Guidelines for verbal presentations of medical decision analyses. Queens, NY; 1999. 04-12-1 A.D. Doctor of Education. Available
Med Decis Making. 1997;17:228–30. at: http://wwwlib.umi.com/dissertations.
13. Dunn R. Understanding the Dunn and Dunn learning styles
7. Sonnenberg FA, Beck JR. Markov models in medical decision
making: a practical guide. Med Decis Making. 1993;13:322–38. model and the need for individual diagnosis and prescription.
Reading Writing Learn Disabil. 1990;6:223–47.
8. Instruction manual for Treeage Pro v 2006. Chapter 34. Build-
ing and Analyzing Markov Models. Williamstown (MA): TreeAge 14. Decision Maker for Windows operating manual. Page 39. New
Software; 2006. Brunswick: University of Medicine and Dentistry of New Jersey;
2007.
9. Dunn R. Capitalizing on college students’ learning styles:
theory, practice, and research. In: Dunn R, Griggs SA, eds. Practical 15. Instruction manual for Treeage Pro v 2007. Chapter 35, Page
Approaches to Using Learning Styles in Higher Education. Despot 476. Chapter 35: Markov modeling tools and techniques. Williams-
(CT): Bergin & Garvey; 2000. town (MA): TreeAge Software; 2007.
10. Ingham J. An experimental investigation of the relationships
among learning style perceptual strength, instructional strategies,