Drugmart: Heroin Epidemics as Complex

Adaptive Systems

MICHAEL H. AGAR 1
1
Friends Social Research Center and Ethknoworks, PO Box 5804, Takoma Park MD 20913; e-mail:
magar@anth.umd.edu

DWIGHT WILSON 2
2
Department of Computer Science, Johns Hopkins University, and D. S. Wilson Consulting, 2221 St. Paul
Street, Baltimore, MD 21218; e-mail: wilson@cs.jhu.edu

Received April 15, 2002; accepted June 12, 2002

Illicit drug epidemics are infamous for their unexpected arrival and their speed of onset. Based on ethnographic
work with youthful heroin experimenters in the Baltimore metropolitan area, an explanation was constructed
based on circulating stories of drug reputation. An agent-based model was built in SWARM to evaluate the
explanation with good and bad stories about the drug passed among agents, which changed their initial
attitudes. On repeated runs with different initial attitudes the model shows wide variation in outcomes and a
dampening effect of increased social connections, contrary to epidemiologic expectations. The conclusion spells
out implications for drug intervention and social research that relies on single case studies. 䉷 2002 Wiley
Periodicals, Inc.

Key Words: social research; epidemiology; illicit drug use; artificial societies; complex adaptive systems
modeling

C
omplex worlds have been the object of social research
since social research began. In the case of illicit drug
epidemics, no one has yet explained why they appear
so unexpectedly and generate steeply rising incidence
curves. Based on ethnographic research with youthful ex-
perimenters, an explanation was constructed: epidemics are
driven by rapidly circulating stories—both good and bad—
about a drug and its effects. In this article, we build and run
an adaptive agent model to evaluate this explanation.
Here is the case study we mean to use for the test: Agar
has been working in substance abuse research, off and on,
for decades. One thing is clear: Illicit drug epidemics always
cluster in a population. Yet no one has figured out an an-
swer to the fundamental epidemiologic question: why these
people in this place at this time? Currently he works on an
NIH/NIDA funded project to look at different cases of illicit
drug epidemics—“epidemic” in the sense of a steeply rising
incidence curve—to try and answer that question.

44 COMPLEXITY © 2002 Wiley Periodicals, Inc., Vol. 7, No. 5

 One case, researched with his associate Heather Schacht
Reisinger, involved an upsurge in heroin use among white
suburban youth in the late 1990s [1]. The full analysis of this
and other cases led to what is called “trend theory” [2]. We
will not go into the details of the theory here. At the moment
we just want to point out that full trend theory requires an
analysis of production and distribution systems and their
interactions with a population under certain historical con-
ditions. The work we report here only involves populations
who are more or less disposed to try illicit drugs, with no
explanation of why those differences in disposition might
exist.
Ethnographic research with the 1990s youth revealed a
most interesting result. According to the youth, the primary
engine that drove the heroin epidemic curve—both its ac-
celeration and its flattening out—was the local reputation of
the drug. Reputation was conveyed from person to person
in the form of stories, advice, observations, etc. The “buzz”
around the drug, good and bad, was mostly learned from
experience, either of the person doing the talking or of an-
other youth that the person doing the talking knew or had
heard about.
The importance of reputation in spread of illicit drug
use, based on one’s own experience or the experiences of
others, is not a novel argument. In fact, it is compatible with
other kinds of studies. A model developed by RAND makes
the same basic claim, and epidemiological work during the
1960s/70s heroin epidemic showed that first use typically
occurred when a peer who had recently experimented ini-
tiated another in the same social network [3–5]. One robust
finding of the national U.S. survey of adolescents, called
“Monitoring the Future,” is that frequency of use of illicit
drugs varies monotonically with peer attitudes toward
them [6].
Drug education programs and media coverage also con-
vey drug reputation. They clearly can have an effect as well,
though typically not until after the corner has already been
turned on the rising incidence curve. By the time the media
and the educators notice, things are already well underway.
Besides, media and education often feature only the bad
news; when they do that, then they lose credibility, because
positive as well as negative stories are circulating through
the networks.
The local story, in other words, looks like it is more im-
portant than adult-provided drug education, media, or any
other outsider messages. And the local story is never con-
sistently good or consistently bad, at least not for heroin.
Both positive and negative stories circulate. On the one
hand, you hear that heroin is the best high, ever. On the
other, you hear that it turns you into a zombie and makes
you throw up. The youth taught us that at first the stories
are more likely to be positive than negative, but with time
the balance changes and negative stories predominate, es-
pecially when addicts appear and are observed first hand.
© 2002 Wiley Periodicals, Inc.
This dynamic of good and bad stories at the local level was,
in fact, the key to understanding how a heroin epidemic
curve first takes off and then flattens out.
The question is, if we build a world of adaptive agents
who circulate good and bad stories about an illicit drug,
what sorts of incidence curves will that world produce? Will
circulation of good and bad stories produce an emergent
incidence curve at the system level? And if so, can the way
the model works teach us anything new about how epidem-
ics work? Or what we might do about them?
We want to report on the current state of the model we
have built that takes these questions seriously. We call the
model Drugmart. To foreshadow what is to come, we want
to show three results in this article.
The first result is obvious—raising or lowering the initial
attitude of the agents toward illicit drugs affects the curve in
predictable ways. If Drugmart begins with antidrug agents,
then fewer of those agents will use in the end. Less obvious,
though, is that the simple circulation of stories is, in fact,
sufficient to produce curves of the sort we see in real
epidemics.
The second result is not so obvious—when we run the
program numerous times under the same parameters, with
only the random seed varying, it produces a wide variety of
incidence curves. Sensitivity to initial conditions is of course
to be expected, but given the general image of heroin—once
it lands, massive use ensues—the range of outcomes is strik-
ing. In other words, the phase space of Drugmart epidemics
is larger than we imagined.
The third result, however, is less obvious still: more
densely connected social networks result in less use, in con-
trast to the epidemiologic expectation that denser networks
would produce more, and more rapid, “infection.” The rea-
son, as we will see, lies in the circulation of both positive
and negative stories, together with the higher credibility of
stories from “friends,” i.e., those with whom an agent has a
social connection.
We had no idea, before the fact, what kind of curve the
shifting balance of stories would produce. In the end, what
the youth taught us did, more or less, produce just what
they said it would. The stories turned the curve upward, and
then, with time, flattened it out.
ARTIFICIAL SOCIETIES
At a general level, complexity theory is already useful to
social research as a metaphor, because its nonlinear dy-
namic assumptions provide a better intuitive fit with the
kinds of problems most social researchers study. Several
social researchers have discovered this intuitive fit. In fact,
applications are a varied and interesting lot. Complexity and
chaos are clearly growth industries. Sometimes reading the
literature gets downright disorienting. For example, com-
plexity is offered as a solution to both “quantitative” and
“postmodern” social research. Byrne [7] argues that com-
COMPLEXITY 45
 plexity reframes the quantitative sociological paradigm; for
example, he thinks of statistical clusters as system “attrac-
tors” that then need to be explained. Cilliers [8], on the
other hand, shows how neural nets clarify and make more
precise the postmodern program, with shifting patterns of
distributed knowledge and power that adapt to changing
circumstance.
Drugmart fits into the more specific category of “artificial
societies.” This application of complexity to social research
goes beyond the general idea of introducing complexity
ideas at a metaphorical level. Most useful for the problem at
hand is Epstein and Axtell’s work on “artificial societies” [9].
They developed a model called “Sugarscape” consisting of a
world of sugar resources and agents who want to eat it.
Epstein and Axtell build their model in layers, chapter by
chapter. Initially there are just agents and sugar, agents
who, by the way, rapidly generate inequitable distributions
of “wealth.” Then complications are added, bit by bit: re-
production and death, ethnic groups, trade, new commodi-
ties, conflict, and so on.
Here we are less interested in the details of Sugarscape
and more in their comments on the general nature of “ar-
tificial societies.”
Indeed, the defining feature of an artificial society
model is precisely that fundamental social structures
and group behaviors emerge from the interaction of
individual agents operating on artificial environments
under rules that place only bounded demands on each
agent’s information and computational capacity. The
shorthand for this is that we “grow” the collective
structures “from the bottom up” [9, p. 6].
In a nutshell, agents—who know some things and can do
other things—interact with an environment—which also
knows some things and can do other things. Agents and
environments will both change as a result of these interac-
tions. The “macrostructures” that these local interactions
produce over time—maybe they’re surprising, certainly
they’re variable from time to time—are the phenomena that
we are trying to explain. The program, then, represents a
candidate explanation for the phenomenon.
Clearly, agent-based social science does not seem to
be either deductive or inductive in the usual senses.
But then what is it? We think generative is an appro-
priate term. The aim is to provide initial microspeci-
fications (initial agents, environments, and rules) that
are sufficient to generate the macrostructures of inter-
est. We consider a given macrostructure to be “ex-
plained” by a given microspecification when the
latter’s generative sufficiency has been established
[9, p. 177].
46 COMPLEXITY
It is worth remembering that the purpose is not to create
a “simulation” of a world in the sense, say, of a flight simu-
lator or computer game. Axelrod, a political scientist who
has also used agent-based models, talks about a program as
a way to conduct thought experiments, to tinker with intui-
tions. One shears away the detail of particular instances to
“enrich our understanding of fundamental processes that
may appear in a variety of applications” [10, p. 5]. Artificial
societies are meant to test explanations, not provide lifelike
replicas.
Axelrod uses the prisoner’s dilemma game as one ex-
ample. He conducted a tournament among models for
game strategy. Anatol Rappaport’s simple “tit-for-tat” won,
a strategy where the player first cooperates and then does
whatever the other player did on its last move [11]. Axelrod
then used Holland’s “genetic algorithm” [12] to see if agents
would figure this out for themselves. With a couple of im-
portant exceptions that are neglected here, they in fact did.
The model helped understand both “foraging fish and di-
vorcing,” not because it “represented” either, but because it
captured a feature of many situations, the conflict between
“the advantages of selfishness in the short run versus the
need to elicit cooperation from the other player to be suc-
cessful in the longer run” [10, p. 6].
Epstein and Axtell, and Axelrod, talk about how agent-
based modeling requires a “transdisciplinary” approach to
social research, an issue that recalls the writings of the an-
thropologist Eric Wolf, though he was more of a “top down”
theorist.
When such multifaceted agents are released into an en-
vironment in which (and with which) they interact, the re-
sulting society will—unavoidably—couple demography,
economics, cultural change, conflict, and public health. All
these spheres of social life will emerge—and merge—
naturally and without top-down specification, from the
purely local interactions of the individual agents. Because
the individual is multifaceted, so is the society” [9, p. 158].
Epstein and Axtell present a “gedanken experiment” to
show how artificial it is to isolate phenomena. Consider
population: as one turns various rules on or off that have to
do with movement, sex, culture, trade, disease, etc., the
curve moves. Treating the curve in isolation would be ri-
diculous.” …is it sensible to study long-range population
dynamics as though economic structure were irrelevant?”
[9, p. 159].
So, one casts aside disciplinary boundaries and zeroes in
on a “macrostructure” of interest, in our case a wave of
heroin experimentation. Why does it happen in different
ways among different people at different times? Can we dig
into the case studies and come up with a few simple char-
acteristics of agents and their worlds and build an agent-
based model? Will that model run on repeated occasions
and produce an epidemic S-curve? What else might it
produce?
© 2002 Wiley Periodicals, Inc.
 DRUGMART
Drugmart, implemented in SWARM, is very much a work in
progress. It creates 500 randomly placed agents in a simple
world defined by an 80 ⳯ 80 grid. One cell is randomly
selected, and heroin is placed there. Each agent is assigned
a risk value between 0 and 1, based on a random-normal
distribution. The use of this distribution is justified by the
many decades of research in “diffusion of innovation” [13]
that shows that willingness to try something new is distrib-
uted in just that way in numerous populations for numer-
ous kinds of innovations.
An attitude value, also varying between 0 and 1, is set as
a parameter. That value is assigned to all 500 agents. This
parameter allows us to experiment with different orienta-
tions to illicit drug use when heroin first appears. Agent
worlds, in other words, can range from strongly anti- to
strongly pro-illicit drug use.
The first basic mechanism of the model is this: If an
agent lands on the cell where heroin is located or on a cell
where another agent has heroin (more on this in a mo-
ment), then the agent checks the current value of its atti-
tude. If its attitude is greater than one minus its risk, it takes
heroin. If its attitude is less than one minus its risk, it does
not.
But an agent’s attitude changes as the program ticks
along. Therein lies a long story. Before embarking on the
story, though, let us outline the basics. Agents change their
attitudes for two good reasons. The first reason is, the way
attitude changes depends on whether an agent has a good
or bad experience with a drug. Agents also change their
attitudes depending on the “buzz,” what they hear from
other agents as they move through the world.
Those are the basics; now for the longer story. First, what
determines whether an agent has a good or bad experience?
The answer is simple, on first glance. We define a table of
probabilities for good or bad experiences and an associated
amount by which the value of the attitude increases or
decreases.
But the answer is more complicated, on second glance:
The probability of a good or bad experience depends, in
turn, on what kinds of prior experiences an agent has had
with the drug. All agents are “clean” at the beginning of the
program, “clean” being the usual informal term for one who
does not use illicit drugs. But with time, the experience of
different agents will have varied. So we will find users, or
nonusers, of several different types:
1. A clean agent has never tried heroin.
2. An experimenter has tried heroin at least once.
3. A user uses “regularly.” A random number is generated
for each agent, between 0 and its risk value. This is its
“user threshold.” If attitude becomes higher than one
minus the user threshold, then the agent becomes a user.
If the number drops below that value, it stops being a
© 2002 Wiley Periodicals, Inc.
user. A user also carries heroin, and uses at intervals
ranging from n1 to n2, those numbers being set as
parameters.
4. An addict is a user who has used to a certain threshold
value, a value also set as a parameter. The threshold
value decays by a certain proportion each time heroin is
not used, a proportion that is also set as a parameter. An
addict, like a user, carries heroin.
5. An exaddict is an addict who has stopped using, though
there is a possibility of relapsing into heroin use. This is
an aspect of Drugmart that we have not worked with yet,
so it is neglected here.
So, for example, if a clean agent tries heroin, it has a
probability of a good experience, and it has a probability of
a bad experience. If the experience is good, attitude will
increase by a certain amount. If the experience is bad, atti-
tude will decrease. The same is true for experimenters, us-
ers, addicts, and exaddicts, although for those other types of
agents, probabilities and amounts of change will be differ-
ent. But the same mechanism operates: If an agent uses,
there are probabilities of good and bad experiences, and
depending on whether the experiences are good or bad, the
value of their attitudes will change.
Table 1 shows the numbers we actually used in the runs
reported here. The left-hand side represents the probabili-
ties that the experience of using heroin is good or bad, that
is, whether or not it will increase or decrease the agent’s
attitude. The right-hand side represents how much the at-
titude will increase or decrease as a proportion on a scale
between 0 and 1.
Walking the reader through every number here, never
mind the matrices to come, is beyond the scope of an ar-
ticle. Notice, as one quick example, how probability of a
TABLE 1
Look-up table for agents
Experience probability Experience amount
increase increase
0, 75, 0, 80, 50, 10, 0, 20, 0, 10, 10, 5,
10, 75 5, 20
Experience probability Experience amount
decrease decrease
0, 15, 75, 15, 15, 0, 20, 20, 10, 10, 5,
10, 10, 5 5, 0
Values are the probability that attitude toward heroin will increase
or decrease based on first-hand experience and amount of that
increase or decrease. The eight values are, respectively, for clean
agents, experimenters, experimenters with a bad experience, us-
ers, users with a bad experience, addicts, addicts with a bad expe-
rience, and former addicts. The types are defined in the article.
COMPLEXITY 47
 good experience is quite high for an experimenter—the sec-
ond figure, 0.75, in the top-left quadrant. This corresponds
to what we learned in ethnographic interviews. But once an
experimenter has had a bad experience with the drug, the
probability of an increase goes to zero—the third figure in
the top-left quadrant. Again, this is the story we frequently
heard, “I tried it, bad things happened, and that was the end
of that.”
To offer one more example, note that probabilities of
both increase and decrease are low for an addict, with or
without bad experiences. Values of 0.10 are placed in the
seventh and eighth positions on the left-hand side of the
Table 1. By the time one is an addict, heroin is primarily
keeping one from getting sick, because one is physically
dependent on the drug. Good/bad experiences will happen,
but by then they are beside the point.
And finally, notice how the amount of attitude change
(Table 1, right-hand side) is higher at the less-experienced
end. The further one moves along the clean-to-addicted
scale—the more one physically needs the drug—the less
relevant the rational risk/attitude comparison. Again, this
corresponds to what we know about heroin addiction, but
notice the last column, the “exaddict” position, where prob-
ability of a good experience and amount of attitude change
is once again high. Once one has been addicted and then
cleans up, it is very difficult to stay clean, because one
knows how positive the experience can be and knows that
in the “clean” state effects will be strong.
The probabilities and amounts we have just discussed in
Table 1 tell us what an agent does when the agent itself uses
heroin, which is a guide for the direct experience of the
agent. However, direct encounters with the drug are not the
only way the agent changes its attitude. It also hears about
the drug—the good, the bad, and the ugly—from other
agents it encounters as it wanders around its world.
To model how an agent picks up on “the buzz” around
the drug, we have each agent check with all other agents
immediately adjacent to it or with other agents that share
the cell where it lands, after each tick of the program. Those
other agents might or might not influence its attitude. The
chance that those other agents will or will not influence it?
It depends on two things.
First of all, both agents, the one who is checking for
influence and the one who is doing the influencing, are
classified as a certain type of user. The categories were de-
scribed earlier: experimenter (with or without a bad expe-
rience), user (with or without a bad experience), and so on.
The probability of influence and the amount of attitude
change will depend on what kind of agent is trying to influ-
ence what other kind of agent. For example, two experi-
menters neither of whom have had a bad experience will be
likely to influence in a positive direction. If the influencing
agent has had a bad experience, then there is a strong like-
lihood that it will influence the other experimenter in a
48 COMPLEXITY
negative direction. On the other hand, there is a strong like-
lihood that an addict will influence experimenters or users
in a negative direction.
So, to guide how the “buzz” influences agents, the model
relies on an N ⳯ N matrix, with kind of user in the rows
influencing kind of user in the columns. And, as we said
earlier about Table 1, going over each individual matrix en-
try is beyond the scope of this article, so we neglect that
exercise here. The general idea remains the same—an agent
looks up the probability of good or bad news about the drug,
depending on its own category and the category of the agent
that is in a position to influence it. If the influence occurs, it
looks up the amount of attitude change that corresponds to
that agent/agent pairing.
The second thing that guides how buzz flows through
Drugmart—Some agents are connected by a social link, and
others are not. We will call those that are connected
“friends” and those that are not “strangers.” The number of
connections is set as a parameter, and this parameter will be
important in the next section when we report results. For
this article, we ran Drugmart under three conditions. In the
first, there are no connections among agents; Drugmart is a
world of strangers. In the second, we specify 200 connec-
tions. The program draws two agents at random, checks to
see if there is already a connection, and if not, puts one in
place. It does this until it has successfully connected 200
agents. In the third condition, the parameter is increased to
400 connections.
If two agents land in a position where influence on atti-
tude is possible and those two are socially connected, then
a different matrix comes into play. If two agents are adjacent
and the two are “friends,” then the agent who is looking up
the chances of being influenced has to use different matri-
ces. Friend have more influence than strangers in both posi-
tive and negative directions.
COSMIC NUMBERS
The heart of Drugmart is now in place. Before we show
preliminary results, though, some discussion of an issue we
have mentioned and then avoided is in order, particularly
for those who, like the authors, view numerical representa-
tions of the human situation with a healthy skepticism.
How in the world do we assign numbers to probabilities
and attitude changes? Consider a simple case from Table 1
discussed earlier. The table shows that, for an experimenter,
the probability of a good experience is 0.75, of a bad expe-
rience, 0.15. However, if an experimenter has already had a
bad experience with the drug, the probability of a good ex-
perience goes to 0 and of a bad experience, to 0.75.
These numbers are not complete fabrications. On the
basis of everything we know from the drug field, we can say
that an initial experience with heroin does tend to be pow-
erful and pleasant—not always, but more often than not. So
0.75 as “sort of a lot” and 0.15 as “not all the time” are at
© 2002 Wiley Periodicals, Inc.
 least in the range of plausible translations. And, if an ex-
TABLE 2
perimenter has already had a bad experience and tries again
anyway, probably the same reasons it was a bad experience
Ranking of English quantifiers on a scale of 100
the first time will work again, whatever those reasons might
have been. Even if an experimenter tries again and does not
Always 99 100
have a bad experience, he/she probably will not improve Very often 88 87
their attitude. Perhaps there should be a slight chance of an Usually 85 79
increase in attitude here, but the number is set to 0 because Often 78 74
one so often hears or reads about the “I tried it once and Generally 78 72
Frequently 73 74
didn’t like it” story when a new illicit drug sweeps through Rather often 65 74
a population. Often as not 50 50
The numbers are related to reasonable assumptions Now/then 20 34
about how the world works, but clearly they are guesses at Sometimes 20 29
Occasionally 20 28
best. To pin them down a little more, we used results of two
Once awhile 15 22
studies as guidelines, as reported in an overview of fuzzy Not often 13 16
logic [14]. The researchers drew the usual academic sample Usually not 10 16
of college students, who were asked to listen to ordinary Seldom 10 9
Hardly ever 7 8
language terms that represented frequency. They then
Very seldom 6 7
placed each term on a scale of 0 to 100. Numbers represent Rarely 5 5
the median scores from both studies [15,16]. Almost never 3 2
Does using such results improve the situation any? It is Never 0 0
reassuring to see that the figure used in Table 1, 0.75, is in
the “often” and “generally” range, whereas 0.15 translates as
“not often” and “usually not.” At least the numbers sound
DRUGMART IN ACTION
like a reasonable estimate.
To summarize, then, we have described the heart of
For present purposes, we will assume that the numbers
Drugmart. Five hundred agents begin with normal, ran-
we put in the matrices are at the right “order of magnitude,”
domly distributed risk and a shared attitude set to some
but it bothers us that the production of these numbers
number with a parameter. The agents move around and, if
raises so many critical questions from both qualitative and
they encounter heroin, they compare attitude to risk. If at-
quantitative points of view, critical questions that we ne-
titude is higher than T-risk, they try the heroin. And, if they
glect here. Can we derive any reasonable summary of how a
try it, they have good or bad experiences, with some prob-
moment of influence, one agent to another, will always go? ability, and those experiences, should they occur, change
And if we can do that, can we represent the force of that their attitude by some amount.
general influence in a numerically sensible way, in a field Agents also change their attitude, depending on the
where numbers on that same scale represent other kinds of “buzz” around the drug that they pick up as they move
influence that may be qualitatively different as well? Re- around the world. After a tick of the model, any adjacent
search aimed at obtaining these numbers, should they turn agent might influence their attitude. The chances they do
out to be reasonable snapshots of complex interactions at so, and the amount of the influence, will depend on the
all, would take a long time to do. We would never get to the combined effect of both agents’ experiences. Chances and
model. amount also depend on whether the two agents are “strang-
This is, and will continue to be, such a recurrent problem ers” or “friends.”
in adaptive agent model-building for social research that we So what, in the end, do we get for all this trouble?
should name it something—like the complexity conundrum. We decided, for this first exploration, to look at the “at-
Any complex model involving human agents will require tractor space” of Drugmart under varying conditions. Such
translations of contingent and socially constructed “quali- exploration is, of course, a major attraction of complex
ties” of life into static numerical form. Usually (which trans- models for those who begin their research with detail in a
lates to 0.85 to 0.79 in the list of terms in Table 2), the single case study. Such researchers can model the explana-
translation will be problematic, in that neither data nor tion of that single case and run it several times to see what
guidelines will exist for accomplishing it. More problematic kinds of variations on the theme might also occur, thus
still, the translation will be buried in program details, likely achieving a kind of “generalization” that otherwise would
to be passed over as the general operation of the model is not be possible. Sampling in silico we might call it.
viewed and evaluated. We ran Drugmart under three conditions, corresponding
This is not a good thing, and yet we proceed. to the way we set the attitude parameter: a strong antidrug

© 2002 Wiley Periodicals, Inc. COMPLEXITY 49

 attitude, a strong prodrug attitude, and an attitude midway
between the two. For each of these conditions, we ran the
program with no social connections, with 200 social con-
nections, and with 400 social connections.
Under each set of conditions, the program was run 100
times. Each run ended when no significant changes were
occurring or after 500 ticks, whichever came first. The first
results are shown in Figure 1.
For all runs shown in Figure 1, attitude was initially set to
strongly antidrug, or 0.2 on a scale of 0 to 1, and three sets
of runs are shown. For the first set the social connection
parameter is set to 0; for the second set, to 200; and for the
third set, to 400.
The graph shows the end results of each of the 100 runs:
first for the no social connection runs, then for the 200
social connections runs, and finally for the 400 social con-
nections runs. The red line indicates the total number of
agents, out of 500, who tried heroin at least once, the so-
called “lifetime prevalence.” The green line shows the total
number of agents who used heroin frequently and repeat-
edly enough to cross the threshold and become addicts.
The results of the 100 runs, under each set of conditions
of social connectedness, were sorted by total number of
agents who had used heroin at least once. Thus, the red line
always increases in Figure 1 and in subsequent figures. Then
the results were plotted on a simple line graph, with total
number of agents on the y-axis and the three sets of 100
runs representing different degrees of social connection on
the x-axis.
A first impression from Figure 1 is that a complex model
based on circulation of stories that convey heroin reputa-
tion is sufficient to generate epidemiologic curves of the sort
we find in actual illicit drug epidemics. A second impression
is that the variation in outcomes is striking under the same
initial conditions or at least striking given the usual stereo-
type of heroin epidemics. A third impression is that, al-
though the number of addicts increases with number who
have ever used, it is not a simple linear relationship, and a
final impression is that an increase in the number of social
connections has a dampening effect on both ever used and
addiction, contrary to what infectious disease models from
epidemiology would lead us to expect.
Before discussing these results let’s examine some oth-
ers. The first graph (Figure 1), with its strong antidrug atti-
tude, did not produce many agents who had ever used or
who became addicts. The maximum ever used was less than
30 of 500. The story changes when the attitude parameter
increases to 0.5, midway between the possible values of 0
and 1.
Numbers rose dramatically in the second graph (Figure
2). Now the maximum number of experimenters is around
150 of 500, and the maximum number of addicts is just
under 100. In fact, looking across all three sets, it is worth
noting that research usually shows that between 10% and
50 COMPLEXITY
20% of a population become addicted during a heroin epi-
demic [17–19]. However, although the numbers are differ-
ent, the general pattern seen in the first graph holds up here
as well: striking variation in ever-used and addict numbers,
general increase in addiction with ever-used but not linear,
and dampening effect of increasing social connections.
Now let’s turn the attitude up to strongly prodrug with a
value of 0.8 on a 0 to 1 scale.
All hell broke loose, with maximums of ever used above
400 of 500 agents and addiction upwards of 250, or fully half
the agents (Figure 3). Notice at this extreme attitude level
that variation is less when compared to the first two graphs
(Figures 1 and 2) and that the dampening effect of social
connections does not have much influence on ever-used
numbers. At the same time, connection still reduces the
numbers of addicted.
DRUGMART AND ITS FUTURE
In the end, what do we have? Many things, and many things
to do. First, by Epstein and Axell’s argument, described ear-
lier, we have established that circulating stories and expe-
riences of heroin are sufficient to generate epidemic inci-
dence curves. As they note, this establishes a candidate ex-
planation, though it does not rule others out.
The dynamics of Drugmart depend on the numbers as-
signed to the matrices of good/bad experiences and influ-
ence by “strangers” and “friends.” We already discussed our
concerns about those numbers, though we believe they are
at least the right order of magnitude. In future work, we
want to design two other separate systems of matrices and
run the model again. The two systems would represent a
“benign drug” and a “disaster drug.” Though not without its
problematic aspects, marijuana, for example, would be
closer to the “benign drug” end of the scale, whereas some-
thing like PCP would be closer to the “disaster drug” end.
Drugmart should generate the appropriate different inci-
dence curves.
If reputation does explain the incidence curve, then
there are several implications for policy and prevention.
1. Illicit drug epidemics may be self-limiting, if the dynamic
of increasing negative “buzz” holds. By the time policy
changes and prevention efforts begin, the epidemic may
already have peaked. In one case study, it was shown that
by the time President Nixon’s “war on drugs” began, the
1960s wave of heroin had already peaked [20].
2. Early appearance of negative buzz suggests that preven-
tion material is available in the world of users. If inter-
veners know what is happening as it happens, simply
making space available for experimenters and early users
to talk about the experience among themselves, using
their own stories, might accelerate the braking effect of
negative stories.
© 2002 Wiley Periodicals, Inc.
 FIGURE 1
Drugmart runs with attitude set at 0 2.
3. This strategy is limited by U.S. “war on drugs ideology,”
which in its stereotypic form holds that any use of an
illicit drug is “abuse,” and that no positive effects can
result from illicit drug use. The reasonable argument be-
hind this premise is that interveners should send consis-
tently negative messages to youth about illicit drug use.
The problem here is that positive stories are part of the
dynamic as well and that experimentation does not nec-
essarily lead to problem use.
4. The preoccupation with “forecasting” drug epidemics
may be misplaced. The same initial conditions can lead
FIGURE 2
Drugmart runs with attitude set at 0.5.
© 2002 Wiley Periodicals, Inc.
FIGURE 3
Drugmart runs with attitude set at 0.8.
to wildly different outcomes: more appropriate, if
Drugmart’s lessons are correct, would be leading rather
than lagging indicators (currently the norm) that enable
early and rapid interventions of the type outlined in no.
3 above.
5. Initial attitude is obviously critical, as many studies dis-
cussed earlier have shown. Drugmart is neutral on this
issue, simply taking attitude as an initial parameter to be
set. This area is a topic for another article. We just note
two things here. First, attempts to influence attitude only
in a negative direction will, for many illicit drugs, be con-
tradicted by those stories that circulate that are positive.
Credibility of the source of the exclusively negative mes-
sages will be damaged early on. Second, much of our
work on trend theory, alluded to earlier, looks at histori-
cal conditions around using populations that increases
the value of the subjective effects of a particular illicit
drug. Attitude is an extremely complicated issue, but
Drugmart does not contribute in this area.
Drugmart also has some implications for those who con-
duct social research that involves a quest to learn complex
local historical patterns among some group of people, wheth-
er ethnographers, community psychologists, investigative
journalists, or intelligence analysts. This kind of research
involves an approach quite different from positivist scien-
tific traditions, as recent and abundant debates in the lit-
erature illustrate. The limitation here is that research of ne-
cessity focuses on the rich details in a particular case study.
Critical reactions of such researchers to earlier presen-
tations of Drugmart hold that the very richness of detail that
results from such studies is not fully represented in a com-
plexity model of this type. The criticism is of course correct.
COMPLEXITY 51
 The counterargument is this: When ethnographers, for ex-
ample, present research results, they typically abstract from
that rich detail and offer a core problem and describe a
basic process that explains it. The identification of that
problem and the construction of that explanation presup-
pose a prior investigation of those details to correctly iden-
tify and represent them.
So we add a caveat: A good complexity model requires a
prior detailed study of the world to be modeled. This is
hardly a radical proposal, because it harks back to Bacon’s
emphasis on extensive observation before any experimen-
tation. With that caveat, though, a model like Drugmart can
help test a bottom-up explanation characteristic of ethnog-
raphy, by running it repeatedly to examine its attractor
space and by changing conditions basic to the explanation
to see if properties emerge that follow expectations. The
potential research strategy is a powerful one.
At the same time, problems remain when numbers must
be assigned to qualitative propositions that constitute the
problem description and the explanatory system. We have
discussed this here in several different places. This problem
may, in the end, set limits on the value of adaptive agent
models for this type of research. At a minimum, though, the
problem casts the eternal and often fruitless “quantitative/
qualitative” debate in a more interesting light. It is no longer
a question of what a Pearson’s R between two experience-
distant decontextualized variables tells us about how the
world works. Rather, it is a more interesting question—
namely, is it possible to translate an explanatory proposi-
tion about how the social world works into a mathematical
representation?
We end on an optimistic note. Adaptive agent models
promise to help social research in their quest for theories
about how the world works, but it is a tempered optimism,
born of the historical waxing and waning of formal models
that ultimately proved inadequate to the task. Rather than
rounding up all the usual clichés about the future, we’ll just
end it with that and get back to work.
ACKNOWLEDGMENTS
Support by NIDA grant RO1 10735 is gratefully acknowl-
edged.

REFERENCES
1. Agar, M. H.; Reisinger, H. S. Explaining heroin use trends: Suburban heroin use in Baltimore County, In: Illicit Drugs: Patterns of Use—Patterns
of Response; Springer, A.; Uhl, A., Eds.; STUDIENVerlag: Innsbruck, 2000; pp 143–165.
2. Agar, M. H.; Reisinger, H. S. Trend theory: Explaining heroin use trends, J Psychoactive Drugs 2001, 33, 203–212.
3. Hughes, P. H.; Crawford, G. A. A contagious disease model for researching and intervening in heroin epidemics, Arch Gen Psychiatry 1972, 27,
149–155.
4. Hunt, L. G. Recent spread of heroin use in the United States, Am J Pub Health 1974, 64, 16–23.
5. Hunt, L. G.; Chambers, C. D. The Heroin Epidemics: A Study of Heroin Use in the United States; 1965–1976; Spectrum: New York, 1976.
6. Bachman, J. G.; Johnston, L. D.; O’Malley, P. P. Explaining the recent increases in students’ marijuana use, Am J Public Health 1998, 88,
887–892.
7. Byrne, D. Complexity Theory and the Social Sciences: An Introduction; Routledge: London, 1998.
8. Cilliers, P. Complexity and Postmodernism: Understanding Complex Systems; Routledge: London, 1998.
9. Epstein, J. M.; Axell, R. Growing Artificial Societies: Social Science from the Bottom Up; MIT Press: Cambridge, 1996.
10. Axelrod, R. The Complexity of Cooperation: Agent-Based Models of Competition and Collaboration; Princeton University Press: Princeton, 1997.
11. Rapoport, A.; Chammah, A. W. Prisoner’s Dilemma; University of Michigan Press: Ann Arbor, 1965.
12. Holland, J. H. Hidden Order: How Adaptation Builds Complexity; Perseus Books: Reading, MA, 1995.
13. Rogers, E. M. Diffusion of Innovations, 4th ed.; The Free Press: New York, 1995.
14. McNeill, D.; Freiberger, P. I. Fuzzy logic: The computer technology that is changing our world; Touchstone: New York, 1993.
15. Hakel, M. D. How often is often? Am Psychologist 1968, 23, 437–445.
16. Simpson, R. Q. The specific meaning of certain terms indicating differing degrees of frequency, J Speech 1944, 30, 328–330.
17. DuPont, R. L. The rise and fall of heroin addiction, Nat Hist 1974, 83, 66–71.
18. Jonnes, J. Hep-Cats, Narcs, and Pipe Dreams: A History of America’s Romance with Illegal Drugs; Scribner: New York, 1996.
19. Robins, L. N.; Murphy, G. E. Drug use in a normal population of young negro men, Am J Pub Health 1967, 57, 1580–1596.
20. Agar, M. H.; Reisinger, H. S. A heroin epidemic at the intersection of histories, Culture, Medicine and Psychiatry; forthcoming, 2000.

52 COMPLEXITY © 2002 Wiley Periodicals, Inc.