Imperfect Information and Drug
Diffusion
By
Jon Atwell and Martha G. Alatriste Contreras
For
this project we investigate the relationship between drug efficacy and the adoption
of a new drug in a network structure. We assume that most drug treatments are unable
to address all of the underlying causes of a medical condition but nonetheless
might still be effective treatments. This less than perfect correspondence is a
source of imperfect information in that the drug manufacturers are only able to
communicate the efficacy of their drug for a subset of symptoms related to the
condition.
Thus,
when a person exhibits a condition and desires treatment, he/she lacks full information
about the appropriateness of the drug and therefore the potential for side
effects. Augmenting publicly available information about the drug are the
recommendations of network neighbors. These recommendations are based on
successful past experiences, defined as when a person who seeks treatment has
more symptoms successfully addressed than the number of side effects they
experience.
To
pursue this simple setup, we assumed that there exists a list of
characteristics with some possible relevance to the condition. One can think of
this list as a portion of the genome of individuals. Only a subset of these
characteristics or genes are effected by the condition or disease but for the
most part drug manufactures donŐt know the whole set of characteristics that
define the condition. This in turn implies that the new drug may not be fully
effective. The efficacy of the drug is a parameter defined in each simulation.
We explored scenarios in which the efficacy of the drug varied from 0.6 to 1,
meaning that the drug cures 60 percent of the symptoms or all of the symptoms,
respectively. This drug has an opportunity to diffuse across a small-world
network (Newman variation of the Watts-Strogatz
model; 4 neighbors, .1 edge-probability) of either 500 or 1000 potential
patients.
In
this setting, the new drug will diffuse to new persons if either of the
following two conditions are fulfilled: 1) the would-be patients symptoms are
very close (Hamming distance <=2) to those of the condition described in the
drug manual ; or 2) 20 percent of the would-be patientŐs nearest neighbours had
a successful experience and recommended the drug and the symptoms presented are
close to those pointed out by the drug description (Hamming distance <=4).
We
ran the model 50 times for parameter combinations of 500 and 1000 agents and
drug efficacy of .60, .75, .9 and 1 (total sample = 400) and analysed the runs
for possible relationships between drug efficacy, the number of recommendations
and the percentage of the population that successfully used the drug.
First
set of results are presented according to the efficacy of the drug. One can
observe from the graph below that when the efficacy is low there is a lower
percentage of patients successfully treated, and this success relies on
recommendation coming from the agentŐs nearest neighbours. When the efficacy of
the drug increases to 0.75, a combination of efficacy and neighbor
recommendations brings about a high percentage of successful treatments.
Finally, when the efficacy of the drug is high (0.9 and 1.0), the percentage of
patients successfully treated is also high but market penetration is lower
because recommendations are far fewer.
Driving
this dynamic appears to be the instances of unsuccessful treatments with the
drugs of low efficacy. This can be seen in the following graph which plots the ratio
of successful to unsuccessful treatments and the percent of the population
having been treated. One can see that imperfect information about the
appropriateness of the drug encourages experimentation with it as a possible
treatment, but more cases are unsuccessful for efficacy of .60. A slightly high
efficacy makes use of this tendency toward experimentation but minimizes the
number of unsuccessful treatments, thereby making the drug very successful. For
very high efficacy, there is little possibility for experimentation and
therefore market penetration is far less.
The
last graph shows the relationship between the number of side effects and the
success of the drug.
There
is not a one-to-one trade off between side effects and successful treatment
because the more effective drugs still address underlying symptoms while
producing side-effects. Drugs with an effectiveness of .75 still produce as
many side effects as an effectiveness of .60, but remain more successful in the
market.
In
conclusion, we observe that there might be an optimal efficacy for drug that is
not fully effective. This is because there is a trade-off between the potential
for recommendations from neighbors and the rate of successful treatments. This
suggests a positive but diminishing rate of return to neighbor recommendations
and a positive but diminishing rate of return to drug efficacy, which while
stronger than the return to recommendation, suggests an optimal level of drug
efficacy for a given recommendation mechanism.
There
are obviously limits to such a simple model and fully understanding the
dynamics and sensitivity to assumptions would take significant effort.
Nonetheless, the basic insight might prove valuable for future studies. Extensions
of the model include modeling feedback of successful
treatment to medical doctors that prescribe the new drug. This feedback could
potentially reinforce the successful treatments and have an effect on the
importance of recommendations. This type of model could also be applied to the
diffusion of other products where the mechanism of adoption relies on
recommendations and proximity to individual characteristics.