Drug prices are steep and unpredictable even during stable periods

Why does it matter if drug development is inherently unpredictable and not simply high risk?

The question strikes many in an industry with an end-to-end success rate of between 2% and 3% as (a) obvious and (b) irrelevant. Yet I would argue that the failure to distinguish between high and inherently unpredictable risk is a principal cause of pharma’s potentially catastrophic decline in productivity .

Many of the best practitioners persist in operating as if drug development were simply high-risk--with lamentable consequences. Through skill or scale they believe that they can tip the odds in their favor. More insight and effort, new tools, smarter researchers—e.g. someone who had a blockbuster drug approved in the past--will produce the growth the industry needs.

They are wrong. After 40 years, the innovation crisis shows no signs of abating, despite remarkable advances in both knowledge and technology.

What evidence is there that drug development is inherently unpredictable?

The performance of a business is measured by return-on-investment (ROI), sales minus costs. Returns in pharma are approaching zero and headed negative. Investors provide capital in the expectation of returns. The probability of achieving those returns is risk. To understand risk in an industry, one must look not only at the magnitude of uncertainty but its nature. Different types of risk call for different operating strategies.

A frequency distribution or probability density function (PDF) of annual sales enables researchers to characterize investment risk mathematically. [Graph Ommited]

Blockbusters exert an out-sized effect on industry returns:

Compare average [expected returns in a normal system] and median [mid-point on the X-axis with equal numbers of drugs above and below] sales. In 2012, they were $1.06 billion vs. $546 million respectively. Three-fourths of sales were below the average, which underscores the fact that the “expected returns” are distorted by a few large outliers and not representative of the population.

Other years show the same shape, with many small drugs and an extreme skew to the right (figure 2). [Graph Ommited]

With the average cost of bringing a new drug to market in 2014 running at $2.6 billion, more than twice the average or expected return, pharma is not going to solve its productivity problem by producing more “average” drugs. Without the possibility of blockbusters, the pharmaceutical industry could not afford to develop new drugs.

But why is the occurrence of these rare-event products inherently unpredictable, rather than simply low-probability, a problem that could theoretically be managed by scaling-up?

Examining the PDFs of sales for the years 1986 to 2012 in Figure 2, Dr. De Vany focuses on a parameter called alpha (α), which is a measure of the “skewness” of the chart and has specific implications for the expected (average) returns and the probability (risk) of achieving them. He observes,

For each year in our data set, α < 1, which implies that pharmaceutical sales have an indeterminate mean [expected return] and an indeterminate variance [level of risk]. This reflects the inherent unpredictability of drug sales and explains the impossibility of forecasting them.

To understand generally the implications of a PDF, scholars look for the best match among a selection of well-characterized standard curves, as Dr. De Vany demonstrates below. [Graph Ommited]

For comparison purposes, …we look at three different reference curves for best fit—normal (gaussian), lognormal and Pareto... The normal distribution is a poor fit. The lognormal and stable [Pareto] are much better fits but differ in important ways.

The best overall fit is the lognormal curve, which would reflect a low-probability, as opposed to an unpredictable, process. However, as explained above, the portion of interest to drug developers is the upper tail or far right section of the curve where blockbusters are found. Here the best fit is one called a Stable Pareto curve.

By amplifying the all-important upper tail, Dr. De Vany shows the critical difference between the lognormal curve (best overall fit) and the Pareto curve (best fit for the upper the tail). [Graph Ommited]

Although the lognormal distribution is a reasonable fit for the data, it is not an appropriate model for pharmaceutical sales. …While a lognormal process converges toward the lognormal distribution as the sample size increases, under the Stable Paretian distribution, the scale of deviations [uncertainty] of drug sales continues to grow with sample size, which is consistent with the data…

The Stable Paretian distribution has important implications for understanding pharmaceutical market dynamics. Scale and diversification do not improve the probability of success [i.e. returns from larger portfolios are less certain—riskier—than those from smaller portfolios, though in a good way, since the uncertainty is caused by the higher probability of a blockbuster], and therefore cannot be used to mitigate risk. It is impossible to forecast drug sales from past sales, which are a function of portfolio size and are disproportionately influenced by outliers…

The drug developer’s dilemma: To increase productivity, he doubles his portfolio and his costs. Since blockbusters are extremely rare, it is not likely that either the old or the new halves, taken separately, contain one. When combined, the average or expected return increases slightly, because the probability of finding a blockbuster in the larger portfolio goes up, but the gain is small relative to the step-up in costs. The increased possibility of a blockbuster brings greater uncertainty, because their occurrence is rare and unpredictable. Expenses have gone up substantially; expected returns increased slightly; and the outcome is less certain. Pharmaceutical portfolios do not scale up in a “normal” or predictable manner.