Here are summaries of my recent research projects:
- Clinical trials: forecasted and achieved error rates: Do trials achieve the false-positive and false-negative rates statisticians size them to produce? What percentage of each phase's compounds are safe and effective? I am trying to answer these questions for phases II and III, traditional and adaptive trials, therapeutic areas, small and large molecules, NMEs and nonNMEs, and other categories by analyzing historical data with data science.
- The optimizer's curse: Even when project evaluations are unbiased, project selection, by exploiting imprecise project evaluations, reduces and overestimates portfolio value. The problem can cause simulation optimization, a portfolio optimization technique many companies are adopting, to fail. Bayes' law can mitigate these problems.
- Estimating probability of technical success for a compound's multiple indications: To sequence the trials of compounds with multiple indications, decision analysis provides powerful tools for maximizing value, including decision trees, spreadsheet models, and Monte Carlo analysis. However, these techniques require conditional probabilities that describe the correlated results of trials. Here is a method of estimating these probabilities.
- Portfolio size, return, risk, and almost stochastic dominance: If (a) projects' results exceeds a (low) minimum expected payoff and (b) are poorly coorelated, adding projects to a portfolio always improves a portfolio's risk and return characteristics, a phenomenon called almost stochastic dominance. Partnering with other companies, increasing the number of compounds in your portfolio, exploits this property to produce greater returns with less risk.
- Risk Management for clinical trials: Common methods of managing risk, such as red-yellow-green scales and risk matrices, can poorly prioritize risk and lead to poor (irrational) decisions. Decision analysis offers more powerful tools for identifying, assessing, and managing risks, including framing (identifying facts, assumptions, hypotheses to test, risks, decisions, and alternatives) and analysis (decision trees, tornado diagrams, Monte Carlo analysis, and value of information). Decision-making under deep uncertainty (DMDU) provides techniques, like annealing plans, for addressing black swan events and situations where future scenarios and contingencies are unknowable.
Fuzzy front-end analysis (nonpharma): How do you fill product development with high-value, big-winner projects? The equation is:
Good Choices + Good Choosing = Profitable Portfolio
By applying data science to small data situations, you can estimate the quality of your project proposals (choices) and of project evaluation and selection (choosing). These new KPIs for the front-end of product development can identify your organization's strengths, find opportunities for improvement, and improve your ability to fill product development with high-value projects.