Essays on contest experiments and supervised learning in the pharmaceutical industry

The field of economics witnesses a growing interest to betterunderstand how individuals form decisions and how thesedecisions can be supported with the help of sophisticated dataminingtools. The first half of the thesis analyzes investmentdecisions in a competitive environment that is free of confoundingfactors by means of experimental data. The secondhalf focuses on how supervised learning algorithms can beapplied to predict the outcome and perceived success probabilityof pharmaceutical projects for aiding decisions of policymakers, managers and financial investors.More specifically, chapter two contrasts repeated individualexpenditure decisions in different contest treatments by varyingthe uncertainty of the outcomes and the number of contestopponents. Contests with probabilistic outcomes showdecreasing over-expenditures and a higher rate of “drop out”.If outcomes become deterministic, expenditures quickly convergencetowards equilibrium predictions and a near to fullparticipation. These results are robust to changes in the numberof opponents. A learning parameter estimation usingthe experience-weighted attraction model suggests that subjectsadopt different learning modes across different conteststructures and helps to explain expenditure patterns deviatingfrom theoretical predictions.Chapter three explores the presence of latent contestant typesby applying the classifier Lasso to two versions of contest experiments,one that keeps the grouping of contestants fixedand one that randomly regroups contestants after each round.Results suggest that there exist three distinct types of players in both contest regimes. The majority of contestants in fixedgroups behaves reciprocal to opponents’ previous choices. Forexperiments in which contestants are regrouped, the share of“reciprocators” is significantly lower. In both cases the reactionof other player types seems to differ from what is expectedfrom a myopic best-reply.Pharmaceutical drug development can be seen as a real-worldexample for lottery contest. Assessing drug candidates’ oddsof success often relies on methods based on historical successrates. However, machine learning offers a more data-drivenapproach to identify promising projects. To evaluate its usefulness,chapter four assesses the performance of several supervisedlearning algorithms that are trained and validatedon a large database of projects. Using a sizeable list of projectcharacteristics as input data, classification via state-of-the-artsupervised learning methods is more accurate compared tomore simplistic methods. The chapter aids stakeholders inthe pharamceutical industry to make more informed decisionsregarding stage-specific project outcomes.Chapter five extends the study of project outcomes by assessingthe relationship between product innovation announcementsof bio-pharmaceutical companies and their stock reactionsusing an event study approach. We hypothesize that financialreturns that follow news on product innovation areshaped by a “probability effect”, that depends on how investorsperceive the product’s likelihood of success, and a“portfolio effect” that depends on the relative importance ofa product within a company’s portfolio. To test for the probabilityeffect, project specific success probabilities are estimatedvia supervised learning methods. The portfolio effect is measuredby the share of the product’s net present value. Marketreactions are found to be higher when assosciated to projectswith high portfolio importance but lower when associated toprojects with high expected success probability.