Nova Acta Leopoldina Band 110 Nummer 377

to a wide range of Applications: “customer acquisition and retention for financial services, patient care improvement for the healthcare industry, crime prevention for the public sector and ideal store location for retailers and manufacturers.” Clearly, the power of statistical analysis has not escaped the attention of some industrial sectors. As the press release reveals, they consider the major advantage statistics can help re- alize not only through the evaluation of what works where, but also through the discovery of causal effects: bottlenecks and enhancers of better, healthier, more environmentally friendly, more sociable or cheaper outcomes. To produce these analyses a great number of data analysts will be needed. If any lessons have been learned from the HRT saga however, it is that complex models with many predictors and smaller residual error by no means guarantee a better or more reliable causal result. In fact, it can even launch one on a path in the opposite direction of progress. Careful thinking of scientists with a deep understanding of causal structure is what is needed to arrive at well justified/understood robust conclusions. 8. In Conclusion Monitoring of results is essential to good governance and reliability of products or services. Understanding associations within one’s process is even better. It suggests where to look when things go wrong and allows for more accurate prediction and hence better planning. Some- times, masses of data become a mountain in which the needle is to be found. Any pattern is more likely to show up if there are plenty of paths to walk on. Adjusting for multiplicity is one key necessity to avoid naïve conclusions when reproducibility and prediction is at stake. But there is more to gain than by prediction for stationary systems. Causal inference aims not merely to predict or explain what happens where, but to discover causes of improved or worse outcome. Such knowledge forms the key to unlocking interventions that may improve the sys- tem. Not surprisingly the art of causal inference has been at the center of scientific investigation for centuries. With the digital revolution statistical analysis under such fancy names as data mining, business analytics, and the like is seen as a gateway to greater wealth. Massive in- vestments are being directed that way. While it seems wise to invest in such promising tool, it is equally unwise to suggest causal findings are easy and quick to reap. The complication with causal inference lies in the fact that new tools are concerned with causal parameters ex- pressed in terms of latent variables and rely on untestable assumptions cast in those terms. Any statistical causal analysis therefore requires deeper thinking and more information than a standard (association) analysis would require. Should an army of statistical data analysts be hired by large companies trading in ‘business intelligence’ then at least a sizeable regiment of highly and specifically trained statisticians and co-scientists will be needed to guide the thinking and the implementation process. Blind modeling is doomed to fail, as is the adagio ‘more is (automatically) better’ when it comes to correcting an observed association for covariates. The newly developed causal methods as outlined above form a great step forward in methodology. They require more sophisticated programming, but this is nothing compared to the sophistication of the causal modeling and inference per se. We will need people too who can bridge the gap between the few specialist forerunners and the large numbers of programmers and scientists. Without a large rollout in this sense, we are bound to see many more stork epidemiologists fly and many more dollars spent on (HRT) treatments that are actually harmful rather than helping. Nova Acta Leopoldina NF 110, Nr. 377, 47–64 (2011) Els Goetghebeur 62