On 2016 Jun 03, David C. Norris commented:

The ‘spam’ clinical messaging applications advanced by Schneeweiss will be Medicine’s just deserts for adopting “analytic approaches that embrace the data turmoil”¹ of a routine clinical practice which eschews scientific method. Only scientific practice, capturing in computable form exquisite clinical observation together with detailed clinical reasoning, holds the promise of individualized care within a true learning system.² Whatever obstacles Medicine presents to scientific reform, equally fearsome technical and social-engineering problems bedevil ‘big data’.

These problems arise from Judea Pearl’s “golden rule of causal analysis: No causal claim can be established by a purely statistical method, be it propensity scores, regression, stratification, or any other distribution-based design.”^3(p350) Sound causal inference in non-experimental settings apparently requires graphical methods to support correct reasoning about confounding.⁴ When deconfounding variables are missing—surely typical wherever inference relies on ‘big data’ scavenged from the waste bins of routine clinical practice—then quantitative bias analysis⁵ becomes essential for credibly evaluating the uncertainty engendered by residual confounding. Would the software automation Schneeweiss envisions be grounded in such methods? If so, will “users with little training”¹ adopt methods at which most professionals balk?

[1] Schneeweiss S, 2014

[2] Weed LL, Weed L. Medicine in Denial. Charleston, SC: CreateSpace; 2011.

[3] Pearl, J. Causality: Models, Reasoning and Inference. 2nd edition. New York: Cambridge University Press, 2009.

[4] Howards PP, 2012

[5] Greenland, S. Multiple-Bias Modelling for Analysis of Observational Data (with Discussion). Journal of the Royal Statistical Society: Series A (Statistics in Society) 168, no. 2 (March 2005): 267–306.

On 2016 Jun 03, David C. Norris commented:

The ‘spam’ clinical messaging applications advanced by Schneeweiss will be Medicine’s just deserts for adopting “analytic approaches that embrace the data turmoil”¹ of a routine clinical practice which eschews scientific method. Only scientific practice, capturing in computable form exquisite clinical observation together with detailed clinical reasoning, holds the promise of individualized care within a true learning system.² Whatever obstacles Medicine presents to scientific reform, equally fearsome technical and social-engineering problems bedevil ‘big data’.

These problems arise from Judea Pearl’s “golden rule of causal analysis: No causal claim can be established by a purely statistical method, be it propensity scores, regression, stratification, or any other distribution-based design.”^3(p350) Sound causal inference in non-experimental settings apparently requires graphical methods to support correct reasoning about confounding.⁴ When deconfounding variables are missing—surely typical wherever inference relies on ‘big data’ scavenged from the waste bins of routine clinical practice—then quantitative bias analysis⁵ becomes essential for credibly evaluating the uncertainty engendered by residual confounding. Would the software automation Schneeweiss envisions be grounded in such methods? If so, will “users with little training”¹ adopt methods at which most professionals balk?

[1] Schneeweiss S, 2014

[2] Weed LL, Weed L. Medicine in Denial. Charleston, SC: CreateSpace; 2011.

[3] Pearl, J. Causality: Models, Reasoning and Inference. 2nd edition. New York: Cambridge University Press, 2009.

[4] Howards PP, 2012

[5] Greenland, S. Multiple-Bias Modelling for Analysis of Observational Data (with Discussion). Journal of the Royal Statistical Society: Series A (Statistics in Society) 168, no. 2 (March 2005): 267–306.