Our goal in this blog is to promote a Triple Approach to the Earth Citizens' Health. This Triple Approach consists of : 1) Quaternary Prevention (Preventing Unnecessary Interventions in Clinical Medicine and Public Health ), 2) Prevention of Diseases and Treatment of Patients by Improving Living Conditions and Social Prescribing, 3) Preconception Prevention of Diseases by Preconception Planning.
Sunday, March 24, 2019
John P. A. Ioannidis: What Have We (Not) Learnt from Millions of Scientific Papers with P Values?
A paper published in PLOS ONE recently fully supports these views: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0212302. It shows that (amongst many other things) the ‘idealistic’ probability of getting a ‘true’ result less extreme than a null hypothesis is approximately 1-P. It is idealistic because it assumes impeccable methodology (eg with pre-registration of the study). Failing this one has to assess the probability that it was impeccable (no publication bias etc) and use this to estimate the ‘realistic’ probability of replication, which may be much lower.
The next step is to use the realistic probability distribution of replication in combination with the prior probability of a scientific hypothesis to estimate a posterior probability of that hypothesis. This can only reliably show that a hypothesis is improbable, so that the remaining hypothesis (or hypotheses) or a hypothesis not yet considered become more probable (as taught by Karl Popper).
All this is based on the recognition that a Bayesian prior probability distribution can be regarded as the result of combining a uniform prior distribution with a Bayesian (or ‘subjective’) likelihood distribution. This means that the Bayesian posterior probability could be regarded as combining two statistically independent likelihood distributions with a uniform prior. It is all explained with detailed examples in the above PLOS ONE paper.
Huw Llewelyn Aberystwyth University hul2@aber.ac.uk
1 comment:
A paper published in PLOS ONE recently fully supports these views: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0212302. It shows that (amongst many other things) the ‘idealistic’ probability of getting a ‘true’ result less extreme than a null hypothesis is approximately 1-P. It is idealistic because it assumes impeccable methodology (eg with pre-registration of the study). Failing this one has to assess the probability that it was impeccable (no publication bias etc) and use this to estimate the ‘realistic’ probability of replication, which may be much lower.
The next step is to use the realistic probability distribution of replication in combination with the prior probability of a scientific hypothesis to estimate a posterior probability of that hypothesis. This can only reliably show that a hypothesis is improbable, so that the remaining hypothesis (or hypotheses) or a hypothesis not yet considered become more probable (as taught by Karl Popper).
All this is based on the recognition that a Bayesian prior probability distribution can be regarded as the result of combining a uniform prior distribution with a Bayesian (or ‘subjective’) likelihood distribution. This means that the Bayesian posterior probability could be regarded as combining two statistically independent likelihood distributions with a uniform prior. It is all explained with detailed examples in the above PLOS ONE paper.
Huw Llewelyn
Aberystwyth University
hul2@aber.ac.uk
Post a Comment