Collider stratification bias or index event bias This excellent paper covers a very interesting topic: collider stratificaiton bias.
In short, lots of observational studies from medical literature have reported paradoxes such as obesity being protective of progression of osteoarthritis (even though obesity is a risk factor for osteoarthritis) and smoking being protective of rheumatoid arthritis progression (even though smoking is a risk factor for rheumatoid arthritis).
These can be explained by an artefactual inverse association between the risk factor and unmeasured confounders, brought by conditioning on the outcome (OA/RA in these cases).
However the paper does not go into detail of how this can be remediated. Can anyone offer any wisdom or personal experience?
 A: First you need to check whether such claims were due to ecological fallcy. Then this may be due to survivorship bias. Another major reason is the use of cross-sectional rather than prospective designs. In other words, you need to measure the potential risk factor before the condition occurs (even years before), as simultaneous assessment does not really inform on causation.
A: The obesity paradox is not well understood, at least from a statistical perspective, and thus is hotly debated especially in the causal modelling world. 
It should be noted that confounding is a separate issue from collider bias. Confounding is the lack of (counterfactual) control of one or more factors which is causally related to the predictor of interest as well as the outcome of interest. Collider bias is the unnecessary stratification of comparisons by a factor which is a causal result of both the exposure and outcome of interest.
The following paper from Hernan (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3653612/) discusses the strange "stochastic" behavior of relative risk when a prospective cohort study modelled an outcome over time segments. It should be noted that the study of interest, the Women's Health Initiative, was a randomized cohort, so the randomization assumption indicates that women who were recruited were otherwise balanced between HRT and control in terms of the distribution of risk factors (at baseline). 
What was observed was that when increasing durations of follow-up were considered for analyses, the hazard ratio behaved unexpectedly: it tended to taper off, attenuating the relative risk of CVEs for HRT vs control. After one year, the women who were at highest risk of CVEs were most likely to die of them, and women undergoing HRT were elevated moreso. At the next segment of time, the distribution of risk factors in HRT group was less variable and favored healthier CVE outcomes. It very much resembles what population biologists observe among species as survival of the fittest. 
I suspect this is the trend also noted in the obesity paradox: that being, this trend can be so pronounced as to completely reverse hazard ratios (given a long enough duration of follow-up). If you don't die because you are fat, you will eventually die because you are skinny/frail and be outlived still by those who are fatter than you. The bathtub shaped hazard for all-cause mortality among the obese is steeper up front and shallower in the tails than for their healthy-weight counterparts. This is a violation of the proportional hazards assumption.
Hernan suggests no solution to this problem in a general sense exists. I know however one less-than-satisfactory work-around: when reporting risk ratios, age-period-and-cohort (APC) should be considered and modeled in a life table with log-linear models, and then tested for interactions to show whether there is in the cohort, this analogous sense of "genetic drift".
A: Properly considering the causal pathways, such as using direct acyclic graphs (DAGs), can prevent collider stratification and other biases.
This paper in BMC methodology (Reducing bias through directed acyclic graphs, by Shrier and Platt) describes the process in detail. Collider stratification bias occurs when a shared descendent of two variables in the causal map is adjusted for.
