When using linear regression, when should you log-transform your data? Many people seem to think that any non-Gaussian, continuous variables should be transformed so that the data “look more normal.” Linear regression does in fact assume the errors are normally distributed, but it is fairly robust to violations of this assumption, and there are no such assumptions regarding the predictor variables. What is often ignored or misunderstood is the impact that variable transformations have on the linearity assumption of regression models, and on coefficient interpretation.
To get a job, you have to get an interview. Who knows what will happen at that point, but securing an interview is at least partly in your own hands. Having just finished shortlisting the applicants to a postdoctoral research post I recently advertised, here are my thoughts on how to get your foot further in the door. None of this is original. It’s just fresh on my mind.
Professors Nancy Krieger (NK) and George Davey Smith (GDS) recently published an editorial in the IJE titled The tale wagged by the DAG: broadening the scope of causal inference and explanation for epidemiology. In it, they argue that causal inference in epidemiology is dominated by an approach characterised by counterfactuals (or potential outcomes) and directed acyclic graphs (DAGs); and that this hegemony is limiting the scope of our field, and preventing us from adopting a more useful, pluralistic view of causality.
Registration of clinical trials, prior to any patient recruitment, is now common. Though trial registrations often omit the important details regarding their proposed analyses (despite advice to the contrary), most trialists seem to agree, at least in principle, that you should transparently describe your plans for clinical trial data before they are collected. Unfortunately, this remains a foreign concept in other areas of clinical and public health research.
I have, for some time, wanted to respond to Sandro Galea’s essay on a Consequentialist Epidemiology, published in 2013 (!) in the American Journal of Epidemiology. It is hard to argue against the importance of consequentialism for academic epidemiology. It is equally hard to dismiss Galea’s concerns about our lack of influence with funders and policy makers. However, my views beyond this diverge from Professor Galea’s, and in the spirit of his “provocation,” I would like to respectfully offer an alternate perspective.
Is it ever ok to conduct an exploratory data analysis, for the purpose of “generating hypotheses”?
As of today I am a Senior Lecturer at the Clinical Research Facility Cork. I am feeling reflective. In the 5 short years following my PhD, I went from the USA, to England, to Brazil, to Ireland. I have gone from Lecturer, to Post-Doc, to Senior Lecturer, in that order. I started in pursuit of a career in public health nutrition, but now work primarily as a biostatistician conducting clinical research. It’s been a strange path, though strange paths in academia are starting to feel awfully normal. Apologies for my narcissism and any name dropping. I hope this is helpful.