Biomarkers of exposure typically follow a skewed frequency distribution typified by a lognormal one. It is common to log-transform such measurements in analysis and to plan studies (as we do in our calculators) using normally distributed exposures that are derived from log-transformation of the raw biomarker data. An alternative approach is to use untransformed biomarkers in their original units. Which approach do you think is the best practice in epidemiology studies that employ biomarkers to estimate for each participant their own exposure (individual-based exposure assessment, as is the focus of our calculators)? (please explain)

Answer Explanations

• Log-transform to ensure normality

  Expert 3
   Environmental exposures are typically log-normally distributed, i.e., high exposures are rare, and low exposures are common. Hence, it is expected that biomarkers will also follow a log-normal distribution. However, log-transforming data is not always a choice—it depends on the type of analysis being used. For example, a simple OLS regression analysis assumes that the data is normally distributed. If the data is log-normally distributed, the results of the regression analysis will be incorrect unless the data is log-transformed. In contrast, logistic regression does not assume normally distributed data, so log transformation is not necessary. 
• Present results with and without log-transformation of biomarkers

  Expert 7
  Examine frequency distribution of a biomarker and then decide on approach.
• Present results with and without log-transformation of biomarkers

  Expert 1
  I think providing different options is acceptable and provides scientists with alternatives. Some may wish to use log-transformations, others may not, if both are available that would be optimal.
• Present results with and without log-transformation of biomarkers

  Expert 9
  I would prefer to first examine the frequency distribution of the biomarker and then decide what to do based on that.  If it exhibits a roughly log-normal distribution, then proceed with that transformation.  But depending on your study population, it might be something else.  For example, it might be sufficiently normal and not require transformation at all.  It might be something quite different, like bimodal,  For odd distributions, I would consider non-parametrics, ranking/categorizing, or other approaches.    
  If exposure data are transformed, I like to be transparent and present results with and without the transformation so that the reader can form their own interpretations and appreciate the impact of the transformation.   
• Present results with and without log-transformation of biomarkers

  Expert 8
  In general, I'm very much in favor of reporting multiple types of estimates, e.g. by different transformations vs not, different units/categories, different levels of adjustment, other analytical approaches, etc. Although this can make tables large (can be presented in supplemental files) and interpretation sometimes challenging, it facilitates comparison between studies and meta-analyses of findings. 
  
  More specifically:  log-transformation may result in a more normal distribution and therefore meeting assumptions of many statistical tests. I would present these as main results. But adding the untransformed values and estimates will help in interpreting effect sizes by different audiences and help guide e.g policy change. I would see analyses of untransformed values also as a sort of sensitivity analyses, if results are greatly different, one might want to check the underlying data, models and assumptions again  (e.g. maybe logtransformation obscured non-linear associations). 
• Present results with and without log-transformation of biomarkers

  Expert 5
  I would not limit myself to log-transformations only. We would typically evaluate different transformations (i.e. sq root, x2 etc.) to normalize the residuals in the analysis (note, we are generally looking for multivariate normal distribution).  In addition, we almost always analyze the ranked data (non-parametric) and/or ordinal categories and compare these results.  Depending on the analyses, we may only publish some of the results depending on the Journal.  I like to include these other analyses in appendices if possible, so that readers/reviewers can evaluate the assumptions/decisions made.       
• Present results with and without log-transformation of biomarkers

  Expert 4
  I don't think there exists a single best solution so I suggest to present both results.
• Log-transform to ensure normality

  Expert 2
  In epidemiology studies that use biomarkers for individual-based exposure assessment, log-transforming biomarker data is generally the recommended practice, especially when the data follow a skewed distribution, as is common with exposure biomarkers. Log transformation normalizes the distribution, making it more symmetrical and suitable for parametric statistical models that assume normality. It also stabilizes variance across different exposure levels, improving the precision and consistency of effect estimates. Additionally, log-transformed data can be seen as a more precise version of ranks, preserving the order of data while providing detailed information about the magnitude of differences between values. This makes the interpretation of associations more meaningful in public health contexts. Furthermore, log-transformed data enhances model performance, as statistical methods like linear regression perform better with normally distributed data. While using original units or ranks with non-parametric methods may be considered, it can result in skewed residuals and reduced statistical power. Therefore, I prefer log transformation for yielding more accurate, interpretable, and statistically robust results, though the interpretation of the results should always take into account the original data distribution and the study context.