0
$\begingroup$

Is there any biological basis for excluding outliers in a dataset of blood cytokine levels (e.g. values outside 3 standard deviations from the mean for each cytokine) as measured by multiplex technology? How may I distinguish values reflecting the potential heterogeneity of my sample cohort from real outliers (e.g. measurement or coding errors)?

$\endgroup$
  • 1
    $\begingroup$ Could you please give the biological context of your study, explain the analysis you're undertaking, & explain the distinction you make between "real value" & "real outlier". $\endgroup$ – Scortchi - Reinstate Monica Jan 2 '19 at 10:46
  • $\begingroup$ @Scortchi There are two sources of uncertainty for problems of this type. The first is the patient population values, which one can imagine to be "true" values, and the second is measurement system errors, which can be thought of as values that distort the "true" values. $\endgroup$ – Carl Jan 2 '19 at 19:36
  • $\begingroup$ There is no reason to assume that the mean value is an optimal measurement of data location, nor that standard deviations are a useful measure of dispersion of data. Outliers are more typically defined by nonparametric criteria; based on quartile distances from the median. Identification of how the data are distributed would be a first step. There is software that is helpful for this problem, but comments are not suited for long answers. $\endgroup$ – Carl Jan 2 '19 at 19:43
  • 1
    $\begingroup$ Now that you have clarified the question, it seems better suited for a biology site than a statistics one. Perhaps some here are experts on cytokines, but you're more likely to run into experts on that somewhere else. If you know how the cytokine values ought to be distributed, then that could be used. $\endgroup$ – Peter Flom Jan 3 '19 at 9:57
  • $\begingroup$ Yes, I don't think the distribution of cytokine levels over individuals (healthy or infected?) is general knowledge readers can be expected to have. The same for the properties of multiplex technologies as a measuring system. For what it's worth, I'd guess that one or two infected individuals in a larger sample of healthy ones might well be flagged as outliers by the method you propose - which wouldn't of course imply a measurement error. $\endgroup$ – Scortchi - Reinstate Monica Jan 3 '19 at 10:12