0

I have plotted the total consumption (g) of a solution by my experimental organisms. However, when I plotted the error bars (s.e.m) a few crossed zero (see image).

(1) is this correct/plausible? (2) How should I interpret this?

thanks in advance.

[![Standard-error bar plot][1]][1]

Tom Oliver
  • 1
  • 1
  • 4
  • Your error bars appear to be symmetrical, so this is entirely plausible. What are these point estimates (histogram is not the right plot for mean / proportion) and how were the "error bars" estimated? – user2974951 Oct 22 '18 at 09:22
  • (1) Do you want to analyse this data using a model or are you only planning on doing descriptive statistics? (2) Is your consumption bound between zero and one (percentage) or can it be any positive number? In any case, I would simply visualize the data using [boxplots](https://en.wikipedia.org/wiki/Box_plot) instead since the arithmetic mean (and hence the standard error of the mean) isn't the best choice for your type of data. – Stefan Oct 22 '18 at 13:37
  • 1
    @Stefan (1) I am only planning to do descriptive statistics. (2) I thought this might be an option, I'll go ahead and create a boxplot. Thanks! – Tom Oliver Oct 23 '18 at 14:06
  • I'm not able to see the image. (Is it there?) But also, I'm having trouble coming up with a toy data set that will exhibit this behavior. (Can anyone offer one?) Are you sure you haven't made a mathematical error? – Sal Mangiafico Oct 25 '19 at 12:48

1 Answers1

1

I assume the quantity you are looking at should never be less than zero. Thus, this indicates that with the limited amount of data you have your model is a bit problematic and predicts that values could be below zero. Potential ways of dealing with this could include transformation of the data for analysis. One obvious option is log-transformation (especially if you never observe an actual zero value). If you then back-transform the numbers in each group after analysis, you get a geometric mean (instead of the arithmetic mean) with CIs/error bars that would not overlap zero.

Björn
  • 21,227
  • 2
  • 26
  • 65
  • Thanks for the help! However, there is actual zero values within this data set. Could you suggest another type of transformation I could use? – Tom Oliver Oct 22 '18 at 09:48
  • Are these true zeros or just below some lowest measurable level? – Björn Oct 22 '18 at 10:19
  • These are true zeros – Tom Oliver Oct 22 '18 at 10:21
  • @TomOliver: [This thread](https://stats.stackexchange.com/questions/1444/how-should-i-transform-non-negative-data-including-zeros) and the linked article by Hyndman discusses a few options. – AkselA Oct 22 '18 at 15:07