An apparently unconventional significance indicator The following is a statistic used by a water quality consultancy in assessing contaminant concentrations over time:
quality indicator $= \exp(\text{mean}(\ln C_i)+2\,\text{SD}(\ln C_i))$
where $C_i$ is $i$th concentration and $\text{SD}$ is standard deviation
The consultants don't provide the formulation: the reader is left to extract it from their description of the indicator as "the mean plus two standard deviations calculated using log transformed data collected during the baseline period".
The indicator is calculated over a ‘baseline’ period and then subsequent increases in concentration are gauged against this statistic to assess whether a significant change has occurred.
As far as I can tell this doesn’t correspond to a recognised/conventional statistic providing a measure of deviation from a mean (e.g. in the sense of $\text{mean}(\ln C_i)+2\,\text{SD}(\ln C_i)$. 
As I read it the quality indicator $= \exp(\text{mean}(C_i))\exp(2\, \text{SD}(\ln C_i))$.
My question is, does it matter?
I should add that the baseline data is close to Normal and applying the log transform results in a slightly more ‘distorted’ distribution. 
So it is unclear why they don't simply use $\text{mean}(C_i)+2\,\text{SD}(C_i)$. 
 A: If we assume that concentrations are reasonably close to lognormally distributed, so the log is reasonably close to normally distributed, mean + 2 sd is approximately an upper end of a 95% prediction interval (for a future observation) or approximately the limit of a 97.5% one tailed prediction interval.
In particular, if this is being used as a one-tailed test (an observation outside the prediction limit leads us to reject the hypothesis that the new observation comes from the same distribution as that of the data used to generate the prediction limit, it's (notionally) an approximate 2.5% test. 
[We can consider this either as approximating 1.96 by 2 in a large sample test or as approximating the corresponding t-critical value; in both cases assuming the sample size is so large that the width of the prediction interval is almost entirely due to the sample standard deviation. Various assumptions are required for this to hold, but it's doubtful several of them will]
However, all that aside, it sounds like there's a sequence of such tests to be performed, which changes things from that interpretation; arguably we get into sequential testing.
