If it is posssible -i use a calibration curve

First to talk of reference range, it should be based on the plasma levels of apparantly healthy subjects. Second you should have adequate number of observations say more than 30. Third, you can use log transformation to find out the mean and sd. From the log mean, calculate the antilog of the mean that is expressed it as power of e call it transformed mean. The Coefficient of variation (SD*100/mean) seen based on the log values to be used for finding out the SD for antilog. Say, the CV is 20% in log values then your SD in the case of transformed mean should be CV*mean/100.  I hope I have answered your question adequately. 

Ramnath is on the right way:

The calculation as Mean±1.96xSD should give the range of values covering about 95% of the values that one may expect. This means: the reference range is calculated so that from really many many measured values about 95% of them will fall within this range, 2.5% will be lower than the lower limit, and 2.5% will be larger than the upper limit.

The calculation is based on a model about the distribution of such values, and this model is known as the "normal distribution" or Gaussian distribution. This model may or may not be sensible for a particular case. One needs to check this. If you have a lot of data (some 100s or better 1000s of values) then you can for instance plot a histrogram and compare the shape to the shape of a normal distribution (with mean and variance calculated from the data). If you have less data you can use a normal-quantile-quantile plot to check if the Gaussian distribution model is adequate (see http://en.wikipedia.org/wiki/Q%E2%80%93Q_plot).

In your case the Gaussian model is likely inadequate, because concentrations of biological molecules usually do not follow a Gaussian distribution. Their distribution is often right-skewed.  Examples:

http://www.thelancet.com/cms/attachment/2001070221/2004074948/gr1.jpg

http://web.med.unsw.edu.au/QMP/QMPTut2_2012/Skew.gif

http://ehp.niehs.nih.gov/wp-content/uploads/2013/10/ehp.1206115.g001.png

http://www.itrcweb.org/gsmc-1/Content/Resources/Images/5-1%20GLHistogram%20nn%20skewed.png

Sometimes the logarithms of such right.skewed values have a distribution that is approximately Gaussian. So you can calculate the reference range for the logarithms. The (lower and upper) limits of this range can be simply antiloged again to get the range on the original scale. This is what Ramnath proposed.

Say your values are called "y" (y1, y2, y3, ...)

Then first calculate the logs: y' = log(y)

Now check again if the distribution of the y' values is approximately gaussian (histogram or QQ-plot).

If this is the case, you calculate the mean and the SD of these y' values.

The reference range for y' is then mean(y') ±1.96xSD(y'):

the lower limit is L' = mean(y') -1.96xSD(y'):

the upper limit is U' = mean(y') +1.96xSD(y'):

The range for y is obtained by antiloging these limits:

L = exp(L')

U = exp(U')

This is it.

If the distribution of y' is not approximately gaussian, then you need to claculate the appropriate quantiles directly from the data you have. This will give you possibly quite bad estimates for the reference range when you have only few values. If you have some 100s or 1000s of values, this method works good and always - independent of the actual distribution of the values.

The lower limit of the range is simply the 0.025-quantile of the values,

the upper limit of the range is simply the 0.975-quantile of the values.

Done.

You can use a statistics software (even Excel will do!) to get these quantiles. You can alternatively sort all the values and use the 2.5%-smallest and 97.5%-smallest value as the limits (these are the searched quantiles). Example: if you have 1000 values in a sorted list (from small to large), the 0.025*1000 = 25th value will be the value of the lower limit and the (1000-25)=975th value in this list will be the upper limit.

Thank you for your help, i just have 80 controls. I will use a box cox transformation and the standard method to estimate the normal  range and to eliminate the outliers from my data. I hope that i am doing it right. 

Yes, the box-cox procedure will determine the transformation (from a family of transformations) that will give the closest approximation to the Gaussian distribution of the transformed values. You should check their distribution anyway (QQ-plot!). If ok, then you can determine the limits of the transformed values with mean±1.96xSD. To get the limits on the original scale you need to undo the box-cox transformation on these limits.

Thank you Joshen for your answer.

When the distribution is not Gaussian, is not ok to use Median and SD at 95% confidence rather than Mean and SD...is that correct?