The ANZECC Guidelines provide an overview of managing and analysing monitoring data in Management Guidelines Chapter 6 and this Chapter also provides guidance on the use of common statistical methods for the analysis of water quality data.
"The information is at an introductory level that will help the monitoring team identify suitable methods of analysis and assist with the interpretation of results. Much of the technical detail and the more advanced statistical procedures have been relegated to Appendix 5 where, in most cases, they are illustrated with the help of worked examples that demonstrate options available, methods of implementation and interpretation of results. The information provided in this chapter is not exhaustive; complex studies may require a greater level of statistical sophistication than is presented here, and for these studies the monitoring team is advised to consult a professional statistician."
Data analysis should be viewed as an integral component of the water quality management process.
Make sure the analysis fits the data you are collecting
"Those who use statistical methods in environmental studies tend to fall into one of two distinct groups regarding the assumptions underlying the methods they use. Either they ignore the fact that there are assumptions at all or they are paranoid about them and rely on nonparametric methods. What is argued here is that
the assumptions of the method should be understood at the time it is chosen
the likelihood and consequences of violation should be assessed (with the aid of data from preliminary sampling ), and then
use of the method should proceed with awareness of the risks and the possible remedies.
.... Non parametric methods are rarely necessary." (Green, p.44)
Standard methods are regarded as being quite robust but if in doubt the first thing to do is plot the data to assess skewness and variability. Common situations that may require nonparametric methods are for samples below the detection limit or where the data is known to be one sided; these should be known before the (full) study begins and can be accounted for. A more critical issue arises from "pseudoreplication" when inferential statistics are not appropriate and simple comparison of means may be all that is required and justified.
Green (1979, p43ff) discusses in detail the kind of transformations that may be useful but you need to be careful of spurious correlation where a transformation makes the dependent variable a function of the independent variable (often with a spectacular increase in correlation!). Green suggests a decision sequence for considering the justification for transforming data:
Are there serious violations of [statistical] assumptions?
Is the assumption of homogeneous error variances a tenable one for the data in question?
If H0: "homogeneity of variance" is rejected, will a transformation  any transformation  reduce the heterogeneity? We must ask, what can a transformation do?
Transform the data according to whichever model is chosen and proceed with the statistical analysis.
Given that "the word 'transformation' should be interpreted broadly to include more than, say, taking the logarithm or the square root of each value " (Green, 1979, p47) any manipulation of the data should be considered for its implications for the interpretation of the data. It is important also not to divorce the analysis from the problem. As StewartOaten et al. (1986) note, "...all statistical procedures require assumptions, and these assumptions must be justified by reference both to the data (by plots and formal tests) and to a prior knowledge of the physical and biological system generating the observations."

