For a more detailled description of what Francois Massol has just said see page 71 of Zuur et al (2009) Mixded Effects Models and Extensions in Ecology with R. It is all in there.

If you want to use an ANCOVA test, take in count the special use for repeated measures (like time). Look at it: http://www.statisticshell.com/docs/ancova.pdf

Maybe a GEE GLM could be considered to take into account the correlation between observations at time t, t-1 and t+1

if you are having huge data set you can go for ANN method.

Thankyou to everyone.

Just to be clear. I have 8 different sampling times (in the three sites) and the measurements at each sampling are independent (as sites are offshore...).

I don't think that is an huge data set, so maybe the ANN method is not the best, as well as repeated measures.

I want to check the Zuur book and also I want to try to understand the GEE GLM suggestion...

The geeglm function (in the "geepack" package) is functioning like classical glm in R. The additional option is "corstr" in which you have to specify the correlation structure. For additional info, see chap. 12 in the Zuur book.

I have a very naive question about this. Why is it necessary to perform a statistical analysis to see if there is a difference or not ? I mean, the apparatus you use give you the exact temperature, conductivity and so on. The sole error of measurement is the accuracy of the apparatus. There is no experimental error. So if the difference between the sites is superior to the accuracy, it means that the parameter takes significantly different values between your sites. For instance, the accuracy for temperature is more or less 0.1°C in general. So, if you have a difference between your sites superior to 0.1°C, it means that temperature is different between the sites. If it has biological relevance is another matter. As I said, it is only a naive comment. What is your opinion about this ?