How to analyse a set of data without normality and homoscedasticity?


I'm a postgraduate student in Biology. From some time on, I wanted to publish a paper regarding the affectance of temperature to the shell length of a species of marine gastropod. To make things harder, this gastropod is exctinct, so I've got the fossils of the shells from 26 localities. Those localities can be from two different interglacial periods (one cooler than the other), but we don't know which fossil deposits are from each one (only 4 have absolute datation). From some localities I have as many as 250 samples, while in others I have just 4.

The thing is, larger shells correspond to the cooler one, while smaller shells are supposed to be from the warmer one. That's great, but there's some sort of a gradient, since we rarely find big shells in "warm" localities (but we still do). I wanted a test to make groups and, if possible, separate the 26 localities in 2 (or 3) groups, meaning temperature. For this purpose, I tried to perform an ANOVA with one factor (length of the shell) at 26 levels (the localities). Some of them have +200 replicas, whlie others have 4.

But when I compared them (multiple range test), 3 groups appeared, more or less defined. The problem is that I later checked for both normality and homoscedasticity, and it appears that none of the tests was successful. So, I think I should perform an analysis that is able to overcome both the lack of normality and the heteroscedasticity. A teacher of mine did mention somthing like a "permutation test", but wouldn't help me further.

Do any of you know how to do it, or have any alternatives for this set of data? It is really giving me headache... :p

Just for the record, I dislike Kruskal-Wallis.

Thank you very much in advanced!

P.S. If you are wondering why I would be happy if 3 groups appeared, it's because we have 2 populations of very small shells, that could represent the peak of the interglacial, when the temperature was the warmest.

P.P.S. If you would like me to post (or send you) the data in case you want to try stuff on it I'll be very happy to forward it to you.