I have a dataset of fish species from different sites within a harbour collected over 17 years. The dataset consists of 1,042 sampling transects collected at 6 specific locations where fish were captured and identified to species. There are two sites in the east, two in the west and two in the south which are suspected to have slightly different communities. I want to analyze the community structure and see if it has changed at the sites where artificial reefs were constructed part way through the study.
To analyze this I'm pooling my catches across year and running a PERMANOVA with location (East, West, South) and presence of a reef (Yes, No) and the interaction effect included as fixed effects. What I'm wondering is how I should include the sampling year to account for all of the repeated sampling. My first impulse was to include it as a mixed effect to control for the repeated samples which wouldn't be independent from each other; but then I thought that including it as a factor would allow me to determine its effect and see if there were any interesting interactions. But then I thought that including it as a covariate would be more powerful, as it would preserve the order of the years which would potentially be more powerful.
My question is: If I include Year as a fixed effect versus a random effect versus a covariate what does that mean about my conclusions? My understanding is that if I include it as a fixed effect I can only draw inferences about the years included in my dataset and I can't extrapolate the effect of the reef construction itself; if I include it as a random effect I control for the multiple repeated samples; and I'm unsure what it means if I include it as a covariate as I haven't encountered anyone who's done that.