Structuring a linear mixed model in R with nesting My ecological question is: "What are the trends in percent coral cover by island and depth across the state of Hawaii from 1999 to 2012?"  
I am trying to analyze this hierarchical data set using R with 10 transects at each depth, 2 depths per site, and site nested in island.
Data structure:
Fixed effects:
 Island: Hawaii, Maui, Molokai, Kahoolawe, Oahu, Kauai.
 DepthCat: S = Shallow, D = Deep.
 WYear: 0-13. It was suggested that I use this factor as a covariate for years.

Random effects:
 Site: 34 sites across the 6 islands with 2 depths per site.
 Transect: 10 permanent transects at each depth.
 Year: 1999 – 2012 (14 years)

Dependent variable: PercentCover

Currently, I am using the lmer function in the lmerTest package and this is the model that I've constructed.
fit1 <- lmer(PercentCover ~ WYear*Island*DepthCat +
             (1+WYear|Island/Site/DepthCat/Transect) + (1|Year), data=Benthic)

Unfortunately, the data are spotty (i.e., missing data in multiple years for a number of sites) so the model returns [1] "Asymptotic covariance matrix A is not positive!", even using arcsin transformed data. I can still run the summary statistics to get results, but I don't feel comfortable with the error message. Perhaps I have not structured the model correctly in terms of organizing the nested factors, but the number of observations for each of the levels in the summary stats seems correct. I tried different and simpler iterations of the model such as:
 fit1 <- lmer(PercentCover ~ WYear + Island + DepthCat + (1+WYear|Transect/Site) + 
              (1|Year), data=Benthic)

which works, but doesn't give me the interaction information and returns a larger AIC suggesting that the model does not fit the data as well.
To deal with all of the missing data, I tried another approach by using the regression slope of percent cover over time as the dependent variable for each site X depth combination.
Data structure:

Fixed effects:
 Island: Hawaii, Maui, Molokai, Kahoolawe, Oahu, Kauai.
 DepthCat: S = Shallow, D = Deep.

Random effects:
 Site: 34 sites across the 6 islands with 2 depths per site.
 Transect: 10 permanent transects at each depth.

Dependent variable: Trend

I used the following model, but the summary results did not make much sense, even after transforming the data.
 fit1<-lmer(Trend ~ Island*DepthCat + (1| Island/Site/DepthCat/Transect), data=Benthic)

Any suggestions on improving my analytical approach would be appreciated.
 A: The fixed term WYear*Island*DepthCat allows for two-way interactions among island and depth category in the intercept and slope over time.  Fine so far, but the key thing here is that you can't also include these interactions as random effects, which your current specification does. So you want Site and Transect as random effects, but you need the interactions with the higher-level grouping variables in order to make sure they're uniquely identified (e.g., you're not confusing Site 1 on one island with Site 1 on another island).  Probably the easiest way to specify this is as:
lmer(PercentCover ~ WYear*Island*DepthCat + 
      (1+WYear|Island:Site) + (1+WYear|Island:Site:DepthCat:Transect)+
      (1|Year), data=Benthic)

This will only work if you have multiple observations per transect per year: otherwise the middle random effect term is confounded with the residual variation, and you should just leave it out of the model.  On the other hand, if you have
multiple samples per site and (1) you don't have interesting covariates that vary
within transects (2) you're not explicitly interested in within-transect variation (3) you have approximately the same amount of information per transect, you're probably better off aggregating them anyway (see Murtaugh 2007 Ecology).
