Object types #

There are two fundamental types of objects created by the model API. The objects are distinguished by their prefix – either hat or pred. hat-type objects involve an evaluation of the mean over time, but holding any other covariates at their mean (or at zero, in the case of indicators / dummy variables). pred-type objects use the known values of each covariate, at each site and timestep, while evaluating the mean. This critical difference between object types produces outputs that we have colloquially referred to as “smooth” (hat) vs. “bumpy” (pred) results. The hat results are smooth because the only terms in play are the intercept and time slope. The pred results can be bumpy if: 1) the covariates included are influential, and 2) they vary from year-to-year.

Site, stratum, and park level inference #

At the site-scale, we make predictions – whether hat or pred – only to sites that were sampled. At a stratum scale, things work a bit differently.

Modifying the default behavior of pred-type means #

Recall that pred-type means include all of the covariates used in a model. There are times when some of our covariates might produce results that are less-than-interesting from an ecological perspective. This is often the case for a nuisance variable. Let’s say we have two botanists, one of which was less experienced than the other, and tended to identify fewer species on average than the more experienced botanist. In a model of species richness, the default behavior of pred-type objects would produce means that “jump” up at sites / times in which the more experienced botanist was making observations, and down otherwise. These changes in the mean aren’t being caused by ecological factors, they’re entirely a result of the observer. We can change the way pred-type means are evaluated using the trend conditions block, described here. This block of the analysis config file can also be used to “fix” continuous variables at a specific value.