forest | A Virtual Reality forest implemented with Three.js

Nice catch.

Although undocumented, this is due to the bootstrap sampling taking place by default in a Random Forest model (see my answer in Why is Random Forest with a single tree much better than a Decision Tree classifier? for more on the RF algorithm details and its difference from a mere "bunch" of decision trees).

Let's see an example with the iris data:

If you have a model like this:

What you can use is something like:

There's nothing in xarray that interprets raster data as geometries or point collections, or to calculate the centroid of such features.

Your approach is clever - you could certainly use weighted averaging to get a simple centroid, e.g.:

coefplot automatically excludes coefficients that are flagged as "omitted" or as "base levels"---as in the case of your coefficients. To include all coefficients in the plot, you should specify the "omitted" and "baselevels" options. So it would look like something along these lines:

note: although your question is about the lmer() function, this answer also applies to lm() and other R functions that fit linear models.

The way that coefficient estimates from linear models in R are presented can be confusing. To understand what's going on, you need to understand how R fits linear models when the predictor is a factor variable.

Before we look at factor variables, let's look at the more straightforward situation where the predictor is continuous. In your example dataset, one of the predictors is wind speed (continuous variable). The estimated coefficient is about -0.35. It's easy to interpret this: averaged across the other predictors, for every increase of 1 km/h in wind speed, your response value is predicted to decrease by 0.35.

But what about if the predictor is a factor? A categorical variable cannot increase or decrease by 1. Instead it can take several discrete values. So what the lmer() or lm() function does by default is automatically code your factor variable as a set of so-called "dummy variables." Dummy variables are binary (they can take values of 0 or 1). If the factor variable has n levels, you need n-1 dummy variables to encode it. The reference level or control group acts like an intercept.

In the case of your habitat variable, there are only 2 levels so you have only 1 dummy variable which will be 0 if habitat is not Forest and 1 if it is Forest. Now we can interpret the coefficient estimate of -68.8: the average value of your response is expected to be 68.8 less in forest habitat relative to the reference level of grassland habitat. You don't need a second dummy variable for grassland because you only need to estimate the one coefficient to compare the two habitats.

If you had a third habitat, let's say wetland, there would be a second dummy variable that would be 0 if not wetland and 1 if wetland. The coefficient estimate there would be the expected difference between the value of the response variable in wetland habitat compared to grassland habitat. Grassland will be the reference level for all the coefficients.

Now to directly address your question of why habitatForest is the coefficient name.

Because by default no reference level or control group is specified, the first one in the factor level ordering becomes the reference level to which all other levels are compared. Then the coefficients are named by appending the variable's name to the name of the level being compared to the reference level. Your factor is ordered with grassland first and forest second. So the coefficient is the effect of the habitat being forest habitat, compared to the reference level, which is grassland in this case. If you switched the habitat factor level ordering, Forest would be the reference level and you would get habitatGrassland as the coefficient instead. (Note that default factor level ordering is alphabetical, so without specifically ordering the factor levels as you seem to have done, Forest would be the reference level by default).

Incidentally, the two links you give in your question (guides to mixed models from Phillip Alday and Tufts) do in fact have the same kind of output as you are getting. For example in Alday's tutorial, the factor recipe has 3 levels: A, B, and C. There are two coefficients in the fixed effects summary, recipeB and recipeC, just as you would expect from dummy coding using A as reference level. You may be confusing the fixed effects summary with the ANOVA table presented elsewhere in his post. The ANOVA table does only have a single line for recipe which gives you the ratio of variance due to recipe (across all its levels) and the total variance. So that would only be one ratio regardless of how many levels recipe has.

This is not the place for a full discussion of contrast coding in linear models in R. The dummy coding (which you may also see called one-hot encoding) I described here is just one way to do it. These resources may be helpful:

• Documentation for the base R function contrasts()
• description of different types of categorical variable coding in R from UCLA IDRE Stats
• Marissa Barlaz' tutorial on R contrast coding

You use variable "ssetTerm" on second step, it's misspelled.

Variable "setTerm" would be undefined, so it's not a function.

You could specify both site and datetime in the on argument of the rolling join with nearest option:

Looks pretty straightforward to translate it directly: