FAQ.Rmd

---
title: "FSSgam FAQs"
author: "Rebecca Fisher"
date: "04/05/2020"
output: html_document
---

```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = TRUE)
```

## How do you choose which variables to include in **test.fit** (e.g. Model1)

The **generate.model.set** function must be supplied **test.fit**, which is a gam(m) model fitted to the desired response data, with the appropriate random effects structure (unless supplied as an **re** agument to **gam**, see below) and **family**. This **test.fit** is simply updated with new model formula by the **fit.model.set** function, so anything supplied in the formula of the **test.fit** is not actually used in the full subsets fitting procedure - rather just the underlying structure of the model set. This means it doesn't actually matter what variables are included in **test.fit** as predictors, they must just be present in the data.frame supplied as **use.dat**. Fitting a sensible **test.fit** can be helpful to examine the assumptions of your model fits before you even get started, as this can save a lot of time in the long run compared to waiting for your "top model" or model set to examine the validity of the underlying model assumptions. I highly recommending fitting a reasonable **test.fit** and exploring this thoroughly to be confident the model has fit as you intend.

There are two examples of generating **test.fit** in the supporting material. The first is on line 102  <https://github.com/beckyfisher/FSSgam/blob/master/case_study1_reef_fish.R>). This example shows a **test.fit** calling the gam function in mgcv directly, which is necessary to use the **tw()** distribution in this example. The main things that need to be specified are the response variable (here called simply **response** - but this would be whatever variable you want to build the model set to predict), family - in this case **tw()**, and the data to use. Here we have fit one of the predictors that we think will be important (although that was not strictly necessary), as well as the two predictors that we are going to use in our null model (again, also not necessary, but **test.fit** is a good place to make sure your gam models are fitting as intended). Note our null model in this example consists of a random effect of **site** specified through **s(site, bs="re")**. Note that specifying the null terms in **test.fit** does not automatically mean the null terms are carried into the full subsets gam. They must also be specified as **null.terms** (see line 111 of the case_study1_reef_fish.R example).

At line 48 of <https://github.com/beckyfisher/FSSgam/blob/master/extra_examples.R> you can see an alternative specification of **test.fit** using the **uGamm** function from the **MuMIn**. **uGamm** is wrapper function that allows **gamm4** model fits to be updateable and therefore usable by **FSSgam**. Here we are fitting a binomial model using the **cbind** format for **successes** and **failures** which must be as labelled in the **use.dat** data.frame. With uGamm the random effects are specified outside the model formula, meaning they will not be updated by **FSSgam** and must be specified exactly in the **test.fit**. The random effect terms will not need to be specified explicitly as **null.terms**, but they will need to appear in **use.dat**.

## What exactly is the r2.vals metric, and how is it derived.

Some information on the R-squuared values reported by **FSSgam** can be found in the help file, type **?fit.model.set** into the R console. The default value supplied is an approximation based on the R-square value calculated for a model between the fitted values and the observed data. This approximation was used because R.sq is not always provided by the underlying **gam(m)** functions used, and may cause errors. If you set the argument **r2.type=”r2”** **FSSgam** will return the R-square as calculated by the underlying gam function, which in most cases is what should be reported. If you want an **R.sq** that represents that left over after random effects are removed, you can also set **report.unique.r2=TRUE** which will subtract the null model **R2** from each of the fitted models in the set. The **report.unique.r2** as defined can sometimes yield negative R.sq values, so may or may not be useful for reporting. There are more formal methods for calculating the R.sq for the fixed component of models as these cannot be easily generalised across all the model types handled by **FSSgam** this is not currently implemented in the package. Subtracting the null model R.sq is a simplification, but potentially still informative.

## Do I need to include a Null intercept only model in **null.terms**?

FSSgam automatically includes a null, intercept only model in the model set. This model will add any additional terms specified as **null.terms** to the model formula as well. The null model is useful to assess if any of the included predictors have explanatory value, thus if your null model is in the top set (say within 2AICc), it may be inferred that none of the predictors are useful beyond simply a mean value of the response.

## When I use lm4=TRUE, so I need to include my random effect in **null.terms**?

The FSSgam package only updates the model formula, everything else in the model remains the same and becomes part of the null model. When a gamm4 is fit using lme4 = TRUE, this is a separate argument to gamm4 ("random") - and so the specified random effect will be retained as part of the null model. SO the answer is NO, in this case you do not need to add your random effect to **null.terms**. The only reason random effects sometimes need to be specified in **null.terms** is when the formulation bs='re' is used to specify them in gam from mgcv - because in that case they are part of the model formula and will be overwritten by the set of model formula as they are fit by FSSgam.