gam_predict_newdata: Testing predictive power for GAMs
In jpwrobinson/funk: James Robinson's functions, attributed where stolen

Function builds GAM and dredges saturated models, fixed to maximum of 3 covariates to prevent overfitting. Then predicts response against new predictor covariates 1. the top-ranked model (AICc > 2 lower than candidate models) OR 2. across a top-ranking model set (AICc < 7 of top-ranked model)

gam_predict_newdata(
  dataset,
  newdataset,
  exp.names,
  indicator,
  family,
  base_k = -1,
  smoother = "cr",
  n.param.max = 3
)

`dataset`	= dataset containing y and all x covariates. should be scaled and centered (mean = 0, sd = 1)
`newdataset`	= generating predictions with 'newdata'. should be scaled + centered (mean = 0, sd = 1)
`exp.names`	= explanatory covariate names, passed as vector of characters
`indicator`	= y variable of interest (character)
`family`	= GAM family distribution, takes 'gaussian' or 'Gamma'
`base_k`	= number of knots for each smoother term. Default = -1, mgcv uses generalized cross-validation
`smoother`	= cubic regression spline method (cr is default)
`n.param.max`	= maximum number of parameters included in each reduced model