Function to identify covariates that are correlated in terms of Spearman's rank

Source: R/DataPreparation.R

Correlation engender problems of identifiability. Correlated parameters in the dataset will be separated in tested models. Test for the Spearman factor for non linear correlation between covariates of all the stations. Complete the test with a visual test if needed.

Param_corr(
  x,
  rm = NULL,
  visual = FALSE,
  thd = 0.7,
  plot = TRUE,
  saveWD = NULL,
  figname = "Covariate_correlation",
  img.size = 12
)

Arguments

x

dataframe of covariates only, on which to test correlation of covariates. This can also be dataset as issued from Prepare_dataset.
rm

vector of column numbers to be removed from the analysis. Default to NULL. If you specify your complete dataset as x, you may define rm=1 to remove your observations from the columns.
visual

logical. Whether to define visually if data are considered correlated or not. See details. Be careful, all previous figures are closed with graphics.off() before running visual analysis.
thd

numeric. Correlation (absolute) value above which to consider that covariates are correlated and should not remain in the same model. This value is necessary when visual = FALSE. See details.
plot

logical. Whether to plot the figure of correlation values between covariates. Set to FALSE if using a MPI cluster.
saveWD

path to directory where to save a jpg figure file. If NULL and plot = TRUE, then figure is shown on screen.
figname

character. The name (w/o extension) of the figure to be saved in saveWD.
img.size

size in cm of the output (square) image. May increase labels size as well.

Details

  • Correlation of covariates is to be tested in the dataset itself. It is not important if covariates are correlated in real life. What affect model fits is the correlation within the dataset. Therefore is this function...

  • Spearman's rank correlation coefficient has been chosen as it allows to test for correlation between categorical and continuous variables. This may look as a non-sense to test for correlation with categorical covariates. In some cases, the order of classes inside a category may have a sense, e.g. continuous variable that has been categorised for any reason. In this function classes are temporary turned into numbers to calculate correlation. This implies that the alphanumeric order of classes has a sense. The Spearman test is also less sensitive to non-gaussian distributions of data

  • In some cases, correlation value is high because of one extreme rank value. Visual verification allows to define for each couple of covariates if the user may consider correlation or not. Figures are grouped by correlation values which accelerates the visual verification.

  • It is suggested to do a visual verification the first time data are processed or to do a complete real exploration of your dataset before running any model. This exploration may suggest to remove or combine covariates before starting the modelling procedure. After that, you may be able to define a thd value for covariate correlation limit.

  • Defining a high thd value may let some correlated covariates to appear in the same model in the following of the procedure. Nevertheless, because the procedure of this package is based on cross-validation, if two variables are correlated they may likely have a low score as they will be as efficient as covariate alone. Thus, if you hesitate between two thd values, choose the higher one, as it will allow more models to pass through the selection procedure.