Vikki Lowe what does the missingness look like? E.g., 5% for 1 variable, 20% for 10? Are the missing values at random or is there a pattern? E.g., all samples with a depth > 30 m are missing variable X. More detail about the missingness of your data will be helpful.

One potential solution: if the missingness is small and at random, then you can use predictive mean matching to replace the NAs with realistic values (https://stats.idre.ucla.edu/r/faq/how-do-i-perform-multiple-imputation-using-predictive-mean-matching-in-r/).

Typically you have some monotone missing values or a series of values that are missing caused as soon as a measurement units fails. Thus you definitely look beforehand the reason of missingness. You may also visualise them to learn about the missing value structure, e.g. using R-package VIM and its missing values plots.

Missing values are due to ocean sites not having the full range of water column density. For example, sites further south will have higher water densities ranging from 27.4 to 29.4, while sites further north will have a range from 26.2 to 27.5. If I am using the depth of water density 27.6 as a variable, there will be sites that do not have this density, and so the value is missing. I still need to be able to assess the importance of these density ranges. If I am to assess the importance of a range of density depths in an RDA, I need to deal with these missing values in order to run the forward selection for parsimony. That's where I am having trouble. Is there a package or work-around that allows for omitting NA's from the dataframe when forward selecting so that ONLY that NA is omitted, not the whole sample/site?

You may constitute an imputation strategy to replace missing values as a preprocessing for the forward selection procedure. For this, you have a wide range of univariate/multivariate statistical and machine learning alternatives. You just have to choose the method you prefer depending on the constraints (e.g., Gaussian distribution function, compositional nature, etc.) and assumptions (e.g., spatial similarity) on your dataset.

For practical examining of these, you may refer to R packages such as "accelmissing", "ForImp", "imputeMissings", "missRanger", and "VIM".

Hi Vikki, I am running into a similar issue. Have you worked around that?