binomial logistic regression response variables

[Shan-square]

Hello, Is there a way to code the response variable when it is a two-column matrix of counts when using binomial logistic regression with the following code:

run_ml(data, method="glmnet", outcome_colname="y", kfold=5, cv_times=100, training_frac=0.8)

I might be wrong, but it seems the outcome_colname option only allows one column as the input. I would like to use a "y" which includes the counts of individuals with disease and without disease. An example of "y" would be:

disease no_disease
2 3
5 8
3 9
....

Thanks.

[kelly-sovacool]

Hi @Shan-square, mikropml and caret require a data frame where the outcome_colname is exactly one column, and all remaining columns are used as features (a.k.a predictor variables), with each row as an individual/observation. What are the features / predictors you wish to train the model on?

[kelly-sovacool]

Binomial logistic regression is for outcomes/responses that are dichotomous, e.g. yes/no, pass/fail, whether any fruits from a plant were diseased -- in which case you would still have only one outcome column. I would recommend transforming your response variable to fractions and using logistic (correction: linear) regression, e.g.

library(dplyr)
set.seed(20231208)
# create example data frame with dummy otu abundances
dat <- data.frame(disease=c(2,5,3), 
                  no_disease=c(3,8,9),
                  otu1 = runif(3),
                  otu2 = runif(3),
                  otu3 = runif(3))
dat
#>   disease no_disease       otu1      otu2       otu3
#> 1       2          3 0.77708972 0.8929019 0.74143156
#> 2       5          8 0.03412326 0.6826952 0.08643496
#> 3       3          9 0.12043893 0.1737634 0.78740345

# calculate fraction of fruits that were diseased for each plant
dat_frac <- dat %>% 
  mutate(frac_diseased = disease / (disease + no_disease), .before = otu1) %>%
  select(-c(disease, no_disease))
dat_frac
#>   frac_diseased       otu1      otu2       otu3
#> 1     0.4000000 0.77708972 0.8929019 0.74143156
#> 2     0.3846154 0.03412326 0.6826952 0.08643496
#> 3     0.2500000 0.12043893 0.1737634 0.78740345

^{Created on 2023-12-08 with reprex v2.0.2}

[Shan-square]

Thanks Kelly! Do I need to do add anything to
run_ml (data, method="glmnet", outcome_colname="frac_disease", kfold=5, cv_times=100, training_frac=0.8)
to show that it is a logistic regression?

[kelly-sovacool]

Nope, that will work as-is as long as data is your reformatted dataframe. glmnet works generally for regression and will detect the type based on the outcome column. Do take a look at our vignettes if you haven't yet. https://www.schlosslab.org/mikropml/articles/introduction.html#the-methods-we-support

[kelly-sovacool]

Oh by the way, this will perform linear regression since the outcome column will be a continuous variable. Logistic regression is for categorical variables.

[Shan-square]

That sounds good. Thanks for your help Kelly!