CODEWITHSUNDEEP
rdplyr
nathaneastwood
nathaneastwood

Reputation: 3764

Fitting multiple nls functions with dplyr

I wish to fit multiple nls functions using a group_by from the dplyr package but I am unsure how I can pass multiple starting values. Let's take a simpler example (see ?nls for the inspiration).

DNase1 <- subset(DNase, Run == 1)
modelDNase1 <- DNase1 %>% 
  do(model = nls(density ~ 1/(1 + exp((xmid - log(conc))/scal)),
             data = .,
             start = list(xmid = 0, scal = 1),
             algorithm = "plinear"))

So here I am fitting a single model. But what if I want to extend this so I am fitting the following:

DNase$Run <- factor(DNase$Run)
modelDNase <- DNase %>%
  group_by(Run) %>% 
  do(model = nls(density ~ 1/(1 + exp((xmid - log(conc))/scal)),
             data = .,
             start = list(xmid = 0, scal = 1),
             algorithm = "plinear"))

How would I pass on multiple start parameters? Would the purrr package be of any use?

Upvotes: 3

Views: 1302

Answers (1)

Gregor Thomas
Gregor Thomas

Reputation: 145745

(Comment to answer.) My first guess was correct, the .$ syntax seems to work.

As a convenient way of picking starting values, create a table with the unique group values and the desired starting values in new columns. Knowing nothing about this data, I assigned them randomly:

starts = data.frame(Run = unique(DNase$Run),
           xmid_start = rnorm(11, sd = 0.1),
           scale_start = 1 + rnorm(11, sd = 0.1))

We can then join this to the data and proceed, pulling the first starting value from each grouping to give to the model:

mods = DNase %>% 
    left_join(starts) %>%
    group_by(Run) %>%
    do(model = nls(density ~ 1/(1 + exp((xmid - log(conc))/scal)),
             data = .,
             start = list(xmid = first(.$xmid_start),
                          scal = first(.$scale_start)),
             algorithm = "plinear"))

Upvotes: 4

Related Questions