R: Summarize inside an nlsLM() statement

Question

I\'m using nlsLM() to make a model of a power function, but I need to summarize my data within the function call to find the appropriate coefficient and exponen

Answer 1

What you want isn't clear from the question. Do you want to fit a separate model for every unique month in your data? Or do you want to fit one model for all of the data and then take monthly averages of the value of MW^b ?

Here's one way on how to do the latter case.

require(minpack.lm)
require(tidyverse)
require(broom)

dat <- structure(...) # provided in the question

predictions <- 
    dat %>% 
    ungroup %>%
    mutate(row = row_number()) %>%
    do(augment(nlsLM(
                formula = value ~ a * MW^b + 0*row, 
                data = .,
                start = list(a = 100000, b=1/3),
                upper = c(Inf, 1), 
                lower = c(0, 1/5)
               )
           )
       )

joined <- 
    dat %>%
    mutate(row = row_number()) %>%
    left_join(predictions, by=c('MW', 'value', 'row')) %>%
    select(-row)

joined %>%
    group_by(mon) %>%
    mutate(monthly_avg_prediction = mean(.fitted))

Notes:

1. This kind of stuff is much easier with the broom package. This is because broom converts the results of model-finding functions like lm, nls, or nlsLM etc. into dataframes. So you don't have to memorize or re-lookup the function-specific structure of the model object (e.g. model$params[['estimate']][[1]]) or similar stuff; the model results are already in R-standard dataframe format.
2. My solution uses ideas from this StackOverflow answer on how to join broom-generated dataframes of predictions with original data. That's why the stuff with row_number() and left_join() are in there. Otherwise, in the general case, augment will throw away data from the original dataframe that is not used in the model prediction, and it will not work well if there are repeated values in the data that is used.
3. The .fitted column is generated by broom's augment function. It is the model prediction at the indicated datapoint.
4. The result (that I think you might want) is in the monthly_avg_prediction column of the joined dataframe. But that represents a single, global model, fit on all the data, and predictions from that model are averaged by month.