I would like to apply a function across columns of a nested grouped tibble as in the example below.
library(tidyverse)
df <- swiss %>%
group_by(Catholic > 20) %>%
nest()
Which results in a tibble that looks like:
> df
# A tibble: 2 x 2
# Groups: Catholic > 20 [2]
`Catholic > 20` data
<lgl> <list>
1 FALSE <tibble [26 × 6]>
2 TRUE <tibble [21 × 6]>
Now I make some function to build a model
fit <- function(df, modL = NA){
if (modL == 1) {fit <- lm(Fertility ~ Education, data = df)}
if (modL == 2) {fit <- lm(Fertility ~ Education + Examination, data = df)}
fit
}
Now I map that model to columns of the grouped data and make two new variables to store the model fits.
df <- df %>%
mutate(model1 = map(data, fit, modL = 1)) %>%
mutate(model2 = map(data, fit, modL = 2))
Which produces a tibble with two new columns that contain the model fits
> df
# A tibble: 2 x 4
# Groups: Catholic > 20 [2]
`Catholic > 20` data model1 model2
<lgl> <list> <list> <list>
1 FALSE <tibble [26 × 6]> <lm> <lm>
2 TRUE <tibble [21 × 6]> <lm> <lm>
What I want to achieve is a purr-type map function that does the same thing as the following code.
anova(df$model1[[1]], df$model2[[1]])
anova(df$model1[[2]], df$model2[[2]])
I though the following code would work, but it does not.
map(df[,3:4], anova)
Gurus, how do I map a function across columns of a nested and grouped dataset to give one result per row using the columns of that row as input?
Brant
df %>%
mutate(anova = map2(model1, model2, ~ anova(.x,.y)))%>%
mutate(pvalue = map_dbl(anova, ~.$`Pr(>F)`[2]))
I think this is what you want? Can you clarify please! Second mutate will pull out the p-value of the anova for each pairwise comparison.
Related
I have a regular Surv object from the survival package;
s <- Surv(sample(100:150, 5), sample(c(T, F), 5, replace = T))
And a matrix of multiple variables;
df <- data.frame(var1 = rnorm(5),
var2 = rnorm(5),
var3 = rnorm(5))
I need to fit a Cox-PH model for each variable separately. My code currently uses a loop as follows:
for (v in colnames(df)) {
coxph(s ~ df[[v]])
}
Of course, in reality there are thousands of variables and this process takes a bit. I wanted to follow the answer given here to try and do it all with tidyr but I'm kinda stumped because the predictand isn't a factor, it's a survival object, so I don't quite know how to handle it as part of a tibble.
Assuming your response is s for the survival model, you can use a nested dataframe similar to the answer you link to, then map the model to the different variables:
library(tidyverse)
df_nested <- df %>% pivot_longer(cols = var1:var3) %>% group_by(name) %>% nest()
surv_model <- function(df) {
coxph(s ~ df$value)
}
df_nested <- df_nested %>% mutate(model = map(data, surv_model))
df_nested
# A tibble: 3 x 3
# Groups: name [3]
name data model
<chr> <list> <list>
1 var1 <tibble [5 x 1]> <coxph>
2 var2 <tibble [5 x 1]> <coxph>
3 var3 <tibble [5 x 1]> <coxph>
I'm trying to figure out a purrr approach to iteratively map over columns within a list of data frames to fit univariate GLMs. Using map2, the first element, .x, would be the three pred columns, and the second element, .y, would be the list of data frames (or vice-versa). map2 seems to be able to do this, but I recognize that I need to cross the .x and .y elements first, so I use tidyr::crossing first to do this. From here, I am unsure how to properly reference the columns to select within the data frames. Example code is below:
#Sample data
set.seed(100)
test_df <- tibble(pred1 = sample(40:80, size = 1000, replace = TRUE),
pred2 = sample(40:80, size = 1000, replace = TRUE),
pred3 = sample(40:80, size = 1000, replace = TRUE),
resp = sample(100:200, size = 1000, replace = TRUE),
group = sample(c('a','b','c'), size = 1000, replace = TRUE))
#Split into list
test_ls <- test_df %>%
group_by(group) %>%
{df_groups <<- .} %>%
group_split()
#Obtain keys and name list elements
group_keys <- df_groups %>%
group_keys() %>%
pull()
test_ls <- test_ls %>% setNames(nm = group_keys)
#Cross all combinations of pred columns and list element names
preds <- c('pred1','pred2','pred3')
map_keys <- crossing(preds, group_keys)
#.y = list of data frames; iterate over data frames
#.x = three pred columns; iterate over columns
#Use purrr to fit glm of each .x columns within each of .y dfs
#Example structure - does not work
map2(.x, .y, .f = ~glm(resp ~ .x, data = .y))
#Workaround that does work
lapply(test_ls, function(x) {
x %>%
select(pred1, pred2, pred3) %>%
map(.f = ~glm(resp ~ .x, data = x))
})
There's something I'm missing, and I can't seem to figure it out. I've gotten a variety of errors with a few approaches, but I think it's coming down to not properly referencing the .x columns within the .y data frames. My approaches don't seem to recognize that .x is a column within .y. The workaround does the trick, but I'd prefer to avoid using both lapply and map.
My suggestion would be to NOT split the data before fitting models, since you are considering all possible combinations of variables that are already available directly in your original dataset. Instead, consider converting the original data frame to the "long" format, and then grouping by the necessary variables:
test_df %>% gather( pred, value, pred1:pred3 ) %>%
nest( -c(group, pred) ) %>%
mutate( models = map(data, ~glm(resp ~ value, data=.x)) )
# # A tibble: 9 x 4
# group pred data models
# <chr> <chr> <list> <list>
# 1 b pred1 <tibble [340 x 2]> <glm>
# 2 a pred1 <tibble [317 x 2]> <glm>
# 3 c pred1 <tibble [343 x 2]> <glm>
# 4 b pred2 <tibble [340 x 2]> <glm>
# 5 a pred2 <tibble [317 x 2]> <glm>
# 6 c pred2 <tibble [343 x 2]> <glm>
# 7 b pred3 <tibble [340 x 2]> <glm>
# 8 a pred3 <tibble [317 x 2]> <glm>
# 9 c pred3 <tibble [343 x 2]> <glm>
This substantially simplifies your code, and you can now split the result, if you still need those models in a list.
In my data, I have correlated data (diet and liver) for 50+ different compounds (simplified here).
library(tidyverse)
Sigma <- matrix(.7, nrow=6, ncol=6) + diag(6)*.3
vars_tr <- data.frame(MASS::mvrnorm(n=10, mu=c(2:7), Sigma=Sigma))
tr<-tibble(
compound=c(rep("A", 10), rep("B", 10), rep("C",10)),
diet=c(vars_tr$X1, vars_tr$X2, vars_tr$X3),
liver=c(vars_tr$X4, vars_tr$X5, vars_tr$X6))
Following the guidance on doing regressions for multiple models, I created a nested data frame and stored the output (learning this method this week was a lifesaver!).
model<-function(df){lm(data=df, liver~diet)}
mods<- tr %>%
group_by(compound) %>%
nest() %>%
mutate(model=map(data, model))
Now I have new 'diet' data for which no 'liver' data exists.
new<-tibble(
compound=c(rep("A", 10), rep("B", 10), rep("C",10)),
diet=c(rnorm(10, 4), rnorm(10, 5), rnorm(10,6)))
What I would like to do is take advantage of purrr generate a liver concentration for each diet concentration using the correct model for the compound. My best attempt looks like:
preds<-function(c, x){
add_predictions(tibble(diet=x), filter(mods, compound==c)$model[[1]], 'liver')$liver
}
new%>%
mutate(liver=map2(compound, diet, preds))
which returns an error.
I would greatly appreciate any help!
EDIT 6/4/2020:
Based on the helpful comments from Bruno and Ronak Shah below, I've made some progress but haven't found the solution. Both suggest joining the models to the existing table, which makes way more sense than what I was doing.
Based on that, it is relatively simple to do the following:
new_mods<-
new%>%
group_by(compound)%>%
nest()%>%
left_join(., select(mods_d, compound, model), , by='compound')%>%
mutate(predicts = map2(data, model, add_predictions))%>%
unnest(predicts)
You can use a join operation and keep working on tibbles
library(tidyverse)
library(MASS)
Sigma <- matrix(.7, nrow=6, ncol=6) + diag(6)*.3
vars_tr <- data.frame(mvrnorm(n=10, mu=c(2:7), Sigma=Sigma))
tr<-tibble(
compound=c(rep("A", 10), rep("B", 10), rep("C",10)),
diet=c(vars_tr$X1, vars_tr$X2, vars_tr$X3),
liver=c(vars_tr$X4, vars_tr$X5, vars_tr$X6))
model<-function(df){lm(data=df, liver~diet)}
mods<- tr %>%
nest_by(compound) %>%
mutate(model = list(model(data)))
new<-tibble(
compound=c(rep("A", 10), rep("B", 10), rep("C",10)),
diet=c(rnorm(10, 4), rnorm(10, 5), rnorm(10,6)))
new_nest <- new %>%
nest_by(compound)
results <- mods %>%
left_join(new_nest,by = "compound") %>%
mutate(predicts = list(predict(model,data.y)))
You can create a function for prediction :
preds<-function(data, mod){
modelr::add_predictions(data, mod)$liver
}
nest the dataframe for each compound, join with mods and apply the respective model for each group of data.
library(dplyr)
new %>%
tidyr::nest(data = diet) %>%
left_join(mods, by = 'compound') %>%
mutate(liver = purrr::map2(data.y, model, preds))
# A tibble: 3 x 5
# compound data.x data.y model liver
# <chr> <list> <list> <list> <list>
#1 A <tibble [10 × 1]> <tibble [10 × 2]> <lm> <dbl [10]>
#2 B <tibble [10 × 1]> <tibble [10 × 2]> <lm> <dbl [10]>
#3 C <tibble [10 × 1]> <tibble [10 × 2]> <lm> <dbl [10]>
Based on the requirement you can select relevant columns and unnest the results if needed.
In the spirit of purr, broom, modelr, I am trying to create a "meta" data.frame in which each row denotes the dataset (d) and the model parameters (yvar, xvars, FEvars). For instance:
iris2 <- iris %>% mutate(Sepal.Length=Sepal.Length^2)
meta <- data.frame(n=1:4,
yvar = c('Sepal.Length','Sepal.Length','Sepal.Length','Sepal.Length'),
xvars= I(list(c('Sepal.Width'),
c('Sepal.Width','Petal.Length'),
c('Sepal.Width'),
c('Sepal.Width','Petal.Length'))),
data= I(list(iris,iris,iris2,iris2)) )
Now, I would like to run a model for each column of "meta". And then add a list column "model" with the model output object. To run the model I use an auxiliary function that uses a dataset, a y variable and a vector of x variables:
OLS_help <- function(d,y,xvars){
paste(y, paste(xvars, collapse=" + "), sep=" ~ ") %>% as.formula %>%
lm(d)
}
y <- 'Sepal.Length'
xvars <- c('Sepal.Width','Petal.Length')
OLS_help(iris,y,xvars)
How can I execute OLS_help for all the rows of meta and adding the output of OLS_help as a list column in meta? I tryed the following code, but it did not work:
meta %>% mutate(model = map2(d,yvar,xvars,OLS_help) )
Error: Can't convert a `AsIs` object to function
Call `rlang::last_error()` to see a backtrace
OBS: The solution to when only the "data" (nested) list column (corvered in Hadley's book here) is:
by_country <- gapminder %>% group_by(country, continent) %>% nest()
country_model <- function(df) { lm(lifeExp ~ year, data = df) }
by_country <- by_country %>% mutate(model = map(data, country_model))
We can use pmap in the following way
df <- meta %>%
as_tibble() %>%
mutate_if(is.factor, as.character) %>%
mutate(fit = pmap(
list(yvar, xvars, data),
function(y, x, df) lm(reformulate(x, response = y), data = df)))
## A tibble: 4 x 5
# n yvar xvars data fit
# <int> <chr> <I<list>> <I<list>> <list>
#1 1 Sepal.Length <chr [1]> <df[,5] [150 × 5]> <lm>
#2 2 Sepal.Length <chr [2]> <df[,5] [150 × 5]> <lm>
#3 3 Sepal.Length <chr [1]> <df[,5] [150 × 5]> <lm>
#4 4 Sepal.Length <chr [2]> <df[,5] [150 × 5]> <lm>
Explanation: pmap iterates over multiple arguments simultaneously (similar to base R's Map); here we simultaneously loop throw entries in column yvar, xvar and data, then use reformulate to construct the formula to be used within lm. We store the lm fit object in column fit.
I am trying to add a set of extrapolated "observations" to a matrix in R. I know how to do this using normal programming techniques (read; bunch of nested loops and functions) but I feel this must be possible in a much more clean way by using build in R-functionality.
The code below illustrates the point, and where it breaks down
Many thanks in advance for your help!
With kind regards
Sylvain
library(dplyr)
# The idea is that i have a table of observations for e.g. x=5, 6, 7, 8, 9 and 10. The observations (in this example 2)
# conform fairly decently to sets of 2nd order polynomials.
# Now, I want to add an extrapolated value to this table (e.g. x=4). I know how to do this programmically
# but I feel there must be a cleaner solution to do this.
#generate dummy data table
x <- 5:10
myData <- tibble(x, a = x^2 * 2 + x * 3 + 4 + rnorm(1,0,0.01), b = x^2 * 3 + x * 4 + 5 + rnorm(1,0,0.01) )
#Gather (put in Data-Key format)
myDataKeyFormat <- gather(myData,key = "someLabel", value = "myObservation", -x)
fitted_models <- myDataKeyFormat %>% group_by(someLabel) %>% do(model = lm(myObservation ~ poly(x,2), data = .))
myExtrapolatedDataPointx <- tibble(x = 4)
#Add the x=4 field
fitted_points <- fitted_models %>% group_by(someLabel) %>% do(predict(.$model,myExtrapolatedDataPointx)) #R really doesnt like this bit
#append the fitted_points to the myDataKeyFormat
myDataKeyFormatWithExtrapolation <- union(myDataKeyFormat,fitted_points)
#use spread to
myDataWithExtrapolation <- myDataKeyFormatWithExtrapolation %>% spread(someLabel,myObservation)
Here is a solution in the tidyverse, and using purrr to create the different models. The idea is to nest (using tidyr::nest) and then purrr::map to train the model. I will then add new values and compute the predictions using modelr::add_predictions. Here you have all the data in the same place : training data, models, testing data and prediction, by your variable someLabel. I also give you a way to visualise the data.
You can check R for Data Science by Hadley Wickham & Garrett Grolemund, and especially the part about models for more information.
library(dplyr)
library(tibble)
library(tidyr)
library(purrr)
library(modelr)
library(ggplot2)
set.seed(1) # For reproducibility
x <- 5:10
myData <- tibble(x,
a = x^2 * 2 + x * 3 + 4 + rnorm(1,0,0.01),
b = x^2 * 3 + x * 4 + 5 + rnorm(1,0,0.01))
#Gather (put in Data-Key format)
myDataKeyFormat <- gather(myData,key = "someLabel", value = "myObservation", -x)
myModels <- myDataKeyFormat %>%
nest(-someLabel) %>%
mutate(model = map(data, ~lm(myObservation ~ poly(x,2), data = .x)))
Here is the result at this point : you have a model for each value of someLabel.
# A tibble: 2 × 3
someLabel data model
<chr> <list> <list>
1 a <tibble [6 × 2]> <S3: lm>
2 b <tibble [6 × 2]> <S3: lm>
I'll add some data points in a new column (map is to create it as a tibble for each line of the data frame).
# New data
new_data <- myModels %>%
mutate(new = map(data, ~tibble(x = c(3, 4, 11, 12))))
I add the predictions: add_predictions take a data frame and a model as argument, so I use map2 to map over the new data and the models.
fitted_models <- new_data %>%
mutate(new = map2(new, model, ~add_predictions(.x, .y)))
fitted_models
# A tibble: 2 × 4
someLabel data model new
<chr> <list> <list> <list>
1 a <tibble [6 × 2]> <S3: lm> <tibble [4 × 2]>
2 b <tibble [6 × 2]> <S3: lm> <tibble [4 × 2]>
There you go: you have for each label the data and model trained on this data, and the new data with predictions.
In order to plot it, I use unnest to take the data back to the data frame, and I bind the rows to have the "old" data and the new values together.
my_points <- bind_rows(unnest(fitted_models, data),
unnest(fitted_models, new))
ggplot(my_points)+
geom_point(aes(x = x, y = myObservation), color = "black") +
geom_point(aes(x = x, y = pred), color = "red")+
facet_wrap(~someLabel)