Rolling window slider::slide() with grouped data - r

In the following example I try to compute the first coefficient from a linear model for time t = 1 until t. It's an expanding rolling window.
It works well with ungrouped data, but when grouped by case, I get the error Error: Columncoef1must be length 10 (the group size) or one, not 30.
How can I handle grouped data?
library(dplyr)
library(slider)
get_coef1 <- function(data) {
coef1 <- lm(data = data, r1 ~ r2 + r3) %>%
coef() %>%
.["r2"] %>%
unname()
return(coef1)
}
data <- tibble(t = rep(1:10, 3),
case = c(rep("a", 10), rep("b", 10), rep("c", 10)),
r1 = rnorm(30),
r2 = rnorm(30),
r3 = rnorm(30))
data %>%
# ungroup() %>%
group_by(case) %>%
mutate(coef1 = slider::slide_dbl(., ~get_coef1(.x),
.before = Inf, .complete = T))

You have to first tidyr::nest the cases. Within the nested tibbles (accessed via purrr::map) you can then apply slide (same technique as with purrr::map). The important point is that you do not want to slide across cases, but only within cases.
library(dplyr)
library(tidyr)
library(purrr)
library(slider)
get_coef1 <- function(data) {
coef1 <- lm(data = data, r1 ~ r2 + r3) %>%
coef() %>%
.["r2"] %>%
unname()
return(coef1)
}
data <- tibble(t = rep(1:10, 3),
case = c(rep("a", 10), rep("b", 10), rep("c", 10)),
r1 = rnorm(30),
r2 = rnorm(30),
r3 = rnorm(30))
data %>%
# ungroup() %>%
group_by(case) %>% nest() %>%
mutate(rollreg = map(data, ~ .x %>% mutate(coef1 = slider::slide_dbl(., ~get_coef1(.x), .before = Inf, .complete = TRUE)))) %>%
select(-data) %>% unnest(rollreg)
I have been trying for a while to use the new dplyr::nest_by() from dplyr 1.0.0 trying to use summarise in combination with the rowwise cases but couldn't get that to work.

I realize this is an old post, but for the sake of completeness, I offer another solution. Is this what you're looking for? Two subtle changes to the arguments to slide_dbl. The code runs.
data %>%
# ungroup() %>%
group_by(case) %>%
mutate(coef1 = slider::slide_dbl(.x = cur_data(), # use cur_data() instead of .; arg .x
.f = ~get_coef1(.x), # arg .f
.before = Inf, .complete = T))
See the slider() documentation for underlying reasons.

Related

How can I pass change arguments into character in a function?

Give a minimum example.
df <- data.frame("Treatment" = c(rep("A", 2), rep("B", 2)), "Price" = 1:4, "Cost" = 2:5)
I want to summarize the data by treatments for all the variables I have, and put them together, so I define a function to do this for each variable first, and then rbind them later on.
SummarizeFn <- function(x,y,z) {
x %>% group_by(Treatment) %>%
summarize(n = n(), Mean = mean(y), SD = sd(y)) %>%
cbind("Var" = rep(y, 3)) # add a column to show which variable those statistics belong to.
}
SumPrice <- SummarizeFn(df, df$Price, "Price")
However, R tells me that object "Price" is not found. How to solve this problem?
Also, how to make y as a character indicating the mean and sd are of price?
Price isnt a variable, you need SummarizeFn(df,df$Price) because Price is just defined in your list df
SummarizeFn <- function(x,y,z)
{
df1<-(x %>% group_by(Treatment)
%>% summarize(n = n(), Mean = mean(y), SD = sd(y))
)
df1<- df1 %>% mutate ("Var" = z)
return(df1)
}
SumPrice <- SummarizeFn(df, df$Price,"Price")

Apply different data to a function in R

I have the following data frame:
library(tidyverse)
set.seed(1234)
df <- data.frame(
x = seq(1, 100, 1),
y = rnorm(100)
)
Where I apply a smooth spline using different knots:
nknots <- seq(4, 15, 1)
output <- map(nknots, ~ smooth.spline(x = df$x, y = df$y, nknots = .x))
What I need to do now is to apply the same function using 2-point and 3-point averages:
df_2 <- df %>%
group_by(., x = round(.$x/2)*2) %>%
summarise_all(funs(mean))
df_3 <- df %>%
group_by(., x = round(.$x/3)*3) %>%
summarise_all(funs(mean))
In summary, I need to apply the function I used in output with the following data frames:
df
df_2
df_3
Of course, this is a minimal example, so I am looking for a efficient way of doing it. Preferably with the purrr package.
Using lapply, and the library zoo to calculate the moving average in a more simple and elegant manner:
library(zoo)
lapply(1:3,function(roll){
dftemp <- as.data.frame(rollmean(df,roll))
map(nknots, ~ smooth.spline(x = dftemp$x, y = dftemp$y, nknots = .x))
})
Here's one possible solution:
library(tidyverse)
set.seed(1234)
df <- data.frame(x = seq(1, 100, 1),
y = rnorm(100))
# funtion to get v-point averages
GetAverages = function(v) {
df %>%
group_by(., x = round(.$x/v)*v) %>%
summarise_all(funs(mean)) }
# specify nunber of knots
nknots <- seq(4, 15, 1)
dt_res = tibble(v=1:3) %>% # specify v-point averages
mutate(d = map(v, GetAverages)) %>% # get data for each v-point
crossing(., data.frame(nknots=nknots)) %>% # combine each dataset with a knot
mutate(res = map2(d, nknots, ~smooth.spline(x = .x$x, y = .x$y, nknots = .y))) # apply smooth spline
You can use dt_res$res[dt_res$v == 1] to see all results for your original daatset, dt_res$res[dt_res$v == 2] to see results for your 2-point estimate, etc.

How to get formulas of multiple regressions by vectorizing

Suppose I have the following code that makes multiple regressions and stores the lm and lm with stepwise selection models in tibbles:
library(dplyr)
library(tibble)
library(MASS)
set.seed(1)
df <- data.frame(A = sample(3, 10, replace = T),
B = sample(100, 10, replace = T),
C = sample(100, 10, replace = T))
df <- df %>% arrange(A)
formula_df <- as.tibble(NA)
aic_df <- as.tibble(NA)
for (i in unique(df$A)){
temp <- df %>% filter(A == i)
formula_df[i, 1] <- temp %>%
do(model = lm(B ~ C, data = .))
aic_df[i, 1] <- temp %>%
do(model = stepAIC(formula_df[[1,1]], direction = "both", trace = F))
}
Is it possible to vectorize to make it faster, for example using the *pply functions? The loop becomes extremely slow when the data gets larger. Thank you in advance.
You could try something like:
model <- df %>% group_by(A) %>%
summarise(formula_model = list(lm(B ~ C))) %>%
mutate(aic_model = list(stepAIC(.[[1,2]], direction = "both", trace = F)))

Mutate within nested data frame

I would like to perform kmeans within groups and add to my data information about cluster number and center which an observation was assigned to (still, within groups so cluster 1 is not the same for group A and group B). I thought that I can pluck cluster assignment and centroid from kmeans and then maybe join these two with each other and finally, with original data. To do the former I wanted to add a row number to data frames with centers and then join by the number of cluster. But how can I add row number within nested data frames? The following code works well until the last, 'nested' mutate.
my_data <- data.frame(group = c(sample(c('A', 'B', 'C'), 20, replace = TRUE)), x = runif(100, 0, 10), y = runif(100, 0, 10))
my_data %>%
group_by(group) %>%
nest() %>%
mutate(km_cluster = map(data, ~kmeans(.x, 3) %>% pluck('cluster')),
km_centers = map(data, ~kmeans(.x, 3) %>% pluck('centers') %>% mutate(cluster = row_number())))
#Luke.sonnet provided an answer that works well with map, but interestingly not with map2, see below:
my_data %>%
group_by(group) %>%
nest() %>%
mutate(number = sample(3:7, 3)) %>%
mutate(km_cluster = map2(data, number, ~kmeans(.x, .y) %>% pluck('cluster')),
km_centers = map2(data, number, ~kmeans(.x, .y) %>% pluck('centers') %>% as_tibble() %>% mutate(cluster = row_number())))
Any ideas how to solve the issue in that case? And equally important, what is the cause of such behaviour?
The problem is that pluck() is returning a matrix. Cast to a tibble first and number differently.
library(tidyverse)
my_data <- data.frame(group = c(sample(c('A', 'B', 'C'), 20, replace = TRUE)), x = runif(100, 0, 10), y = runif(100, 0, 10))
my_data %>%
group_by(group) %>%
nest() %>%
mutate(number = sample(3:7, 3)) %>%
mutate(km_cluster = map2(data, number, ~kmeans(.x, .y) %>% pluck('cluster')),
km_centers = map2(data, number, ~kmeans(.x, .y) %>% pluck('centers') %>% as_tibble() %>% mutate(cluster = seq_len(nrow(.)))))
Note you can also do mutate(cluster = row_number(x)))) and this provides different numbers (note that just using row_number() uses the rows from the parent df). I think given kmeans that the matrix of centers is ordered row-wise by cluster number that the answer in the main chunk is correct.

working with paired data across groups in the tidyverse

I have multiple observations from each of a few groups and I'd like to make a matrix of QQ plots (or another type of plot), comparing each group to every other group.
Here's an example of what I'm talking about:
library(tidyverse)
set.seed(27599)
n <- 30
d <- data_frame(person = c(rep('Alice', n),
rep('Bob', n),
rep('Charlie', n),
rep('Danielle', n)),
score = c(rnorm(n = n),
rnorm(n = n, mean = 0.1),
rnorm(n = n, sd = 2),
rnorm(n = n, mean = 0.3, sd = 1.4)))
by_hand <- data_frame(a = sort(d$score[d$person == 'Alice']),
b = sort(d$score[d$person == 'Bob']),
c = sort(d$score[d$person == 'Charlie']),
d = sort(d$score[d$person == 'Danielle']))
pairs(x = by_hand,
lower.panel = function(x, y) { points(x, y); abline(0, 1);})
Here, I've manipulated the data by hand and used graphics::pairs() to make the plot. Can the same be done inside the tidyverse?
Here's what I've tried.
d %>%
group_by(person) %>%
mutate(score = sort(score)) %>%
glimpse()
This seems promising.
d %>%
group_by(person) %>%
mutate(score = sort(score)) %>%
spread(key = person, value = score)
This gives the 'duplicate identifiers' error.
Maybe reshape2 would be better to use here?
d %>%
group_by(person) %>%
mutate(score = sort(score)) %>%
dcast(formula = score ~ person)
This creates a data.frame with 120 rows, and most of the values (90 per person) are NA. How can I create a wide data.frame without introducing so many NA?
You need a variable that links the row position for each person. Try
by_tidyverse <- d %>%
group_by(person) %>%
mutate(rowID=1:n(),
score=sort(score)
) %>%
spread(key = person, value = score) %>%
select(-rowID)
pairs(x = by_tidyverse, lower.panel = function(x, y) { points(x, y); abline(0, 1);})

Resources