[This article was first published on R on The broken bridge between biologists and statisticians, and kindly contributed to R-bloggers]. (You can report issue about the content on this page here)
Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.

I have been involved with data crunching for 30 years, and, due to my age, I see myself as a dinosaur within the R-users community. I must admit, I’m rather slow to incorporate new paradigms in my programming workflow … I’m pretty busy and the time I save today is often more important than the time I could save in the future, by picking up new techniques. However, resisting to progress is not necessarily a good idea and, from time to time, also a dinosaur feels like living more dangerously and exploring new ideas and views.

Today, I want to talk about column-wise tasks in relatively big datasets. These tasks are rather common with agricultural and biological experiments, where we have several subjects in different treatment groups and we record, for each subject, a high number of traits. In these conditions, the very first step to data analyses is to calculate several descriptive stats for each group of subjects and each variable. What is the best method to write simple and reusable code? As I will show later, the most natural approach to me may not necessarily be the most fashionable one.

# The example

A few days ago, I met a colleague from the plant pathology group, who asked my co-operation for an experiment where he studied the content in 62 mycotoxins in wheat caryopses, as recorded in 70 samples coming from different Italian regions. The dataset was structured as 70 x 64 table, as shown in the scheme below: there is one row for each wheat sample, while the first column represents the sample id, the second column represents the region from where that sample was collected and the other columns represent the concentrations of the 62 toxins (one column per toxin). That colleague wanted me to calculate, for each toxin and for each region, the following stats:

1. number of collected samples
2. mean concentration value
3. maximum concentration value
4. standard error
5. number of contaminated samples (concentration higher than 0)
6. percentage of contaminated samples
7. mean value for contaminated samples
8. standard error for contaminated samples

He also wanted me to return to him a list of 62 tables (one per toxin), with all the regions along with their descriptive stats.

# I have already done something like this!

When facing this task, my first reaction was to search in my ‘computer’s attic’, looking for previously written codes to accomplish the very same task. As I said, this is a fairly common task in my everyday work and I could find several examples of old codes. Looking at those, I realised that, when the number of variables is high, I have so far made an extensive use of for loops to repeat the same task across columns. This is indeed no wonder, as I came across R in 2000, after an extensive usage of general purpose languages, such as Quick Basic and Visual Basic.

In practise, most of my earliest R codes were based on the tapply() function to calculate statistics for grouped data and for loops to iterate over columns. In the box below I show an example of such an approach, by using a factitious dataset, with the very same structure as my colleague’s dataset.

rm(list = ls())
# str(dataset)
# 'data.frame': 70 obs. of  64 variables:
#  $Sample : int 1 2 3 4 5 6 7 8 9 10 ... #$ Region : chr  "Lombardy" "Lombardy" "Lombardy" "Lombardy" ...
#  $DON : num 8.62 16.2 18.19 27.08 10.97 ... #$ DON3G  : num  21.7 28.4 34.7 26.9 26.4 ...
#  $NIV : num 23.5 25.3 22.6 27.8 29.4 ... #$ NIVG   : num  38.6 26.2 13.5 21.4 15.4 ...
#  $T2 : num 23.9 23.6 19.7 25.7 21.7 ... #$ HT2    : num  19 37.6 32.1 22.3 25.6 ...
#  $HT2G : num 13.7 25.7 25.8 33.4 32.7 ... #$ NEO    : num  29.06 7.56 27.52 32.91 24.49 ...

returnList <- list()
for(i in 1:(length(dataset[1,]) - 3)){
y <- as.numeric( unlist(dataset[, i + 3]) )
Count <- tapply(y, dataset$Region, length) Mean <- tapply(y, dataset$Region, mean)
Max <- tapply(y, dataset$Region, max) SE <- tapply(y, dataset$Region, sd)/sqrt(Count)
nPos <- tapply(y != 0, dataset$Region, sum) PercPos <- tapply(y != 0, dataset$Region, mean)*100
muPos <- tapply(ifelse(y > 0, y, NA), dataset$Region, mean, na.rm = T) muPos[is.na(muPos)] <- 0 sdPos <- tapply(ifelse(y > 0, y, NA), dataset$Region, sd, na.rm = T)
SEpos <- sdPos/sqrt(nPos)
returnList[[i]] <- data.frame(cbind(Count, Mean, Max, SE, nPos, PercPos, muPos, SEpos))
names(returnList)[[i]] <- colnames(dataset)[i + 3]
}
print(returnList$CRV, digits = 2) ## Count Mean Max SE nPos PercPos muPos SEpos ## Abruzzo 4 28 30 0.85 4 100 28 0.85 ## Apulia 9 25 40 2.67 9 100 25 2.67 ## Campania 2 20 21 0.74 2 100 20 0.74 ## Emilia Romagna 8 23 33 2.80 8 100 23 2.80 ## Latium 7 20 33 2.76 7 100 20 2.76 ## Lombardy 4 25 32 5.12 4 100 25 5.12 ## Molise 1 18 18 NA 1 100 18 NA ## Sardinia 6 25 38 3.32 6 100 25 3.32 ## Sicily 6 21 30 2.94 6 100 21 2.94 ## The Marche 5 19 23 1.58 5 100 19 1.58 ## Tuscany 5 30 34 1.22 5 100 30 1.22 ## Umbria 9 21 32 2.83 9 100 21 2.83 ## Veneto 4 23 28 1.99 4 100 23 1.99 I must admit that the above code is ugly: first, for loops in R are not very efficient and, second, reusing that code requires quite a bit of editing and it is prone to errors. Therefore, I asked myself how I could write more efficient code … First of all, I thought I should use apply() instead of the for loop. Thus I wrote a function to calculate the required stats for each column and apply()-ed that function to all columns of my data-frame. funBec <- function(y, group){ # y <- as.numeric( unlist(dataset[, i + 3]) ) Count <- tapply(y, group, length) Mean <- tapply(y, group, mean) Max <- tapply(y, group, max) SE <- tapply(y, group, sd)/sqrt(Count) nPos <- tapply(y != 0, group, sum) PercPos <- tapply(y != 0, group, mean)*100 muPos <- tapply(ifelse(y > 0, y, NA),group, mean, na.rm = T) muPos[is.na(muPos)] <- 0 sdPos <- tapply(ifelse(y > 0, y, NA), group, sd, na.rm = T) SEpos <- sdPos/sqrt(nPos) data.frame(cbind(Count, Mean, Max, SE, nPos, PercPos, muPos, SEpos)) } returnList2 <- apply(dataset[3:length(dataset[1,])], 2, function(col) funBec(col, dataset$Region))
# kable(returnList2$CRV, digits = 2) I am pretty happy with the above approach; it feels ‘natural’, to me and the coding appears to be pretty clear and easily reusable. However, it makes only use of the base R implementation and, therefore, it looks a bit out-fashioned… a dinosaur’s code… What could I possibly do to write more ‘modern’ code? A brief survey of posts from the R world suggested the ultimate solution: “I must use the ‘tidyverse’!”. Therefore, I tried to perform my column-wise task by using dplyr and related packages and I must admit that I could not immediately find an acceptable solution. After some careful thinking, it was clear that the tidyverse requires a strong change in programming habits, which is not always simple for those who are over a certain age. Or, at least, that is not simple for me… Eventually, I thought I should switch from a for attitude to a split-apply-combine attitude; indeed, I discovered that, if I ‘melted’ the dataset so that the variables were stacked one above the other, I could, consequently: 1. split the dataset in several groups, consisting of one toxin and one region, 2. calculate the stats for each group, and 3. combine the results for all groups. I’d like to share the final code: I used the pivot_longer() function to melt the dataset, the group_by() function to create an ‘internal’ grouping structure by regions and toxins, the summarise() and across() functions to calculate the required stats for each group. In the end, I obtained a tibble, that I split into a list of tibbles, by using the group_split() function. The code is shown in the box below. library(tidyverse) returnList6 <- dataset %>% select(-Sample) %>% pivot_longer(names_to = "Toxin", values_to = "Conc", cols = c(3:length(dataset[1,]) - 1)) %>% group_by(Toxin, Region) %>% summarise(across("Conc", .fns = list(Mean = mean, Max = max, SE = function(df) sd(df)/sqrt(length(df)), nPos = function(df) length(df[df > 0]), percPos = function(df) length(df[df > 0])/length(df)*100, muPos = function(df) mean(df[df > 0]), SEpos = function(df) sd(df[df > 0])/sqrt(length(df[df > 0])) ))) %>% ungroup() %>% group_split(Toxin, .keep = F) names(returnList6) <- names(dataset)[3:64] # returnList6$CRV

Well, I can say that with the above code I have, at least, tried not to be a dinosaur… but, I am sure that many young scientists out there can suggest better solutions. If so, please drop me a line at the email address below. Thanks for reading!

Andrea Onofri
Department of Agricultural, Food and Environmental Sciences
University of Perugia (Italy)
email: [email protected]