When I first encountered the plyr package for R, I thought it was needless frippery. There was nothing that a well-written loop or subset couldn't do. Then I read Hadley Wickham's Split-Apply-Combine paper, taught plyr as a TA, and tried it out in my own code. I've been converted, it's a great tool for expressive R code. It's best to see this by example.

A simple example

For my ADA work I'm working with power spectra from a large number of recordings of neural activity in mice brains. When I'm doing my analysis the first part of my code is loading the data. My first approach is below.

dfs <- list()

for(ii in seq_along(files)) {
    file <- files[ii]
    mouse <- mice[ii]
    label <- labels[ii]

    dfs[[ii]] <- cbind(read_spectrum(file), mouse = mouse, label = label)
}

df <- do.call(rbind, dfs)

This is a fairly straightforward approach. I pass a list of filenames, mouse identifiers, and disease labels (using the wonderful argparse package) and load the spectra as dataframes which I eventually rbind together.

Here's the corresponding plyr version

inputs <- data.frame(file = files, mouse = mice, label = labels)

df <- mdply(inputs, function(file, mouse, label) {
    return(cbind(read_spectrum(file), mouse = mouse, label = label))
})

Looking at this code, it's more compact at 5 lines compared to 11 for the for loop.

A more complicated example

Now, once I've loaded the spectra as a dataframe, I need to aggregate them, first within mice, and then between label.

The resulting R code is a little messy.

average_specs <- function(df) {
    freqs <- unique(df$frequency)
    avepower <- rep(NA, length(freqs))

    for(ii in seq_along(freqs)) {
        freq <- freqs[ii]
        nRecordings <- nrow(df) / length(freqs)
        avepower[ii] <- sum(df$power[df$frequency == freq]) / nRecordings
    }

    data.frame(frequency = freqs, power = avepower)

    return(df)
}

## Aggregate by mouse
micedfs <- list()
uniqMice <- unique(mice)
for(ii in seq_along(uniqMice)) {
    mouse <- uniqMice[ii]
    label <- df$label[df$mouse == mouse][1]

    mousedf <- averageSpecs(df[df$mouse == mouse])
    mousedf$label <- label
    micedfs[ii] <- mousedf
}

micedf <- do.call(rbind, micedfs)

## Aggregate by condition
labeldfs <- list()
uniqLabels <- unique(labels)
for(ii in seq_along(uniqLabels)) {
    label <- uniqLabels[ii]

    labeldf <- averageSpecs(micedf[micedf$label == label])
    labeldf$label <- label
    labeldfs[ii] <- labeldf
}

df <- do.call(rbind, labeldfs)

I don't want to talk about this. It's just ugly.

This was actually the example which lead to my plyr epiphany. Just look at the plyr version.

average_specs <- function(df) {
    df <- ddply(df, .(frequency), summarize, power <- mean(power))
    return(df)
}

df <- ddply(specdf, .(label), function(labelsub) {
    df <- ddply(specdf, .(mouse), function(mousesub) {
        df <- average_specs(sub)
        return(df)
    })
    return(df)
})

Wow, that just looks nice. It's concise with very little boilerplate. I can easily extend it to calculate other summary statistics such as standard deviations.

Icing on the cake

Now that I've gotten the summary statistics I like to plot them.

for(label in uniqLabels) {
    outfile <- paste0(outdir, "/", label, ".pdf")

    select <- df$label == label
    subdf <- df[select, ]
    fig <- ggplot(subdf, aes(x = frequency, y = power)) + geom_line()
    ggsave(outfile, fig)
}

The plyr version is very similar.

d_ply(df, .(label), function(df) {
    label <- df$label[1]
    outfile <- paste0(outdir, "/", label, ".pdf")

    fig <- ggplot(df, aes(x = frequency, y = power)) + geom_line()
    ggsave(outfile, fig)
})

That's nice, but not really that different. But what if I got data from a several location in the brain and wanted to plot those locations individually?

for(label in uniqLabels) {
    for(location in uniqLocations) {
        select <- df$label == label & df$location == location
        if(any(select)) { # avoid empty df error
            outfile <- paste0(outdir, "/", label, "-", location, ".pdf")

            subdf <- df[select, ]
            fig <- ggplot(subdf, aes(x = frequency, y = power)) + geom_line()
            ggsave(outfile, fig)
        }
    }
}

I dislike the look of this. The first four lines are too long and I'm already at three levels of indentation.

By contrast, plyr handles this very gracefully.

d_ply(df, .(label, location), function(df) {
    label <- df$label[1]
    location <- df$location[1]
    outfile <- paste0(outdir, "/", label, , "-", location, ".pdf")

    fig <- ggplot(df, aes(x = frequency, y = power)) + geom_line()
    ggsave(outfile, fig)
}

Conclusion

Experimenting with plyr is a manifestation of my effort to write more idiomatic R. Like the examples above, I find that most of the R code that stands out to me as ugly is the result of failing to use the correct R idiom. I haven't compared the speed of plyr. I assume it would compare well enough, but that's somewhat beside the point. Expressiveness and readability are the metrics that I'm trying to optimize here.

A similar process occured with ggplot2. At first I thought it was completely unnecessary. Now I feel pain at the thought of showing people base R plots. The syntax is just so much easier and expressive.

The common thread is that both packages are part of the "Hadleyverse". I'll have to explore further, dplyr and tidyr seem like the logical next steps.

Oh, and just in case you're wondering, you can do some of the same things in julia.