Author: Bob Muenchen

Stata’s Academic Growth Nearly as Fast as R’s

by Bob Muenchen

Analytics tools take significant effort to master, so once learned people tend to stick with them for much of their careers. This makes the tools used in academia of particular interest in the study of future trends of market share. I’ve been tracking The Popularity of Data Analysis Software regularly since 2010, and thanks to an astute reader, I now have a greatly improved estimate of Stata’s academic growth. Peter Hedström, Director of the Institute for Analytical Sociology at Linköping University, wrote to me convinced that I was underestimating Stata’s role by a wide margin, and he was right.

Fig_2e_ScholarlyImpactBig6

Two things make Stata’s popularity difficult to guage: 1) Stata means “been” in Italian, and 2) it’s a common name for the authors of scholarly papers and those they cite. Peter came up with the simple, but very effective, idea of adding Statacorp’s headquarter, College Station, Texas, to the search. That helped us find far more Stata articles while blocking the irrelevant ones. Here’s the search string we came up with:

("Stata" "College Station") OR "StataCorp" OR "Stata Corp" OR 
"Stata Journal" OR "Stata Press" OR "Stata command" OR 
"Stata module"

The blank between Stata and College Station is an implied logical “and”. This string found 20% more articles than my previous one. This success motivated me to try and improve some of my other search strings. R and SAS are both difficult to search for due to how often those letters stand for other things. I was able to improve my R search string by 15% using this:

"r-project.org" OR "R development core team" OR "lme4" OR 
"bioconductor" OR "RColorBrewer" OR "the R software" OR 
"the R project" OR "ggplot2" OR "Hmisc" OR "rcpp" OR "plyr" OR 
"knitr" OR "RODBC" OR "stringr" OR "mass package"

Despite hours of effort, I was unable to improve on the simple SAS search string of “SAS Institute.” Google Scholar’s logic seems to fall apart since “SAS Institute” OR “SAS procedure” finds fewer articles!  If anyone can figure that out, please let me know in the comments section below. As usual, the steps I use to document all searches are detailed here.

The improved search strings have affected all the graphs in the Scholarly Articles section of The Popularity of Data Analysis Software. At the request of numerous readers, I’ve also added a log-scale plot there which shows the six most popular classic statistics packages:

Fig_2f_ScholarlyImpactAllStat

If you’re interested in learning R, DataCamp.com offers my 16-hour interactive workshop,
R for SAS, SPSS and Stata Users for $25. That’s a monthly fee, but it definitely won’t take you a month to take it!  For students & academics, it’s $9. I also do training on-site but I’m often booked about 8 weeks out.

I invite you to follow me on this blog and on Twitter.

Fastest Growing Software for Scholarly Analytics: Python, R, KNIME…

In my ongoing quest to “analyze the world of analytics”, I’ve added the following section below to The Popularity of Data Analysis Software:

It would be useful to have growth trend graphs for each of the analytics packages I track, but collecting such data is too time consuming since it must be re-collected every year (since search algorithms change). What I’ve done instead is collect data only for the past two complete years, 2013 and 2014. Figure 2e shows the percent change from 2013 to 2014, with the “hot” packages whose use is growing shown in red. Those whose use is declining or “cooling” are shown in blue. Since the number of articles tends to be in the thousands or tens of thousands, I have removed any software that had fewer than 100 articles in 2013. Going from one to five articles may represent 500% growth, but it’s not of much interest.

Figure 2e. Change in the number of scholarly articles using each software in the most recent two complete years (2013 to 2014). Packages shown in red are "hot" and growing, while those shown in blue are "cooling down" or declining.
Figure 2e. Change in the number of scholarly articles using each software in the most recent two complete years (2013 to 2014). Packages shown in red are “hot” and growing, while those shown in blue are “cooling down” or declining.

The three fastest growing packages are all free and open source: Python, R and KNIME. All three saw more than 25% growth. Note that the Python figures are strictly for analytics use as defined here. At the other end of the scale are SPSS and SAS, both of which declined in use by around 25%. Recall that Fig. 2a shows that despite recent years of decline, SPSS is still extremely dominant for scholarly use.

Three of the packages whose use is growing implement the powerful and easy-to-use workflow or flowchart user interface: KNIME, RapidMiner and SPSS Modeler. As useful as that approach is, it’s not sufficient for success as we see with SAS Enterprise Miner, whose use declined nearly 15%.

It will be particularly interesting to see what the future holds for KNIME and RapidMiner. The companies were two of only four chosen by the Gartner Group as having both a complete vision of the future and the ability to execute that vision (Fig. 7a). Until recently, both were free and open source. RapidMiner then started charging for its current version, leaving its older version as the only free one. Recent offers to make it free for academic use don’t include use on projects with grant funding, so I expect KNIME’s higher rate of growth to remain faster than RapidMiner’s. However, in absolute terms, scholarly use of RapidMiner is currently almost twice that of KNIME, as shown in Fig. 2b.

Google Scholar Finds Far More SPSS Articles; Analytics Forecast Updated

Only last August I wrote that among scholars, the use of R had probably exceeded that of SPSS to become their most widely used software for analytics. That forecast was based on Google Scholar searches focused on one year at a time, from 1995 through 2014. Each year from 2010 through 2014, I re-collected that entire data set just in case Google changed the search algorithm enough to affect the overall pattern. The data stayed roughly the same for those four years, but Google Scholar now finds almost twice as many articles for SPSS (at its peak year of 2008) than it found last year and 12% more articles for SAS. Changes in search results for articles that used R varied slightly with fewer in the early years and more in the latter ones. So R did not become the most widely used analytics software among academics in 2014. It’s unlikely to become so for another two years, unless present trends change.

So what happened? We’re looking back across many years, so while it’s possible that SPSS suddenly became much more popular in 2014, that could not account for lifting the whole trend line. It’s possible Google Scholar improved its algorithm to find articles that existed previously. It’s also possible that new journal archives have opened themselves up to being indexed by Google. Why would it affect SPSS more than SAS or R? SPSS is menu-driven so it’s easy to install with its menus and dialog boxes translated into many languages. Since SAS and R are much more frequently used via their English-based languages, they may not be as popular in non-English speaking countries. Therefore, one might see a disproportionate impact on SPSS by new non-English archives becoming available. If you have an alternate hypothesis, please leave it in the comments below.

The remainder of this post is the complete updated section on this topic from The Popularity of Data Analysis Software:

Scholarly Articles

The more popular a software package is, the more likely it will appear in scholarly publications as a topic or as a tool of analysis. The software that is used in scholarly articles is what the next generation of analysts will graduate knowing, so it’s a leading indicator of where things are headed. Google Scholar offers a way to measure such activity. However, no search of this magnitude is perfect and will include some irrelevant articles and reject some relevant ones. The details of the search terms I used are complex enough to move to a companion article, How to Search For Data Science Articles. Since Google regularly improves its search algorithm, each year I re-collect the data for all years.

Figure 2a shows the number of articles found for each software package for the most recent complete year, 2014. SPSS is by far the most dominant package, likely due to its balance between power and ease-of-use. SAS has around half as many, followed by MATLAB and R. The software from Java through Statgraphics show a slow decline in usage from highest to lowest. Note that the general purpose software C, C++, C#, MATLAB, Java and Python are included only when found in combination with analytics terms, so view those as much rougher counts than the rest.

Fig_2a_ScholarlyImpact2014
Figure 2a. The number of scholarly articles that use each software package during the most recent complete year, 2014.

From RapidMiner on down, the counts appear to be zero. That’s not the case, the counts are just very low compared to the more popular packages, used in tens of thousands articles. Figure 2b shows the software only for those packages that have fewer than 825 articles (i.e. the bottom part of Fig. 2a), so we can see how they compare. RapidMiner, KNIME, SPSS Modeler and SAS Enterprise Miner are packages that all use the powerful and easy-to-use workflow interface, but their use has not yet caught on among scholars. BMDP is one of the oldest packages in existence. Its use has been declining for many years, but it’s still hanging in there. The software in the bottom half of this figure contain the newcomers, with the notable exception of Megaputer, whose Polyanalyst software has been around for many years now.

Fig_2b_ScholarlyImpact2014
Figure 2b. The number of scholarly articles for software that was used by fewer than 825 scholarly articles (i.e. the bottom part of Fig. 2a, rescaled.)

I’m particularly interested in the long-term trends of the classic statistics packages. So in Figure 2c I’ve plotted the same scholarly-use data for 1995 through 2014, the last complete year of data when this graph was made. As in Figure 2a, SPSS has a clear lead, but now you can see that its dominance peaked in 2008 and its use is in sharp decline. SAS never came close to SPSS’ level of dominance, and it also peaked around 2008. Note that the decline in the number of articles that used SPSS or SAS is not balanced by the increase in the other software shown. This is likely due to the fact that those two leaders faced increasing competition from many more software packages than can be shown in this type of graph (such as those shown in Figure 2a).

Fig_2a_ScholarlyImpactBig6
Figure 2c. The number of scholarly articles found in each year by Google Scholar. Only the top six “classic” statistics packages are shown.

Since SAS and SPSS dominate the vertical space in Figure 2c by such a wide margin, I removed those two packages and added the next two most popular statistics packages, Systat and JMP, with the result shown in Figure 2d. Freeing up so much space in the plot now allows us to see that the use of R is experiencing very rapid growth and is pulling away from the pack, solidifying its position in third place. If the current trends continue, the use of R may pass that of SPSS and SAS by the end of 2016. Note that current trends have shifted before as discussed here.

Stata has moved into fourth place, crossing above Statistica in 2014. The growth in the use of Stata is more rapid than all the classic statistics packages except for R. The use of Statistica, Minitab, Systat and JMP are next in popularity, respectively, with their growth roughly parallel to one another. [Note that in the plots from previous years, Statistica was displayed as a flat line at the very bottom of the graph. That turned out to be a search-related artifact. Many academics who use Statistica don’t mention the package by software name but rather say something like, “we used the statistics package by Statsoft.”]

Fig_2a_ScholarlyImpactLittle6
Figure 2d. The number of scholarly articles found in each year by Google Scholar for classic statistics packages after market leaders SPSS and SAS have been removed.

I’ll announce future update on Twitter, where you can follow me as @BobMuenchen.

 

It’s Analytics Survey Time!

Every other year Rexer Analytics surveys Data Analysts, Predictive Modelers, Data Scientists, Data Miners, and all other types of analytic professionals, students, and academics regarding the software they use.  I then update the main results in The Popularity of Data Analysis Software. Here’s where you can take the survey:

Survey Link:  www.rexeranalytics.com/Data-Miner-Survey-2015-Intro.html

Access Code:  R8M4E2  

Survey results will be unveiled at the Fall-2015 Boston Predictive Analytics World event.

Rexer Analytics has been conducting the Data Miner Survey since 2007.  Each survey explores the analytic behaviors, views and preferences of data miners and analytic professionals.  Over 1,200 people from around the globe participated in the 2013 survey.  Summary reports (40 page PDFs) from previous surveys are available FREE to everyone who requests them by emailing DataMinerSurvey@RexerAnalytics.com.  Also, highlights of earlier Data Miner Surveys are available at www.rexeranalytics.com/Data-Miner-Survey-Results-2013.html, including best practices shared by respondents on analytic success measurement, overcoming data mining challenges, and other topics.  The FREE Summary Report for this 2015 Data Miner Survey will be available to everyone Fall-2015.

Please help spread the word.

Rexer Analytics is a consulting firm focused on providing data mining and analytic CRM solutions.  Recent solutions include customer loyalty analyses, customer segmentation, predictive modeling to predict customer attrition and to target direct marketing, fraud detection, sales forecasting, market basket analyses, and complex survey research.  More information is available at www.RexerAnalytics.com or by calling +1 617-233-8185.

Helping Your Organization Migrate to R

by Bob Muenchen

As the R programming environment has grown in capability and popularity, so have the number of organizations planning to migrate to it from proprietary tools. I’ve helped members of various organizations transition from SAS, SPSS and/or Stata to R (see Workshop Participants), and the process typically involves the following steps:

1) Begin with the most important question: who should you migrate to R? Learning R is not a trivial task (see Why R is Hard to Learn). However, once mastered by people who use it regularly, I think it’s easier to use than other software. But if you have some people who use something like SAS only occasionally and view it as hard to use, you might consider getting them something other than R. Menu-based solutions such as SPSS or
R Commander may be a better fit for them. If they want to continue using SAS while lowering your licensing costs, you might consider the SAS implementation used in WPS (see World Programming).

2) —Motivate people to migrate. Discussing your current software budget may help. Showing your staff the growth of R’s capabilities and popularity may also help (see The Popularity of Data Analysis Software). Keep in mind that attempting to motivate people to change by criticizing their current choice is likely to backfire. People’s choice of software is very personal and criticizing it is like telling them they have the wrong religion.

—3) Use training & documentation that leverages what they already know, that speaks their language. A trainer who knows both your existing environment and R can convert what the analysts currently know rather than simply starting from scratch (note that this self-serving advice!) There are two parts to this process: learning the new R code and learning to interpret the new R output. Choosing to use R packages that provide output similar to your current software choice will help smooth the transition.  Good sources for training are listed here: my own on site training, Training and Consulting Partners – RStudio and here: R for SAS, SPSS and STATA Users | DataCamp. Books that help with the conversion process include:

—R for SAS and SPSS Users, Muenchen
—SAS and R, Kleinman & Horton
—R for Stata Users, Muenchen & Hilbe
—R Through Excel, Heiberger & Neuwirth (for those who use the SAS Excel plug-in)

4) —Provide in-house tech support. Before training a whole team, get one in-house expert trained to act as a consultant to others. Make sure this person is well known by everyone and has time freed up to provide help.

—5) Match your staff’s current work style, work flow and output. This is a particularly complex topic. Some examples: if your people are running SAS from the Excel plug-in, get the the R plug-in; if they’re using Enterprise Miner, consider a similar interface that controls R such as the KNIME Analytics Platform. If Microsoft Word is their main word processor, don’t complicate the conversion by switching them to LaTeX text processor at the same time (LaTeX is wonderful and very popular among R users, but it’s a mess to learn that and R at the same time).  Instead, use an approach that generates Word output.

—6) Migrate one step at a time if possible. For example, if you use SAS/ETS for forecasting, consider replacing just that one piece. When finished and successful, choose the next product, saving SAS/Base for last.

7) —Convert your programs or use conversion services. If your programs are all in production, this could be a huge job. However, if you mostly use SAS for new research tasks, you may not need to convert old code from which you just needed a solution.  Be careful to avoid line-by-line conversion; think like R (e.g. avoid for- and while-loops in R). When using external conversion services, make sure to involve your own staff in the process so you don’t end up with code that’s almost impossible to maintain.

I have found that following these steps helps during conversions to R. It’s a big job, though, so allocate plenty of time to it. Good luck!

 

r4stats.com 2014 in review

The WordPress.com stats helper monkeys prepared a 2014 annual report for this blog.

Here’s an excerpt:

The Louvre Museum has 8.5 million visitors per year. This blog was viewed about 260,000 times in 2014. If it were an exhibit at the Louvre Museum, it would take about 11 days for that many people to see it.

Click here to see the complete report.

SAS is #1…In Plans to Discontinue Use

I’ve been tracking The Popularity of Data Analysis Software for many years now, and a clear trend is the decline of the market share of the bigger analytics firms, notably SAS and SPSS. Many people have interpreted my comments as implying the decline in the revenue of those companies. But the fields involved in analytics (statistics, data mining, analytics, data science, etc.) have been exploding in popularity, so having a smaller slice of a much bigger pie still leaves billions in revenue for the big players.

Each year, the Gartner Group, “the world’s leading information technology research and advisory company”, collects data in a survey of the customers of 42 business intelligence firms. They recently released the data on the customers’ plans to discontinue use of their current software in one to three years. The results are shown in the figure below. Over 16% of the SAS Institute customers surveyed reported considering discontinuing their use of the software, the highest of any of the vendors shown. It will be interesting to see if this will actually lead to an eventual decline in revenue. Although I have helped quite a few organizations migrate from SAS to R, I would be surprised to see SAS Institute’s revenue decline. They offer excellent software and service which I still use, though not anywhere near as much as R.

The full Gartner report is available here.

SAS Attrition Plot

Learning R: Live Webinar, Interactive Self-Paced, or Site Visit?

My recent blog post, Why R is Hard to Learn, must have hit a nerve as it was read by over 6,000 people in its first two days online.  If you’re using R to augment your work in SAS, SPSS or Stata or you’re considering switching to R, my workshops can help minimize many of those headaches by pointing out the commands and options that frustrate users of those packages the most. Also find out which of the thousands of R packages will give you the output you’re most used to.

My next two live webinars done in partnership with Revolution Analytics are in January:
R for SAS, SPSS and Stata Users
Managing Data with R (updated to include dplyr, broom, tidyr, etc.)
Course outlines and registration for both is here.

My R for SAS, SPSS and Stata Users workshop is also now available as a self-paced interactive video workshop at DataCamp.com.

I do site visits in partnership with RStudio.com, whose software I recommend and use in every form of my training.  If your company does its training through Xerox Learning Services, I also partner with them. For further details or to arrange a site visit, you can reach me at muenchen.bob@gmail.com.

Updated: Why R is Hard to Learn

I’ve updated one of my most widely read blog posts, Why R is Hard to Learn. It focuses on the aspects of R which tend to trip up beginners. The new version is over twice as long as the original and it is located under the Articles menu, making it easier to find. Of course my new interactive workshop on DataCamp.com and my up-coming webinars with Revolution Analtyics cover these trouble spots thoroughly.

 

Adding the SPSS MEAN.n Function to R

SPSS contains a very useful set of functions that R lacks. If you’re lucky enough to have access to SPSS, you can use SPSS and R very well together. If not, it’s easy to add these functions to R. The functions perform calculations across values within each observation. Rather than limit you to removing missing values or not, they let you specify how many valid values you want before setting the result to missing. For example in SPSS,
MEAN.5(Q1 TO Q10) asks for the mean only if at least five of the ten variables have valid values. Otherwise the result will be a missing value. This “.n” extension is also available for SPSS’ SUM, SD, VARIANCE, MIN and MAX functions.

Let’s now take a look at how to do this in R. First we’ll create some data with different numbers of missing values for each observations.

> q1 <- c(1, 1, 1)
> q2 <- c(2, 2, NA)
> q3 <- c(3, NA, NA)
> df <- data.frame(q1, q2, q3)
> df
q1 q2  q3
1 1 2   3
2 1 2  NA
3 1 NA NA

R already has a mean function, but it lacks a function to count the number of valid values. A common way to do this in R is to use the is.na() function to generate a vector of TRUE/FALSE values for missing or not, respectively, then sum them. As with many software packages, R views TRUE as having the value 1 and FALSE as having a value of 0, so this approach gets us the number of missing values. The “!” symbol means “not” in R so !is.na() will find the number non-missing values. Here’s a function that does this:

> nvalid <- function(x) sum(!is.na(x))
> nvalid(q2)
[1] 2

So it has found that there are two valid values for q2. This nvalid() function obviously works on vectors, but we need to apply it to the rows of our data frame. We can select the first three variables using df[1:3] and then pass the result into as.matrix() to make the rows easily accessible by R’s apply() function. The apply() function’s second argument is 1 indicating that we would like to compute the mean across rows (the value 2 would indicate columns). The final arguments are the functions to apply and any arguments they need.

> means  <- apply(as.matrix(df[1:3]), 1, mean, na.rm = TRUE)
> counts <- apply(as.matrix(df[1:3]), 1, nvalid)
> means
[1] 2.0 1.5 1.0
> counts
[1] 3 2 1

We have our means and the counts of valid values, so all that remains is to choose our desired value of counts and accept the mean if the data have that value or greater, but return a missing value (NA) if not. This can be done using the ifelse() function, whose first argument is the logical condition, followed by the value desired when TRUE, then the value when false.

> means <- ifelse(counts >= 2, means, NA)
> means
[1] 2.0 1.5 NA

We’ve seen all the parts work, so all that remains is to put them together into a single function that has two arguments, one for the data frame and one for the n required.

mean.n   <- function(df, n) {
  means <- apply(as.matrix(df), 1, mean, na.rm = TRUE)
  nvalid <- apply(as.matrix(df), 1, function(df) sum(!is.na(df)))
  ifelse(nvalid >= n, means, NA)
}

Let’s test our function requiring 1, 2 and 3 valid values.

> df$mean1 <- mean.n(df[1:3], 1)
> df$mean2 <- mean.n(df[1:3], 2)
> df$mean3 <- mean.n(df[1:3], 3)
> df
q1 q2  q3 mean1 mean2 mean3
1 1 2   3   2.0   2.0     2
2 1 2  NA   1.5   1.5    NA
3 1 NA NA   1.0   NA     NA

That looks good. You could apply this same idea to various other R functions such as sd() or var(). You could also apply it to sum() as SPSS does, but I rarely do that. If you were creating a scale score from a set of survey Likert items measuring agreement and a person replied “strongly agree” (a value of 5), to only half the items but skipped the others, would you want the resulting score to be a neutral value as the sum would imply, or “strongly agree” as the mean would indicate? The mean makes much more sense in most situations. Be careful though as there are standardized tests that require use of the sum.

If you’re an SPSS user looking to learn just enough R to use the two together, you might want to read this, or to learn more you could take one of my workshops. If you really want to dive into the details, you might consider reading my book, R for SAS and SPSS Users.