Uncategorized | r4stats.com

Group-By Modeling in R Made Easy

There are several aspects of the R language that make it hard to learn, and repeating a model for groups in a data set used to be one of them. Here I briefly describe R’s built-in approach, show a much easier one, then refer you to a new approach described in the superb book, R for Data Science, by Hadley Wickham and Garrett Grolemund.

For ease of comparison, I’ll use some of the same examples in that book. The gapminder data set contains a few measurements for countries around the world every five years from 1952 through 2007.

> library("gapminder")
> gapminder

# A tibble: 1,704 × 6
 country continent year lifeExp pop gdpPercap
 <fctr> <fctr> <int> <dbl> <int> <dbl>
1 Afghanistan Asia 1952 28.801 8425333 779.4453
2 Afghanistan Asia 1957 30.332 9240934 820.8530
3 Afghanistan Asia 1962 31.997 10267083 853.1007
4 Afghanistan Asia 1967 34.020 11537966 836.1971
5 Afghanistan Asia 1972 36.088 13079460 739.9811
6 Afghanistan Asia 1977 38.438 14880372 786.1134
7 Afghanistan Asia 1982 39.854 12881816 978.0114
8 Afghanistan Asia 1987 40.822 13867957 852.3959
9 Afghanistan Asia 1992 41.674 16317921 649.3414
10 Afghanistan Asia 1997 41.763 22227415 635.3414
# ... with 1,694 more rows

Let’s create a simple regression model to predict life expectancy from year. We’ll start by looking at just New Zealand.

> library("tidyverse")
> nz <- filter(gapminder, 
+              country == "New Zealand")
> nz_model <- lm(lifeExp ~ year, data = nz)
> summary(nz_model)

Call:
lm(formula = lifeExp ~ year, data = nz)

Residuals:
 Min 1Q Median 3Q Max 
-1.28745 -0.63700 0.06345 0.64442 0.91192

Coefficients:
 Estimate Std. Error t value Pr(>|t|) 
(Intercept) -307.69963 26.63039 -11.55 4.17e-07 ***
year 0.19282 0.01345 14.33 5.41e-08 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 0.8043 on 10 degrees of freedom
Multiple R-squared: 0.9536, Adjusted R-squared: 0.9489 
F-statistic: 205.4 on 1 and 10 DF, p-value: 5.407e-08

If we had just a few countries, and we wanted to simply read the output (rather than processing it further) we could write a simple function and apply it using R’s built-in by() function. Here’s what that might look like:

my_lm <- function(df) {
  summary(lm(lifeExp ~ year, data = df))
}
by(gapminder, gapminder$country, my_lm)
...
----------------------------------------------- 
gapminder$country: Zimbabwe

Call:
lm(formula = lifeExp ~ year, data = df)

Residuals:
 Min 1Q Median 3Q Max 
-10.581 -4.870 -0.882 5.567 10.386 

Coefficients:
 Estimate Std. Error t value Pr(>|t|)
(Intercept) 236.79819 238.55797 0.993 0.344
year -0.09302 0.12051 -0.772 0.458

Residual standard error: 7.205 on 10 degrees of freedom
Multiple R-squared: 0.05623, Adjusted R-squared: -0.03814 
F-statistic: 0.5958 on 1 and 10 DF, p-value: 0.458

Since we have so many countries, that wasn’t very helpful. Much of the output scrolled out of sight (I’m showing only the results for the last one, Zimbabwe). But in a simpler case, that might have done just what you needed. It’s a bit more complex than how SAS or SPSS would do it since it required the creation of a function, but it’s not too difficult.

In our case, it would be much more helpful to save the output to a file for further processing. That’s when things get messy. We could use the str() function to study the structure of the output, then write another function to extract the pieces we need, then apply that function, then continue to process the result until we get what we finally end up with a useful data frame of results. Altogether, that is a lot of work! To make matters worse, what you learned from all that is unlikely to generalize to a different function. The output’s structure, parameter names, and so on, are often unique to each of R’s modeling functions.

Luckily, David Robinson made a package called broom that simplifies all that. It has three ways to “clean up” a model, each diving more deeply into its details. Let’s see what it does with our model for New Zealand.

> library("broom")
> glance(nz_model)

  r.squared adj.r.squared sigma statistic p.value df
1 0.9535846 0.9489431 0.8043472 205.4459 5.407324e-08 2

   logLik      AIC    BIC   deviance df.residual
1 -13.32064 32.64128 34.096 6.469743     10

The glance() function gives us information about the entire model, and it puts it into a data frame with just one line of output. As we head towards doing a model for each country, you can imagine this will be a very convenient format.

To get a bit more detail, we can use broom’s tidy() function to clean up the parameter-level view.

> tidy(nz_model)
        term     estimate  std.error statistic p.value
1 (Intercept) -307.699628 26.63038965 -11.55445 4.166460e-07
2 year           0.192821  0.01345258  14.33339 5.407324e-08

Now we have a data frame with two rows, one for each model parameter, but getting this result was just as simple to do as the previous example.

The greatest level of model detail is provided by broom’s augment() function. This function adds observation-level detail to the original model data:

> augment(nz_model)

 lifeExp year  .fitted   .se.fit     .resid
1 69.390 1952 68.68692 0.4367774 0.70307692
2 70.260 1957 69.65103 0.3814859 0.60897203
3 71.240 1962 70.61513 0.3306617 0.62486713
4 71.520 1967 71.57924 0.2866904 -0.05923776
5 71.890 1972 72.54334 0.2531683 -0.65334266
6 72.220 1977 73.50745 0.2346180 -1.28744755
7 73.840 1982 74.47155 0.2346180 -0.63155245
8 74.320 1987 75.43566 0.2531683 -1.11565734
9 76.330 1992 76.39976 0.2866904 -0.06976224
10 77.550 1997 77.36387 0.3306617 0.18613287
11 79.110 2002 78.32797 0.3814859 0.78202797
12 80.204 2007 79.29208 0.4367774 0.91192308

        .hat    .sigma      .cooksd .std.resid
1 0.29487179 0.8006048 0.2265612817 1.04093898
2 0.22494172 0.8159022 0.1073195744 0.85997661
3 0.16899767 0.8164883 0.0738472863 0.85220295
4 0.12703963 0.8475929 0.0004520957 -0.07882389
5 0.09906760 0.8162209 0.0402635628 -0.85575882
6 0.08508159 0.7194198 0.1302005662 -1.67338100
7 0.08508159 0.8187927 0.0313308824 -0.82087061
8 0.09906760 0.7519001 0.1174064153 -1.46130610
9 0.12703963 0.8474910 0.0006270092 -0.09282813
10 0.16899767 0.8451201 0.0065524741 0.25385073
11 0.22494172 0.7944728 0.1769818895 1.10436232
12 0.29487179 0.7666941 0.3811504335 1.35014569

Using those functions was easy. Let’s now get them to work repeatedly for each country in the data set. The dplyr package, by Hadly Wickham and Romain Francois, provides an excellent set of tools for group-by processing. The dplyr package was loaded into memory as part of the tidyverse package used above. First we prepare the gapminder data set by using the group_by() function and telling it what variable(s) make up our groups:

> by_country <- 
+   group_by(gapminder, country)

Now any other function in the dplyr package will understand that the by_country data set contains groups, and it will process the groups separately when appropriate. However, we want to use the lm() function, and that does not understand what a grouped data frame is. Luckily, the dplyr package has a do() function that takes care of that problem, feeding any function only one group at a time. It uses the period “.” to represent each data frame in turn. The do() function wants the function it’s doing to return a data frame, but that’s exactly what broom’s functions do.

Let’s repeat the three broom functions, this time by country. We’ll start with glance().

> do(by_country, 
+    glance( 
+       lm(lifeExp ~ year, data = .)))

Source: local data frame [142 x 12]
Groups: country [142]

      country r.squared adj.r.squared sigma
    <fctr>      <dbl>      <dbl>    <dbl>
1 Afghanistan 0.9477123     0.9424835 1.2227880
2 Albania     0.9105778     0.9016355 1.9830615
3 Algeria     0.9851172     0.9836289 1.3230064
4 Angola      0.8878146     0.8765961 1.4070091
5 Argentina   0.9955681     0.9951249 0.2923072
6 Australia   0.9796477     0.9776125 0.6206086
7 Austria     0.9921340     0.9913474 0.4074094
8 Bahrain     0.9667398     0.9634138 1.6395865
9 Bangladesh  0.9893609     0.9882970 0.9766908
10 Belgium    0.9945406     0.9939946 0.2929025

# ... with 132 more rows, and 8 more variables:
# statistic <dbl>, p.value <dbl>, df <int>,
# logLik <dbl>, AIC <dbl>, BIC <dbl>,
# deviance <dbl>, df.residual <int>

Now rather than one row of output, we have a data frame with one row per country. Since it’s a data frame, we already know how to manage it. We could sort by R-squared, or correct the p-values for the number of models done using p.adjust(), and so on.

Next let’s look at the grouped parameter-level output that tidy() provides. This will be the same code as above, simply substituting tidy() where glance() had been.

> do(by_country, 
+    tidy( 
+      lm(lifeExp ~ year, data = .)))

Source: local data frame [284 x 6]
Groups: country [142]

country term estimate std.error
    <fctr>      <chr>        <dbl>     <dbl>
1 Afghanistan (Intercept) -507.5342716 40.484161954
2 Afghanistan year           0.2753287  0.020450934
3 Albania     (Intercept) -594.0725110 65.655359062
4 Albania     year           0.3346832  0.033166387
5 Algeria     (Intercept) -1067.8590396 43.802200843
6 Algeria     year            0.5692797  0.022127070
7 Angola      (Intercept)  -376.5047531 46.583370599
8 Angola      year            0.2093399 0.023532003
9 Argentina   (Intercept)  -389.6063445 9.677729641
10 Argentina  year            0.2317084 0.004888791

# ... with 274 more rows, and 2 more variables:
# statistic <dbl>, p.value <dbl>

Again, this is a simple data frame allowing us to do whatever we need without learning anything new. We can easily search for models that contain a specific parameter that is significant. In our organization, we search through salary models that contain many parameters to see if gender is an important predictor (hoping to find none, of course).

Finally, let’s augment the original model data by adding predicted values, residuals and so on. As you might expect, it’s the same code, this time with augment() replacing the tidy() function.

> do(by_country, 
+ augment( 
+ lm(lifeExp ~ year, data = .)))
Source: local data frame [1,704 x 10]
Groups: country [142]

   country   lifeExp year .fitted  .se.fit
    <fctr>     <dbl> <int> <dbl>    <dbl>
1 Afghanistan 28.801 1952 29.90729 0.6639995
2 Afghanistan 30.332 1957 31.28394 0.5799442
3 Afghanistan 31.997 1962 32.66058 0.5026799
4 Afghanistan 34.020 1967 34.03722 0.4358337
5 Afghanistan 36.088 1972 35.41387 0.3848726
6 Afghanistan 38.438 1977 36.79051 0.3566719
7 Afghanistan 39.854 1982 38.16716 0.3566719
8 Afghanistan 40.822 1987 39.54380 0.3848726
9 Afghanistan 41.674 1992 40.92044 0.4358337
10 Afghanistan 41.763 1997 42.29709 0.5026799

# ... with 1,694 more rows, and 5 more variables:
# .resid <dbl>, .hat <dbl>, .sigma <dbl>,
# .cooksd <dbl>, .std.resid <dbl>

If we were to pull out just the results for New Zealand, we would see that we got exactly the same answer in the group_by result as we did when we analyzed that country by itself.

We can save that augmented data to a file to reproduce one of the residual plots from R for Data Science.

> gapminder_augmented <-
+ do(by_country, 
+   augment( 
+     lm(lifeExp ~ year, data = .)))
> ggplot(gapminder_augmented, aes(year, .resid)) +
+   geom_line(aes(group = country), alpha = 1 / 3) + 
+   geom_smooth(se = FALSE)

`geom_smooth()` using method = 'gam'

This plots the residuals of each country’s model by year by setting “group=country” then it follows it with a smoothed fit (geom_smooth) for all countries (blue line) by leaving out “group=country”. That’s a clever approach that I haven’t thought of before!

The broom package has done several very helpful things. As we have seen, it contains all the smarts needed to extract the important parts of models at three different levels of detail. It doesn’t just do this for linear regression though. R’s methods() function will show you what types of models broom’s functions are currently capable of handling:

> methods(tidy) 
 [1] tidy.aareg* 
 [2] tidy.acf* 
 [3] tidy.anova* 
 [4] tidy.aov* 
 [5] tidy.aovlist* 
 [6] tidy.Arima* 
 [7] tidy.betareg* 
 [8] tidy.biglm* 
 [9] tidy.binDesign* 
[10] tidy.binWidth* 
[11] tidy.boot* 
[12] tidy.brmsfit* 
[13] tidy.btergm* 
[14] tidy.cch* 
[15] tidy.character* 
[16] tidy.cld* 
[17] tidy.coeftest* 
[18] tidy.confint.glht* 
[19] tidy.coxph* 
[20] tidy.cv.glmnet* 
[21] tidy.data.frame* 
[22] tidy.default* 
[23] tidy.density* 
[24] tidy.dgCMatrix* 
[25] tidy.dgTMatrix* 
[26] tidy.dist* 
[27] tidy.ergm* 
[28] tidy.felm* 
[29] tidy.fitdistr* 
[30] tidy.ftable* 
[31] tidy.gam* 
[32] tidy.gamlss* 
[33] tidy.geeglm* 
[34] tidy.glht* 
[35] tidy.glmnet* 
[36] tidy.glmRob* 
[37] tidy.gmm* 
[38] tidy.htest* 
[39] tidy.kappa* 
[40] tidy.kde* 
[41] tidy.kmeans* 
[42] tidy.Line* 
[43] tidy.Lines* 
[44] tidy.list* 
[45] tidy.lm* 
[46] tidy.lme* 
[47] tidy.lmodel2* 
[48] tidy.lmRob* 
[49] tidy.logical* 
[50] tidy.lsmobj* 
[51] tidy.manova* 
[52] tidy.map* 
[53] tidy.matrix* 
[54] tidy.Mclust* 
[55] tidy.merMod* 
[56] tidy.mle2* 
[57] tidy.multinom* 
[58] tidy.nlrq* 
[59] tidy.nls* 
[60] tidy.NULL* 
[61] tidy.numeric* 
[62] tidy.orcutt* 
[63] tidy.pairwise.htest* 
[64] tidy.plm* 
[65] tidy.poLCA* 
[66] tidy.Polygon* 
[67] tidy.Polygons* 
[68] tidy.power.htest* 
[69] tidy.prcomp* 
[70] tidy.pyears* 
[71] tidy.rcorr* 
[72] tidy.ref.grid* 
[73] tidy.ridgelm* 
[74] tidy.rjags* 
[75] tidy.roc* 
[76] tidy.rowwise_df* 
[77] tidy.rq* 
[78] tidy.rqs* 
[79] tidy.sparseMatrix* 
[80] tidy.SpatialLinesDataFrame* 
[81] tidy.SpatialPolygons* 
[82] tidy.SpatialPolygonsDataFrame*
[83] tidy.spec* 
[84] tidy.stanfit* 
[85] tidy.stanreg* 
[86] tidy.summary.glht* 
[87] tidy.summary.lm* 
[88] tidy.summaryDefault* 
[89] tidy.survexp* 
[90] tidy.survfit* 
[91] tidy.survreg* 
[92] tidy.table* 
[93] tidy.tbl_df* 
[94] tidy.ts* 
[95] tidy.TukeyHSD* 
[96] tidy.zoo* 
see '?methods' for accessing help and source code
>

Each of those models contain similar information, but often stored in a completely different data structure and named slightly different things, even when they’re nearly identical. While that covers a lot of model types, R has hundreds more. David Robinson, the package’s developer, encourages people to request adding additional ones by opening an issue here.

I hope I’ve made a good case that doing group-by analyses in R can be done easily through the combination of dplyr’s do() function and broom’s three functions. That approach handles the great majority of group-by problems that I’ve seen in my 35-year career. However, if your needs are not met by this approach, then I encourage you to read Chapter 25 of R for Data Science (update: in the printed version of the book, it’s Chapter 20, Many Models with purrr and broom.) But as the chapter warns, it will “stretch your brain!”

If your organization is interested in a hands-on workshop that covers many similar topics, please drop me a line. Have fun with your data analyses!

Keeping Up with Your Data Science Options

The field of data science is changing so rapidly that it’s quite hard to keep up with it all. When I first started tracking The Popularity of Data Science Software in 2010, I followed only ten packages, all of them classic statistics software. The term data science hadn’t caught on yet, data mining was still a new thing. One of my recent blog posts covered 53 packages, and choosing them from a list of around 100 was a tough decision!

To keep up with the rapidly changing field, you can read the information on a package’s web site, see what people are saying on blog aggregators such as R-Bloggers.com or StatsBlogs.com, and if it sounds good, download a copy and try it out. What’s much harder to do is figure out how they all relate to one another. A helpful source of information on that front is the book Disruptive Analtyics, by Thomas Dinsmore.

I was lucky enough to be the technical reviewer for the book, during which time I ended up reading it twice. I still refer to it regularly as it covers quite a lot of material. In a mere 262 pages, Dinsmore manages to describe each of the following packages, how they relate to one another, and how they fit into the big picture of data science:

Alluxio
Alpine Data
Alteryx
APAMA
Apex
Arrow
Caffe
Cloudera
Deeplearning4J
Drill
Flink
Giraph
Hadoop
HAWQ
Hive
IBM SPSS Modeler
Ignite
Impala
Kafka
KNIME Analytics Platform
Kylin
MADLib
Mahout
MapR
Microsoft R Aerver
Phoenix
Pig
Python
R
RapidMiner
Samza
SAS
SINGA
Skytree Server
Spark
Storm
Tajo
Tensorflow
Tez
Theano
Trafodion

As you can tell from the title, a major theme of the book is how open source software is disrupting the data science marketplace. Dinsmore’s blog, ML/DL: Machine Learning, Deep Learning, extends the book’s coverage as data science software changes from week to week.

I highly recommend both the book and the blog. Have fun keeping up with the field!

Jobs for “Data Science” Up 7-fold, for “Statistician” Down by Half

The Bureau of Labor Statistics projects that jobs for statisticians will grow by 34% between 2014 and 2024. However, according to the nation’s largest job web site, the number of companies looking for “statisticians” is actually in sharp decline. Those jobs are likely being replaced by postings for “data scientists.”

I regularly monitor the Popularity of Data Science Software, and as an offshoot of that project, I collected data that helps us understand how the term “data science” is defined. I began by finding jobs that required expertise in software used for data science such as R or SPSS. I then examined the tasks that the jobs entailed, such as “analyze data,” and looked up jobs based only on one task at a time. I switched back and forth between searching for software and for the terms used to describe the jobs, until I had a comprehensive list of both. In the end, I had searched for over 50 software packages and over 40 descriptive terms or tasks. I had also skimmed thousands of job advertisements. (Additional details are here).

Search Terms	2/26/2017	2/17/2014	Ratio
Big Data	20,646	10,378	1.99
Data analytics	15,774	6,209	2.54
Machine learning	12,499	3,658	3.42
Statistical analysis	11,397	9,719	1.17
Data mining	9,757	7,776	1.25
Data Science	6,873	973	7.06
Quantitative analysis	4,095	3,365	1.22
Business analytics	4,043	2,867	1.41
Advanced Analytics	3,479	1,497	2.32
Data Scientist	3,272	974	3.36
Statistical software	2,835	2,102	1.35
Predictive analytics	2,411	1,497	1.61
Artificial intelligence	2,404	794	3.03
Predictive modeling	2,264	1,804	1.25
Statistical modeling	2,040	1,462	1.40
Quantitative research	1,837	1,380	1.33
Research analyst	1,756	1,722	1.02
Statistical tools	1,414	1,121	1.26
Statistician	904	1,711	0.53
Statistical packages	784	559	1.40
Survey research	440	559	0.79
Quantitative modeling	352	322	1.09
Statistical research	208	174	1.20
Statistical computing	153	108	1.42
Research computing	133	97	1.37
Statistical analyst	125	141	0.89
Data miner	34	19	1.79

Many terms were used outside the realm of data science. Other terms were used both in data science jobs and in jobs that require little analytic skill. Terms that could not be used to specifically find data science jobs were: analytics, data visualization, graphics, data graphics, statistics, statistical, survey, research associate, and business intelligence. One term, econometric(s), required deep analytical skills, but was too focused on one field.

The search terms that were well-focused on data science, but not overly focused in a single field are listed in the following table. The table is sorted by the number of jobs found on Indeed.com on February 26, 2017. While each column displays counts taken on a single day, the large size of Indeed.com’s database of jobs keeps its counts stable. The correlation between the logs of the two counts is quite strong, r=.95, p= 4.7e-14.

During this three-year period, the overall unemployment rate dropped from 6.7% to 4.7%, indicating a period of job growth for most fields. Three terms grew very rapidly indeed with “data science” growing 7-fold, and both “data scientist” and “artificial intelligence” tripling in size. The biggest surprise was that the use of the term “statistician” took a huge hit, dropping to only 53% of its former value.

That table covers a wide range of terms, but only on two dates. What does the long-term trend look like? Indeed.com has a trend-tracking page that lets us answer that question. The figure below shows solid the growth in the percentage of advertisements that used the term “data scientist” (blue, top right), while those using the term “statistician” (yellow, lower right) are steadily declining.

The plot on the company’s site is interactive (the one shown here is not) allowing me to see that the most recent data points were recorded on December 27, 2016. On that date, the percentage of jobs for data scientist were 474% of those for statistician.

As an accredited professional statistician, am I worried about this trend? Not at all. Statistical analysis software has broadened its scope to include many new capabilities including: machine learning, artificial intelligence, Structured Query Language, advanced visualization techniques, interfaces to Python, R, and Apache Spark. The software has changed because the job known as “statistician” has changed. Statisticians aren’t going away, their jobs are evolving into what we now know as data science. And that field is growing quite nicely!

Data Science Job Report 2017: R Passes SAS, But Python Leaves Them Both Behind

I’ve just updated another section of The Popularity of Data Science Software. It is reproduced below to save you the trouble of reading the entire article. If your organization is looking for training in the R language, you might consider my books, R for SAS and SPSS Users or R for Stata Users, or my on-site workshops.

Job Advertisements

One of the best ways to measure the popularity or market share of software for data science is to count the number of job advertisements for each. Job advertisements are rich in information and are backed by money so they are perhaps the best measure of how popular each software is now. Plots of job trends give us a good idea of what is likely to become more popular in the future.

Indeed.com is the biggest job site in the U.S., making its collection the best around. As their co-founder and former CEO Paul Forster stated, Indeed.com includes “all the jobs from over 1,000 unique sources, comprising the major job boards – Monster, Careerbuilder, Hotjobs, Craigslist – as well as hundreds of newspapers, associations, and company websites.” Indeed.com also has superb search capabilities and it includes a tool for tracking long-term trends.

Searching for jobs using Indeed.com is easy, but searching for software in a way that ensures fair comparisons across packages is tricky. Some software is used only for data science (e.g. SPSS, Apache Spark) while others are used in data science jobs and more broadly in report-writing jobs (e.g. SAS, Tableau). General-purpose languages (e.g. C, Java) are heavily used in data science jobs, but the vast majority of jobs that use them have nothing to do with data science. To level the playing field I developed a protocol to focus the search for each software within only jobs for data scientists. The details of this protocol are described in a separate article, How to Search for Data Science Jobs. All of the graphs in this section use those procedures to make the required queries.

I collected the job counts discussed in this section on February 24, 2017. One might think that a sample of on a single day might not be very stable, but the large number of job sources makes the counts in Indeed.com’s collection of jobs quite consistent. The last time I collected this data was February 20, 2014, and those that were collected using the same protocol (the general purpose languages) yielded quite similar results. They grew between 7% and 11%, and correlated r=.94, p=.002.

Figure 1a shows that SQL is in the lead with nearly 18,000 jobs, followed by Python and Java in the 13,000’s. Hadoop comes next with just over 10,000 jobs, then R, the C variants, and SAS. (The C, C++, and C# are combined in a single search since job advertisements usually seek any of them). This is the first time this report has shown more jobs for R than SAS, but keep in mind these are jobs specific to data science. If you open up the search to include jobs for report writing, you’ll find twice as many SAS jobs.

Next comes Apache Spark, which was too new to be included in the 2014 report. It has come a long way in an incredibly short time. For a detailed analysis of Spark’s status, see Spark is the Future of Analytics, by Thomas Dinsmore.

Tableau follows, with around 5,000 jobs. The 2014 report excluded Tableau due to its jobs being dominated by report writing. Including report writing will quadruple the number of jobs for Tableau expertise to just over 2o,ooo.

Figure 1a. The number of data science jobs for the more popular software (those with 250 jobs or more, 2/2017).

Apache Hive is next, with around 3,900 jobs, then a very diverse set of software comes next, with Scala, SAP, MATLAB, and SPSS, each having just over 2,500 data science jobs. After those, we see a slow decline from Teradata on down.

Much of the software had fewer than 250 job listings. When displayed on the same graph as the industry leaders, their job counts appear to be zero; therefore I have plotted them separately in Figure 1b. Alteryx comes out the leader of this group with 240 jobs. Microsoft was a difficult search since it appears in data science ads that mention other Microsoft products such as Windows or SQL Server. To eliminate such over-counting, I treated Microsoft different from the rest by including product names such as Azure Machine Learning and Microsoft Cognitive Toolkit. So there’s a good chance I went from over-emphasizing Microsoft to under-emphasizing it with only 157 jobs.

Figure 1b. The number of analytics jobs for the less popular software (under 250 jobs, 2/2017).

Next comes the fascinating new high-performance language Julia. I added FORTRAN just for fun and was surprised to see it still hanging in there after all these years. Apache Flink is also in this grouping, which all have around 125 jobs.

H2O follows, with just over 100 jobs.

I find it fascinating that SAS Enterprise Miner, RapidMiner, and KNIME appear with a similar number of jobs (around 90). Those three share a similar workflow user interface that make them particularly easy to use. The companies advertise the software as not needing much training, so it may be possible that companies feel little need to hire expertise if their existing staff picks it up more easily. SPSS Modeler also uses that type of interface, but its job count is about half that of the others, at 50 jobs.

Bringing up the rear is Statistica, which was sold to Dell, then sold to Quest. Its 36 jobs trails far behind its similar competitor, SPSS, which has a staggering 74-fold job advantage.

The open source MXNet deep learning framework, shows up next with 34 jobs. Tensorflow is a similar project with a 12-fold job advantage, but these two are both young enough that I expect both will be growing rapidly in the future.

In the final batch that has few, if any, jobs, we see a few newcomers such as DataRobot and Domino Data Labs. Others have been around for years, leaving us to wonder how they manage to stay afloat given all the competition.

It’s important to note that the values shown in Figures 1a and 1b are single points in time. The number of jobs for the more popular software do not change much from day to day. Therefore the relative rankings of the software shown in Figure 1a is unlikely to change much over the coming year. The less popular packages shown in Figure 1b have such low job counts that their ranking is more likely to shift from month to month, though their position relative to the major packages should remain more stable.

Each software has an overall trend that shows how the demand for jobs changes across the years. You can plot these trends using Indeed.com’s Job Trends tool. However, as before, focusing just on analytics jobs requires carefully constructed queries, and when comparing two trends at a time, they both have to fit in the same query limit. Those details are described here.

I’m particularly interested in trends involving R so let’s see how it compares to SAS. In Figure 1c we see that the number of data science jobs for SAS has remained relatively flat from 2012 until February 28, 2017 when I made this plot. During that same period, jobs for R grew steadily and finally surpassed jobs for SAS in early 2016. As noted in a blog post (and elsewhere in this report), use of R in scholarly publications surpassed those for SAS in 2015.

: Figure 1c. Data science job trends for R (blue) and SAS (orange).

A long-standing debate has been taking place on the Internet regarding the relative place of Python and R. Ironically, this debate about data science software has involved very little actual data. However, it is possible now to at least study the job trends. Figure 1a showed us that Python is well out in front of R, at least on that single day the searches were run. What has the data looked like over time? The answer is shown in Figure 1d.

Figure 1d. Jobs trends for R (blue & lower) and Python (orange & upper).

As we see, Python surpassed R in terms of data science jobs back in 2013. These are, of course, very different languages and a quick scan of job descriptions will show that the R jobs are much more focused on the use of existing methods of analysis, while the Python jobs have more of a custom-programming angle to them.

Gartner’s 2017 Take on Data Science Software

In my ongoing quest to track The Popularity of Data Analysis Software, I’ve finally decided to change the title to use the newer term “data science”. The 2017 version of Gartner’s Magic Quadrant for Data Science Platforms was just published, so I have updated my IT Research Firms section, which I repeat here to save you from having to dig through the entire 40+ page tome. If your organization is looking for training in the R language, you might consider my books, R for SAS and SPSS Users or R for Stata Users, or my on-site workshops.

IT Research Firms

IT research firms study software products and corporate strategies, they survey customers regarding their satisfaction with the products and services, and then provide their analysis on each in reports they sell to their clients. Each company has its own criteria for rating companies, so they don’t always agree. However, I find the reports extremely interesting reading. While these reports are expensive, the companies that receive good ratings often purchase copies to give away to potential customers. An Internet search of the report title will often reveal the companies that are distributing such copies.

Gartner, Inc. is one of the companies that provides such reports. Out of the roughly 100 companies selling data science software, Gartner selected 16 which had either high revenue or lower revenue but high growth (see full report for details). After extensive input from both customers and company representatives, Gartner analysts rated the companies on their “completeness of vision” and their “ability to execute” that vision. Figure 3 shows the resulting plot. Note that purely open source software is not rated by Gartner, but nearly all the software in Figure 3 includes the ability to interact with R and Python.

The Leader’s Quadrant is the place for companies who have a future direction in line with their customer’s needs and the resources to execute that vision. The four companies in the Leaders quadrant have remained the same for the last three reports: IBM, KNIME, RapidMiner, and SAS. Of these, they rate IBM as having slightly greater “completeness of vision” due to the extensive integration they offer to open source software compared to SAS Institute. KNIME and RapidMiner are quite similar as the are driven by an easy to use workflow interface. Both offer free and open source versions, but RapidMiner’s is limited by a cap on the amount of data that it can analyze. IBM and SAS are market leaders based on revenue and, as we have seen, KNIME and RapidMiner are the ones with high growth.

Figure 3a. Gartner Magic Quadrant for Data Science Platforms

The companies in the Visionaries quadrant are those that have a good future plans but which may not have the resources to execute that vision. Of these, Microsoft increased its ability to execute compared to the 2016 report, and Alpine, one of the smallest companies, declined sharply in their ability to execute. The remaining three companies in this quadrant have just been added: H2O.ai, Dataiku, and Domino Data Lab.

Those in the Challenger’s quadrant have ample resources but less customer confidence on their future plans. Mathworks, the makers of MATLAB, is new to the report. Quest purchased Statistica from Dell, and it appears in roughly the same position as Dell did last year.

The Niche Players quadrant offer tools that are not as broadly applicable.

In 2017 Gartner dropped coverage of Accenture, Lavastorm, Megaputer, Predixion Software, and Prognoz.

Knoxville, TN: R for Text Analysis Workshop

The Knoxville R Users Group is presenting a workshop on text analysis using R by Bob Muenchen. The workshop is free and open to the public. You can join the group at https://www.meetup.com/Knoxville-R-Users-Group. A description of the workshop follows.

Seeking Cloud

R for Text Analysis

When analyzing text using R, it’s hard to know where to begin. There are 37 packages available and there is quite a lot of overlap in what they can do. This workshop will demonstrate how to do three popular approaches: dictionary-based content analysis, latent semantic analysis, and latent Dirichlet allocation. We will spend much of the time on the data preparation steps that are important to all text analysis methods including data acquisition, word stemming/lemmatization, removal of punctuation and other special characters, phrase discovery, tokenization, and so on. While the examples will focus on the automated extraction of topics in the text files, we will also briefly cover the analysis of sentiment (e.g. how positive is customer feedback?) and style (who wrote this? are they telling the truth?)

The results of each text analysis approach will be the topics found, and a numerical measure of each topic in each document. We will then merge that with numeric data and do analyses combining both types of data.

The R packages used include quanteda, lsa, topicmodels, tidytext and wordcloud; with brief coverage of tm and SnowballC. While the workshop will not be hands-on due to time constraints, the programs and data files will be available afterwards.

Where: University of Tennessee Humanities and Social Sciences Building, room 201. If the group gets too large, the location may move and a notice will be sent to everyone who RSVPs on Meetup.com or who registers at the UT workshop site below. You can also verify the location the day before via email with Bob at muenchen@utk.edu.

When: 9:05-12:05 Friday 1/27/17

Prerequisite: R language basics

Members of UT Community register at: http://workshop.utk.edu under Researcher Focused

Members of other groups please RSVP on your respective sites so I can bring enough handouts.

R Passes SAS in Scholarly Use (finally)

Way back in 2012 I published a forecast that showed that the use of R for scholarly publications would likely pass the use of SAS in 2015. But I didn’t believe the forecast since I expected the sharp decline in SAS and SPSS use to level off. In 2013, the trend accelerated and I expected R to pass SAS in the middle of 2014. As luck would have it, Google changed their algorithm, somehow finding vast additional quantities of SAS and SPSS articles. I just collected data on the most recent complete year of scholarly publications, and it turns out that 2015 was indeed the year that R passed SAS to garner the #2 position. Once again, models do better than “expert” opinion! I’ve updated The Popularity of Data Analysis Software to reflect this new data and include it here to save you the trouble of reading the whole 45 pages of it.

If you’re interested in learning R, you might consider reading my books R for SAS and SPSS Users, or R for Stata Users. I also teach workshops on R, but I’m currently booked through mid October, so please plan ahead.

Scholarly Articles

Scholarly articles are also rich in information and backed by significant amounts of effort. The more popular a software package is, the more likely it will appear in scholarly publications as an analysis tool or even an object of study. 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; each 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, 2015. SPSS is by far the most dominant package, as it has been for over 15 years. This may be due to its balance between power and ease-of-use. For the first time ever, R is in second place with around half as many articles. Although now in third place, SAS is nearly tied with R. Stata and MATLAB are essentially tied for fourth and fifth place. Starting with Java, usage slowly tapers off. Note that the general-purpose software C, C++, C#, MATLAB, Java, and Python are included only when found in combination with data science terms, so view those as much rougher counts than the rest. Since Scala and Julia have a heavy data science angle to them, I cut them some slack by not adding any data science terms to the search, not that it helped them much!

From Spark 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 1,200 articles (i.e. the bottom part of Fig. 2a), so we can see how they compare. Spark and RapidMiner top out the list of these packages, followed by KNIME and BMDP. There’s a slow decline in the group that goes from Enterprise Miner to Salford Systems. Then comes a group of mostly relative new arrivals beginning with Microsoft’s Azure Machine Learning. A package that’s not a new arrival is from Megaputer, whose Polyanalyst software has been around for many years now, with little progress to show for it. Dead last is Lavastorm, which to my knowledge is the only commercial package that includes Tibco’s internally written version of R, TERR.

Fig_2b_ScholarlyImpact2015 — Figure 2b. The number of scholarly articles for software that was used by fewer than 1,200 scholarly articles (i.e. the bottom part of Fig. 2a, rescaled.)

Figures 2a and 2b are useful for studying market share as it is now, but they don’t show how things are changing. It would be ideal to have long-term growth trend graphs for each of the analytics packages, but collecting such data is too time consuming since it must be re-collected every year (since Google’s search algorithms change). What I’ve done instead is collect data only for the past two complete years, 2014 and 2015. Figure 2c shows the percent change across those years, 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 500 articles in 2014.

Figure 2c. 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 2c. Change in the number of scholarly articles using each software in the most recent two complete years (2014 to 2015). Packages shown in red are “hot” and growing, while those shown in blue are “cooling down” or declining.

Python is the fastest growing. Note that the Python figures are strictly for data science use as defined here. The open-source KNIME and RapidMiner are the second and third fastest growing, respectively. Both use the easy yet powerful workflow approach to data science. Figure 2b showed that RapidMiner has almost twice the marketshare of KNIME, but here we see use of KNIME is growing faster. That may be due to KNIME’s greater customer satisfaction, as shown in the Rexer Analytics Data Science Survey. The companies are two of only four chosen by IT advisory firm Gartner, Inc. as having both a complete vision of the future and the ability to execute that vision (Fig. 3a).

R is in fourth place in growth, and given its second place in overall marketshare, it is in an enviable position.

At the other end of the scale are SPSS and SAS, both of which declined in use by 25% or more. Recall that Fig. 2a shows that despite recent years of decline, SPSS is still extremely dominant for scholarly use. Hadoop use declined slightly, perhaps as people turned to alternatives Spark and H2O.

I’m particularly interested in the long-term trends of the classic statistics packages. So in Figure 2d I’ve plotted the same scholarly-use data for 1995 through 2015, 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 in this particular graph. However, if you add up all the other software shown in Figure 2a, you come close. There still seems to be a slight decline in people reporting the particular software tool they used.

Fig_2d_ScholarlyImpact — Figure 2d. 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 2d by such a wide margin, I removed those two curves, leaving only a single point of SAS usage in 2015. The the result is shown in Figure 2e. Freeing up so much space in the plot now allows us to see that the growth in the use of R is quite rapid and is pulling away from the pack (recall that the curve for SAS has a steep downward slope). If the current trends continue, R will cross SPSS to become the #1 software for scholarly data science use by the end of 2017. Stata use is also growing more quickly than the rest. Note that trends have shifted before as discussed here. The use of Statistica, Minitab, Systat and JMP are next in popularity, respectively, with their growth roughly parallel to one another.

Figure 2e. 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. — Figure 2e. The number of scholarly articles found in each year by Google Scholar for classic statistics packages after the curves for SPSS and SAS have been removed.

Using a logarithmic y-axis scales down the more popular packages, allowing us to see the full picture in a single image (Figure 2f.) This view makes it more clear that R use has passed that of SAS, and that Stata use is closing in on it. However, even when one studies the y-axis values carefully, it can be hard to grasp how much the logarithmic transformation has changed the values. For example, in 2015 value for SPSS is well over twice the value for R. The original scale shown in Figure 2d makes that quite clear.

Fig_2f_ScholarlyImpactLogs — Figure 2f. A logarithmic view of the number of scholarly articles found in each year by Google Scholar. This combines the previous two figures into one by compressing the y-axis with a base 10 logarithm.

Rexer Data Science Survey: Satisfaction Results

by Bob Muenchen

I previously reported on the initial results of Rexer Analytics’ 2015 survey of data science tools here. More results are now available, and the comprehensive report should be released soon. One of the more interesting questions on the survey was, “Please rate your overall satisfaction with [your previously chosen software].” Most of the measures I report in my regularly-updated article, The Popularity of Data Analysis Software are raw measures of usage, so it’s nice to have data that goes beyond usage and into satisfaction. The results are show in the figure below for the more popular software (other software had very small sample sizes and so are not shown).

Rexer-2015-Satisfaction — Results from the question, “Please rate your overall satisfaction with [your previously chosen software].” Only software with substantial number of responses shown.

People reported being somewhat satisfied with their chosen tool, which doesn’t come a much of a surprise. If they weren’t at least somewhat satisfied, they would be likely to move on to another tool. What really differentiated the tools was the percent of people who reported being extremely satisfied. The free and open source KNIME program came out #1 with 69% of its users being extremely satisfied. (KNIME is also the 2nd fastest growing data science package among scholarly researchers). IBM SPSS Modeler came in second with 60%, followed closely by R with 57%.

Both of the top two packages use the workflow user interface which has many advantages that I’ve written about here and here. However, RapidMiner and SAS Enterprise Miner also use the workflow interface, and their percent of extremely satisfied customers were less than half at 32% and 29%, respectively. We might wonder if people are more satisfied with KNIME because they’re using the free desktop version, but RapidMiner also has a free version, so cost isn’t a factor on that comparison.

Although both R and SAS have menu-based interfaces, they are predominantly programming languages. R has almost triple the number of extremely satisfied users, which may be the result of its being generally viewed as the more powerful language, albeit somewhat harder to learn. The fact that R is free while SAS is not may also be a factor in that difference.

R Training at Nicholls State University

I’ll be presenting two workshops back-to-back at Nicholls State University in Thibodaux Louisiana June 14-16. The first workshop will cover a broad range of R topics. Each topic will include a brief comparison to how R differs from the popular commercial data science packages, SAS, SPSS, and Stata. If you have a background in any of those packages, you’ll know the things that are likely to trip you up as you learn R. If not, you’ll just come away with a solid intro to R along with how it compares to other software.

The second workshop will focus on the broad range of data management tasks that are usually needed to prepare your data for analysis. Both workshops will be done using the very latest R packages to make things as easy as possible. These packages include tibble, dplyr, magrittr, stringr, lubridate, broom, and more.

Seats are still available at no charge with the Nicholls’ State community and other academics getting priority seating. To register for the workshop, please contact professor Allyse Ferrara, 985/448-4736, or allyse.ferrara@nicholls.edu.

R’s Growth Continues to Accelerate

Each year I update the growth in R’s capability on The Popularity of Data Analysis Software. And each year, I think R’s incredible rate of growth will finally slow down. Below is a graph of the latest data, and as you can see, R’s growth continues to accelerate.

Since I’ve added coverage for many more software packages, I have restructured the main article to reflect the value of each type of data. They now appear in this order:

Job Advertisements
Scholarly Articles
IT Research Firm Reports
Surveys of Use
Books
Blogs
Discussion Forum Activity
Programming Popularity Measures
Sales & Downloads
Competition Use
Growth in Capability

Growth in Capability remains last because I only have complete data for R. To save you from having to dig through all 40+ pages of the article, the updated section is below. I’ll be updating several other sections in the coming weeks. If you’re interested, you can follow this blog, or follow me on Twitter as @BobMuenchen.

If you haven’t yet learned R, I recommend my books R for SAS and SPSS Users and R for Stata Users. I do R training as well, but that’s booked up through the end of August, so please plan ahead.

Growth in Capability

The capability of analytics software has grown significantly over the years. It would be helpful to be able to plot the growth of each software package’s capabilities, but such data are hard to obtain. John Fox (2009) acquired them for R’s main distribution site http://cran.r-project.org/ for each version of R. To simplify ongoing data collection, I kept only the values for the last version of R released each year (usually in November or December), and collected data through the most recent complete year.

These data are displayed in Figure 10. The right-most point is for version 3.2.3, released 12/10/2015. The growth curve follows a rapid parabolic arc (quadratic fit with R-squared=.995).

Fig_9_CRAN — Figure 10. Number of R packages available on its main distribution site for the last version released in each year.

To put this astonishing growth in perspective, let us compare it to the most dominant commercial package, SAS. In version, 9.3, SAS contained around 1,200 commands that are roughly equivalent to R functions (procs, functions, etc. in Base, Stat, ETS, HP Forecasting, Graph, IML, Macro, OR, and QC). In 2015, R added 1,357 packages, counting only CRAN, or approximately 27,642 functions. During 2015 alone, R added more functions/procs than SAS Institute has written in its entire history.

Of course while SAS and R commands solve many of the same problems, they are certainly not perfectly equivalent. Some SAS procedures have many more options to control their output than R functions do, so one SAS procedure may be equivalent to many R functions. On the other hand, R functions can nest inside one another, creating nearly infinite combinations. SAS is now out with version 9.4 and I have not repeated the arduous task of recounting its commands. If SAS Institute would provide the figure, I would include it here. While the comparison is far from perfect, it does provide an interesting perspective on the size and growth rate of R.

As rapid as R’s growth has been, these data represent only the main CRAN repository. R has eight other software repositories, such as Bioconductor, that are not included in Fig. 10. A program run on 4/19/2016 counted 11,531 R packages at all major repositories, 8,239 of which were at CRAN. (I excluded the GitHub repository since it contains duplicates to CRAN that I could not easily remove.) So the growth curve for the software at all repositories would be approximately 40% higher on the y-axis than the one shown in Figure 10.

As with any analysis software, individuals also maintain their own separate collections available on their web sites. However, those are not easily counted.

What’s the total number of R functions? The Rdocumentation site shows the latest counts of both packages and functions on CRAN, Bioconductor and GitHub. They indicate that there is an average of 19.78 functions per package. Given the package count of 11,531, as of 4/19/2016 there were approximately 228,103 total functions in R. In total, R has approximately 190 times as many commands as its main commercial competitor, SAS.