intubate <||> Roberto Bertolusso
2016-07-26 - 2016-08-27 (GPL >= 2)
The aim of intubate
(<||>
) is to offer a painless way to
add R functions that that are not pipe-aware
to data science pipelines implemented by magrittr
with the
operator %>%
, without having to rely on workarounds. It also
offers three extensions (implemented but still considered experimental),
called intubOrders
, intuEnv
, and intuBags
.
Installation
- the latest released version from CRAN (0.99.3) with
install.packages("intubate")
- the latest development version from github with
# install.packages("devtools")
devtools::install_github("rbertolusso/intubate")
In a nutshell
If you like magrittr
pipelines (%>%) and you are looking
for an alternative to perfoming a statistical analysis
in the following way:
fit <- lm(sr ~ ., LifeCycleSavings)
summary(fit)
intubate
let's you do it in these other ways:
library(intubate)
library(magrittr)
## 1) Using interface (provided by intubate OR user defined)
LifeCycleSavings %>%
ntbt_lm(sr ~ .) %>% ## ntbt_lm is the interface
summary()
## 2) Calling the non-pipe-aware function directly using ntbt
LifeCycleSavings %>%
ntbt(lm, sr ~ .) %>% ## ntbt calls lm without need to define interface
summary()
intubate
currently implements 461 interfaces (below you will find a list of packages
containing the interfaced functions) that can be related to data science
methodologies and other disciplines. For now the focus is on interfacing
non-pipe-aware functions having "formula" and "data" (in that order),
but the non-formula variants should also work (even cases currently lacking interfaces).
As a proof of concept of this, two libraries that contain non-formula variants only (glmnet
and lars
) have also been interfaced.
Interfaces "on demand"
intubate
also allows to create your own interfaces "on demand",
right now, giving you full power of decision regarding which functions
to interface. It also lets you call the non-pipe-aware functions directly,
without the need of defining an interface.
The ability of being able to amplify the scope of intubate
may prove to be particularly welcome in case you are related to a particular
field that may, in the long run, continue to lack interfaces due to my
unforgivable, but unavoidable, ignorance.
As an example of creating an interface "on demand", suppose the interface to
cor.test
was lacking in the current version of intubate
and suppose
(only for a moment) that you want to
create yours because you are searching for a pipeline-aware alternative to
any of the following styles of coding (which, by the way, are perfectly fine
as they are and you shouldn't change yours if you like it and serves your purposes):
data(USJudgeRatings)
## 1)
cor.test(USJudgeRatings$CONT, USJudgeRatings$INTG)
## 2)
attach(USJudgeRatings)
cor.test(CONT, INTG)
detach()
## 3)
with(USJudgeRatings, cor.test(CONT, INTG))
## 4)
USJudgeRatings %>%
with(cor.test(CONT, INTG))
To be able to create an interface to cor.test
"on demand", the only thing you
need to do is to add the following line of code somewhere before its use
in your pipeline:
ntbt_cor.test <- intubate ## intubate is the helper function
## Note the lack of parentheses
Nothing else is required.
The only thing you need to remember is that that the names of all interfaces
must start with ntbt_
followed by the name of the interfaced function
(cor.test
in this particular case), no matter which function you want to
interface.
Now you can use your "still hot" interface in any pipeline. A pipeline alternative to the above code may look like this:
USJudgeRatings %>%
ntbt_cor.test(CONT, INTG) ## Use it right away
USJudgeRatings %>%
ntbt_cor.test(~ CONT + INTG) ## Also the formula variant
Calling non-pipe-aware functions directly
Moreover, you do not have to create an interface if you do not
want to. You can call the non-pipe-aware function directly with ntbt
,
in the following way:
USJudgeRatings %>%
ntbt(cor.test, CONT, INTG) ## Use it right away
USJudgeRatings %>%
ntbt(cor.test, ~ CONT + INTG) ## Also the formula variant
You can potentially use ntbt
with any function, also the ones without an interface
provided by intubate
. In principle,
the functions you would like to call are the ones you cannot use directly in
a pipeline (because data
is in second place instead of first).
Example showing different techniques
The link below is to Dr. Sheather's website where code was extracted (around June 2015). In the link there is also information about the book. This code could be used to produce Figure 3.1 on page 46, containing 4 plots. Different strategies are illustrated.
http://www.stat.tamu.edu/~sheather/book/
par(mfrow=c(2,2)) # Produce 4 panels to accommodate each plot.
## 1) As in the book (without using pipes and attaching data):
attach(anscombe)
plot(x1, y1, xlim = c(4, 20), ylim = c(3, 14), main = "Data Set 1")
abline(lsfit(x1, y1))
detach()
## You needed to attach so variables are visible locally.
## If not, you should have used anscombe$x1 and anscombe$y1.
## You could also have used 'with'.
## Spaces were added for clarity and better comparison with code below.
## 2) Alternative using magrittr pipes (%>%) and intubate (no need to attach):
anscombe %>%
ntbt_plot(x2, y2, xlim = c(4, 20), ylim = c(3, 14), main = "Data Set 2") %>%
ntbt(lsfit, x2, y2) %>% # Call non-pipe-aware function directly with `ntbt`
abline() # No need to interface 'abline'.
## * 'ntbt_plot' is the interface to 'plot' provided by intubate.
## As 'plot' returns NULL, intubate forwards (invisibly) its input
## automatically without having to use %T>%, so 'lsfit' gets the
## original data (what it needs) and everything is done in one pipeline.
## * 'ntbt' let's you call the non-pipe-aware function 'lsfit' directly.
## You can use 'ntbt' *always* (you do not need to use 'ntbt_' interfaces
## if you do not want to), but 'ntbt' is particularly useful to interface
## directly a non-pipe-aware function for which intubate does not provide
## an interface.
## 3) Alternatively, if intubate does not provide an interface to a given
## function (as currently happens with lsfit), you can create your own
## interface "on demand" and use it right away in your pipeline.
## You only need to include the following line of code before its use:
ntbt_lsfit <- intubate # NOTE: we are *not* including parentheses.
## That's it. Just remember that:
## 1) intubate interfaces *must* start with 'ntbt_' followed by the
## name of the function to interface.
## 2) parentheses are *not* used in the definition of the interface.
## You can now use 'ntbt_lsfit' in your pipeline:
anscombe %>%
ntbt_plot(x3, y3, xlim = c(4, 20), ylim = c(3, 14), main = "Data Set 2") %>%
ntbt_lsfit(x3, y3) %>% # Using just created "on demand" interface
abline()
## 4) Alternatively, you can use the formula variant (original aim of intubate):
anscombe %>%
ntbt_plot(y4 ~ x4, xlim = c(4, 20), ylim = c(3, 14), main = "Data Set 4") %>%
ntbt_lm(y4 ~ x4) %>% # We use 'ntbt_lm' instead of 'ntbt_lmfit'
abline()
This is the plot produced:
Interfaced libraries
The R packages that have interfaces implemented so far are:
adabag
: Multiclass AdaBoost.M1, SAMME and BaggingAER
: Applied Econometrics with Raod
: Analysis of Overdispersed Dataape
: Analyses of Phylogenetics and Evolutionarm
: Data Analysis Using Regression and Multilevel/Hierarchical Modelsbetareg
: Beta Regressionbrglm
: Bias reduction in binomial-response generalized linear modelscaper
: Comparative Analyses of Phylogenetics and Evolution in Rcar
: Companion to Applied Regressioncaret
: Classification and Regression Trainingcoin
: Conditional Inference Procedures in a Permutation Test FrameworkCORElearn
: Classification, Regression and Feature Evaluationdrc
: Analysis of Dose-Response Curvese1071
: Support Vector Machinesearth
: Multivariate Adaptive Regression SplinesEnvStats
: Environmental Statistics, Including US EPA GuidancefGarch
: Rmetrics - Autoregressive Conditional Heteroskedastic Modellingflexmix
: Flexible Mixture Modelingforecast
: Forecasting Functions for Time Series and Linear Modelsfrontier
: Stochastic Frontier Analysisgam
: Generalized Additive Modelsgbm
: Generalized Boosted Regression Modelsgee
: Generalized Estimation Equation Solverglmnet
: Lasso and Elastic-Net Regularized Generalized Linear Modelsglmx
: Generalized Linear Models Extendedgmnl
: Multinomial Logit Models with Random Parametersgplots
: Various R Programming Tools for Plotting Datagss
: General Smoothing Splinesgraphics
: The R Graphics Packagehdm
: High-Dimensional MetricsHmisc
: Harrell Miscellaneousipred
: Improved PredictorsiRegression
: Regression Methods for Interval-Valued Variablesivfixed
: Instrumental fixed effect panel data modelkernlab
: Kernel-Based Machine Learning Labkknn
: Weighted k-Nearest NeighborsklaR
: Classification and Visualizationlars
: Least Angle Regression, Lasso and Forward Stagewiselattice
: Trellis Graphics for RlatticeExtra
: Extra Graphical Utilities Based on Latticeleaps
: Regression Subset Selectionlfe
: Linear Group Fixed Effectslme4
: Linear Mixed-Effects Models using 'Eigen' and S4lmtest
: Testing Linear Regression ModelsMASS
: Robust Regression, Linear Discriminant Analysis, Ridge Regression, Probit Regression, ...MCMCglmm
: MCMC Generalised Linear Mixed Modelsmda
: Mixture and Flexible Discriminant Analysismetafor
: Meta-Analysis Package for Rmgcv
: Mixed GAM Computation Vehicle with GCV/AIC/REML Smoothness Estimationminpack.lm
: R Interface to the Levenberg-Marquardt Nonlinear Least-Squares Algorithm Found in MINPACK, Plus Support for Boundsmhurdle
: Multiple Hurdle Tobit Modelsmlogit
: Multinomial logit modelmnlogit
: Multinomial Logit Modelmodeltools
: Tools and Classes for Statistical Modelsnlme
: Linear and Nonlinear Mixed Effects Modelsnlreg
: Higher Order Inference for Nonlinear Heteroscedastic Modelsnnet
: Feed-Forward Neural Networks and Multinomial Log-Linear Modelsordinal
: Regression Models for Ordinal Dataparty
: A Laboratory for Recursive Partytioningpartykit
: A Toolkit for Recursive Partytioningplotrix
: Various Plotting Functionspls
: Partial Least Squares and Principal Component RegressionpROC
: Display and Analyze ROC Curvespscl
: Political Science Computational Laboratory, Stanford Universitypsychomix
: Psychometric Mixture Modelspsychotools
: Infrastructure for Psychometric Modelingpsychotree
: Recursive Partitioning Based on Psychometric Modelsquantreg
: Quantile RegressionrandomForest
: Random Forests for Classification and RegressionRchoice
: Discrete Choice (Binary, Poisson and Ordered) Models with Random Parametersrminer
: Data Mining Classification and Regression Methodsrms
: Regression Modeling Strategiesrobustbase
: Basic Robust Statisticsrpart
: Recursive Partitioning and Regression TreesRRF
: Regularized Random ForestRWeka
: R/Weka InterfacesampleSelection
: Sample Selection Modelssem
: Structural Equation ModelsspBayes
: Univariate and Multivariate Spatial-temporal Modelingstats
: The R Stats Package (glm, lm, loess, lqs, nls, ...)strucchange
: Testing, Monitoring, and Dating Structural Changessurvey
: Analysis of Complex Survey Samplessurvival
: Survival AnalysisSwarmSVM
: Ensemble Learning Algorithms Based on Support Vector Machinessystemfit
: Estimating Systems of Simultaneous Equationstree
: Classification and Regression Treesvcd
: Visualizing Categorical Datavegan
: Community Ecology Package
The aim is to continue adding interfaces to most methodologies used in data science or other disciplines.
Experimental features, not for general or production use
intubate
includes two experimental features: intubOrders, and intuBags.
These experimental features are already in place but are still under development, are not considered for general use, and are not documented (yet). I want to make sure first, to the best of my abilities, that they are as general as possible, and that eventual future extensions will be backward compatible. You can play with them if you like (I need to play with them more...), but if you use them in production code be prepared to have to change it if I decide to modify the architecture while in the experimental phase.
- intubOrders allow to forward the input without using %T>%, and to run
print
,
summary
, anova
, plot
, and such, in place. This may prove to be interesting to non-pipeline
oriented people too. intubOrders are also needed by intuBags
. If you want to have an
idea of what I mean, please run (using the last version committed to github):
## Note: this is *only* to demonstrate what you *can* do with intubOrders,
## not a suggestion of what you *should* do with them.
## 1) Using interface
LifeCycleSavings %>%
ntbt_lm(sr ~ pop15 + pop75 + dpi + ddpi,
"< head(#, n=10); tail(#, n=3); str; dim; summary; View
|f|
print; summary; anova; plot(#, which=1); plot(#, which=2);
par(mfrow=c(2,2)); plot(#, which=3:6) >") %>%
head()
## 2) Calling function directly
LifeCycleSavings %>%
ntbt(lm, sr ~ pop15 + pop75 + dpi + ddpi,
"< head(#, n=10); tail(#, n=3); str; dim; summary; View
|f|
print; summary; anova; plot(#, which=1); plot(#, which=2);
par(mfrow=c(2,2)); plot(#, which=3:6) >") %>%
head()
Below there is a more involved case, transforming the code in
https://cran.r-project.org/web/packages/survey/vignettes/survey.pdf
library(survey)
data(api)
## First, the original code from the vignette
vars<-names(apiclus1)[c(12:13,16:23,27:37)]
dclus1 <- svydesign(id = ~dnum, weights = ~pw, data = apiclus1, fpc = ~fpc)
summary(dclus1)
svymean(~api00, dclus1)
svyquantile(~api00, dclus1, quantile=c(0.25,0.5,0.75), ci=TRUE)
svytotal(~stype, dclus1)
svytotal(~enroll, dclus1)
svyratio(~api.stu,~enroll, dclus1)
svyratio(~api.stu, ~enroll, design=subset(dclus1, stype=="H"))
svymean(make.formula(vars),dclus1,na.rm=TRUE)
svyby(~ell+meals, ~stype, design=dclus1, svymean)
regmodel <- svyglm(api00~ell+meals,design=dclus1)
logitmodel <- svyglm(I(sch.wide=="Yes")~ell+meals, design=dclus1, family=quasibinomial())
summary(regmodel)
summary(logitmodel)
## Now using intubOrders. I will use ntbt. Alternatively,
## the interfaces could be defined and used.
## Strategy 1: long pipeline, light use of intubOrders.
apiclus1 %>%
ntbt(svydesign, id = ~dnum, weights = ~ pw, fpc = ~ fpc, "<|C| summary >") %>%
ntbt(svymean, ~ api00, "<|Cf| print >") %>%
ntbt(svyquantile, ~ api00, quantile = c(0.25,0.5,0.75), ci = TRUE, "<|Cf| print >") %>%
ntbt(svytotal, ~ stype, "<|Cf| print >") %>%
ntbt(svytotal, ~ enroll, "<|Cf| print >") %>%
ntbt(svyratio, ~ api.stu, ~ enroll, "<|Cf| print >") %>%
ntbt(svyratio, ~ api.stu, ~ enroll, design=subset("#", stype=="H"), "<|Cf| print >") %>%
ntbt(svymean, make.formula(vars), na.rm = TRUE, "<|Cf| print >") %>%
ntbt(svyby, ~ ell + meals, ~ stype, svymean, "<|Cf| print >") %>%
ntbt(svyglm, api00 ~ ell + meals, "<|Cf| summary >") %>%
ntbt(svyglm, I(sch.wide=="Yes") ~ ell + meals, family = quasibinomial(), "<|Cf| summary >") %>%
summary() ## We have forwarded the result from svydesign (line 2),
## so we could still continue using it downstream.
## Strategy 2: short pipeline, heavy use of *one* intubOrder.
apiclus1 %>%
ntbt(svydesign, id = ~dnum, weights = ~pw, fpc = ~fpc,
"<|Cf|
summary;
svymean(~api00, #);
svyquantile(~api00, #, quantile = c(0.25, 0.5, 0.75), ci = TRUE);
svytotal(~stype, #);
svytotal(~enroll, #);
svyratio(~api.stu,~enroll, #);
svyratio(~api.stu, ~enroll, design = subset(#, stype == 'H'));
svymean(make.formula(vars), #, na.rm = TRUE);
svyby(~ell+meals, ~stype, #, svymean);
summary(svyglm(api00~ell+meals, #));
summary(svyglm(I(sch.wide == 'Yes')~ell+meals, #, family = quasibinomial())) >") %>%
head() ## We have forwarded the original dataset,
## so we could continue using it downstream.
intubOrders
are under heavy development, so this is just an idea that is searching
for a final form.
- intuBags allow to run one pipeline containing several sources. intuBags can be
dynamically populated by result(s?) at each step of the pipeline. Results can be, for example, modifications of an original source that can be replaced by the modification or saved as new object. They can also be from a statistical procedure, such as lm
, or other things.
Each step of the pipeline can choose which source(s?) to use, and has the choice to add its
result(s?) to the intuBag, and these added or modified results can be used downstream,
or once the pipeline is ended. If you save the end result, you will have a single object
containing all the sources, their modifications (or replacements of sources by their
modifications), and results of processing sources by the different steps (if you
wanted to save them). You can also stop the pipeline at any point (saving your iBag
perhaps with a %<>%), do something else, and then restarting your pipeline from the
point you stopped.
This means you could potentially have one object only, the intuBag, containing all the
objects (sources and products). One possible use of intuBags could be to process a whole database (several tables, data.frames, or tibbles) in one pipeline. I am not sure yet,
but maybe it is a good idea to intubate pipe-aware functions too (I have played with
intubating subset
). Maybe not. Of course pipe-aware functions can always get the
whole intuBag at any point of the pipeline, do their magic without being intubated,
and then let the pipe continue.
This means intubate
will have three modes of operations:
- as interface only,
- as interface + intubOrders, and
- as interface + intubOrders + intuBags.
More about interfaced libraries
intubate
core depends only on base
, stats
, and utils
libraries.
To keep it as lean as possible, and to be able to continue to include more interfaces without bloating your machine, starting from version 0.99.3 intubate
will not
install the packages that contain the functions that are interfaced.
You will need to install them yourself,
and load the corresponding libraries before using them in your pipelines. This
also applies to magrittr
(in case you want to use intubate
without pipelines).
Then, if you are only interested in a given field, say: bio-statistics, bio-informatics, environmetrics, econometrics, finance, machine learning, meta-analysis, pharmacokinetics, phylogenetics, psychometrics, social sciences, surveys, survival analysis, ..., you will not have to install all the packages for which interfaces are provided if you intend to use only a subset of them. You only need to install the subset of packages you intend to use (which are probably already installed in your machine).
Also, there are cases where some packages are in conflict if loaded simultaneously, leading
to a segmentation fault (for example, kernlab functions fail when testing the whole
examples provided with intubate
, but not when testing kernlab only examples
in a clean environment. I ignore which is/are the other(s) package(s) conflicting with it.
The only thing I know is that the package name is alphabetically ordered prior to kernlab)
I make no personal judgement (mostly due to personal ignorance)
about the merit of any interfaced function.
I have used only a subset of what is provided, and I am happy to include others,
that I am currently unaware of, down the line. In principle I plan on including
packages that are listed as reverse depends, imports, or suggest on package Formula
(I am missing still quite a bit of them). Adding interfacings is easy (and can be
boring...) so I will appreciate if you want to contribute (and you will be credited
in the help of the interfaced package). Also is welcome the improvement of the
provided examples (such as making sure the data used is correct for the statistical
technique used).
I do not claim to be a data scientist (I am barely a statistician and I still have almost no clue of what a data scientist is or is not, and my confusion about the subject only increases with time), nor someone entitled to tell you what to use or not.
As such, I am not capable of engaging in disputes of what is relevant or not, or, if there are competing packages, which to use. I will leave that to you to decide.
Please keep in mind that intubate
will not install any packages correponding to the interfaces
that are provided. You can install only those that you need
(or like) and disregard the rest. Also please remember that you can create your
own interfaces (using helper function intubate
),
or call non-pipe-aware functions directly (using ntbt
).
The original aim of intubate
was to be able to include functions that
have formula and data (in that order) in a magrittr
pipeline using %>%
.
As such, my search so far has been concentrated in packages containing formulas and
misplaced (from pipes point of view) data (with the exeption of a couple of packages
with non-formula variants interfaced as proofs of concept).
For example, this was my first implementation of ntbt_lm
ntbt_lm <- function(data, formula, ...)
lm(formula, data, ...)
This was supposed to be repeated for each interface.
Soon after I realized that intubate
could have a few helper functions (that was
version 0.99.2), later that only one helper function was needed (intubate
), and
later that you could call non-pipe-aware functions directly without defining
interfaces (ntbt
) and
that the interfaces and ntbt
could also be successfully used in cases
where non-formula variants are implemented.
However, my starting point inevitably led the way. I did not see the big picture
(well, what today I think the big picture is...),
so the current version only addresses packages containing functions that use formula
variant, even
if in those cases you can also use the non-formula variants (you can see the
examples corresponding to pROC
, where both cases for formula and non-formula
are demonstrated. You should be able to use that technique also for the rest of
the packages).
I am brewing some ideas about a general approach to packages that
do not use formula interface, but I leave that for a future iteration of intubate
.
This means that there are three possibilities to the eventual lack of inclusion of your favorite package for the time being:
- The package only uses matrices or x- y- like notation (and not formulas)
- (more likely reason) I should know better, but I missed it
(truth is that by implementing the supplied interfaces I realized how little I knew, and still know, about a field in which I am supposed to be an expert), and I apologize for that. 3. I got to the point I need to take a rest (this reason is competing with 2. with increasing strength as time passes by)
Also, please keep in mind you can always create your own interfaces
(with the helper function intubate
), or call the non-pipe-aware functions
directly (with ntbt
).
Bugs and Feature requests
The robustness and generality of the interfacing machinery still needs to be further verified (and very likely improved), as there are thousands of potential functions to interface and certainly some are bound to fail when interfaced. Some have already been addressed when implementing provided interfaces (as their examples failed).
The goal is to make intubate
each time more robust by
addressing the peculiarities of newly discovered failing functions.
For the time being, only cases where the
interfaces provided with intubate
fail will be considered as bugs.
Cases of failing user defined interfaces or when using ntbt
to call functions
directly that do not have interfaces provided with released versions of intubate
,
will be considered feature requests.
Of course, it will be greatly appreciated, if you have some coding skills and can follow the code of the interface, if you could provide the proposed solution, that shouldn't break anything else, together with the feature request.
Logo of intubate
The logo of intubate
is: <||>
. It corresponds to an intuBorder. I have not
found it in a Google search as of 2016/08/08. I intend to use it as a visual
identification of intubate
. If you know of it having being in use before this date
in any software related project, please let me know, and I will change it.
Names used
intuBorder(s) and intubOrder(s), as of 2016/08/08, only has been found, on Google,
in a snippet of code for the name of a variable (intUBorder
) (http://www.office-loesung.de/ftopic246897_0_0_asc.php) that would mean something
like an "integer upper border". There is also an intLBorder
for the lower border.
intuBag(s), as of 2016/08/08, seems to be used for a small bag for bikes (InTuBag,
meaning Inner Tub Bag)
(https://felvarrom.com/products/intubag-bike-tube-bag-medium-blue-inside?variant=18439367751),
but not for anything software related. If intubate
succeeds, they may end selling
more InTuBags!
intubate, as of 2016/08/08, seems to be used related to the medical procedure, perhaps also by the oil pipeline industry (at least "entubar" in Spanish is more general than the medical procedure), but not for software related projects.
intuEnv, as of 2016/08/18, was found only in some latin text.
I intend to use "intubate", "<||>", "intuBorder", "intubOrder(s)", "intuBag(s)", "intuEnv(s)"and other derivations starting with "intu", in relation to the use and promotion of "intubate" for software related activities.
Down the line I intend to register the names and logo as trademarks.
See also
- The setter
package contains mutators to set attributes of variables,
that work well in a pipe (much like stats::setNames())
.
- The srvyr package allows
for analysis of complex surveys using the pipe-friendly syntax of dplyr.
Entries by date
At this point you may have an idea if intubate
is or not for you. If you elect
to continue reading, please be warned that my style or writing may or may not be
of your liking. Most of what follows is for my personal amusement.
2016/08/05
- Core of the interface function (now called
intubate
) should be finished.
Please torture test with as many cases as possible to see how robust it is. I will be out for the rest of the week.
(I want to point out that I have delivered on my promise made on 2016/08/03 to
further reduce the labor involved while defining interfaces. I am happy to
report that using intubate
instead of ntbt_function_data
represents a
55.55% reduction of the amount of typing required, which will significantly
increase the production of interfaces per minute. I am pleased with the
results and consider this a huge success! Congratulations to Everybody!)
Please know that there is much more to come about which I will not comment for now, but that should potentially make intubate
useful even if you
do not want to use it in pipelines (remember that you should not be forced to
use pipelines if you like other alternatives better). It should also take
pipelines to a different dimension, in case you like them.
The Evil Plan
What I will (in fact, have to) publicly inform is that intubate
features an
Undisclosed Evil Plan of World Domination that will be unveiled in the time to come
according to my Organization's Evil Master Plan.
Please also be notified that we are thinking about achieving our goals in three stages (aka The Stages):
- Set foundations needed for the ones to come later. This will be already World Domination but nobody will notice it.
- Using Deception, lay The Bait to attract you to The Trap.
- Convert you into our slave, without you being able to do anything about it.
Appendix: Legal aspects related to Evil Plans
If you find yourself at loss regarding the legality of the above and/or wonder why I am disclosing this information, please know that someone (I cannot tell the name but trust me that he/she is very well connected and knows everything about these things), informed me recently that: "to have an Undisclosed Evil Plan of World Domination" is OK (well... he/she really said "is acceptable") "as long as you disclose that you have one and clearly specify The Stages", which can be non-specific but need to include wording that "warns about the existence of The Bait, The Trap, Deception and Coercion in case any or all of these standard resources are planned to be utilized at some point".
The rationale is that "even if the Regulatory Agencies share the public's concern and are fully aware of the common misconception according to which they should protect, to the full extent of their abilities, the unaware citizen from being abused, full disclosure of The Evil Plan could prove unacceptably counterproductive to the successful fulfillment of World Domination by an Evil Organization that, we cannot forget, also pays taxes and thus is entitled to some level of protection to conduct its business in a satisfactory way".
This solution, albeit non perfect, is considered "a fair compromise that protects, to some degree, the rights of everybody involved".
You have been served!
Have fun (while you can...)
2016/08/03
- Now all interfaces derive from only one helper function called,
for now,
ntbt_function_data
.
### Create some non implemented interfaces
ntbt_legend <- ntbt_cat <-
ntbt_function_data ## One helper function only to create interfaces
(Two steps to create an interface seemed excessive, too much, no-way dude, especially with one of them requiring a rational decision. You need one step only now, so you can create your interfaces when you drink your coffee and text while driving to work.
This will be particularly appreciated down the line, when the technology of self-driving cars is definitely polished and established, and you will be forced to work while commuting to work after working at home... I mean you will have even more free time available to relax and enjoy.
Anyway, I am still not satisfied with the amount of labor required on the user's side and will put my best effort in trying to further reduce it.)
- In addition to formula versions:
library(magrittr)
library(intubate)
USJudgeRatings %>%
ntbt_cor.test(~ CONT + INTG)
you can also use, for example, the x
y
versions:
USJudgeRatings %>%
ntbt_cor.test(CONT, INTG)
All the examples in the documentation run (but they are formula-only versions). Tests to see if this works as expected are welcome.
Of course it would be pure magic if it just simply works no matter what you throw at it, and I simply have no clue of all the possible behaviors of the interfaced functions.
My goal is that it works reasonably in cases you would normally use in data science pipelines.
I anticipate limitations. I will try to address the ones that can be solved in a general and easy way and that really represent a need, because the helper function machinery has to stay powerful yet simple.
- If interfaced function returns NULL, the interface function forwards invisibly the input, so you can use the data downstream.
library(dplyr)
CO2 %>%
mutate(color=sample(c("green", "red", "blue"),
length(conc), replace = TRUE))%>%
ntbt_plot(conc, uptake, col = color) %>% ## plot returns NULL
ntbt_lm(conc ~ uptake) %>% ## data passes through ntbt_plot
summary()
within(warpbreaks, {
time <- seq_along(breaks)
W.T <- wool:tension
}) %>%
ntbt_plot(breaks ~ time, type = "b") %>%
ntbt_text(breaks ~ time, label = W.T,
col = 1 + as.integer(wool)) %>%
ntbt_cat("And now we write a legend.") %>%
ntbt_legend("top",
legend = levels(wool),
col = 1 + as.integer(wool)) %>%
invisible()
- You can also do things like
ntbt_cat <- ntbt_print <- ntbt_View <-
ntbt_function_data
CO2 %>%
ntbt_cat("The first row has uptake -", uptake[1],
"- and concentration", conc[1],"\n") %>%
ntbt_cat("The mean uptake is -", mean(uptake),
"- with standard deviation", sd(uptake),"\n") %>%
ntbt_cat("uptake observations", uptake, sep="\n") %>%
ntbt_print() %>%
ntbt_View() %>%
ntbt_lm(conc ~ .) %>%
summary()
2016/07/30
Pipelines
Pipelines in R are made possible by the package magrittr
,
by Stefan Milton Bache and Hadley Wickham.
dplyr
, by Hadley Wickham, Romain Francois, and RStudio,
is used here to illustrate data transformation.
## Packages needed
library(dplyr) ## Does data transformation
library(magrittr) ## Implements pipelines
## Data used
# devtools::install_github("hadley/yrbss")
library(yrbss)
data(survey)
This machinery allows to perform data transformations using pipelines in the following way:
survey %>%
group_by(year) %>%
summarise(count = n(),
countNA = sum(is.na(stheight)),
propNA = mean(is.na(stheight))) %>%
knitr::kable()
Pipelines seem to be the preferred way, these days, of doing data transformation. If you want an introduction about pipelines, and/or to learn more about them, please follow this link (http://r4ds.had.co.nz/transform.html) to the chapter on data transformation of the forthcoming book "R for Data Science" by Garrett Grolemund and Hadley Wickham.
R statistical functions and pipelines
Suppose you want to perform a regression analysis of the weight on height of males corresponding to the year 2013 (assuming for the sake of argument that it is a valid analysis to perform. See at the very end of the document for more on this).
As most R functions are not pipeline-aware, you should do something like the following.
First, you perform your data science transformations
and save the result to a temporary object (tmp
in this case).
survey %>%
filter(!is.na(stheight) & !is.na(stweight) &
year == 2013 & sex == "Male"
) %>%
select(stheight, stweight) ->
tmp
Then, you perform your regression analysis on the transformed
data stored in tmp
.
fit <- lm(stweight ~ stheight, tmp)
summary(fit)
(There is nothing wrong in this approach. In fact it is good. Jolly good. Splendid!
If you are absolutely happy with doing things this way then there is no need
to continue to devote your efforts in reading this document.
intubate
is not for you. I am happy we could establish this in such little time.)
But what if, in addition to the data transformation, you would also like to perform your data modeling/analysis under the same pipeline paradigm (by adding lm to it), which would impart notation consistency and would avoid the need of creating the temporary object?
survey %>%
filter(!is.na(stheight) & !is.na(stweight) &
year == 2013 & sex == "Male"
) %>%
select(stheight, stweight) %>%
lm(stweight ~ stheight) %>% ## Using the original function
summary()
You get an error.
The reason of this failure is that pipeline-aware functions (such as the ones
in dplyr
that were specifically designed to work in pipelines) receive the data as
the first parameter, and most
statistical procedures that work with formulas to specify the model,
such as lm
and lots of other rock solid reliable functions that implement
well established statistical procedures, receive the data as the second
parameter.
This minor detail can make a difference, actually a huge one. In fact, it may create a division line of two clearly separated cultures, that I will call, for the lack of better names, the "traditionalists" and the "modernists".
(They could be the "modeldatas" and the "datamodels". Whichever you prefer that does not offend anyone)
The aim of intubate
is to provide an easy alternative so nobody has to change the way they do things.
If you are a "traditionalist" and you want to create your new
statistical package in the traditional way
(first model
and then data
), you will not potentially find yourself at a crossroad if you think you need to decide which community to serve, when in fact you can serve both communities without having to do anything differently to what you have done so far. You can just keep doing it in the traditional way. In fact, (please...) keep doing it in the traditional way!
Why? Because Everybody will benefit.
EXT. BUCOLIC PASTURE - EARLY MORNING
Background music initially inaudible slowly increases in
volume while the panning camera, starting from a small
and fragile flower, reveals legions of smiling people
holding hands. Half dress t-shirts with a capital T, the
other half with a capital M.
Everybody raises their arms to the sky - still holding hands -
as if trying to embrace the universe.
Camera slowly raises, zooming out and tilting down, making
sure Everybody is included in the frame, while Everybody
mantains eye contact with the camera.
Sun rays break through heavy pure snow-white clouds.
Everybody opens their mouth and slowly inhales while closing
their eyes as if they really mean what comes next.
(This is critical. Make sure it looks credible.)
Music at full volume.
Everybody sing.
Everybody
We aaaaare the Woooorld - ta ta ta ta taaaa...
We aaaare the chiildreeeen - ta ta ta ta taaaa...
For "traditionalist" users (as I was until a couple of months ago), nothing will have changed. In fact, they will be completely unaware of anything different happening at all. Just business as usual and another fantastic statistical procedure to add to their bag of resources.
For "modernist" users, intubate
will do a couple of tricks behind the scenes so
they will be able to run, right at the end of any required data transformation,
your statistical procedures without any hassle using their preferred paradigm of pipelines.
There are alternatives that allow
to include lm
(and others) in the pipeline without errors and without intubate
.
They require workarounds
of varying levels of complexity and are illustrated later.
If you choose intubate
is because you do not want to bother about workarounds when working with pipelines that include statistical procedures.
intubate
The solution intubate
proposes is to provide an interface to lm
,
called ntbt_lm
, that can be used directly in the pipeline without error and
without losing any of lm
's capabilities.
library(intubate)
survey %>%
filter(!is.na(stheight) & !is.na(stweight) &
year == 2013 & sex == "Male"
) %>%
select(stheight, stweight) %>%
ntbt_lm(stweight ~ stheight) %>% ## Using the interface function
summary()
By using
the interface the error vanishes, as the interface receives data
as its first parameter and formula
second, performs some function transformations, and then calls
lm
in the way it expects to receive the parameters (formula
in first place,
and data
in second place). Now lm
can
continue to do all the good things we are used to.
(It is as if lm
couldn't take anymore being accused by some
of looking old. So it went to the beauty parlor, had a hair
cut, and suddently looks "modern" and now is popular again.)
All the interfaces start with ntbt_
followed by the name of the interfaced function.
Just in case, worry not! The interfaces do not perform any statistical computations (you should be very suspicious if they would). The interfaced functions (those that are already well tested, or should be) are the ones performing the computations.
Interfaces "on demand"
(It used to be "on the fly", but "on demand" sounds more marketable. Right?)
What if you would like to have an interface to a non pipeline-aware function
that is not currently implemented by intubate
?
In a vast majority of cases of R functions (I would like to say all but I still do not know), you can create your own interface "on demand".
To help you in the process, intubate
exposes one helper function,
called intubate
, to assist you.
(Well... something had to do this lazy package...)
Steps to create interfaces "on demand"
For the sake of argument, suppose ntbt_lm
(the interface to lm
)
is not implemented by intubate
, and
you want to create an interface for it.
All you need is adding the following one line of code somewhere before the code that uses it:
Hard?
(You see, it is not enough that intubate
is a lazy package. It also
promotes laziness).
Is it confusing?
No?
Really?
OK, let's see how you do in a quiz, under the Honor Code.
Name:
Honor Code Statement:
Question: Which helper function would you use to construct
your interface for the function `t.test`? [5]
How should you name your interface? [5]
(Psst!... here... don't look!... I told you not to look!... yes, yes, play dumb... you don't have to worry... someone told me there is a solution manual somewhere in the net...)
Remember: names of interfaces must start with
ntbt_
followed by the name of the function
(lm
in this case) you want to interface.
You can now use the interface function in any context
in which you would use the original function.
If you do not want to name the parameters, just remember to switch
the order of formula
and data
arguments when using the interface
(first data
argument and then formula
argument).
As usual, you can put them in any order if you name the arguments
(actually, there are cases in which this is not true. More on this
later)
fit <- ntbt_lm(tmp, stweight ~ stheight)
summary(fit)
Of course you should want to use the interface in a pipeline context.
Otherwise, intubate
is virtually worthless.
tmp %>%
ntbt_lm(stweight ~ stheight) %>%
summary()
Adding interfaces to the intubate
package also represent one-liners for me.
(Did you think I would work more than you?)
The time consuming part on my side is to prepare the documentation, that certainly needs improvement, and make sure the examples work.
Discussion
Disclaimer:
I have a vested interest in making intubate
a success
for egotistical purposes. As such, I may
be overstating the strengths and understating the
weaknesses (weaknesses?? which weaknesses??) of intubate
.
More than a discussion, this can be easily considered
like a sales pitch for a product of dubious quality.
You have been warned. Continue at your own risk.
What can you do if you do not want to use intubate
and you still want to use these kind of functions in pipelines?
Example 1:
lm
can still be added directly to the pipeline,
without error, by specifying the name of the parameter
associated with the model (formula
in this case).
tmp %>%
lm(formula = stweight ~ stheight)
(So what is the big fuss about intubate
?)
The drawback of this approach is that not all functions
use formula
to specify the model.
So far I have encountered 5 variants:
formula
x
object
model
, andfixed
This means you will have to know and remember (yes, also some months from now) which name has been assigned to the model by each particular function.
(OK, OK, you don't need to remember. You can go back to the documentation... over and over again!)
In fact, the following are examples of functions using the other variants.
Example 2:
Using xyplot
directly in a data pipeline will raise an error
library(lattice)
iris %>%
xyplot(Sepal.Length + Sepal.Width ~ Petal.Length + Petal.Width | Species,
scales = "free", layout = c(2, 2),
auto.key = list(x = .6, y = .7, corner = c(0, 0)))
unless x
is specified.
iris %>%
xyplot(x = Sepal.Length + Sepal.Width ~ Petal.Length + Petal.Width | Species,
scales = "free", layout = c(2, 2),
auto.key = list(x = .6, y = .7, corner = c(0, 0)))
Example 3:
Using tmd
(a different function in the same package)
directly in a data pipeline will raise an error
library(lattice)
iris %>%
tmd(Sepal.Length + Sepal.Width ~ Petal.Length + Petal.Width | Species,
scales = "free", layout = c(2, 2),
auto.key = list(x = .6, y = .7, corner = c(0, 0)))
unless object
is specified.
iris %>%
tmd(object = Sepal.Length + Sepal.Width ~ Petal.Length + Petal.Width | Species,
scales = "free", layout = c(2, 2),
auto.key = list(x = .6, y = .7, corner = c(0, 0)))
Example 4:
Using gls
directly in a data pipeline
will raise an error
library(nlme)
Ovary %>%
gls(follicles ~ sin(2*pi*Time) + cos(2*pi*Time),
correlation = corAR1(form = ~ 1 | Mare))
unless model
is specified.
Ovary %>%
gls(model = follicles ~ sin(2*pi*Time) + cos(2*pi*Time),
correlation = corAR1(form = ~ 1 | Mare))
Example 5:
Using lme
directly in a data pipeline
will raise an error
library(nlme)
try(Orthodont %>%
lme(distance ~ age))
geterrmessage()
unless fixed
(!) is specified.
Orthodont %>%
lme(fixed = distance ~ age)
(Subliminal message 1:
there may be many much more different variants - possibly thousands -
lurking around in the darkness. They may hunt you and hurt you... badly...
intubate
will keep you warm and safe, in a happy place where
everybody loves you and nothing wrong can happen to you.)
I find that having to remember the name of the parameter associated to the model in each case is unfortunate, error prone, and gives an inconsistent look and feel to an otherwise elegant interface.
Moreover, it is consider good practice in R to not specify the name of the first two parameters, and name the remaining.
(Note the lack of citation to such categorical statement... ughh, sheer desperation).
Not having to specify the name of the model argument completely hides the heterogeneity of names that can be associated with it. You only write the model and completely forget which name has been assigned to it.
(Subliminal message 2:
- not using
intubate
=> uncool, - using
intubate
=> extremely cool)
Other nightmares around the corner
If you are still not convinced (well, why should you?),
be aware that there are functions that rely on the order of the parameters
(such as aggregate
, cor.test
and other 28 I found so far) that will still
raise an error even if you name the model.
Did you know that there are cases (and for very good reasons), where it is not true that if in a function call you name the parameters you can write them in any order you want?
You don't believe it? How about the following
examples corresponding to cor.test
?
1) Unnamed parameters in the natural order. Works
cor.test(~ CONT + INTG, USJudgeRatings)
2) Named parameters in the natural order. Works
cor.test(formula = ~ CONT + INTG, data = USJudgeRatings)
3) Named parameters with the order changed. Doesn't work!
cor.test(data = USJudgeRatings, formula = ~ CONT + INTG)
(Convinced?)
So let's see what happens if we want to add these cases to the %>%
pipeline.
Example of error 1: cor.test
Using cor.test directly in a data pipeline will raise an error
USJudgeRatings %>%
cor.test(~ CONT + INTG)
even when specifying formula
(as it should be according to
the documentation).
USJudgeRatings %>%
cor.test(formula = ~ CONT + INTG)
Was it y
then?
USJudgeRatings %>%
cor.test(y = ~ CONT + INTG)
Nope...
Was it x
then?
USJudgeRatings %>%
cor.test(x = ~ CONT + INTG)
Nope...
Example of error 2: aggregate
Using aggregate
directly in a data pipeline
will raise an error
ToothGrowth %>%
aggregate(len ~ ., mean)
even when specifying formula
ToothGrowth %>%
aggregate(formula=len ~ ., mean)
or other variants.
Example of error 3: lda
Using lda
directly in a data pipeline
will raise an error
library(MASS)
Iris <- data.frame(rbind(iris3[,,1], iris3[,,2], iris3[,,3]),
Sp = rep(c("s","c","v"), rep(50,3)))
Iris %>%
lda(Sp ~ .)
even when specifying formula
.
Iris %>%
lda(formula = Sp ~ .)
or other variants.
Let's try another strategy. Let's see if the %$% operator, that expands the names of the variables inside the data structure, can be of help.
Iris %$%
lda(Sp ~ .)
Still no...
One last try, or I give up!
Iris %$%
lda(Sp ~ Sepal.L. + Sepal.W. + Petal.L. + Petal.W.)
Finally! But... we had to specify all the variables
(and they may be a lot), and use %$%
instead of %>%
.
There is still another workaround that allows
these functions to be used directly in a pipeline.
It requires the use of another function (with
)
encapsulating the offending function. Here it goes:
Iris %>%
with(lda(Sp ~ ., .))
In the case of aggregate
it goes like
ToothGrowth %>%
with(aggregate(len ~ ., ., mean))
(Do you like it? Do you consider it safe for your children? Really? Come on! What kind of father are you??? You must be one of those that feed unpasteurized milk to them... Shame on you!)
In addition, there is the added complexity of
interpreting the meaning of each of those .
(unfortunately they do not mean the same)
which may cause confusion, particularly at a future
time when you may have to remember why you had to
do this to yourself.
(Hint: the first is specifying to include in the
rhs of the model all the variables in the data but len
,
the second is the name of the data
structure passed by the pipe. Yes, it is called .
!)
Undoubtedly, there may be more elegant workarounds that I am unaware of. But the point is that, no matter how elegant, they will be, well, still workarounds. You want to force unbehaving functions into something that is unnatural to them:
- In one case you had to name the parameters,
- in the other you had to use
%$%
instead of%>%
and where not allowed
to use .
in your model definition,
- if you wanted to use
%>%
you had to use
also which
and include .
as the second parameter.
Does this sound right to you?
I certainly do not want to be distracted implementing workarounds when I am supposed to concentrate in producing the right statistical analysis.
The idea of avoiding such "hacks"
motivated me to write intubate
.
(That was low, please! Were you really that desperate that you had to use the word motivation to try to make a sell? Come on! What is the first thing they teach you in Trickery 101?)
Which was, again, the intubate
alternative?
(Well... if you insist...)
For Example 1:
No need to specify formula
.
tmp %>%
ntbt_lm(stweight ~ stheight)
For Example 2:
No need to specify x
.
iris %>%
ntbt_xyplot(Sepal.Length + Sepal.Width ~ Petal.Length + Petal.Width | Species,
scales = "free", layout = c(2, 2),
auto.key = list(x = .6, y = .7, corner = c(0, 0)))
For Example 3:
No need to specify object
.
iris %>%
ntbt_tmd(Sepal.Length + Sepal.Width ~ Petal.Length + Petal.Width | Species,
scales = "free", layout = c(2, 2),
auto.key = list(x = .6, y = .7, corner = c(0, 0)))
For Example 4:
No need to specify model
.
Ovary %>%
ntbt_gls(follicles ~ sin(2*pi*Time) + cos(2*pi*Time),
correlation = corAR1(form = ~ 1 | Mare))
For Example 5:
No need to specify fixed
.
Orthodont %>%
ntbt_lme(distance ~ age)
For Example of error 1:
It simply works.
USJudgeRatings %>%
ntbt_cor.test(~ CONT + INTG)
For Example of error 2:
It simply works.
ToothGrowth %>%
ntbt_aggregate(len ~ ., mean)
For Example of error 3:
It simply works.
Iris %>%
ntbt_lda(Sp ~ .)
I think the approach intubate
proposes
looks consistent, elegant, simple and clean,
less error prone, and easy to follow
(But, remember, I have a vested interest).
After all, the complication should be in the analysis you are performing, and not in how you are performing it.
Conclusions
(making a huge effort to seem unbiased):
- You can use
intubate
to provide a consistent
look and feel that allows to use non-pipe-aware functions in data science pipelines without having to rely on hacks of different levels of complexity.
- You can use the machinery provided
to create "on demand" interfaces that have not been implemented (and may never be).
- The real thing is done by the interfaced function,
not by the interface function.
intubate
simply sells air and takes all the credit by providing only a refined syntactic
sugar (and let's not start on the potential consequences that the consumption of refined sugars may, or may not, have on your health...) that will be more or less palatable according to your individual taste.
- Perhaps the fact that the documentation
provides working examples of statistical and machine learning procedures suitable for data science is a plus.
Bottom line
At the end of the day, you can certainly be a
very successful data scientist (yes, what
in the past may have been referred to as
a statistician with some
computer skills) if you already
feel confident with your ways and decide
not to use intubate
. After all, it is just
another tool.
Final thoughts for new data scientists, unrelated to intubate
.
(Please skip it if you still believe in Santa)
I am not stating that, in the particular case presented at the beginning of the document, doing a regression analysis is the correct thing to do. I am neither stating it is wrong. I simply didn't put any effort in establishing the merit or not of doing a regression analysis on that particular data. This would entail asking a lot of questions about how that data was collected. I just didn't do that. It was a little on purpose so I can have the discussion that follows.
That example and the rest provided in this document are only for illustration purposes related strictly to the computational techniques described, and have nothing to do with the validity or not of using any particular statistical methodology on any particular data.
(Well, the rest of the examples were taken from the help of the interfaced functions. I want to believe that the authors put some effort in establishing that they were doing what they were supposed to do.)
Sure, you may argue that, after all, in the case of linear regression the fitting procedure is no other thing than an optimization technique (minimization of the sum of squared residuals) completely unrelated to statistics, so why can't you just find the line of best fit for the sake of it, even if it doesn't make any sense and does not represent anything? Touché, you got me! (You are good!) You can, and nobody is going to throw you in jail for that.
But do you want to make population parameter estimation, statistical inferences such as testing if the population parameters are different than zero, or confidence intervals of the population parameters, or confidence bands of the population regression line, or prediction intervals for an observation?
Then you should probably start asking yourself, before doing anything, if the assumptions of the linear regression model are met or not, and if not, how bad is the violation of such assumptions (what about independence of the Y? Are you sure they don't form a time series? Are the errors iid normally distributed with mean zero and constant variance? Are you sure the error variance doesn't change when you change the values of the independent variable?).
The statistical function (every statistical function) will only perform calculations and spit out a collection of numerical values whose interpretation will only make sense provided the data - and you are responsible for the data you use on the statistical analysis - reasonably follows the assumptions of the particular statistical model you are entertaining. The more you depart from the assumptions, the less interpretable your results become.
Always remember that there is no provision coded in that black box that is the statistical function that will protect you from doing a statistical analysis on the wrong data. Nothing will be corrected or compensated on your behalf. For example, if the methodology is expecting a random sample (the vast majority do) and you are not providing one, no magical trick will convert your non-random-sample into one that is random and satisfies all the assumptions.
(Remember I warned you to skip this section if you still believed in Santa?)
Maybe you shouldn't run a statistical procedure just because you can, and then report results with interpretations that may make no meaning at all.
If you are not truly confident on what you are doing, perhaps your best move should be to consult first with your PCS (Primary Care Statistician) before doing something you may regret down the line.