This is a guide to importing and exporting data to and from R.
The current version of this document is 2.15.1 (2012-06-22).
The relational databases part of this manual is based in part on an earlier manual by Douglas Bates and Saikat DebRoy. The principal author of this manual was Brian Ripley.
Many volunteers have contributed to the packages used here. The principal authors of the packages mentioned are
CORBA Duncan Temple Lang DBI David A. James dataframes2xls Guido van Steen foreign Thomas Lumley, Saikat DebRoy, Douglas Bates, Duncan Murdoch and Roger Bivand gdata Gregory R. Warnes hdf5 Marcus Daniels ncdf, ncdf4 David Pierce rJava Simon Urbanek RJDBC Simon Urbanek RMySQL David James and Saikat DebRoy RNetCDF Pavel Michna RODBC Michael Lapsley and Brian Ripley ROracle David A, James RPostgreSQL Sameer Kumar Prayaga RSPerl Duncan Temple Lang RSPython Duncan Temple Lang RSQLite David A, James SJava John Chambers and Duncan Temple Lang WriteXLS Marc Schwartz XLConnect Mirai Solutions GmbH xlsReadWrite Hans-Peter Suter XML Duncan Temple Lang
Brian Ripley is the author of the support for connections.
Reading data into a statistical system for analysis and exporting the results to some other system for report writing can be frustrating tasks that can take far more time than the statistical analysis itself, even though most readers will find the latter far more appealing.
This manual describes the import and export facilities available either in R itself or via packages which are available from CRAN or elsewhere.
Unless otherwise stated, everything described in this manual is (at least in principle) available on all platforms running R.
In general, statistical systems like R are not particularly well suited to manipulations of large-scale data. Some other systems are better than R at this, and part of the thrust of this manual is to suggest that rather than duplicating functionality in R we can make another system do the work! (For example Therneau & Grambsch (2000) commented that they preferred to do data manipulation in SAS and then use package survival in S for the analysis.) Database manipulation systems are often very suitable for manipulating and extracting data: several packages to interact with DBMSs are discussed here.
There are packages to allow functionality developed in languages such as
python to be directly integrated
with R code, making the use of facilities in these languages even
more appropriate. (See the rJava package from CRAN and
the SJava, RSPerl and RSPython packages from the
Omegahat project, http://www.omegahat.org.)
It is also worth remembering that R like S comes from the Unix
tradition of small re-usable tools, and it can be rewarding to use tools
perl to manipulate data before import or
after export. The case study in Becker, Chambers & Wilks (1988, Chapter
9) is an example of this, where Unix tools were used to check and
manipulate the data before input to S. The traditional Unix tools
are now much more widely available, including for Windows.
The easiest form of data to import into R is a simple text file, and
this will often be acceptable for problems of small or medium scale.
The primary function to import from a text file is
scan, and this
underlies most of the more convenient functions discussed in
However, all statistical consultants are familiar with being presented by a client with a memory stick (formerly, a floppy disc or CD-R) of data in some proprietary binary format, for example `an Excel spreadsheet' or `an SPSS file'. Often the simplest thing to do is to use the originating application to export the data as a text file (and statistical consultants will have copies of the most common applications on their computers for that purpose). However, this is not always possible, and Importing from other statistical systems discusses what facilities are available to access such files directly from R. For Excel spreadsheets, the available methods are summarized in Reading Excel spreadsheets. For ODS spreadsheets from Open Office, see the Omegahat package1 ROpenOffice.
In a few cases, data have been stored in a binary form for compactness and speed of access. One application of this that we have seen several times is imaging data, which is normally stored as a stream of bytes as represented in memory, possibly preceded by a header. Such data formats are discussed in Binary files and Binary connections.
For much larger databases it is common to handle the data using a database management system (DBMS). There is once again the option of using the DBMS to extract a plain file, but for many such DBMSs the extraction operation can be done directly from an R package: See Relational databases. Importing data via network connections is discussed in Network interfaces.
Unless the file to be imported from is entirely in ASCII, it is usually
necessary to know how it was encoded. For text files, a good way to
find out something about its structure is the file
command-line tool (for Windows, included in
reports something like
text.Rd: UTF-8 Unicode English text text2.dat: ISO-8859 English text text3.dat: Little-endian UTF-16 Unicode English character data, with CRLF line terminators intro.dat: UTF-8 Unicode text intro.dat: UTF-8 Unicode (with BOM) text
Modern Unix-alike systems, including Mac OS X, are likely to produce
UTF-8 files. Windows may produce what it calls `Unicode' files
UCS-2LE or just possibly
UTF-16LE2). Otherwise most files will be in a
8-bit encoding unless from a Chinese/Japanese/Korean locale (which have
a wide range of encodings in common use). It is not possible to
automatically detect with certainty which 8-bit encoding (although
guesses may be possible and file may guess as it did in the
example above), so you may simply have to ask the originator for some
clues (e.g. `Russian on Windows').
`BOMs' (Byte Order Marks,
http://en.wikipedia.org/wiki/Byte_order_mark) cause problems for
Unicode files. In the Unix world BOMs are rarely used, whereas in the
Windows world they almost always are for UCS-2/UTF-16 files, and often
are for UTF-8 files. The file utility will not even recognize
UCS-2 files without a BOM, but many other utilities will refuse to read
files with a BOM and the IANA standards for
UTF-16BE prohibit it. We have too often been reduced to
looking at the file with the command-line utility od or a hex
editor to work out its encoding.
Exporting results from R is usually a less contentious task, but there are still a number of pitfalls. There will be a target application in mind, and normally a text file will be the most convenient interchange vehicle. (If a binary file is required, see Binary files.)
cat underlies the functions for exporting data. It
file argument, and the
append argument allows a
text file to be written via successive calls to
especially if this is to be done many times, is to open a
connection for writing or appending, and
cat to that connection,
The most common task is to write a matrix or data frame to file as a
rectangular grid of numbers, possibly with row and column labels. This
can be done by the functions
write just writes out a matrix or vector in a specified
number of columns (and transposes a matrix). Function
write.table is more convenient, and writes out a data frame (or
an object that can be coerced to a data frame) with row and column
There are a number of issues that need to be considered in writing out a data frame to a text file.
Most of the conversions of real/complex numbers done by these functions
is to full precision, but those by
write are governed by the
current setting of
options(digits). For more control, use
format on a data frame, possibly column-by-column.
R prefers the header line to have no entry for the row names, so the file looks like
dist climb time Greenmantle 2.5 650 16.083 ...
Some other systems require a (possibly empty) entry for the row names, which
write.table will provide if argument
col.names = NA
is specified. Excel is one such system.
write.csvprovides appropriate defaults. In some locales the comma is used as the decimal point (set this in
dec = ",") and there CSV files use the semicolon as the field separator: use
write.csv2for appropriate defaults. There is an IETF standard for CSV files (which mandates commas and CRLF line endings, for which use
eol = "\r\n"), RFC4180 (see http://tools.ietf.org/html/rfc4180), but what is more important in practice is that the file is readable by the application it is targeted at.
Using a semicolon or tab (
sep = "\t") are probably the safest
NA, but this may be changed by argument
na. Note that
NaNs are treated as
write.table, but not by
quotecontrols if character and factor variables are quoted: some programs, for example Mondrian, do not accept quoted strings (which are the default).
Some care is needed if the strings contain embedded quotes. Three useful forms are
> df <- data.frame(a = I("a \" quote")) > write.table(df) "a" "1" "a \" quote" > write.table(df, qmethod = "double") "a" "1" "a "" quote" > write.table(df, quote = FALSE, sep = ",") a 1,a " quote
The second is the form of escape commonly used by spreadsheets.
fileEncodingargument to make this easier.
The hard part is to know what file encoding to use. For use on Windows,
it is best to use what Windows calls `Unicode'3, that is
"UTF-16LE". Using UTF-8 is a good way
to make portable files that will not easily be confused with any other
encoding, but even Mac OS X applications (where UTF-8 is the system
encoding) may not recognize them, and Windows applications are most
unlikely to. Apparently Excel:mac 2004/8 expects
.csv files in
"macroman" encoding (the encoding used in much earlier versions
of Mac OS).
write.matrix in package MASS provides a
specialized interface for writing matrices, with the option of writing
them in blocks and thereby reducing memory usage.
It is possible to use
sink to divert the standard R output to
a file, and thereby capture the output of (possibly implicit)
options(width) setting may need to be increased.
write.foreign in package foreign uses
write.table to produce a text file and also writes a code file
that will read this text file into another statistical package. There is
currently support for export to
When reading data from text files, it is the responsibility of the user to know and to specify the conventions used to create that file, e.g. the comment character, whether a header line is present, the value separator, the representation for missing values (and so on) described in Export to text files. A markup language which can be used to describe not only content but also the structure of the content can make a file self-describing, so that one need not provide these details to the software reading the data.
The eXtensible Markup Language – more commonly known simply as XML – can be used to provide such structure, not only for standard datasets but also more complex data structures. XML is becoming extremely popular and is emerging as a standard for general data markup and exchange. It is being used by different communities to describe geographical data such as maps, graphical displays, mathematics and so on.
XML provides a way to specify the file's encoding, e.g.
<?xml version="1.0" encoding="UTF-8"?>
although it does not require it.
The XML package provides general facilities for reading and writing XML documents within R. A description of the facilities of the XML package is outside the scope of this document: see the package's Web page at http://www.omegahat.org/RSXML for details and examples. Package StatDataML on CRAN is one example building on XML.
NB: XML is available as a binary package for Windows, normally from the `CRAN extras' repository (which is selected by default on Windows).
In Export to text files we saw a number of variations on the format of a spreadsheet-like text file, in which the data are presented in a rectangular grid, possibly with row and column labels. In this section we consider importing such files into R.
read.table is the most convenient way to read in a
rectangular grid of data. Because of the many possibilities, there are
several other functions that call
read.table but change a group
of default arguments.
read.table is an inefficient way to read in
very large numerical matrices: see
Some of the issues to consider are:
If the file contains non-ASCII character fields, ensure that it is read in the correct encoding. This is mainly an issue for reading Latin-1 files in a UTF-8 locale, which can be done by something like
Note that this will work in any locale which can represent Latin-1 strings, but not many Greek/Russian/Chinese/Japanese ... locales.
We recommend that you specify the
header argument explicitly,
Conventionally the header line has entries only for the columns and not
for the row labels, so is one field shorter than the remaining lines.
(If R sees this, it sets
header = TRUE.) If presented with a
file that has a (possibly empty) header field for the row labels, read
it in by something like
read.table("file.dat", header = TRUE, row.names = 1)
Column names can be given explicitly via the
names override the header line (if present).
Normally looking at the file will determine the field separator to be
used, but with white-space separated files there may be a choice between
sep = "" which uses any white space (spaces, tabs or
newlines) as a separator,
sep = " " and
sep = "\t". Note
that the choice of separator affects the input of quoted strings.
If you have a tab-delimited file containing empty fields be sure to use
sep = "\t".
sep = "\n"the default is changed to
quote = "".
If no separator character is specified, quotes can be escaped within quoted strings by immediately preceding them by `\', C-style.
If a separator character is specified, quotes can be escaped within quoted strings by doubling them as is conventional in spreadsheets. For example
'One string isn''t two',"one more"
can be read by
read.table("testfile", sep = ",")
This does not work with the default separator.
NAto represent missing values, but this can be changed by the argument
na.strings, which is a vector of one or more character representations of missing values.
Empty fields in numeric columns are also regarded as missing values.
In numeric columns, the values
It is quite common for a file exported from a spreadsheet to have all
trailing empty fields (and their separators) omitted. To read such
fill = TRUE.
If a separator is specified, leading and trailing white space in
character fields is regarded as part of the field. To strip the space,
strip.white = TRUE.
read.table ignores empty lines. This can be changed
blank.lines.skip = FALSE, which will only be useful in
fill = TRUE, perhaps to use blank rows to
indicate missing cases in a regular layout.
Unless you take any special action,
read.table reads all the
columns as character vectors and then tries to select a suitable class
for each variable in the data frame. It tries in turn
complex, moving on if any
entry is not missing and cannot be converted.4
If all of these fail, the variable is converted to a factor.
as.is provide greater control.
as.is = TRUE suppresses conversion of character
vectors to factors (only). Using
colClasses allows the desired
class to be set for each column in the input: it will be faster and use
as.is are specified per
column, not per variable, and so include the column of row names
read.table uses `#' as a comment character,
and if this is encountered (except in quoted strings) the rest of the
line is ignored. Lines containing only white space and a comment are
treated as blank lines.
If it is known that there will be no comments in the data file, it is
safer (and may be faster) to use
comment.char = "".
Many OSes have conventions for using backslash as an escape character in text files, but Windows does not (and uses backslash in path names). It is optional in R whether such conventions are applied to data files.
scan have a logical argument
allowEscapes. This is false by default, and backslashes are then
only interpreted as (under circumstances described above) escaping
quotes. If this set to be true, C-style escapes are interpreted, namely
the control characters
\a, \b, \f, \n, \r, \t, \v and octal and
hexadecimal representations like
other escaped character is treated as itself, including backslash. Note
that Unicode escapes such as
\uxxxx are never interpreted.
This can be specified by the
fileEncoding argument, for example
fileEncoding = "UCS-2LE" # Windows 'Unicode' files fileEncoding = "UTF-8"
If you know (correctly) the file's encoding this will almost always work. However, we know of one exception, UTF-8 files with a BOM. Some people claim that UTF-8 files should never have a BOM, but some software (apparently including Excel:mac) uses them, and many Unix-alike OSes do not accept them. So faced with a file which file reports as
intro.dat: UTF-8 Unicode (with BOM) text
it can be read on Windows by
read.table("intro.dat", fileEncoding = "UTF-8")
but in a UTF-8 locale on a Unix-alike is likely to need
con <- file("intro.dat", "r") seek(con, 3) # step over the 3 initial bytes, the BOM foo <- read.table(con) close(con)
Another problem with this (real-life) example is that whereas file-5.03 reported the BOM, file-4.17 found on Mac OS 10.5 (Leopard) did not.
read.table appropriate for CSV and tab-delimited
files exported from spreadsheets in English-speaking locales. The
read.delim2 are appropriate for
use in those locales where the comma is used for the decimal point and
read.csv2) for spreadsheets which use semicolons to separate
If the options to
read.table are specified incorrectly, the error
message will usually be of the form
Error in scan(file = file, what = what, sep = sep, : line 1 did not have 5 elements
Error in read.table("files.dat", header = TRUE) : more columns than column names
This may give enough information to find the problem, but the auxiliary
count.fields can be useful to investigate further.
Efficiency can be important when reading large data grids. It will help
comment.char = "",
colClasses as one of the
atomic vector types (logical, integer, numeric, complex, character or
perhaps raw) for each column, and to give
nrows, the number of
rows to be read (and a mild over-estimate is better than not specifying
this at all). See the examples in later sections.
Sometimes data files have no field delimiters but have fields in pre-specified columns. This was very common in the days of punched cards, and is still sometimes used to save file space.
read.fwf provides a simple way to read such files,
specifying a vector of field widths. The function reads the file into
memory as whole lines, splits the resulting character strings, writes
out a temporary tab-separated file and then calls
This is adequate for small files, but for anything more complicated we
recommend using the facilities of a language like
pre-process the file.
read.fortran is a similar function for fixed-format files,
using Fortran-style column specifications.
An old format sometimes used for spreadsheet-like data is DIF, or Data Interchange format.
read.DIF provides a simple way to read such files. It takes
arguments similar to
read.table for assigning types to each of the columns.
On Windows, spreadsheet programs often store spreadsheet data copied to
the clipboard in this format;
read.DIF("clipboard") can read it
from there directly. It is slightly more robust than
read.table("clipboard") in handling spreadsheets with empty
scan to read the
file, and then process the results of
scan. They are very
convenient, but sometimes it is better to use
scan has many arguments, most of which we have already
read.table. The most crucial argument is
what, which specifies a list of modes of variables to be read
from the file. If the list is named, the names are used for the
components of the returned list. Modes can be numeric, character or
complex, and are usually specified by an example, e.g.
0i. For example
cat("2 3 5 7", "11 13 17 19", file="ex.dat", sep="\n") scan(file="ex.dat", what=list(x=0, y="", z=0), flush=TRUE)
returns a list with three components and discards the fourth column in the file.
There is a function
readLines which will be more convenient if
all you want is to read whole lines into R for further processing.
One common use of
scan is to read in a large matrix. Suppose
file matrix.dat just contains the numbers for a 200 x 2000
matrix. Then we can use
A <- matrix(scan("matrix.dat", n = 200*2000), 200, 2000, byrow = TRUE)
On one test this took 1 second (under Linux, 3 seconds under Windows on the same machine) whereas
A <- as.matrix(read.table("matrix.dat"))
took 10 seconds (and more memory), and
A <- as.matrix(read.table("matrix.dat", header = FALSE, nrows = 200, comment.char = "", colClasses = "numeric"))
took 7 seconds. The difference is almost entirely due to the overhead
of reading 2000 separate short columns: were they of length 2000,
scan took 9 seconds whereas
read.table took 18 if used
efficiently (in particular, specifying
colClasses) and 125 if
Note that timings can depend on the type read and the data. Consider reading a million distinct integers:
writeLines(as.character((1+1e6):2e6), "ints.dat") xi <- scan("ints.dat", what=integer(0), n=1e6) # 0.77s xn <- scan("ints.dat", what=numeric(0), n=1e6) # 0.93s xc <- scan("ints.dat", what=character(0), n=1e6) # 0.85s xf <- as.factor(xc) # 2.2s DF <- read.table("ints.dat") # 4.5s
and a million examples of a small set of codes:
code <- c("LMH", "SJC", "CHCH", "SPC", "SOM") writeLines(sample(code, 1e6, replace=TRUE), "code.dat") y <- scan("code.dat", what=character(0), n=1e6) # 0.44s yf <- as.factor(y) # 0.21s DF <- read.table("code.dat") # 4.9s DF <- read.table("code.dat", nrows=1e6) # 3.6s
Note that these timings depend heavily on the operating system (the basic reads in Windows take at least as twice as long as these Linux times) and on the precise state of the garbage collector.
Sometimes spreadsheet data is in a compact format that gives the covariates for each subject followed by all the observations on that subject. R's modelling functions need observations in a single column. Consider the following sample of data from repeated MRI brain measurements
Status Age V1 V2 V3 V4 P 23646 45190 50333 55166 56271 CC 26174 35535 38227 37911 41184 CC 27723 25691 25712 26144 26398 CC 27193 30949 29693 29754 30772 CC 24370 50542 51966 54341 54273 CC 28359 58591 58803 59435 61292 CC 25136 45801 45389 47197 47126
There are two covariates and up to four measurements on each subject. The data were exported from Excel as a file mr.csv.
We can use
stack to help manipulate these data to give a single
zz <- read.csv("mr.csv", strip.white = TRUE) zzz <- cbind(zz[gl(nrow(zz), 1, 4*nrow(zz)), 1:2], stack(zz[, 3:6]))
Status Age values ind X1 P 23646 45190 V1 X2 CC 26174 35535 V1 X3 CC 27723 25691 V1 X4 CC 27193 30949 V1 X5 CC 24370 50542 V1 X6 CC 28359 58591 V1 X7 CC 25136 45801 V1 X11 P 23646 50333 V2 ...
unstack goes in the opposite direction, and may be
useful for exporting data.
Another way to do this is to use the function
> reshape(zz, idvar="id",timevar="var", varying=list(c("V1","V2","V3","V4")),direction="long") Status Age var V1 id 1.1 P 23646 1 45190 1 2.1 CC 26174 1 35535 2 3.1 CC 27723 1 25691 3 4.1 CC 27193 1 30949 4 5.1 CC 24370 1 50542 5 6.1 CC 28359 1 58591 6 7.1 CC 25136 1 45801 7 1.2 P 23646 2 50333 1 2.2 CC 26174 2 38227 2 ...
reshape function has a more complicated syntax than
stack but can be used for data where the `long' form has more
than the one column in this example. With
reshape can also perform the opposite transformation.
Some people prefer the tools in packages reshape, reshape2 and plyr.
Displaying higher-dimensional contingency tables in array form typically
is rather inconvenient. In categorical data analysis, such information
is often represented in the form of bordered two-dimensional arrays with
leading rows and columns specifying the combination of factor levels
corresponding to the cell counts. These rows and columns are typically
“ragged” in the sense that labels are only displayed when they change,
with the obvious convention that rows are read from top to bottom and
columns are read from left to right. In R, such “flat” contingency
tables can be created using
which creates objects of class
"ftable" with an appropriate print
As a simple example, consider the R standard data set
UCBAdmissions which is a 3-dimensional contingency table
resulting from classifying applicants to graduate school at UC Berkeley
for the six largest departments in 1973 classified by admission and sex.
> data(UCBAdmissions) > ftable(UCBAdmissions) Dept A B C D E F Admit Gender Admitted Male 512 353 120 138 53 22 Female 89 17 202 131 94 24 Rejected Male 313 207 205 279 138 351 Female 19 8 391 244 299 317
The printed representation is clearly more useful than displaying the data as a 3-dimensional array.
There is also a function
read.ftable for reading in flat-like
contingency tables from files.
This has additional arguments for dealing with variants on how exactly
the information on row and column variables names and levels is
represented. The help page for
read.ftable has some useful
examples. The flat tables can be converted to standard contingency
tables in array form using
Note that flat tables are characterized by their “ragged” display of
row (and maybe also column) labels. If the full grid of levels of the
row variables is given, one should instead use
read.table to read
in the data, and create the contingency table from this using
In this chapter we consider the problem of reading a binary data file written by another statistical system. This is often best avoided, but may be unavoidable if the originating system is not available.
In all cases the facilities described were written for data files from specific versions of the other system (often in the early 2000s), and have not necessarily been updated for the most recent versions of the other system.
The recommended package foreign provides import facilities for
files produced by these statistical systems, and for export to Stata. In
some cases these functions may require substantially less memory than
write.foreign (See Export to text files) provides an export mechanism with support currently for
EpiInfo versions 5 and 6 stored data in a self-describing fixed-width
read.epiinfo will read these .REC files into
an R data frame. EpiData also produces data in this format.
read.mtp imports a `Minitab Portable Worksheet'. This
returns the components of the worksheet as an R list.
read.xport reads a file in SAS Transport (XPORT) format
and return a list of data frames. If SAS is available on your system,
read.ssd can be used to create and run a SAS script that
saves a SAS permanent dataset (.ssd or .sas7bdat) in
Transport format. It then calls
read.xport to read the resulting
file. (Package Hmisc has a similar function
running SAS.) For those without access to SAS but running on Windows,
the SAS System Viewer (a zero-cost download) can be used to open SAS
datasets and export them to e.g. .csv format.
read.S which can read binary objects produced by S-PLUS
3.x, 4.x or 2000 on (32-bit) Unix or Windows (and can read them on a
different OS). This is able to read many but not all S objects: in
particular it can read vectors, matrices and data frames and lists
data.restore reads S-PLUS data dumps (created by
data.dump) with the same restrictions (except that dumps from the
Alpha platform can also be read). It should be possible to read data
dumps from S-PLUS 5.x and later written with
If you have access to S-PLUS, it is usually more reliable to
the object(s) in S-PLUS and
source the dump file in R. For
S-PLUS 5.x and later you may need to use
and to read in very large objects it may be preferable to use the dump
file as a batch script rather than use the
read.spss can read files created by the `save' and
`export' commands in SPSS. It returns a list with one
component for each variable in the saved data set. SPSS
variables with value labels are optionally converted to R factors.
SPSS Data Entry is an application for creating data entry
forms. By default it creates data files with extra formatting
read.spss cannot handle, but it is possible to
export the data in an ordinary SPSS format.
Some third-party applications claim to produce data `in SPSS format' but
with differences in the formats:
read.spss may or may not be able
to handle these.
Stata .dta files are a binary file format. Files from versions 5
up to 11 of Stata can be read and written by functions
write.dta. Stata variables with value labels are optionally
converted to (and from) R factors. Stata version 12 by default
writes `format-115 datasets':
read.dta currently may not be able
to read those.
read.systat reads those Systat
SAVE files that are
rectangular data files (
mtype = 1) written on little-endian
machines (such as from Windows). These have extension .sys
or (more recently) .syd.
Octave is a numerical linear algebra system
(http://www.octave.org), and function
package foreign can read in files in Octave text data format
created using the Octave command save -ascii, with support for
most of the common types of variables, including the standard atomic
(real and complex scalars, matrices, and N-d arrays, strings,
ranges, and boolean scalars and matrices) and recursive (structs, cells,
and lists) ones.
There are limitations on the types of data that R handles well. Since all data being manipulated by R are resident in memory, and several copies of the data can be created during execution of a function, R is not well suited to extremely large data sets. Data objects that are more than a (few) hundred megabytes in size can cause R to run out of memory, particularly on a 32-bit operating system.
R does not easily support concurrent access to data. That is, if more than one user is accessing, and perhaps updating, the same data, the changes made by one user will not be visible to the others.
R does support persistence of data, in that you can save a data object or an entire worksheet from one session and restore it at the subsequent session, but the format of the stored data is specific to R and not easily manipulated by other systems.
Database management systems (DBMSs) and, in particular, relational DBMSs (RDBMSs) are designed to do all of these things well. Their strengths are
The sort of statistical applications for which DBMS might be used are to extract a 10% sample of the data, to cross-tabulate data to produce a multi-dimensional contingency table, and to extract data group by group from a database for separate analysis.
Increasingly OSes are themselves making use of DBMSs for these reasons, so it is nowadays likely that one will be already installed on your (non-Windows) OS. Akonadi is used by KDE4 to store personal information and uses MySQL. Several Mac OS X applications, including Mail and Address Book, use SQLite.
Traditionally there had been large (and expensive) commercial RDBMSs (Informix; Oracle; Sybase; IBM's DB2; Microsoft SQL Server on Windows) and academic and small-system databases (such as MySQL, PostgreSQL, Microsoft Access, ...), the former marked out by much greater emphasis on data security features. The line is blurring, with the Open Source MySQL and PostgreSQL having more and more high-end features, and free `express' versions being made available for the commercial DBMSs.
There are other commonly used data sources, including spreadsheets, non-relational databases and even text files (possibly compressed). Open Database Connectivity (ODBC) is a standard to use all of these data sources. It originated on Windows (see http://www.microsoft.com/data/odbc/) but is also implemented on Linux/Unix/Mac OS X.
All of the packages described later in this chapter provide clients to client/server databases. The database can reside on the same machine or (more often) remotely. There is an ISO standard (in fact several: SQL92 is ISO/IEC 9075, also known as ANSI X3.135-1992, and SQL99 is coming into use) for an interface language called SQL (Structured Query Language, sometimes pronounced `sequel': see Bowman et al. 1996 and Kline and Kline 2001) which these DBMSs support to varying degrees.
The more comprehensive R interfaces generate SQL behind the scenes for common operations, but direct use of SQL is needed for complex operations in all. Conventionally SQL is written in upper case, but many users will find it more convenient to use lower case in the R interface functions.
A relational DBMS stores data as a database of tables (or relations) which are rather similar to R data frames, in that they are made up of columns or fields of one type (numeric, character, date, currency, ...) and rows or records containing the observations for one entity.
SQL `queries' are quite general operations on a relational database. The classical query is a SELECT statement of the type
SELECT State, Murder FROM USArrests WHERE Rape > 30 ORDER BY Murder SELECT t.sch, c.meanses, t.sex, t.achieve FROM student as t, school as c WHERE t.sch = c.id SELECT sex, COUNT(*) FROM student GROUP BY sex SELECT sch, AVG(sestat) FROM student GROUP BY sch LIMIT 10
The first of these selects two columns from the R data frame
USArrests that has been copied across to a database table,
subsets on a third column and asks the results be sorted. The second
performs a database join on two tables
school and returns four columns. The third and fourth queries do
some cross-tabulation and return counts or averages. (The five
aggregation functions are COUNT(*) and SUM, MAX, MIN and AVG, each
applied to a single column.)
SELECT queries use FROM to select the table, WHERE to specify a condition for inclusion (or more than one condition separated by AND or OR), and ORDER BY to sort the result. Unlike data frames, rows in RDBMS tables are best thought of as unordered, and without an ORDER BY statement the ordering is indeterminate. You can sort (in lexicographical order) on more than one column by separating them by commas. Placing DESC after an ORDER BY puts the sort in descending order.
SELECT DISTINCT queries will only return one copy of each distinct row in the selected table.
The GROUP BY clause selects subgroups of the rows according to the criterion. If more than one column is specified (separated by commas) then multi-way cross-classifications can be summarized by one of the five aggregation functions. A HAVING clause allows the select to include or exclude groups depending on the aggregated value.
If the SELECT statement contains an ORDER BY statement that produces a unique ordering, a LIMIT clause can be added to select (by number) a contiguous block of output rows. This can be useful to retrieve rows a block at a time. (It may not be reliable unless the ordering is unique, as the LIMIT clause can be used to optimize the query.)
There are queries to create a table (CREATE TABLE, but usually one copies a data frame to the database in these interfaces), INSERT or DELETE or UPDATE data. A table is destroyed by a DROP TABLE `query'.
Kline and Kline (2001) discuss the details of the implementation of SQL in Microsoft SQL Server 2000, Oracle, MySQL and PostgreSQL.
Data can be stored in a database in various data types. The range of data types is DBMS-specific, but the SQL standard defines many types, including the following that are widely implemented (often not by the SQL name).
varchar. Almost always has a limit of 255 chars.
There are variants on
timezone. Other types widely implemented are
blob, for large blocks of text and binary data, respectively.
The more comprehensive of the R interface packages hide the type conversion issues from the user.
There are several packages available on CRAN to help R communicate with DBMSs. They provide different levels of abstraction. Some provide means to copy whole data frames to and from databases. All have functions to select data within the database via SQL queries, and to retrieve the result as a whole as a data frame or in pieces (usually as groups of rows).
All except RODBC are tied to one DBMS, but there has been a
proposal for a unified `front-end' package DBI
(http://developer.r-project.org/db) in conjunction with a
`back-end', the most developed of which is RMySQL. Also on
CRAN are the back-ends ROracle, RPostgreSQL and
RSQLite (which works with the bundled DBMS
http://www.sqlite.org), RJDBC (which uses Java and can
connect to any DBMS that has a JDBC driver) and RpgSQL (a
specialist interface to PostgreSQL built on top of RJBDC).
The BioConductor project has updated RdbiPgSQL (formerly on CRAN ca 2000), a first-generation interface to PostgreSQL.
http://www.joeconway.com/plr/) is a project to embed R into
Package RMySQL on CRAN provides an interface to the
MySQL database system (see http://www.mysql.com and Dubois,
2000.). The description here applies to version
versions had a substantially different interface. The current version
requires the DBI package, and this description will apply with
minor changes to all the other back-ends to DBI.
MySQL exists on Unix/Linux/Mac OS X and Windows: there is a `Community Edition' released under GPL but commercial licenses are also available. MySQL was originally a `light and lean' database. (It preserves the case of names where the operating file system is case-sensitive, so not on Windows.)
dbDriver("MySQL") returns a database connection manager
object, and then a call to
dbConnect opens a database connection
which can subsequently be closed by a call to the generic function
dbDriver("SQLite") with those
DBMSs and packages ROracle, RPostgreSQL or RSQLite
SQL queries can be sent by either
dbGetquery sends the query and retrieves the
results as a data frame.
dbSendQuery sends the query and returns
an object of class inheriting from
"DBIResult" which can be used
to retrieve the results, and subsequently used in a call to
dbClearResult to remove the result.
fetch is used to retrieve some or all of the rows in the
query result, as a list. The function
dbHasCompleted indicates if
all the rows have been fetched, and
dbGetRowCount returns the
number of rows in the result.
These are convenient interfaces to read/write/test/delete tables in the
dbWriteTable copy to and from
an R data frame, mapping the row names of the data frame to the field
row_names in the
> library(RMySQL) # will load DBI as well ## open a connection to a MySQL database > con <- dbConnect(dbDriver("MySQL"), dbname = "test") ## list the tables in the database > dbListTables(con) ## load a data frame into the database, deleting any existing copy > data(USArrests) > dbWriteTable(con, "arrests", USArrests, overwrite = TRUE) TRUE > dbListTables(con)  "arrests" ## get the whole table > dbReadTable(con, "arrests") Murder Assault UrbanPop Rape Alabama 13.2 236 58 21.2 Alaska 10.0 263 48 44.5 Arizona 8.1 294 80 31.0 Arkansas 8.8 190 50 19.5 ... ## Select from the loaded table > dbGetQuery(con, paste("select row_names, Murder from arrests", "where Rape > 30 order by Murder")) row_names Murder 1 Colorado 7.9 2 Arizona 8.1 3 California 9.0 4 Alaska 10.0 5 New Mexico 11.4 6 Michigan 12.1 7 Nevada 12.2 8 Florida 15.4 > dbRemoveTable(con, "arrests") > dbDisconnect(con)
Package RODBC on CRAN provides an interface to database sources supporting an ODBC interface. This is very widely available, and allows the same R code to access different database systems. RODBC runs on Unix/Linux, Windows and Mac OS X, and almost all database systems provide support for ODBC. We have tested Microsoft SQL Server, Access, MySQL, PostgreSQL, Oracle and IBM DB2 on Windows and MySQL, Oracle, PostgreSQL and SQLite on Linux.
ODBC is a client-server system, and we have happily connected to a DBMS running on a Unix server from a Windows client, and vice versa.
On Windows ODBC support is part of the OS. On Unix/Linux you will need an ODBC Driver Manager such as unixODBC (http://www.unixODBC.org) or iOBDC (http://www.iODBC.org: this is pre-installed in Mac OS X) and an installed driver for your database system.
Windows provides drivers not just for DBMSs but also for Excel (.xls) spreadsheets, DBase (.dbf) files and even text files. (The named applications do not need to be installed. Which file formats are supported depends on the the versions of the drivers.) There are versions for Excel and Access 2007/2010 (go to http://download.microsoft.com, and search for `Office ODBC', which will lead to AccessDatabaseEngine.exe), the `2007 Office System Driver' (the latter has a version for 64-bit Windows, and that will also read earlier versions).
On Mac OS X the Actual Technologies (http://www.actualtechnologies.com/product_access.php) drivers provide ODBC interfaces to Access databases (including Access 2007/2010) and to Excel spreadsheets (not including Excel 2007/2010).
Many simultaneous connections are possible. A connection is opened by a
odbcDriverConnect (which on the
Windows GUI allows a database to be selected via dialog boxes) which
returns a handle used for subsequent access to the database. Printing a
connection will provide some details of the ODBC connection, and calling
odbcGetInfo will give details on the client and server.
A connection is closed by a call to
and also (with a warning) when not R object refers to it and at the end
of an R session.
Details of the tables on a connection can be found using
sqlSave copies an R data frame to a table in the
sqlFetch copies a table in the database to an R
An SQL query can be sent to the database by a call to
sqlQuery. This returns the result in an R data frame.
sqlCopy sends a query to the database and saves the result as a
table in the database.) A finer level of control is attained by first
odbcQuery and then
sqlGetResults to fetch the
results. The latter can be used within a loop to retrieve a limited
number of rows at a time, as can function
Here is an example using PostgreSQL, for which the ODBC driver
maps column and data frame names to lower case. We use a database
testdb we created earlier, and had the DSN (data source name) set
up in ~/.odbc.ini under
unixODBC. Exactly the same code
worked using MyODBC to access a MySQL database under Linux or Windows
(where MySQL also maps names to lowercase). Under Windows,
DSNs are set up in the ODBC applet in the Control
Panel (`Data Sources (ODBC)' in the `Administrative Tools' section).
> library(RODBC) ## tell it to map names to l/case > channel <- odbcConnect("testdb", uid="ripley", case="tolower") ## load a data frame into the database > data(USArrests) > sqlSave(channel, USArrests, rownames = "state", addPK = TRUE) > rm(USArrests) ## list the tables in the database > sqlTables(channel) TABLE_QUALIFIER TABLE_OWNER TABLE_NAME TABLE_TYPE REMARKS 1 usarrests TABLE ## list it > sqlFetch(channel, "USArrests", rownames = "state") murder assault urbanpop rape Alabama 13.2 236 58 21.2 Alaska 10.0 263 48 44.5 ... ## an SQL query, originally on one line > sqlQuery(channel, "select state, murder from USArrests where rape > 30 order by murder") state murder 1 Colorado 7.9 2 Arizona 8.1 3 California 9.0 4 Alaska 10.0 5 New Mexico 11.4 6 Michigan 12.1 7 Nevada 12.2 8 Florida 15.4 ## remove the table > sqlDrop(channel, "USArrests") ## close the connection > odbcClose(channel)
As a simple example of using ODBC under Windows with a Excel spreadsheet, we can read from a spreadsheet by
> library(RODBC) > channel <- odbcConnectExcel("bdr.xls") ## list the spreadsheets > sqlTables(channel) TABLE_CAT TABLE_SCHEM TABLE_NAME TABLE_TYPE REMARKS 1 C:\\bdr NA Sheet1$ SYSTEM TABLE NA 2 C:\\bdr NA Sheet2$ SYSTEM TABLE NA 3 C:\\bdr NA Sheet3$ SYSTEM TABLE NA 4 C:\\bdr NA Sheet1$Print_Area TABLE NA ## retrieve the contents of sheet 1, by either of > sh1 <- sqlFetch(channel, "Sheet1") > sh1 <- sqlQuery(channel, "select * from [Sheet1$]")
Notice that the specification of the table is different from the name
sqlFetch is able to map the
Binary connections (Connections) are now the preferred way to handle binary files.
Packages hdf5, h5r, Bioconductor's rhdf5, RNetCDF, ncdf and ncdf4 on CRAN provide interfaces to NASA's HDF5 (Hierarchical Data Format, see http://hdf.ncsa.uiuc.edu/HDF5/) and to UCAR's netCDF data files (network Common Data Form, see http://www.unidata.ucar.edu/packages/netcdf/).
Both of these are systems to store scientific data in array-oriented ways, including descriptions, labels, formats, units, .... HDF5 also allows groups of arrays, and the R interface maps lists to HDF5 groups, and can write numeric and character vectors and matrices.
NetCDF's version 4 format (confusingly, implemented in netCDF 4.1.1 and later, but not in 4.0.1) includes the use of various HDF5 formats. This is handled by package ncdf4 whereas RNetCDF and ncdf handle version 3 files.
The availability of software to support these formats is somewhat limited by platform, especially on Windows.
dBase was a DOS program written by Ashton-Tate and later owned by
Borland which has a binary flat-file format that became popular, with
file extension .dbf. It has been adopted for the 'Xbase' family
of databases, covering dBase, Clipper, FoxPro and their Windows
equivalents Visual dBase, Visual Objects and Visual FoxPro (see
http://www.e-bachmann.dk/docs/xbase.htm). A dBase file contains
a header and then a series of fields and so is most similar to an R
data frame. The data itself is stored in text format, and can include
character, logical and numeric fields, and other types in later versions
write.dbf provide ways to read and
write basic DBF files on all R platforms. For Windows users
odbcConnectDbase in package RODBC provides more
comprehensive facilities to read DBF files via Microsoft's dBase
ODBC driver (and the Visual FoxPro driver can also be used via
Connections are used in R in the sense of Chambers (1998) and Ripley (2001), a set of functions to replace the use of file names by a flexible interface to file-like objects.
The most familiar type of connection will be a file, and file
connections are created by function
file. File connections can
(if the OS will allow it for the particular file) be opened for reading
or writing or appending, in text or binary mode. In fact, files can be
opened for both reading and writing, and R keeps a separate file
position for reading and writing.
Note that by default a connection is not opened when it is created. The
rule is that a function using a connection should open a connection
(needed) if the connection is not already open, and close a connection
after use if it opened it. In brief, leave the connection in the state
you found it in. There are generic functions
close with methods to explicitly open and close connections.
Files compressed via the algorithm used by
gzip can be used as
connections created by the function
gzfile, whereas files
bzip2 can be used via
Unix programmers are used to dealing with special files
stderr. These exist as terminal
connections in R. They may be normal files, but they might also
refer to input from and output to a GUI console. (Even with the standard
Unix R interface,
stdin refers to the lines submitted from
readline rather than a file.)
The three terminal connections are always open, and cannot be opened or
stderr are conventionally used for
normal output and error messages respectively. They may normally go to
the same place, but whereas normal output can be re-directed by a call
sink, error output is sent to
stderr unless re-directed
sink, type="message"). Note carefully the language used here:
the connections cannot be re-directed, but output can be sent to other
Text connections are another source of input. They allow R
character vectors to be read as if the lines were being read from a text
file. A text connection is created and opened by a call to
textConnection, which copies the current contents of the
character vector to an internal buffer at the time of creation.
Text connections can also be used to capture R output to a character
textConnection can be asked to create a new character
object or append to an existing one, in both cases in the user's
workspace. The connection is opened by the call to
textConnection, and at all times the complete lines output to the
connection are available in the R object. Closing the connection
writes any remaining output to a final element of the character vector.
Pipes are a special form of file that connects to another
process, and pipe connections are created by the function
Opening a pipe connection for writing (it makes no sense to append to a
pipe) runs an OS command, and connects its standard input to whatever
R then writes to that connection. Conversely, opening a pipe
connection for input runs an OS command and makes its standard output
available for R input from that connection.
URLs of types `http://', `ftp://' and `file://'
can be read from using the function
url. For convenience,
file will also accept these as the file specification and call
Sockets can also be used as connections via function
socketConnection on platforms which support Berkeley-like sockets
(most Unix systems, Linux and Windows). Sockets can be written to or
read from, and both client and server sockets can be used.
We have described functions
sink as writing to a file, possibly appending to a file if
append = TRUE, and this is what they did prior to R
The current behaviour is equivalent, but what actually happens is that
file argument is a character string, a file connection
is opened (for writing or appending) and closed again at the end of the
function call. If we want to repeatedly write to the same file, it is
more efficient to explicitly declare and open the connection, and pass
the connection object to each call to an output function. This also
makes it possible to write to pipes, which was implemented earlier in a
limited way via the syntax
file = "|cmd" (which can still be
There is a function
writeLines to write complete text lines
to a connection.
Some simple examples are
zz <- file("ex.data", "w") # open an output file connection cat("TITLE extra line", "2 3 5 7", "", "11 13 17", file = zz, sep = "\n") cat("One more line\n", file = zz) close(zz) ## convert decimal point to comma in output, using a pipe (Unix) ## both R strings and (probably) the shell need \ doubled zz <- pipe(paste("sed s/\\\\./,/ >", "outfile"), "w") cat(format(round(rnorm(100), 4)), sep = "\n", file = zz) close(zz) ## now look at the output file: file.show("outfile", delete.file = TRUE) ## capture R output: use examples from help(lm) zz <- textConnection("ex.lm.out", "w") sink(zz) example(lm, prompt.echo = "> ") sink() close(zz) ## now `ex.lm.out' contains the output for futher processing. ## Look at it by, e.g., cat(ex.lm.out, sep = "\n")
The basic functions to read from connections are
readLines. These take a character string argument and open a
file connection for the duration of the function call, but explicitly
opening a file connection allows a file to be read sequentially in
Other functions that call
scan can also make use of connections,
Some simple examples are
## read in file created in last examples readLines("ex.data") unlink("ex.data") ## read listing of current directory (Unix) readLines(pipe("ls -1")) # remove trailing commas from an input file. # Suppose we are given a file `data' containing 450, 390, 467, 654, 30, 542, 334, 432, 421, 357, 497, 493, 550, 549, 467, 575, 578, 342, 446, 547, 534, 495, 979, 479 # Then read this by scan(pipe("sed -e s/,$// data"), sep=",")
For convenience, if the
file argument specifies a FTP or HTTP
URL, the URL is opened for reading via
Specifying files via `file://foo.bar' is also allowed.
C programmers may be familiar with the
ungetc function to push
back a character onto a text input stream. R connections have the
same idea in a more powerful way, in that an (essentially) arbitrary
number of lines of text can be pushed back onto a connection via a call
Pushbacks operate as a stack, so a read request first uses each line
from the most recently pushbacked text, then those from earlier
pushbacks and finally reads from the connection itself. Once a
pushbacked line is read completely, it is cleared. The number of
pending lines pushed back can be found via a call to
A simple example will show the idea.
> zz <- textConnection(LETTERS) > readLines(zz, 2)  "A" "B" > scan(zz, "", 4) Read 4 items  "C" "D" "E" "F" > pushBack(c("aa", "bb"), zz) > scan(zz, "", 4) Read 4 items  "aa" "bb" "G" "H" > close(zz)
Pushback is only available for connections opened for input in text mode.
A summary of all the connections currently opened by the user can be
showConnections(), and a summary of all connections,
including closed and terminal connections, by
The generic function
seek can be used to read and (on some
connections) reset the current position for reading or writing.
Unfortunately it depends on OS facilities which may be unreliable
(e.g. with text files under Windows). Function
seek can change the position on the connection
given by its argument.
truncate can be used to truncate a file opened for
writing at its current position. It works only for
connections, and is not implemented on all platforms.
writeBin read to and write from
binary connections. A connection is opened in binary mode by appending
"b" to the mode specification, that is using mode
reading, and mode
"ab" (where appropriate) for
writing. The functions have arguments
readBin(con, what, n = 1, size = NA, endian = .Platform$endian) writeBin(object, con, size = NA, endian = .Platform$endian)
In each case
con is a connection which will be opened if
necessary for the duration of the call, and if a character string is
given it is assumed to specify a file name.
It is slightly simpler to describe writing, so we will do that first.
object should be an atomic vector object, that is a vector of
raw, without attributes. By default this is
written to the file as a stream of bytes exactly as it is represented in
readBin reads a stream of bytes from the file and interprets them
as a vector of mode given by
what. This can be either an object
of the appropriate mode (e.g.
what=integer()) or a character
string describing the mode (one of the five given in the previous
specifies the maximum number of vector elements to read from the
connection: if fewer are available a shorter vector will be returned.
signed allows 1-byte and 2-byte integers to be
read as signed (the default) or unsigned integers.
The remaining two arguments are used to write or read data for
interchange with another program or another platform. By default binary
data is transferred directly from memory to the connection or vice
versa. This will not suffice if the data are to be transferred to a
machine with a different architecture, but between almost all R
platforms the only change needed is that of byte-order. Common PCs
(`ix86'-based and `x86_64'-based machines), Compaq Alpha
and Vaxen are little-endian, whereas Sun Sparc, mc680x0 series,
IBM R6000, SGI and most others are big-endian. (Network
byte-order (as used by XDR, eXternal Data Representation) is
big-endian.) To transfer to or from other programs we may need to do
more, for example to read 16-bit integers or write single-precision real
numbers. This can be done using the
size argument, which
(usually) allows sizes 1, 2, 4, 8 for integers and logicals, and sizes
4, 8 and perhaps 12 or 16 for reals. Transferring at different sizes
can lose precision, and should not be attempted for vectors containing
Character strings are read and written in C format, that is as a string
of bytes terminated by a zero byte. Functions
writeChar provide greater flexibility.
writeBin will pass missing and
special values, although this should not be attempted if a size change
The missing value for R logical and integer types is
the smallest representable
int defined in the C header
limits.h, normally corresponding to the bit pattern
The representation of the special values for R numeric and complex
types is machine-dependent, and possibly also compiler-dependent. The
simplest way to make use of them is to link an external application
against the standalone
Rmath library which exports double
include the header Rmath.h which defines the macros
If that is not possible, on all current platforms IEC 60559 (aka IEEE
754) arithmetic is used, so standard C facilities can be used to test
for or set
NaN values. On such
NA is represented by the
NaN value with low-word
0x7a2 (1954 in decimal).
Character missing values are written as
NA, and there are no
provision to recognize character values as missing (as this can be done
by re-assigning them once read).
Some limited facilities are available to exchange data at a lower level across network connections.
Base R comes with some facilities to communicate via BSD sockets on systems that support them (including the common Linux, Unix and Windows ports of R). One potential problem with using sockets is that these facilities are often blocked for security reasons or to force the use of Web caches, so these functions may be more useful on an intranet than externally. For new projects it is suggested that socket connections are used instead.
The earlier low-level interface is given by functions
download.file is provided to read a file from a
Web resource via FTP or HTTP and write it to a file. Often this can be
avoided, as functions such as
scan can read
directly from a URL, either by explicitly using
url to open a
connection, or implicitly using it by giving a URL as the
CORBA (Common Object Request Broker Architecture) is similar to DCOM, allowing applications to call methods, or operations, in server objects running in other applications, potentially programmed in different languages and running on different machines. There is a CORBA package available from the Omegahat project (at http://www.omegahat.org/RSCORBA/), currently for Unix but a Windows version looks to be possible.
This package allows R commands to be used to locate available CORBA servers, query the methods they provide, and dynamically invoke methods on these objects. R values given as arguments in these calls are exported in the call and made available to that operation invocation. Primitive data types (vectors and lists) are exported by default, while more complex objects are exported by reference. Examples of using this include communicating with the Gnumeric (http://www.gnumeric.org) spreadsheet, and also interacting with the data visualization system ggobi.
One can also create CORBA servers within R, allowing other applications to call these methods. For example, one might offer access to a particular dataset or to some of R's modelling software. This is done dynamically by combining R data objects and functions. This allows one to explicitly export data and functionality from R.
One can also use the CORBA package to achieve distributed, parallel computing in R. One R session acts as a manager and dispatches tasks to different servers running in other R worker sessions. This uses the ability to invoke asynchronous or background CORBA calls in R. More information is available from the Omegahat project at http://www.omegahat.org/RSCORBA/.
The most common R data import/export question seems to be `how do I read an Excel spreadsheet'. This chapter collects together advice and options given earlier. Note that most of the advice is for pre-Excel 2007 spreadsheets and not the later .xlsx format.
The first piece of advice is to avoid doing so if possible! If you have
access to Excel, export the data you want from Excel in tab-delimited or
comma-separated form, and use
import it into R. (You may need to use
read.csv2 in a locale that uses comma as the decimal point.)
Exporting a DIF file and reading it using
read.DIF is another
If you do not have Excel, many other programs are able to read such
spreadsheets and export in a text format on both Windows and Unix, for
example Gnumeric (http://www.gnome.org/projects/gnumeric/) and
OpenOffice (http://www.openoffice.org). You can also
cut-and-paste between the display of a spreadsheet in such a program and
read.table will read from the R console or, under Windows,
from the clipboard (via
file = "clipboard" or
read.DIF function can also read from
Note that an Excel .xls file is not just a spreadsheet: such files can contain many sheets, and the sheets can contain formulae, macros and so on. Not all readers can read other than the first sheet, and may be confused by other contents of the file.
Windows users (of 32-bit R) can use
package RODBC. This can select rows and columns from any of the
sheets in an Excel spreadsheet file (at least from Excel 97–2003,
depending on your ODBC drivers: by calling
versions back to Excel 3.0 can be read). The version
odbcConnectExcel2007 will read the Excel 2007 formats as well as
earlier ones (provided the drivers are installed, including with 64-bit
Windows R: see RODBC). Mac users can also use RODBC if
they have a suitable driver (e.g. that from Actual Technologies).
Perl users have contributed a module
OLE::SpreadSheet::ParseExcel and a program
convert Excel 95–2003 spreadsheets to CSV files. Package gdata
provides a basic wrapper in its
read.xls function. With suitable
Perl modules installed this function can also read Excel 2007
32-bit Windows package xlsReadWrite from
http://www.swissr.org/ and CRAN has a function
read .xls files (based on a third-party non-Open-Source Delphi
Packages dataframes2xls and WriteXLS each contain a function
to write one or more data frames to an .xls file, using
Python and Perl respectively. Another version of
available in package xlsReadWrite.
Two packages which can read and and manipulate Excel 2007/10 spreadsheets but not earlier formats are xlsx (which requires Java) and the Omegahat package RExcelXML.
Package XLConnect can read, write and manipulate both Excel 97–2003 and Excel 2007/10 spreadsheets, requiring Java.
R. A. Becker, J. M. Chambers and A. R. Wilks (1988) The New S Language. A Programming Environment for Data Analysis and Graphics. Wadsworth & Brooks/Cole.
J. Bowman, S. Emberson and M. Darnovsky (1996) The Practical SQL Handbook. Using Structured Query Language. Addison-Wesley.
J. M. Chambers (1998) Programming with Data. A Guide to the S Language. Springer-Verlag.
P. Dubois (2000) MySQL. New Riders.
M. Henning and S. Vinoski (1999) Advanced CORBA Programming with C++. Addison-Wesley.
K. Kline and D. Kline (2001) SQL in a Nutshell. O'Reilly.
B. Momjian (2000) PostgreSQL: Introduction and Concepts. Addison-Wesley. Also downloadable at http://www.postgresql.org/docs/awbook.html.
B. D. Ripley (2001) Connections. \R News, 1/1, 16–7. \http://www.r-project.org/doc/Rnews/Rnews_2001-1.pdf
T. M. Therneau and P. M. Grambsch (2000) Modeling Survival Data. Extending the Cox Model. Springer-Verlag.
E. J. Yarger, G. Reese and T. King (1999) MySQL & mSQL. O'Reilly.
bzfile: Types of connections
cat: Output to connections
cat: Export to text files
close: Types of connections
close.socket: Reading from sockets
count.fields: Variations on read.table
data.restore: EpiInfo Minitab SAS S-PLUS SPSS Stata Systat
dataframes2xls: Reading Excel spreadsheets
file: Types of connections
format: Export to text files
ftable: Flat contingency tables
gzfile: Types of connections
hdf5: Binary data formats
isSeekable: Listing and manipulating connections
make.socket: Reading from sockets
netCDF: Binary data formats
odbcConnectDbase: dBase files (DBF)
odbcConnectExcel: Reading Excel spreadsheets
odbcConnectExcel2007: Reading Excel spreadsheets
open: Types of connections
pipe: Types of connections
read.csv: Variations on read.table
read.csv: Reading Excel spreadsheets
read.csv2: Variations on read.table
read.dbf: dBase files (DBF)
read.delim: Variations on read.table
read.delim: Reading Excel spreadsheets
read.delim2: Variations on read.table
read.DIF: Data Interchange Format (DIF)
read.DIF: Reading Excel spreadsheets
read.dta: EpiInfo Minitab SAS S-PLUS SPSS Stata Systat
read.epiinfo: EpiInfo Minitab SAS S-PLUS SPSS Stata Systat
read.fortran: Fixed-width-format files
read.ftable: Flat contingency tables
read.fwf: Fixed-width-format files
read.mtp: EpiInfo Minitab SAS S-PLUS SPSS Stata Systat
read.S: EpiInfo Minitab SAS S-PLUS SPSS Stata Systat
read.socket: Reading from sockets
read.spss: EpiInfo Minitab SAS S-PLUS SPSS Stata Systat
read.systat: EpiInfo Minitab SAS S-PLUS SPSS Stata Systat
read.table: Input from connections
read.table: Variations on read.table
read.table: Reading Excel spreadsheets
read.xls: Reading Excel spreadsheets
read.xport: EpiInfo Minitab SAS S-PLUS SPSS Stata Systat
readBin: Binary connections
readChar: Binary connections
readClipboard: Reading Excel spreadsheets
readLines: Input from connections
readLines: Using scan directly
reshape: Re-shaping data
RExcelXML: Reading Excel spreadsheets
scan: Input from connections
scan: Using scan directly
seek: Listing and manipulating connections
showConnections: Listing and manipulating connections
sink: Export to text files
sink: Output to connections
socketConnection: Types of connections
stack: Re-shaping data
stderr: Types of connections
stdin: Types of connections
stdout: Types of connections
Sys.localeconv: Variations on read.table
textConnection: Types of connections
truncate: Listing and manipulating connections
unstack.: Re-shaping data
url: Types of connections
write: Export to text files
write: Output to connections
write.csv: Export to text files
write.csv2: Export to text files
write.dbf: dBase files (DBF)
write.dta: EpiInfo Minitab SAS S-PLUS SPSS Stata Systat
write.foreign: Export to text files
write.matrix: Export to text files
write.socket: Reading from sockets
write.table: Export to text files
write.table: Output to connections
writeBin: Binary connections
writeChar: Binary connections
writeLines: Output to connections
WriteXLS: Reading Excel spreadsheets
XLConnect: Reading Excel spreadsheets
xlsReadWrite: Reading Excel spreadsheets
xlsx: Reading Excel spreadsheets
 Even then,
Windows applications may expect a Byte Order Mark which the
iconv used by R may or may not add depending
on the platform.
 This is normally fast as looking at the first entry rules out most of the possibilities.