Categoricals: Replace single value by a case, so that interactions can be supported

[pspp] / doc / transformation.texi

@node Data Manipulation
@chapter Data transformations
@cindex transformations

The PSPP procedures examined in this chapter manipulate data and
prepare the active dataset for later analyses.  They do not produce output,
as a rule.

@menu
* AGGREGATE::                   Summarize multiple cases into a single case.
* AUTORECODE::                  Automatic recoding of variables.
* COMPUTE::                     Assigning a variable a calculated value.
* COUNT::                       Counting variables with particular values.
* FLIP::                        Exchange variables with cases.
* IF::                          Conditionally assigning a calculated value.
* RECODE::                      Mapping values from one set to another.
* SORT CASES::                  Sort the active dataset.
@end menu

@node AGGREGATE
@section AGGREGATE
@vindex AGGREGATE

@display
AGGREGATE 
        OUTFILE=@{*,'file-name',file_handle@} [MODE=@{REPLACE, ADDVARIABLES@}]
        /PRESORTED
        /DOCUMENT
        /MISSING=COLUMNWISE
        /BREAK=var_list
        /dest_var['label']@dots{}=agr_func(src_vars, args@dots{})@dots{}
@end display

@cmd{AGGREGATE} summarizes groups of cases into single cases.
Cases are divided into groups that have the same values for one or more
variables called @dfn{break variables}.  Several functions are available
for summarizing case contents.

The OUTFILE subcommand is required and must appear first.  Specify a
system file or portable file by file name or file
handle (@pxref{File Handles}), or a dataset by its name
(@pxref{Datasets}).
The aggregated cases are written to this file.  If @samp{*} is
specified, then the aggregated cases replace the active dataset's data.
Use of OUTFILE to write a portable file is a PSPP extension.

If OUTFILE=@samp{*} is given, then the subcommand MODE may also be
specified.
The mode subcommand has two possible values: ADDVARIABLES or REPLACE.
In REPLACE mode, the entire active dataset is replaced by a new dataset
which contains just the break variables and the destination varibles.
In this mode, the new file will contain as many cases as there are
unique combinations of the break variables.
In ADDVARIABLES mode, the destination variables will be appended to 
the existing active dataset.
Cases which have identical combinations of values in their break
variables, will receive identical values for the destination variables.
The number of cases in the active dataset will remain unchanged.
Note that if ADDVARIABLES is specified, then the data @emph{must} be
sorted on the break variables.

By default, the active dataset will be sorted based on the break variables
before aggregation takes place.  If the active dataset is already sorted
or otherwise grouped in terms of the break variables, specify
PRESORTED to save time.
PRESORTED is assumed if MODE=ADDVARIABLES is used.

Specify DOCUMENT to copy the documents from the active dataset into the
aggregate file (@pxref{DOCUMENT}).  Otherwise, the aggregate file will
not contain any documents, even if the aggregate file replaces the
active dataset.

Normally, only a single case (for SD and SD., two cases) need be
non-missing in each group for the aggregate variable to be
non-missing.  Specifying /MISSING=COLUMNWISE inverts this behavior, so
that the aggregate variable becomes missing if any aggregated value is
missing.

If PRESORTED, DOCUMENT, or MISSING are specified, they must appear
between OUTFILE and BREAK.

At least one break variable must be specified on BREAK, a
required subcommand.  The values of these variables are used to divide
the active dataset into groups to be summarized.  In addition, at least
one @var{dest_var} must be specified.

One or more sets of aggregation variables must be specified.  Each set
comprises a list of aggregation variables, an equals sign (@samp{=}),
the name of an aggregation function (see the list below), and a list
of source variables in parentheses.  Some aggregation functions expect
additional arguments following the source variable names.

Aggregation variables typically are created with no variable label,
value labels, or missing values.  Their default print and write
formats depend on the aggregation function used, with details given in
the table below.  A variable label for an aggregation variable may be
specified just after the variable's name in the aggregation variable
list.

Each set must have exactly as many source variables as aggregation
variables.  Each aggregation variable receives the results of applying
the specified aggregation function to the corresponding source
variable.  The MEAN, MEDIAN, SD, and SUM aggregation functions may only be
applied to numeric variables.  All the rest may be applied to numeric
and string variables.

The available aggregation functions are as follows:

@table @asis
@item FGT(var_name, value)
Fraction of values greater than the specified constant.  The default
format is F5.3.

@item FIN(var_name, low, high)
Fraction of values within the specified inclusive range of constants.
The default format is F5.3.

@item FLT(var_name, value)
Fraction of values less than the specified constant.  The default
format is F5.3.

@item FIRST(var_name)
First non-missing value in break group.  The aggregation variable
receives the complete dictionary information from the source variable.
The sort performed by AGGREGATE (and by SORT CASES) is stable, so that
the first case with particular values for the break variables before
sorting will also be the first case in that break group after sorting.

@item FOUT(var_name, low, high)
Fraction of values strictly outside the specified range of constants.
The default format is F5.3.

@item LAST(var_name)
Last non-missing value in break group.  The aggregation variable
receives the complete dictionary information from the source variable.
The sort performed by AGGREGATE (and by SORT CASES) is stable, so that
the last case with particular values for the break variables before
sorting will also be the last case in that break group after sorting.

@item MAX(var_name)
Maximum value.  The aggregation variable receives the complete
dictionary information from the source variable.

@item MEAN(var_name)
Arithmetic mean.  Limited to numeric values.  The default format is
F8.2.

@item MEDIAN(var_name)
The median value.  Limited to numeric values.  The default format is F8.2.

@item MIN(var_name)
Minimum value.  The aggregation variable receives the complete
dictionary information from the source variable.

@item N(var_name)
Number of non-missing values.  The default format is F7.0 if weighting
is not enabled, F8.2 if it is (@pxref{WEIGHT}).

@item N
Number of cases aggregated to form this group.  The default format is
F7.0 if weighting is not enabled, F8.2 if it is (@pxref{WEIGHT}).

@item NMISS(var_name)
Number of missing values.  The default format is F7.0 if weighting is
not enabled, F8.2 if it is (@pxref{WEIGHT}).

@item NU(var_name)
Number of non-missing values.  Each case is considered to have a weight
of 1, regardless of the current weighting variable (@pxref{WEIGHT}).
The default format is F7.0.

@item NU
Number of cases aggregated to form this group.  Each case is considered
to have a weight of 1, regardless of the current weighting variable.
The default format is F7.0.

@item NUMISS(var_name)
Number of missing values.  Each case is considered to have a weight of
1, regardless of the current weighting variable.  The default format is F7.0.

@item PGT(var_name, value)
Percentage between 0 and 100 of values greater than the specified
constant.  The default format is F5.1.

@item PIN(var_name, low, high)
Percentage of values within the specified inclusive range of
constants.  The default format is F5.1.

@item PLT(var_name, value)
Percentage of values less than the specified constant.  The default
format is F5.1.

@item POUT(var_name, low, high)
Percentage of values strictly outside the specified range of
constants.  The default format is F5.1.

@item SD(var_name)
Standard deviation of the mean.  Limited to numeric values.  The
default format is F8.2.

@item SUM(var_name)
Sum.  Limited to numeric values.  The default format is F8.2.
@end table

Aggregation functions compare string values in terms of internal
character codes.  On most modern computers, this is a form of ASCII.

The aggregation functions listed above exclude all user-missing values
from calculations.  To include user-missing values, insert a period
(@samp{.}) at the end of the function name.  (e.g.@: @samp{SUM.}).
(Be aware that specifying such a function as the last token on a line
will cause the period to be interpreted as the end of the command.)

@cmd{AGGREGATE} both ignores and cancels the current @cmd{SPLIT FILE}
settings (@pxref{SPLIT FILE}).

@node AUTORECODE
@section AUTORECODE
@vindex AUTORECODE

@display
AUTORECODE VARIABLES=src_vars INTO dest_vars
        [ /DESCENDING ]
        [ /PRINT ]
        [ /GROUP ]
        [ /BLANK = @{VALID, MISSING@} ]
@end display

The @cmd{AUTORECODE} procedure considers the @var{n} values that a variable
takes on and maps them onto values 1@dots{}@var{n} on a new numeric
variable.

Subcommand VARIABLES is the only required subcommand and must come
first.  Specify VARIABLES, an equals sign (@samp{=}), a list of source
variables, INTO, and a list of target variables.  There must the same
number of source and target variables.  The target variables must not
already exist.

By default, increasing values of a source variable (for a string, this
is based on character code comparisons) are recoded to increasing values
of its target variable.  To cause increasing values of a source variable
to be recoded to decreasing values of its target variable (@var{n} down
to 1), specify DESCENDING.

PRINT is currently ignored.

The GROUP subcommand is relevant only if more than one variable is to be
recoded.   It causes a single mapping between source and target values to
be used, instead of one map per variable.

If /BLANK=MISSING is given, then string variables which contain only 
whitespace are recoded as SYSMIS.  If /BLANK=VALID is given then they
will be allocated a value like any other.  /BLANK is not relevant
to numeric values. /BLANK=VALID is the default.

@cmd{AUTORECODE} is a procedure.  It causes the data to be read.

@node COMPUTE
@section COMPUTE
@vindex COMPUTE

@display
COMPUTE variable = expression.
  or
COMPUTE vector(index) = expression.
@end display

@cmd{COMPUTE} assigns the value of an expression to a target
variable.  For each case, the expression is evaluated and its value
assigned to the target variable.  Numeric and string
variables may be assigned.  When a string expression's width differs
from the target variable's width, the string result of the expression
is truncated or padded with spaces on the right as necessary.  The
expression and variable types must match.

For numeric variables only, the target variable need not already
exist.  Numeric variables created by @cmd{COMPUTE} are assigned an
@code{F8.2} output format.  String variables must be declared before
they can be used as targets for @cmd{COMPUTE}.

The target variable may be specified as an element of a vector
(@pxref{VECTOR}).  In this case, a vector index expression must be
specified in parentheses following the vector name.  The index
expression must evaluate to a numeric value that, after rounding down
to the nearest integer, is a valid index for the named vector.

Using @cmd{COMPUTE} to assign to a variable specified on @cmd{LEAVE}
(@pxref{LEAVE}) resets the variable's left state.  Therefore,
@code{LEAVE} should be specified following @cmd{COMPUTE}, not before.

@cmd{COMPUTE} is a transformation.  It does not cause the active dataset to be
read.

When @cmd{COMPUTE} is specified following @cmd{TEMPORARY}
(@pxref{TEMPORARY}), the @cmd{LAG} function may not be used
(@pxref{LAG}).

@node COUNT
@section COUNT
@vindex COUNT

@display
COUNT var_name = var@dots{} (value@dots{}).

Each value takes one of the following forms:
        number
        string
        num1 THRU num2
        MISSING
        SYSMIS
In addition, num1 and num2 can be LO or LOWEST, or HI or HIGHEST,
respectively.
@end display

@cmd{COUNT} creates or replaces a numeric @dfn{target} variable that
counts the occurrence of a @dfn{criterion} value or set of values over
one or more @dfn{test} variables for each case.

The target variable values are always nonnegative integers.  They are
never missing.  The target variable is assigned an F8.2 output format.
@xref{Input and Output Formats}.  Any variables, including
string variables, may be test variables.

User-missing values of test variables are treated just like any other
values.  They are @strong{not} treated as system-missing values.
User-missing values that are criterion values or inside ranges of
criterion values are counted as any other values.  However (for numeric
variables), keyword MISSING may be used to refer to all system-
and user-missing values.

@cmd{COUNT} target variables are assigned values in the order
specified.  In the command @code{COUNT A=A B(1) /B=A B(2).}, the
following actions occur:

@itemize @minus
@item
The number of occurrences of 1 between @code{A} and @code{B} is counted.

@item
@code{A} is assigned this value.

@item
The number of occurrences of 1 between @code{B} and the @strong{new}
value of @code{A} is counted.

@item
@code{B} is assigned this value.
@end itemize

Despite this ordering, all @cmd{COUNT} criterion variables must exist
before the procedure is executed---they may not be created as target
variables earlier in the command!  Break such a command into two
separate commands.

The examples below may help to clarify.

@enumerate A
@item
Assuming @code{Q0}, @code{Q2}, @dots{}, @code{Q9} are numeric variables,
the following commands:

@enumerate
@item
Count the number of times the value 1 occurs through these variables
for each case and assigns the count to variable @code{QCOUNT}.  

@item
Print out the total number of times the value 1 occurs throughout
@emph{all} cases using @cmd{DESCRIPTIVES}.  @xref{DESCRIPTIVES}, for
details.
@end enumerate

@example
COUNT QCOUNT=Q0 TO Q9(1).
DESCRIPTIVES QCOUNT /STATISTICS=SUM.
@end example

@item
Given these same variables, the following commands:

@enumerate
@item
Count the number of valid values of these variables for each case and
assigns the count to variable @code{QVALID}.

@item
Multiplies each value of @code{QVALID} by 10 to obtain a percentage of
valid values, using @cmd{COMPUTE}.  @xref{COMPUTE}, for details.

@item
Print out the percentage of valid values across all cases, using
@cmd{DESCRIPTIVES}.  @xref{DESCRIPTIVES}, for details.
@end enumerate

@example
COUNT QVALID=Q0 TO Q9 (LO THRU HI).
COMPUTE QVALID=QVALID*10.
DESCRIPTIVES QVALID /STATISTICS=MEAN.
@end example
@end enumerate

@node FLIP
@section FLIP
@vindex FLIP

@display
FLIP /VARIABLES=var_list /NEWNAMES=var_name.
@end display

@cmd{FLIP} transposes rows and columns in the active dataset.  It
causes cases to be swapped with variables, and vice versa.

All variables in the transposed active dataset are numeric.  String
variables take on the system-missing value in the transposed file.

No subcommands are required.  If specified, the VARIABLES subcommand
selects variables to be transformed into cases, and variables not
specified are discarded.  If the VARIABLES subcommand is omitted, all
variables are selected for transposition.

The variables specified by NEWNAMES, which must be a string variable, is
used to give names to the variables created by @cmd{FLIP}.  Only the
first 8 characters of the variable are used.  If
NEWNAMES is not
specified then the default is a variable named CASE_LBL, if it exists.
If it does not then the variables created by FLIP are named VAR000
through VAR999, then VAR1000, VAR1001, and so on.

When a NEWNAMES variable is available, the names must be canonicalized
before becoming variable names.  Invalid characters are replaced by
letter @samp{V} in the first position, or by @samp{_} in subsequent
positions.  If the name thus generated is not unique, then numeric
extensions are added, starting with 1, until a unique name is found or
there are no remaining possibilities.  If the latter occurs then the
FLIP operation aborts.

The resultant dictionary contains a CASE_LBL variable, a string
variable of width 8, which stores the names of the variables in the
dictionary before the transposition.  Variables names longer than 8
characters are truncated.  If the active dataset is subsequently
transposed using @cmd{FLIP}, this variable can be used to recreate the
original variable names.

FLIP honors @cmd{N OF CASES} (@pxref{N OF CASES}).  It ignores
@cmd{TEMPORARY} (@pxref{TEMPORARY}), so that ``temporary''
transformations become permanent.

@node IF
@section IF
@vindex IF

@display
IF condition variable=expression.
  or
IF condition vector(index)=expression.
@end display

The @cmd{IF} transformation conditionally assigns the value of a target
expression to a target variable, based on the truth of a test
expression.

Specify a boolean-valued expression (@pxref{Expressions}) to be tested
following the IF keyword.  This expression is evaluated for each case.
If the value is true, then the value of the expression is computed and
assigned to the specified variable.  If the value is false or missing,
nothing is done.  Numeric and string variables may be
assigned.  When a string expression's width differs from the target
variable's width, the string result of the expression is truncated or
padded with spaces on the right as necessary.  The expression and
variable types must match.

The target variable may be specified as an element of a vector
(@pxref{VECTOR}).  In this case, a vector index expression must be
specified in parentheses following the vector name.  The index
expression must evaluate to a numeric value that, after rounding down
to the nearest integer, is a valid index for the named vector.

Using @cmd{IF} to assign to a variable specified on @cmd{LEAVE}
(@pxref{LEAVE}) resets the variable's left state.  Therefore,
@code{LEAVE} should be specified following @cmd{IF}, not before.

When @cmd{IF} is specified following @cmd{TEMPORARY}
(@pxref{TEMPORARY}), the @cmd{LAG} function may not be used
(@pxref{LAG}).

@node RECODE
@section RECODE
@vindex RECODE

@display
RECODE var_list (src_value@dots{}=dest_value)@dots{} [INTO var_list].

src_value may take the following forms:
        number
        string
        num1 THRU num2
        MISSING
        SYSMIS
        ELSE
Open-ended ranges may be specified using LO or LOWEST for num1
or HI or HIGHEST for num2.

dest_value may take the following forms:
        num
        string
        SYSMIS
        COPY
@end display

@cmd{RECODE} translates data from one range of values to
another, via flexible user-specified mappings.  Data may be remapped
in-place or copied to new variables.  Numeric and
string data can be recoded.

Specify the list of source variables, followed by one or more mapping
specifications each enclosed in parentheses.  If the data is to be
copied to new variables, specify INTO, then the list of target
variables.  String target variables must already have been declared
using @cmd{STRING} or another transformation, but numeric target
variables can
be created on the fly.  There must be exactly as many target variables
as source variables.  Each source variable is remapped into its
corresponding target variable.

When INTO is not used, the input and output variables must be of the
same type.  Otherwise, string values can be recoded into numeric values,
and vice versa.  When this is done and there is no mapping for a
particular value, either a value consisting of all spaces or the
system-missing value is assigned, depending on variable type.

Mappings are considered from left to right.  The first src_value that
matches the value of the source variable causes the target variable to
receive the value indicated by the dest_value.  Literal number, string,
and range src_value's should be self-explanatory.  MISSING as a
src_value matches any user- or system-missing value.  SYSMIS matches the
system missing value only.  ELSE is a catch-all that matches anything.
It should be the last src_value specified.

Numeric and string dest_value's should be self-explanatory.  COPY
causes the input values to be copied to the output.  This is only valid
if the source and target variables are of the same type.  SYSMIS
indicates the system-missing value.

If the source variables are strings and the target variables are
numeric, then there is one additional mapping available: (CONVERT),
which must be the last specified mapping.  CONVERT causes a number
specified as a string to be converted to a numeric value.  If the string
cannot be parsed as a number, then the system-missing value is assigned.

Multiple recodings can be specified on a single @cmd{RECODE} invocation.
Introduce additional recodings with a slash (@samp{/}) to
separate them from the previous recodings.

@node SORT CASES
@section SORT CASES
@vindex SORT CASES

@display
SORT CASES BY var_list[(@{D|A@}] [ var_list[(@{D|A@}] ] ...
@end display

@cmd{SORT CASES} sorts the active dataset by the values of one or more
variables.

Specify BY and a list of variables to sort by.  By default, variables
are sorted in ascending order.  To override sort order, specify (D) or
(DOWN) after a list of variables to get descending order, or (A) or (UP)
for ascending order.  These apply to all the listed variables
up until the preceding (A), (D), (UP) or (DOWN).

The sort algorithms used by @cmd{SORT CASES} are stable.  That is,
records that have equal values of the sort variables will have the
same relative order before and after sorting.  As a special case,
re-sorting an already sorted file will not affect the ordering of
cases.

@cmd{SORT CASES} is a procedure.  It causes the data to be read.

@cmd{SORT CASES} attempts to sort the entire active dataset in main memory.
If workspace is exhausted, it falls back to a merge sort algorithm that
involves creates numerous temporary files.

@cmd{SORT CASES} may not be specified following TEMPORARY.