MEANS [TABLES =] {var_list} [ BY {var_list} [BY {var_list} [BY {var_list} … ]]] [ /{var_list} [ BY {var_list} [BY {var_list} [BY {var_list} … ]]] ] [/CELLS = [MEAN] [COUNT] [STDDEV] [SEMEAN] [SUM] [MIN] [MAX] [RANGE] [VARIANCE] [KURT] [SEKURT] [SKEW] [SESKEW] [FIRST] [LAST] [HARMONIC] [GEOMETRIC] [DEFAULT] [ALL] [NONE] ] [/MISSING = [INCLUDE] [DEPENDENT]]
You can use the MEANS
command to calculate the arithmetic mean and similar
statistics, either for the dataset as a whole or for categories of data.
The simplest form of the command is
MEANS v.
which calculates the mean, count and standard deviation for v. If you specify a grouping variable, for example
MEANS v BY g.
then the means, counts and standard deviations for v after having been grouped by g are calculated. Instead of the mean, count and standard deviation, you could specify the statistics in which you are interested:
MEANS x y BY g /CELLS = HARMONIC SUM MIN.
This example calculates the harmonic mean, the sum and the minimum values of x and y grouped by g.
The CELLS
subcommand specifies which statistics to calculate. The available statistics
are:
MEAN
The arithmetic mean.
COUNT
The count of the values.
STDDEV
The standard deviation.
SEMEAN
The standard error of the mean.
SUM
The sum of the values.
MIN
The minimum value.
MAX
The maximum value.
RANGE
The difference between the maximum and minimum values.
VARIANCE
The variance.
FIRST
The first value in the category.
LAST
The last value in the category.
SKEW
The skewness.
SESKEW
The standard error of the skewness.
KURT
The kurtosis
SEKURT
The standard error of the kurtosis.
HARMONIC
The harmonic mean.
GEOMETRIC
The geometric mean.
In addition, three special keywords are recognized:
DEFAULT
This is the same as MEAN
COUNT
STDDEV
.
ALL
All of the above statistics are calculated.
NONE
No statistics are calculated (only a summary is shown).
More than one table can be specified in a single command. Each table is separated by a ‘/’. For example
MEANS TABLES = c d e BY x /a b BY x y /f BY y BY z.
has three tables (the ‘TABLE =’ is optional). The first table has three dependent variables c, d and e and a single categorical variable x. The second table has two dependent variables a and b, and two categorical variables x and y. The third table has a single dependent variables f and a categorical variable formed by the combination of y and z.
By default values are omitted from the analysis only if missing values
(either system missing or user missing)
for any of the variables directly involved in their calculation are
encountered.
This behaviour can be modified with the /MISSING
subcommand.
Three options are possible: TABLE
, INCLUDE
and DEPENDENT
.
/MISSING = INCLUDE
says that user missing values, either in the dependent
variables or in the categorical variables should be taken at their face
value, and not excluded.
/MISSING = DEPENDENT
says that user missing values, in the dependent
variables should be taken at their face value, however cases which
have user missing values for the categorical variables should be omitted
from the calculation.
The dataset in repairs.sav contains the mean time between failures (mtbf) for a sample of artifacts produced by different factories and trialed under different operating conditions. Since there are four combinations of categorical variables, by simply looking at the list of data, it would be hard to how the scores vary for each category. Example 15.4 shows one way of tabulating the mtbf in a way which is easier to understand.
get file='repairs.sav'. means tables = mtbf by factory by environment. |
The results are shown in Result 15.3. The figures shown indicate the mean,
standard deviation and number of samples in each category.
These figures however do not indicate whether the results are statistically
significant. For that, you would need to use the procedures ONEWAY
, GLM
or
T-TEST
depending on the hypothesis being tested.
|
Note that there is no limit to the number of variables for which you can calculate
statistics, nor to the number of categorical variables per layer, nor the number
of layers.
However, running MEANS
on a large numbers of variables, or with categorical variables
containing a large number of distinct values may result in an extremely large output, which
will not be easy to interpret.
So you should consider carefully which variables to select for participation in the analysis.