)

*Row and column names must be distinct and should start with an
alphabetic character.
Do not enter total scores*

*Characteristics:*

The *Friedman test* is frequently called a *two-way analysis on ranks*.
It is at the same time a generalization of the
\$3)++p;}END{printf("%d",p);}'>&negative=>Sign-Test
and the
>Spearman Rank Correlation Test. The
*Friedman test* models the ratingsof *n* (rows) judges on *k*
(columns) "treatments". A popular example are *wine tastings*, where each
judge rates a colection of wines independently of the other judges. Under
*H0*, the ratings of the judges are not related (e.g., they cannot
distinguish the wines).

The test parameter *W* is called *Kendall's coefficient of
concordance* and is related to the mean *Rank correlation coefficient*.
If the test parameter Q is high (i.e., statistically significant) then the
columns are different *and* the rows are correlated.

Rank ordering is done within rows and each row is used independent of
the others. That is, only the *relative* rank-orders within
each row are used. All entries in a single row can be multiplied by an
arbitrary number without affecting the outcome of the test.

For two columns the *Friedman test* reduces to the
\$3)++p;}END{printf("%d",p);}'>&negative=>Sign-Test.
For two rows, this test reduces to the
>Spearman Rank Correlation Test.
*H0:*

The distributions of ranks within rows are unrelated between rows.

*Assumptions:*

This is a distribution free test, so there are no strong assumptions.

*Scale:*

Ordinal within rows.

*Procedure:*

Start with *n* rows and *k* columns. Rank order the entries of each
row independently of the other rows. Sum the ranks for each column
(*Rct*) and sum the squared column totals ( Sum(*Rct^2*) ).

The test statistics are:

*W* = Sum(*Rct^2*) * 12 / (*n*^2 * *k* * ( *k^2* - 1 ) ) -
3 * ( *k* + 1 ) / ( *k* - 1 )

*Q* = *n* * ( *k* - 1 ) * *W*

*Degrees of Freedom (DoF)* = ( *k* - 1 )

Furthermore, the mean
>Spearman Rank Correlation Coeficient (*Rsm*)
between all the rows is:

*Rsm* = (*n * W* - 1) / ( *n* - 1 ) ( =
for the example above)

*Level of Significance:*

Consult a table of *Q* values for *k* = 3 and *n* <= 15,
or *k* = 4 and *n* <= 8. Otherwise, use the *Chi-square*
approximation below.

*Approximation:*

The probability distribution of *Q* under *H0* can be approximated with
a *X^2* distribution with ( *k* - 1 ) degrees of freedom if the number of
columns, *k*, > 4 or the number of rows *n*, > 15.

*Remarks:*

The *level of significance, p <= *, given above is based on an
approximation of the *Chi-square* distribution. Consult a **table** for the exact
*p* value that belongs to this *X^2 = * value.

Note: A *X^2* approximation of *Q* is not reliable for n="\$1" and k="\$2".");}'>

Limitations

If there are ties, the *p* values are incorrect. There is a correction of Q necessary that can be found here. Thanks to Eben Goodale, University of
Massachusetts, Amherst, for pointing this out to me and finding the correction.

Return to: Statistics