Chapter 9: Nonparametric Techniques

l  Parametric statistics make assumptions about normality and  homogeneity of variance of the distribution

l  Nonparametric statistics is referred to as distribution free, and no assumptions are made about the distribution of scores.

l  They are versatile in that they can deal with ranked scores and categories.

 

l  When variables do not have precise, interval-type or ratio-type data.

l  It could be categories on a questionaire, or subjective rating instruments

l  Numerical counting of events in Qualitative research can be analyzed using nonparametric techniques

l  Main Drawback is that Nonparametric are less Powerful than Parametric statistics

Chi Square: Testing Observed vs. Expected

l  Data sorted into categories, such as sex, age, grade level , treatment groups are nominal (categorical) data. 

l  A researcher may be interested in determining whether the number of cases in each category is different from what would be expected on the basis of chance, utilizing some known source of information

Chi Square

C2 = S[(O – E)2/E]

O = observed frequency; E = expected frequency

Example: Question of whether a tennis court is jinxed.  In several years there were 120 matches lost on four courts numbered 1, 2, 3, 4.  One would expect that there would be 30 losses per court. 

 

                                    Court Number

                                  1     2     3     4    Total

Observed Losses: O =  24   34   22   40    120 

Expected Losses:  E =  30   30   30   30    120

O-E                            -6   +4   -8   +10

(O – E)2                      36   16   64   100

(O – E)2/E                1.20  0.53  2.13  3.33

S[(O – E)2/E] = 7.19

df = n – 1 = 4 – 1 = 3                    

Table A.7 critical value for 3 df is 7.82 at .05 level

 

Example: Is Dr. Niceperson too lenient on her grading practices.  Her observed grades are compared with the department’s prescribed normal curve grade distribution:

 

l  The X2 = 46.46

 

l  In Table A.7 for 4 df at the .01 level of significance a score of 13.28 is needed for significance.

l  What is the interpretation?

l  Dr. Niceperson is giving too many A’s and B’s, and too few D’s and F’s

Contingency Table

l  Often there are two or more categories and two or more groups

l  Questionaires or attitude inventories

l  Example: A group of athletes and nonathletes take a sportsmanship inventory: “A baseball player who traps a fly ball between the ground and his glove should tell the umpire that he did not catch it.”

 

 

In table A.7 for

(r-1)(c-1) df or

(2-1)(3-1)=2 df a chi square of 9.21 is needed at the .01 level of significance.

Restrictions using Chi Square

The observations must be independent

The categories must be mutually exclusive

Should not be used for small samples; the expected frequency for any cell must not be less than one

A 2X2 contingency table should have a correction for continuity call the Yates Correction for Continuity:

     Corrected X2 = S[(O – E - 0.5)2/E]

Total N in a contingency table should not be less than 20.

Rank-Order Tests

Mann-Whitney U test – analogous to the parametric independent t-test

Wilcoxon matched-pairs signed-ranks test – analogous to the dependent t-test

Kruskal-Wallis ANOVA by ranks – analogous to the one-way ANOVA

Friedman two-way ANOVA by ranks – analogous to the repeated-measures ANOVA

Spearman Rank-difference correlation – analogous to the Pearson product moment correlation r.

Rank-Order Tests

These tests are not the best statistical tests. 

These tests have been very popular and are frequently found in the literature.

The better procedure to follow is:

l   Change all data values to ranks

l   Run the same parametric statistical tests

l   Calculate the L-statistic as the significant test and compare to the X2 table

Correlation Example

l  Correlation between biceps and forearm skinfolds on 157 participants

l  Figure 9.1

l  Data not normally distributed

l  Data skewed and kurtotic

l  r = 0.28  using ranks

l  L = (N-1)r2 = (157-1)(.28) = 12.23

l  Compare to X2 table for 1 df = 10.83 for p=.001;  indicating that this r is significant