Bird – Networks & Grades
29
available to fill out the questionnaire. Obviously, this creates distinct differences between
the two versions of the network depending on which way the experiment is bounded.
Fortunately, in Table 1 below it is apparent that the differences between the “entire class”
versions versus the “respondent only” version are relatively minimal. In the “entire class”
model, respondents have indicated interactions with non-respondents. Data was
symmetrized (made equal) in such a way that any interaction was made reciprocal if one
node indicated an interaction.
18
The “respondent only” network also symmetrized data as
described above. However, in this model, all non-respondents were eliminated from the
experiment. The “respondent only” version was created to test for correlations with
motivation and party affiliation (data that was only available from respondents).
T
ABLE
1:
C
OMPARISON OF
E
XPERIMENT
M
ODELS
C
ORRELATIONS BETWEEN
D
EGREE
C
ENTRALITY AND
G
RADE
Correlation
Variables
Bounded Model
(Entire Class or
respondents only)
Centrality
weighting
(if applicable)
Correlation
R Square
Proportion of
variance explained
F Value
Model Fit
One-Tailed
Probability
Significance
Entire class (294)
Weighted ties
0.312
0.097
31.497
< 0.001 %
Entire class
‘1’ ties purged
‘2&3’ ties
dichotomized
0.286
0.082
26.063
< 0.001 %
** Entire class
** all ties
dichotomized
0.323
0.104
33.971
< 0.001 %
Respondents only
(198)
Weighted ties
0.263
0.069
14.590
< 0.001 %
Respondents only
‘1’ ties purged
‘2&3’ ties
dichotomized
0.251
0.063
13.233
< 0.002 %
Degree centrality
and grade
Respondents only
all ties
dichotomized
0.282
0.080
16.942
< 0.001 %
Degree centrality
and grade
(motivation control
variable)
Respondents only
all ties
dichotomized
0.222
0.152
17.504
< 0.001 %
Motivation
and grade
(degree centrality
control variable)
Respondents only
all ties
dichotomized
0.282
0.152
17.504
< 0.001 %
** Best-case model (highest statistical significance)
18
Thus, if student M indicated that they had a low interaction with student P but student P did not indicate
that they had any interaction with student M then they would both be marked with a low interaction.
However, if student M indicated they had a medium interaction with student P and student P indicated they
had a low interaction with student M, the data would not have been changed. The symmetrization process
produces a default version of interaction that is biased towards a respondent who believes an interaction
took place rather than a respondent who might have forgotten the interaction.
Convention