2012 - Questionnaire Evaluation FactAnalysis PDF
2012 - Questionnaire Evaluation FactAnalysis PDF
2012 - Questionnaire Evaluation FactAnalysis PDF
1. Introduction
The pleasure writers experience in writing considerably influences their motivation and consequently
their writing performance (Hayes, 1996). Low-motivated writers perform worse, since they spend less
time on a writing task, are less engaged in a writing task, study less thoroughly instructional material,
and are less willing to attend writing training sessions.
Although someone’s pleasure in writing needs to be taken into account in order to draw valid
conclusions about the factors influencing someone’s writing performance, it is rather difficult to
measure such attitude (O’Keefe, 2002). We cannot look into writers’ minds. We just can ask them to
externalize the attitude we are interested in, but then we probably do not get a truthful answer
(Thurstone, 1977). Assuming that a writing researcher highly values writing, writers probably present
their attitude more positively than it is. Besides, if we just ask writers whether they like writing or not,
we cannot get insight in the aspects that are related to that attitude (O’Keefe, 2002). For example, the
amount of writing experience they have. Since writing is less laborious when you have a lot of
experience, highly experienced writers generally like it more (Hayes, 1996).
To avoid socially preferred answers and be able to receive information about an attitude and aspects
related to an attitude, researchers prefer the use of questionnaires asking for a person’s degree of
agreement with evaluative statements about the object of attitude and related aspects (O’Keefe, 2002).
However, the use of such method does not necessarily mean that reliable and valid indications of
someone’s attitude can be obtained. In the end, some items can measure a completely different
construct than the attitude of interest (Ratray & Jones, 2007).
In this paper, two statistical methods are discussed extensively with which the validity and
reliability of a questionnaire measuring an attitude and attitude related aspects can be tested:
exploratory factor analysis and Cronbach’s alpha (Bornstedt, 1977; Ratray & Jones, 2007). To show
how these tests should be conducted and the results interpreted, a questionnaire used to determine
Dutch seventh graders’ pleasure in writing will be evaluated.
2. Data
In the school years 2010-2011 and 2011-2012, the Centre for Language, Education and
Communication of the University of Groningen has conducted an experiment to test whether writing
instruction in secondary school content courses improves the writing skills, writing attitude and
content knowledge of seventh graders. 114 Seventh graders received instruction in writing an
expository text in the Dutch class. Afterwards, the 57 seventh graders in the experimental group wrote
that genre three times in the history and three times in the science class. The 57 seventh graders in the
control group followed the normal procedure in the history and science classes: making workbook
exercises about the topics of interest. Before and after the intervention, all seventh graders had to write
an expository text about the same subject to test their skill change, had to do a content knowledge
exam to test their knowledge change, and had to fill in a questionnaire to test their attitude change. The
effects of the intervention could be determined by comparing the skill, knowledge and attitude changes
of both groups.
1
p01 I love writing.
p10 I just write, when I can get a good grade for it.
p13 When I have the opportunity to determine on my own what I do in the Dutch class, I
usual do a writing task.
p14 I write even if the teacher does not assign a writing task.
Table 1.
Questionnaire to seventh’ graders pleasure in writing
The attitude measure focused on two attitudes: a writer’s self-efficacy in writing and a writer’s
pleasure in writing. Participants had to indicate their level of agreement with 40 evaluative statements
about writing on a 5-point Likert scale (ranging from strongly agree to strongly disagree). The level of
agreement with the first 20 statements revealed participants’ self-efficacy, the level of agreement with
the other 20 statements participants’ level of pleasure in writing. 7 of each 20 items were formulated
negatively instead of positively to force students to evaluate every statement in its own right. When all
items are formulated in the same direction, people seem to evaluate them equally (Ratray & Jones,
2007). In this paper, just the reliability and validity check of the second part of the questionnaire -
students’ pleasure in writing - is discussed (for the questionnaire see Table 1).
3. Factor analysis
With factor analysis, the construct validity of a questionnaire can be tested (Bornstedt, 1977; Ratray &
Jones, 2007). If a questionnaire is construct valid, all items together represent the underlying construct
2
well. Hence, one’s total score on the twenty items of the questionnaire of interest should represent
one’s pleasure in writing correctly. Exploratory factor analysis detects the constructs - i.e. factors - that
underlie a dataset based on the correlations between variables (in this case, questionnaire items) (Field,
2009; Tabachnik & Fidell, 2001; Rietveld & Van Hout, 1993). The factors that explain the highest
proportion of variance the variables share are expected to represent the underlying constructs. In
contrast to the commonly used principal component analysis, factor analysis does not have the
presumption that all variance within a dataset is shared (Costello & Osborne, 2005; Field, 2009;
Tabachnik & Fidell, 2001; Rietveld & Van Hout, 1993). Since that generally is not the case either,
factor analysis is assumed to be a more reliable questionnaire evaluation method than principal
component analysis (Costello & Osborne, 2005).
3.1. Prerequisites
In order to conduct a reliable factor analysis the sample size needs to be big enough (Costello &
Osborne, 2005; Field, 2009; Tabachnik & Fidell, 2001). The smaller the sample, the bigger the chance
that the correlation coefficients between items differ from the correlation coefficients between items in
other samples (Field, 2009). A common rule of thumb is that a researcher at least needs 10-15
participants per item. Since the sample size in this study is 114 instead of the required 200-300, we
could conclude that a factor analysis should not be done with this data set. Yet, it largely depends on
the proportion of variance in a dataset a factor explains how large a sample needs to be. If a factor
explains lots of variance in a dataset, variables correlate highly with that factor, i.e. load highly on that
factor. A factor with four or more loadings greater than 0.6 “is reliable regardless of sample
size.” (Field, 2009, p. 647). Fortunately, we do not have to do a factor analysis in order to determine
whether our sample size is adequate, the Kaiser-Meyer-Okin measure of sampling adequacy (KMO)
can signal in advance whether the sample size is large enough to reliably extract factors (Field, 2009).
The KMO “represents the ratio of the squared correlation between variables to the squared partial
correlation between variables.” (Field, 2009, p. 647). When the KMO is near 0, it is difficult to extract
a factor, since the amount of variance just two variables share (partial correlation) is relatively large in
comparison with the amount of variance two variables share with other variables (correlation minus
partial correlation). When the KMO is near 1, a factor or factors can probably be extracted, since the
opposite pattern is visible. Therefore,
KMO “values between 0.5 and 0.7 are
5
category (KMO=0.922).
Another prerequisite for factor
3
and to conduct a maximum likelihood Figure 1. Q-Q plot of the variable p06: ‘I make sure that I
factor analysis to determine validly how have to write as less as possible.’
p01 p02 p03 p04 p05 p06 p07 p08 p09 p10 p11 p12 p13 p14 p15 p16 p17 p18 p19 p20
p01 1.00 0.67 0.61 -0.56 0.41 -0.39 0.37 -0.46 0.62 -0.37 -0.52 0.43 0.49 0.57 0.47 -0.41 0.60 -0.47 0.55 0.52
p02 0.66 1.00 0.62 -0.42 0.35 -0.30 0.25 -0.34 0.58 -0.20 -0.34 0.42 0.47 0.53 0.50 -0.29 0.61 -0.30 0.50 0.59
p03 0.61 0.62 1.00 -0.52 0.47 -0.34 0.43 -0.39 0.59 -0.33 -0.40 0.51 0.49 0.60 0.60 -0.39 0.69 -0.40 0.51 0.57
p04 -0.56 -0.42 -0.52 1.00 -0.22 0.51 -0.32 0.47 -0.47 0.46 0.61 -0.44 -0.31 -0.31 -0.35 0.67 -0.43 0.54 -0.36 -0.44
p05 0.41 0.35 0.47 -0.22 1.00 -0.22 0.48 -0.34 0.54 -0.26 -0.27 0.24 0.33 0.46 0.48 -0.21 0.50 -0.23 0.25 0.43
p06 -0.39 -0.30 -0.34 0.51 -0.22 1.00 -0.34 0.56 -0.32 0.44 0.46 -0.31 -0.28 -0.30 -0.38 0.54 -0.40 0.45 -0.40 -0.40
p07 0.37 0.25 0.43 -0.32 0.48 -0.34 1.00 -0.38 0.35 -0.25 -0.24 0.27 0.22 0.34 0.47 -0.31 0.52 -0.20 0.29 0.41
p08 -0.46 -0.34 -0.40 0.47 -0.34 0.56 -0.38 1.00 -0.42 0.46 0.43 -0.26 -0.22 -0.41 -0.38 0.46 -0.40 0.47 -0.39 -0.39
p09 0.62 0.58 0.60 -0.47 0.55 -0.32 0.35 -0.42 1.00 -0.45 -0.41 0.55 0.46 0.61 0.54 -0.34 0.64 -0.45 0.54 0.62
p10 -0.37 -0.21 -0.34 0.46 -0.26 0.44 -0.25 0.46 -0.45 1.00 0.40 -0.26 -0.22 -0.40 -0.33 0.42 -0.38 0.63 -0.29 -0.43
p11 -0.53 -0.39 -0.41 0.61 -0.27 0.46 -0.24 0.43 -0.41 0.40 1.00 -0.38 -0.25 -0.35 -0.24 0.47 -0.33 0.45 -0.25 -0.39
p12 0.42 0.42 0.51 -0.44 0.24 -0.31 0.27 -0.26 0.55 -0.27 -0.38 1.00 0.38 0.49 0.40 -0.38 0.44 -0.42 0.45 0.41
p13 0.49 0.48 0.47 -0.31 0.33 -0.28 0.22 -0.22 0.46 -0.22 -0.25 0.38 1.00 0.50 0.45 -0.23 0.54 -0.18 0.51 0.59
p14 0.57 0.53 0.60 -0.31 0.46 -0.30 0.34 -0.41 0.61 -0.39 -0.35 0.49 0.50 1.00 0.51 -0.28 0.61 -0.38 0.45 0.56
p15 0.47 0.50 0.60 -0.35 0.48 -0.37 0.47 -0.38 0.54 -0.33 -0.24 0.40 0.45 0.51 1.00 -0.44 0.70 -0.28 0.49 0.59
p16 -0.41 -0.29 -0.39 0.67 -0.21 0.54 -0.31 0.46 -0.34 0.42 0.47 -0.38 -0.23 -0.28 -0.44 1.00 -0.39 0.53 -0.38 -0.40
p17 0.60 0.61 0.69 -0.43 0.50 -0.40 0.52 -0.39 0.64 -0.38 -0.33 0.44 0.54 0.61 0.70 -0.39 1.00 -0.35 0.59 0.75
p18 -0.47 -0.30 -0.40 0.53 -0.23 0.45 -0.20 0.47 -0.45 0.63 0.45 -0.42 -0.18 -0.38 -0.28 0.53 -0.35 1.00 -0.35 -0.40
p19 0.55 0.51 0.51 -0.36 0.25 -0.40 0.29 -0.39 0.54 -0.30 -0.25 0.45 0.51 0.45 0.49 -0.38 0.59 -0.35 1.00 0.56
p20 0.52 0.59 0.57 -0.44 0.43 -0.40 0.41 -0.39 0.63 -0.44 -0.40 0.41 0.59 0.56 0.59 -0.40 0.75 -0.40 0.56 1.00
Table 2. Correlation matrix of the dataset. The underlined correlations are too low (-0.3 < r < 0.3).
4
3.2. The factor analysis
3.2.1. Factor extraction
At heart of factor extraction lies complex algebra with the correlation matrix, it reaches beyond the
scope of this paper to explain that comprehensively and in full detail. I refer the interested reader to
chapter 6 and 7 of Statistical Techniques for the Study of Language and Language Behaviour (Rietveld
& Van Hout, 1993).
The algebraic matrix calculations finally end up with eigenvectors (Field, 2009; Tabachnik &
Fidell, 2001; Rietveld & Van Hout, 1993). As can be seen in Figure 2, these eigenvectors are linear
representations of the variance variables share. The longer an eigenvector is, the more variance it
explains, the more important it is (Field, 2009). We can calculate an eigenvector’s value by counting
up the loadings of each variable on the eigenvector. As demonstrated in Figure 3, just a small
proportion of the 20 eigenvectors of the correlation matrix in this study has a considerable eigenvalue:
many reach 0 or have even lower values. We just want to retain the eigenvectors - or factors - that
explain a considerable amount of the variance in the dataset, by which value do we draw the line?
Statistical packages generally retain factors with eigenvalues greater than 1.0 (Costello & Osborne,
2005). Yet, then there is a considerable change that too many factors are retained: in 36% of the
samples Costello and Osborne studied (2005), too many factors were retained.
A more reliable and rather easy method is to look at the scree plot, as the graph in Figure 3 is called
(Costello & Osborne, 2005). The factors with values above the point at which the curve flattens out
should be retained. The factors with values at the break point or below should be eliminated. The
statistical package R helps to determine where the break point is by drawing a straight line at that
point. Thus, looking at Figure 3, two factors should be retained.
However, the best method to determine how many factors to retain is a maximum likelihood factor
analysis, since that measure tests how well a model of a particular amount of factors accounts for the
variance within a dataset (Costello & Osborne, 2005). A high eigenvalue does not necessarily mean
that the factor explains a hugh amount of the variance in a dataset. It could explain the variance in one
cluster of variables, but not in another one. That cluster probably measures another underlying factor
8
6
Eigen values of factors
4
2
0
5 10 15 20
Figure 2. Scatterplot of two variables (x1 and x2). The Figure 3. Screeplot of factors underlying the dataset.
factor number
lines e1 and e2 represent the eigenvectors of the Every point represents one factor.
correlation matrix of variables x1 and x2. The eigenvalue
of an eigenvector is the length of an eigenvector
measured from one end of the oval to the other end.
5
Call:
factanal(x = na.omit(passion), factors = 1)
Loadings:
Factor1
p01 0.767 Factor1
p02 0.710 SS loadings 8.635
p03 0.786 Proportion Var 0.432
p04 -0.637
p05 0.557
p06 -0.542 Test of the hypothesis that 1 factor is sufficient.
p07 0.520 The chi square statistic is 336.46 on 170 degrees of freedom.
p08 -0.572 The p-value is 4.89e-13
p09 0.782
p10 -0.529
p11 -0.545
p12 0.599
p13 0.605
p14 0.719
p15 0.714
p16 -0.558
p17 0.830
p18 -0.559
p19 0.674
p20 0.783
Call:
factanal(x = na.omit(passion), factors = 2, rotation = "oblimin")
Loadings:
Factor1 Factor2
p01 0.547 -0.289 Factor1 Factor2
p02 0.747 SS loadings 6.141 3.534
p03 0.727 Proportion Var 0.307 0.177
p04 0.802 Cumulative Var 0.307 0.484
p05 0.625
p06 0.641
p07 0.463 Factor Correlations:
p08 -0.133 0.558 Factor1 Factor2
p09 0.702 -0.115 Factor1 1.000 -0.642
p10 0.597 Factor2 -0.642 1.000
p11 0.680
p12 0.412 -0.243
p13 0.719 0.122 Test of the hypothesis that 2 factors are sufficient.
p14 0.739 The chi square statistic is 197.76 on 151 degrees of freedom.
p15 0.758 The p-value is 0.00636
p16 0.771
p17 0.917
p18 0.739
p19 0.623
p20 0.771
Call:
factanal(x = na.omit(passion), factors = 3, rotation = "oblimin")
Loadings:
Factor1 Factor2 Factor3
p01 0.565 -0.263 0.293 Factor1 Factor2 Factor3
p02 0.767 0.279 SS loadings 5.963 3.539 0.585
p03 0.720 -0.106 Proportion Var 0.298 0.177 0.029
p04 0.781 -0.121 Cumulative Var 0.298 0.475 0.504
p05 0.600 -0.134
p06 0.672 0.166
p07 0.438 -0.153 -0.336
p08 -0.130 0.575
which should not be ignored. The null hypothesis in a maximum likelihood factor analysis is that the
number of factors fits well the dataset, when the null hypothesis is rejected a model with a larger
amount of factors should be considered. As can be seen in Table 3, a model with one factor is rejected
at an α-level of 0.01, a model with two factors is rejected at an α-level of 0.05 and a model with three
factors at none of the usual α-levels (p=.106). If we would set our α-level at 0.05 as common in social-
scientific research (Field, 2009), a model with three factors would be the best choice. However, the
third factor seems rather unimportant, it just explains 2.9% of the variance in the dataset and is based
on just one variable p07. Variables with loadings lower than 0.3 are considered to have a non-
significant impact on a factor, and need therefore to be ignored (Field, 2009). It seems more
appropriate to set our α-level at 0.01 and assume that two factors should be retained. The second factor
accounts for a considerable amount of the variance in the dataset: 17,7%. All variables load highly on
that factor, except for the ones that load higher on the other factor and therefore seem to make up that
one. Finally, the scree test rendered the same result.
Factor 1 Factor 1
Factor 2 Factor 2
•Figure
Figure 2: graphical
4. Graphical representation
presentation of factor
of factor rotation.
rotation: The left graph
orthogonal represents
rotation (left) andorthogonal rotation(right).
oblique rotation and theThe
axesright one represents
represent oblique
the extracted rotation.
factors, the The
starsstars representvariables
the original the loadings of the
(graph original
source: variables
Field, 2000,on
p. the
439)
factors. (Source for this figure: Field 2000: 439).
7
There are several methods to carry out rotations. SPSS offers five: varimax, quartimax,
equamax, direct oblimin and promax. The first three options are orthogonal rotation; the last
two oblique. It depends on the situation, but mostly varimax is used in orthogonal rotation and
analysis, two factors were retained. One is represented as the x-axis, the other one as the y-axis. The
variables (the stars) get their positions in the graph based on their correlation coefficients with both
factors. It is rather ambiguous to which the circled variable belongs (left graph). It loads just a bit more
on factor 1. However, by rotating both factors, the ambiguity gets solved: the variable loads highly on
factor 1 and lowly on factor 2.
As can be seen in Figure 4, there are two kinds of rotation. The first kind of rotation ‘orthogonal
rotation’ is used, when the factors are assumed to be independent (Field, 2009; Tabachnik & Fidell,
2001; Rietveld & Van Hout, 1993). The second kind of rotation ‘oblique rotation’ is used, when the
factors are assumed to correlate. Since it was assumed that all 20 items in this questionnaire measured
the same construct, we may expect that an oblique rotation is appropriate. This can be checked after
having conducted the factor analysis, since statistical packages always give a correlation matrix of the
factors when you opt an oblique rotation method (oblimin or promax). Therefore, it is highly
recommended to always do a factor analysis with oblique rotation first, even if you are quite sure that
the factors are independent (Costello & Osborne, 2005). The factors in this study certainly correlate
with each other, although negatively: r=-0.64.
4. Cronbach’s alpha
When the questionnaire at issue is reliable, people completely identical - at least with regard to their
pleasure in writing - should get the same score, and people completely different a completely different
score (Field, 2009). Yet, it is rather hard and time-consuming to find two people who are fully equal or
unequal. In statistics, therefore, it is assumed that a questionnaire is reliable when an individual item
or a set of some items renders the same result as the entire questionnaire.
The simplest method to test the internal consistency of a questionnaire is dividing the scores a
participant received on a questionnaire in two sets with an equal amount of scores and calculating the
correlation between these two sets (Field, 2009). A high correlation signals a high internal consistency.
Unfortunately, since the correlation coefficient can differ depending on the place at which you split the
dataset, you need to split the dataset as often as the number of variables in your dataset, calculate a
correlation coefficient for all the different combinations of sets and determine the questionnaire’s
reliability based on the average of all these coefficients. Cronbach came up with a faster and
comparable method to calculate a questionnaire’s reliability:
α = (N²M(Cov))/(∑s² + ∑Cov)
Assumption behind this equation is that the unique variance within variables (s²) should be rather
small in comparison with the covariance between scale items (Cov) in order to have an internal
consistent measure (Cortina, 1993).
8
Reliability analysis
Call: alpha(x = passion)
4.1. Prerequisite
Before the Cronbach’s alpha of a questionnaire can be determined, the scoring of reverse-phrased
items of a questionnaire needs to be reversed (Field, 2009). Hence, a score of 5 on a negatively
formulated item in this questionnaire should be rescored in 1, a score of 4 in 2, et cetera. Assuming
that a seventh grader who really likes writing strongly agrees with the statement ‘I like writing’ and
strongly disagrees with the statement ‘I hate writing’, item scores can differ substantially between
students as long as the scores of the reverse-phrased items are not reversed. Given that covariances
between such scores are negative, the use of reverse-phrased items will finally lead to a lower and
consequently incorrect Cronbach’s alpha, since the top half of the Cronbach’s alpha equation
incorporates the average of all covariances between items. Fortunately, R reverses scores
automatically. Since reverse-phrased items load negatively on an underlying factor, R detects them
easily by determining that underlying factor.
9
Besides, a hugh Cronbach’s alpha should not be interpreted as a signal of unidimensionality (Cortina,
1993; Field, 2009). Since α is a measure of the strength of a factor when there is just one factor
underlying the dataset, many researchers assume that a dataset is unidimensional when the α is rather
high. Yet, the α of this dataset is rather high as well, although the factor analysis revealed that the
dataset is not unidimensional. Thus, if you want to measure with Cronbach’s alpha the strength of a
factor or factors underlying a dataset, Cronbach’s alpha should be applied to all the factors extracted
during a previous factor analysis (Field, 2009). Our factors turn out to be quite strong: the α of the
factor made up of the positively phrased items is 0.93, the α of the factor made up of the negatively
phrased items is 0.86.
The final step in the interpretation of the output of a Cronbach’s alpha analysis is determining how
each item individually contributes to the reliability of the questionnaire (Field, 2009). As can be seen
in Table 4, R also renders the values of the α, when one of the items is deleted. If the α increases a lot
when a particular item is deleted, one should consider deletion. The same counts for items which
decrease the average correlation coefficient a lot, or correlate lower than 0.3 with the total score of the
questionnaire (see the values under r and r.cor in Table 4; r.cor is the item total correlation corrected
for item overlap and scale correlation). In this questionnaire, all items positively contribute to the
reliability of the questionnaire. The α remains the same when an item is deleted, the average r almost
the same, and the correlations between the total score and the item score are moderate to high.
5. Conclusion
The evaluated questionnaire seems reliable and construct valid. The items measure the same
underlying construct. The extraction of two factors in the factor analysis just seems to be a
consequence of the wording of the questionnaire items. After all, the two factors correlate highly with
each other. The result of the reliability measure was high: α=0.93. All items contribute to the reliability
and construct validity of the questionnaire: the items correlate more than 0.4 with the factors that
underlie them, the Cronbach’s alpha does not increase when one of the questionnaire items is deleted,
and the average correlation coefficient sometimes just a bit.
Although a questionnaire is generally accepted as reliable when the Cronbach’s alpha is higher than
0.8, we cannot claim that the questionnaire is valid based on the factor analysis alone (Bornstedt,
1977; Ratray & Jones, 2007). We just know that the items measure the same underlying construct. It
can be expected based on the questionnaire statements that that is one’s pleasure in writing. However,
to prove that it measures one’s pleasure in writing, the results of other measures of one’s pleasure in
writing should be compared (Bornstedt, 1977; Ratray & Jones, 2007). Unfortunately, it is not easy to
invent such measures (O’Keefe, 2002).
Bibliography
Bornstedt, G.W. (1977). Reliability and Validity in Attitude Measurement. In: G.F. Summers (Ed.),
Attitude Measurement (pp. 80-99). Kershaw Publishing Company: London.
Cortina, J.M. (1993). What is coefficient alpha? An examination of theory and applications. Journal of
Applied Psychology, 78, 98-104.
Costello, A.B., & Osborne, J.W. (2005). Best Practices in Exploratory Factor Analysis: Four
Recommendations for Getting the Most From Your Analysis. Practical Assessment, Research and
Evaluation, 10, 1-9.
Field, A. (2000). Discovering Statistics using SPSS for Windows. Sage: London.
Field, A. (2009). Discovering Statistics using SPSS. Sage: London.
Kamoen, N., Holleman, B.C., & Van den Bergh, H.H. (2007). Hoe makkelijk is een niet moeilijke
tekst? Een meta-analyse naar het effect van vraagformulering in tekstevaluatieonderzoek [How
10
easy is a text that is not difficult? A meta-analysis about the effect of question wording in text
evaluation research]. Tijdschrift voor Taalbeheersing, 29, 314–332.
O'Keefe, D.J. (2002). Persuasion. Theory & Research. Sage: Thousand Oaks.
Rattray, J.C., & Jones, M.C. (2007), Essential elements of questionnaire design and development.
Journal of Clinical Nursing, 16, 234-243.
Rietveld, T. & Van Hout, R. (1993). Statistical Techniques for the Study of Language and Language
Behaviour. Berlin - New York: Mouton de Gruyter.
Schmitt, N. & Stults, D.M. (1985). Factors Defined by Negatively Keyed Items: The Result of
Careless Respondents? Applied Psychological Measurement, 9, 367-373.
Tabachnik, B.G., & Fidell, L.S. (2001) Using Multivariate Statistics (4th ed.). Pearson: Needham
Heights, MA.
Thurstone, L.T. (1977). Attitudes can be measured. In: G.F. Summers (Ed.), Attitude Measurement (pp.
127-141). Kershaw Publishing Company: London.
11