Factor Analysis for Political Science

Dalson Figueiredo
Federal University of Pernambuco
Enivaldo Rocha
Federal University of Pernambuco
Mariana Batista
Federal University of Pernambuco
Ranulfo Paranhos José Alexandre
Federal University of Alagoas Federal University of Alagoas
August 15, 2014

Abstract
This paper provides an introductory guide to factor analysis for Polit-
ical Science. On methodological grounds, we use basic simulation to show
how this technique can be applied in empirical research designs. All statis-
tical analyses were performed based on the Statistical Package for Social
Sciences (SPSS). With this paper, we hope to facilitate the understanding
and di¤use the application of factor analysis technique in Political Science
empirical research.
Keywords: factor analysis; quantitative methods; political science.

1 INTRODUCTION1

This paper provides an introductory guide to factor analysis technique. Our tar-
geting audience are both Political Science undergraduate and graduate students
in early stages of training. We focus on the intuitive understanding of factor
analysis rather than its mathematical properties. On methodological grounds,
we use basic simulation to show how this technique can be applied in empirical
research designs. All statistical analyses were performed based on Statistical
Package for Social Sciences (SPSS). With this paper, we hope to facilitate the
understanding of factor analysis and di¤use its application in Political Science.
1 This paper received …nancial support from Universidade Federal de Pernambuco

(UFPE). We also are thankful for Berkeley Initiative for Transparency in the Social Sciences
(BITSS) for special training. All shortcomings are author´s monopoly. Data are available at
http://thedata.harvard.edu/dvn/faces/study/StudyPage.xhtml?globalId=doi:10.7910/DVN/27066

1
 The remainder of the paper is divided as follows. Next section provides
a brief review of factor analysis literature. Then we report software routines and
explain how to interpret the computational outputs. Last section summarizes
the conclusions.

2 WHAT IS FACTOR ANALYSIS AND WHAT IS GOOD FOR?

The di¤erent factor analysis techniques have a common goal: to reduce data di-
mensionality. In other words, to reduce a large number of variables to a smaller
number of factors/components (TABACHINICK AND FIDELL, 2007). Fac-
tors/components are linear combinations of the observed variables that repre-
sent the underlying structure that explain/represent the variance of the original
variables. Therefore, factor analysis is particularly suitable as reduction tech-
nique and as tool to measure phenomena that is not directly observed. For
example, imagine that the researcher aims to measure ideology. As a method-
ological alternative, she can select survey questions that theoretically measure
ideology variation. For example,

V1. Do you agree with marijuana decriminalization?

V2. Do you agree with homosexual relationships?
V3. Do you agree with children adoption by homosexual couples?
V4. Do you agree with abortion?
V5. Do you agree with income redistribution programs?

Using a dichotomy codi…cation (1 = Yes, 0 = No), people who scored …ve

(she agreed with all questions) is considered ideologically di¤erent from people
that disagree in all questions (scored zero). Factor analysis can be employed
to estimate a factor/component of ideology that is not directly observed but
explain/represents the variance in original observed variables (V1, V2, V3, V4
and V5).
Factor analysis techniques also can be applied to estimate indexes. For
example, the Human Development Index (HDI) is based on three observed vari-
ables: education, life expectancy and income. The methodology gives equal
weight for each variable so each one has the same in‡uence to compute the …nal
index. However, many times scholars are interested or need to give di¤erent
weights for variables. In these cases, factor analysis can be very helpful to iden-
tify which variables should be included and which weight they should apply in
estimating the index.

3 PLANNING A FACTOR ANALYSIS

Regarding measurement level, all variables should be continuous or discrete.

However, many times we face ordinal and nominal variables in our dataset.
Hair et al (2009) argue that is possible to include qualitative variables as long

2
as they are dummies (1/0)2 . In this case is natural to observe lower levels of com-
munalities among the original variables and the extracted factors/components.
Another important procedure is to guarantee both reliability and validity of the
original variables (CARMINES AND ZELLER, 1979).
Regarding sample size, Hair et al (2009) suggest a minimum of 50 cases.
Comrey e Lee (1992) de…ne 50 as too little, 100 as little, 200 as acceptable, 300
as good, 500 as very good and 1,000 as excellent. Tabachnick and Fidell (2007)
indicate 300 as a reliable size. Schawb (2007) suggest a minimum rate of …ve
cases for each variable. The higher the correlation pattern among the origi-
nal variables, lower is the sample size requirement to get an acceptable factor
solution. Our recommendation is the bigger the sample size, the better. Fi-
nally, we should observe a consistent pattern of correlation among the variables
(regardless of the direction). Schawb (2007) suggests that most correlation co-
e¢cients should be higher than .3. Our recommendation is that the higher the
correlations among the variables, the better.
After checking sample adequacy the next step is to select an extraction
technique and to determine the number of factors/components that should be
extracted. Comparatively, principal component analysis (PCA) is usually em-
ployed as default in di¤erent statistical software as Statistical Package for Social
Sciences (SPSS) and STATA (COSTELO AND OSBORNE, 2005). According
to Costelo e Osborne (2005), the literature about extraction methods is rare3 .
Fabrigar et al (1999) suggest that maximum likelihood is more suitable when
data follow a multivariate normal distribution. If this assumption is not met,
scholars should employ principal factors method. Tabachnick e Fidell (2007)
argue that when correlation pattern among variables is high and the number of
factors/component and communalities are the same, the …nal solution is very
similar, regardless of the extraction method.
The principal aim of the factor rotation is to facilitate the interpreta-
tion of the …nal solution without changing its original mathematical features
(TABACHNICK AND FIDELL, 2007). There are two general types of ro-
tation: (1) orthogonal and (2) oblique. Orthogonal rotation generates fac-
tors/components statistically independent (r = 0). The loading matrix summa-
rizes the correlation among the original variables and the factors/components
extracted. On the other hand, oblique rotation generates factors/components
that can be correlated. The loading matrix is divided in two: structure matrix
and pattern matrix (TABACHNICK AND FIDELL, 2007).
Factor analysis (FA) generates factors while principal component analy-
sis (PCA) produces components. PCA uses all variance among variables (unique
+ shared) while FA only uses the shares variance. Gorsuch (1977) and McArdlle
(1990) argue that when factors/components are statistically independent and
the communalities are moderate, PCA tends to overestimate the amount of vari-
ance explained by the extracted components. According to Tabachnick e Fidell
2 If all variables are qualitative (ordinal or nominal) the researcher can use or Correspon-

dence Analysis (CA) or Latent Class Analysis (LCA).

3 To get more information about it see Tabachnick and Fidell (2007), Goursuch (1983),

Harman (1976), Mulaik (1972) and Comrey e Lee (1992).

3
(2007), "factors are thought to cause the variables (...) components are sim-
ply aggregates of correlated variables" (TABACHNICK AND FIDELL, 2007:
610). Similarly, Garson (2009) argues that, "the principal component analysis
is usually preferred as data reduction technique (translating the dimensionality
of original variables in optimum space of components), while factor analysis is
usually preferred when the purpose is to detect the underling structure of data
or causal modeling".
In short, in factor analysis the researcher believes that some latent factor
has a causal e¤ect on the observed variables. Therefore, exploratory FA is a
powerful tool for theory development while con…rmatory FA is more adequate
for theory testing (TABACHNICK AND FIDELL, 2007). On the other hand,
PCA is tool for data reduction, without a well developed theoretical reason.
According to Hair et al (2009), both PCA and FA produce very similar results
when the number of variables are higher than 30 and/or if communalities are
higher than .6 for most of observed variables .

4 FACTOR ANALYSIS USING SPSS

This section describes the procedures adopted in order to facilitate the replicabil-
ity of the observed results (KING, 1995; KING, 2003; BOWERS, 2011; LUPIA,
2014; DAFOE, 2014). Our simulation has the following features: 300 observa-
tions; …ve variables and Principal Component Analysis and Maximum likelihood
as extraction techniques. All variables follow a normal distribution with mean
zero and standard deviation one. The …rst step is to observe the correlation
pattern among observed variables. Table 1 summarizes this information4 .

Table 1 - Correlation matrix (n = 300)

X2 X3 X4 X5
X1 .900 .800 .700 .600
X2 1 .720 .630 .540
X3 1 .560 .480
X4 1 .420
X5 1

The correlation matrix should have the majority of correlation coe¢cients

higher than .3 The higher the correlation among original variables, the better.
In our simulated example, as the lowest correlation is .420 (X4, X5) we are
good to continue our analysis. The next step is to analyze the sample adequacy
measures.
4 Analyze–> Dimension Reduction –> Factor. Go in DESCRIPTIVES and chose univariate

statistics, coe¢cients and KMO and Bartlett´s test of sphericity. Go in EXTRACTION and
chose principal component method and display scree plot.

4
 Table 2 - Sample adequacy measures
KMO .827
BTS (chi2 ) 1,127.30
p-value .000

The Kaiser-Meyer-Olkin (KMO) adequacy measure varies from 0 to 1. As

close to 1, the better. We should avoid using factor analysis when KMO measure
is lower than .5. The Bartlett’s Sphericity Test (BTS) also helps the researcher
to evaluate sample adequacy. The null hypothesis states that there is no cor-
relation among observed variables (r = 0). Thus, a statistical signi…cant result
indicates the null hypothesis should be rejected. In our example, we observe
a KMO of .827 with a statistically signi…cant result for BTS (p-value<.001).
Again, data suggest that our analysis is suitable. The next step is to observe
the communalities between factors/components and the observed variables.

Table 3 - Communalities
Variable Initial (PCA) Initial ML
X1 1.000 .918 .883 .999
X2 1.000 .830 .810 .811
X3 1.000 .727 .640 .641
X4 1.000 .612 .490 .490
X5 1.000 .490 .360 .360

Communalities represent the proportion of variance explained/represented

by the factor/component (SCHAWB, 2007). On substantive grounds, the com-
munalities indicate the importance of each observed variable in the estimation of
the index. For the principal component extraction, we observe that all variables
show communalities higher than .4. Costelo and Osborne (2005) argue that "if
an item has a communality of less than .4, it may either: a) not be related
to the other items, or b) suggest an additional factor that should be explored
(COSTELO AND OSBORNE, 2005: 4). The reliability analysis indicates a
Cronbach´s Alpha of .897.
The next step is to observe the total variance explained/represented by the
extracted factors/components.

Table 5 - Total variance explained

Component Total Variance (%) Cumulative (%)
1 3.577 71.540 71.540
2 .605 .12.105 84.646
3 .456 9.110 92.755
4 .282 5.639 98.394
5 .080 1.606 100.00

5
 Following Kaiser´s rule, we should extract only factors/components that
show eigenvalues higher than 1. The …rst solution suggests the extraction of
just one component that represents 71.540% of the total variance of the observed
variables. Hair et al (2009) suggest a minimum of 60%. The scree plot also is
helpful in indicating how many factors/components should be extracted.

The dotted line represents the cuto¤ of Eigenvalue 1. In order to decide how
many factors/components should be extracted, we should observe the variation
of the Eigenvalues across components looking for a big decline between compo-
nents. In this case, there are a substantial di¤erence between components 1 and
2, suggesting that we should extract only one component.
In short, the …st solution model has the following characteristics: (1) all
of the simulated variables display correlations higher than .3; (2) KMO = .827;
(3) all communalities are above .4; (4) Cronbach´s Alpha = .897 and (5) the
model explains 71.54% of all variance. These results suggest that our solution
is technically right and the factor/component was properly extracted.

5 CONCLUSION

Factor analysis is a powerful technique for data reduction and to measure phe-
nomena that is not directly observed. This paper provided an intuitive intro-
duction to factor analysis for Political Science. Our targeting audience is both
undergraduate and graduate students in early training stages. On methodolog-
ical grounds, we employed basic simulation to show how this technique can be
applied in empirical research designs. All statistical analyses were performed
based on the Statistical Package for Social Sciences (SPSS). With this paper,
we hope to facilitate the understanding and di¤use the application of factor
analysis in Political Science.

6
6 APPENDIX

References
[1] ALDENDERFER, M. S. e BLASHFIELD, R. K. "cluster Analysis". Sage
University Paper Series: Quantitative Applications in the Social Science,
1984.
[2] ALQUIST, J. S. and BREUNIG, C. (2011). Model-Based Clustering and
Typologies in the Social Sciences. Political Analysis (Winter 2012) 20 (1):
92-112
[3] BAILEY, K. D. "cluster Analysis". Sociological Methodology, vol. 6, p.
59-128, 1975
[4] COPPEDGE, M.; ALVAREZ, A.; MALDONADO, C. "Two Persistent Di-
mensions of Democracy: Contestation and Inclusiveness". Journal of Poli-
tics, v. 70, n. 3, p. 145, 2008.
[5] DAHL, R. Poliarquia: Participação e Oposição. São Paulo: Edusp, 1971.
[6] DOLNICAR, S. A review of unquestioned standards in used cluster analysis
for datadriven market segmentation. Faculty of Commerce - Papers. 2002.
[7] EVERITT, B.S. Cluster Analysis. Second Edition, London: Heineman Ed-
ucational Books Ltd, 1980.
[8] HAIR, J. et al (2009). Multivariate data analysis. 17a Edição. Prentice Hall.
[9] JANG, J. and HITCHCOCK, D. B. (2012),"Model-based Cluster Analysis
of Democracies", Journal of Data Science, 10, 297-319.
[10] JOHNSON, S. "Hierarchical clustering schemes". Psychometrika, 38, p.241-
254, 1967.
[11] LEWIS-BECK, M. and LEWIS-BECK, Michael. (2008). “Is OLS Dead?”
The Political Methodologist, vol 15, no 2: 2–4.

[12] PETERS, W. S. "cluster Analysis in Urban Demography", Social Forces,

v. 37, n. 1, p. 38-44, 1958.
[13] POHLMANN, M. C. Análise de Conglomerados. In: CORRAR, L. J.;
EDÍLSON, P.; DIAS FILHO, J. M. (Orgs.). Análise Multivariada. São
Paulo: Atlas, 2007.

[14] TABACHNICK, B.; FIDELL, L. (2007). Using multivariate analysis. Need-

ham Heights: Allyn & Bacon.