The Likert scale analysis using parametric based Structural

Equation Modeling (SEM)

Zainudin Awang1, Asyraf Afthanorhan2, Mustafa Mamat3

1,2,3
Faculty of Economics and Management Sciences, Universiti Sultan Zainal Abidin, Kampus Gong Badak, 21300 Kuala
Terengganu, Malaysia.

Abstract
The Likert scale is commonly used in survey research using primary and secondary data to measure the respondent
attitude by asking insofar to which they agree or disagree with a particular questions. In generals, Likert scale would be
preferred in the questionnaire development stage to ascertain the researchers conducting their research needed. However,
the researchers nowadays are abuse to understand the nature of measurement scale in data analysis and thus causes the
finding obtained are meaningless. This article is aimed to compare the performance of two categories of measurement
scales which are 5 point and 10 points of Likert scales using the same sample size and research subject that would pave the
way to understand the real different between both of these ranges using Structural Equation Modeling (SEM). Moreover,
this study also interested to clarify briefly between two types of measurement scale namely ordinal and interval data. The
findings reveal that 10 points of Likert scale is more efficient than 5 points of Likert scale in operating of measurement
model.

Keywords: Likert Scale, Structural Equation Modeling (SEM), Ordinal, Interval, Parametric

1. Introduction
The Likert scale is commonly used in survey research especially from social science, management,
marketing, education, tourism, healthcare and other disciplines to measure the respondents attitude by
asking insofar to which they agree or disagree with a particular question or statement presented. A typical
scale that frequently apply by majority of researchers might be “strongly agree, somewhat agree, not
sure/undecided, somewhat disagree, and strongly disagree. At the outset, survey data using Likert scale
may seem easy to analyze or to identify the factors involve in the study, but there are another important
issues that should be addressed for a data analyst to consider it. This is because the implementation of the
Likert scale in analysis has become one of the main interested technique for each researchers and scholars
lately.
However, most of the researchers from various field abuse the true nature of the Likert scale in the
questionnaire development stage. Therefore, their findings would be troublesome or probability to derive
the true value is risky due to the requirement of the scale needed is limited. Thus, the main problem of
this research is to let the scholars understand the real strength of Likert scale using structural equation
modeling on 5 points and 10 points of Likert scale using the same quantity of data s ets and model.
Generally, the researchers prefer to choose the short ones than the long scales as they believe the result
obtained would not be affected. Consequently, the 5 points of Likert scale was frequently implemented if
it is related with the survey research. Additionally, most of the researchers believe the label for each
measurement scale is required in developing the questionnaire. In fact, their aim of the study was more
prone towards the parametric technique as it is only being rationale for interval and ratio scale. Therefore
labeling term for each scale was exactly as ranking order, but then, it is being ignored by applied
researchers as they thought it can help the respondents to make a choice for each questions presented. By
doing so, the researchers keep analyze the data obtained using parametric technique without concern the
sensitivity of the statistical assumptions.
Lately, Structural Equation Modeling (SEM) has become one of the prominent statistical method as it
is take into account of the multiple variables simultaneously and being free from the measurement error
that associated with every variable. In SEM, the achievement of fitness of measurement model must be
ensured during the Confirmatory Factor Analysis (CFA; Zainudin, 2015). The fitness of measurement
model was very sensitive to the characteristics or pattern of the data sets. Therefore, the good fitness of

corresponding author email: drzainudin888@gmail.com
14 Computational Methods in Social Sciences

measurement model was actually represents the data obtained was compatible with the theories conveyed
(Henseler, Hubona, & Ray, 2016). In order to obtain the good fitness of measurement model, the
researchers must concern with their data sets that is the data was collected based on respondents
perception. In the questionnaire, the Likert scale was considered as the measurement s cale to assess the
degree of the respondents’ opinion. If the measurement scale presented more choices, then, it will
accelerate the researchers to decide their choice.
In the questionnaire development stage, Likert scale might be coding for 5 point, 7 points and 10
points to represent of how much higher of respondent to agree of the particular question. The original idea
for the Likert scale is found in Rensis Likerts 1932 articles in Archive of Psychology titled “A Technique
for the Measurement Attitudes”. In his research, he expand the present knowledge of the procedures
namely Likert scale developed by Thurstone (Edmondson, 2005). In scale of measurement, the 5 points
typically been employed to determine the respondents agree or disagree on questions contemplated.
However, the scale used with 5 points is inadequate at determining of characters of respondents intention
especially when the researchers decide to attempt parametric test in statistical inference. Because 5 points
of Likert scales was represents of 5 choices undermine the strength of parametric techniques.
Consequently, this study interest to make a comparison between 5 point and 10 points of Likert scale
using SEM approach. The outcome of each factors would portray the real different for each strength of
scale of measurement. However, majority of researchers less sensitive on the scale used, instead more
likely concentrate on the analysis adapted. In fact, the probability to attain the true value is closely related
on the scale used in determining of the parameter estimate and hypothesis testing for inference testing. In
addition, the fitness of measurement model and reliability of construct also affected.
The structure of paper is as follows. First, relevant issues on scale of measurement and sampling
technique are discussed. This is followed by methodology and generation of scale items used. Once the
data analysis is focused on CFA approach, path analysis is discussed. Then after psychometric properties
of the scales, results, theoretical, practically, and implication of scale are presented. The paper also
concludes the recommendation, future research and scope of limitations.

2. Likert Scale in the form of rank order

Mostly, the researchers today would using the primary data as a main source for them to carry on the
particular research required. Therefore, they would design their questionnaire based on their previous
literature to support their decision to ask respondents. However, the way they design the questionnaire
can harm the result at the data analysis stage. This is because most of them are still confused to
distinguish between ordinal and interval scale in data analysis. In fact, the wrong identification data can
causes misinterpretation finding.
Ranking procedure require the respondents to order stimuli with respect to some designated property
of interest. For instances, the researchers would classify for each 5 point scales in their questionnaire
design as presented below:
Zainudin Awang, Asyraf Afthanorhan, Mustafa Mamat 15

Figure 1: Likert Scale of Ordinal order

This scale of measurement applied is inappropriate to be applied since the researcher rename of each
scale listed. This step would be recognized as ordinal data due to the rank of each scale. In parametric
test, the researchers usually intend to analyze on how much of attitude of respondent instead to ask rank
of their opinion. Parametric test can be conducted for interval and ratio data, thus, nominal and ordinal is
not eligible to be test as parametric analysis.
Figure 1 is the sample of ordinal scale that classify of each scale. This scale is only useful for the non-
parametric test such as descriptive frequency, mean, Mann Whitney test, Kruskal Wallis and so forth. So,
the major problem here is induced the researchers are inconsiderate to learn the statistical methodology
that is prior in empirical research.
In statistics discipline, there are four scale of measurement that should be understand by researchers
and scholars so that they would know on how to differentiate of all scale provided. Indeed, all scale can
be analyzed using method required to complement their analysis. Nevertheless, many infamous
researchers claim that interval data is more powerful in determining of the attitude of respondent on each
issues faced (Velleman & Wilkinson, 1993; knapp, 1990; korotayev, 2004). Therefore the interval order
often gain attraction of researchers to conduct the parametric analysis as presented Figure 2.

Scale of measurement
Basically, data can be divided into numerical and categorical data. Numerical data contains numbers
that we can manipulate using ordinary arithmetic operations, in a while, categorical data can be sorted
into categories. (Lau, Phang & Zainudin, 2012). Usually, data is classified as nominal, ordinal, interval or
ratio as explained below:
16 Computational Methods in Social Sciences

•Known as categorical data

•The lowest measurement scale
Nominal •No order or structure
•Non-parametric test used: Mode and cross tabulation (Chi-square test)
•Eg: Gender, Race, Yes/No

•Known as ranking scale

•The second lowest measurement scale
•Order structure
Ordinal •Non-parametric test used: Median and Mode, Rank order correlation, kruskal
wallis test, mann whitney test
•E.g: Level of education
•Known as rating scale
•More powerful compare than nominal and ordinal data
•Have equidistance points between each of the scale elements
Interval •Parametric test used: Mean and standard deviation, Correlation, Regression,
Analysis of Variance, Factor Analysis
•E.g: Customer satisfaction, Motivation, job performance

•Known as ratio scale

•The top level of measurement but not available in social research
Ratio •Parametric test used
•Same as interval data
•E.g: Temperature

Source: Lau, Phang & Zainudin (2012), Statistics Edition three

2.1. Treating ordinal scales as interval scales

Today, the use of scale of ordinal scale has become a common practice in statistical analysis. Most of the
researchers do not really understand of requirement for each scale measurement. For instances, researchers today
postulates ordinal scale can be used in parametric test as well such as ANOVA, t-test, regression analysis, and so
forth. In fact, those approach can be well functioned with interval or ratio scale only. Other than that, those
estimation provided for inference testing are vague.
To add, this issue has been continuing controversy since Harvard psychologist S.S. Stevens (1946) advances
his ideas concerning the connection between measurement scales and statistical analysis regarding of traditional
descriptive and inferential statistics. Marcus-Roberts and Roberts (1987) put forth that it is always appropriate to
calculate means for ordinal scales but inappropriate to make certain statements about such means. This statement
justify that the inferential statistics dependent on ordinal is not comprehensive to reflect the actual research that
involving of large sample size.
Knapp (1990) remarked that if the researcher convince that their data applied are ordinal, they should be able
to carry on their research using traditional statistics that leads to fruitful result. This statement also be agreed by
Marcus-Roberts & Roberts, (1987) that if the scale used is ordinal, nonparametric test should be employed in the
inference stage, only because of the distribution-free nature of such test but because they tend to be more
appropriate for hypotheses that are meaningful for ordinal scale.
The appropriate method will provide a meaningful hypothesis as explain by (Marcus-Roberts & Roberts,
1987).Therefore, treating of ordinal scales as interval scales in inferential stage is improper to be applied since
this scales is limited to the nonparametric test. Plus, the meaningless hypothesis would not contribute significant
impact on the social research. Thus, the researchers have to construct their questionnaire as interval scales if
Zainudin Awang, Asyraf Afthanorhan, Mustafa Mamat 17

interested to apply the parametric test. In SEM, the researchers permanently using maximum likelihood estimator
that is one of normality theory. Using this, maximum likelihood is admissible with interval scales. However,
Muthens (1998) offered a new estimator as Weighted Least Square Maximum Variance (WLSMV) in the
integration of SEM was rationale as well with ordinal scale. Such finding enable the researchers to excel their
analysis with SEM. In order to make sure the readers understand distinguish between parametric and non-
parametric approach, we listed several approaches at the Table 1 and different abilities between them at Table 2
as follows:

Table 1: Comparison Parametric and Non-Parametric Approach

Parametric Approach (Means) Non-Parametric Approach (Medians)
Paired t-test McNemar test
1-sample t-test Median test, Wilcoxon sign test
2-independent sample t-test Mann-Whitney test
Pearson Correlation Spearman Correlation, Tau Equivalent, Rank
Correlation, Ordinal Correlation
Analysis of Variance (ANOVA) Kruskal-Wallis test, Mood’s median test
Factorial DOE with one factor and one blocking Friedman test
variable
Multivariate ANOVA (MANOVA) Permutation MANOVA
Binary Logistic Cochran test
Regression Analysis Robust Regression Analysis, Quantile
Regression Analysis, Ordinal Regression

Table 2: Comparison abilities between Parametric and Non-Parametric Test

Parametric test Non-Parametric test
There are numerous statistical assumption or There is no stringent assumption or they relaxes
stringent assumption about the parameter means the statistical assumptions
Parametric center on the mean differences and Non-Parametric center on order or ranking. The
differences between medians data are changes from exact scores to rank and
different signs
The population must have the same variance No same variance (heteroscedasticity)
(homoscedasticity)
High statistical power and efficiency Low statistical power and efficiency
High sensitive to detect the large sample size Less sensitive to detect the large sample
(preferred for small samples)

2.2. Forced Measure (interval scale should be identical and independent)

By using Structural Equation Modeling (SEM) as a main method in this study, the authors would analyze two type
of ranges which is for 5 points and 10 points to obtain the result. Usually, the researchers are interested to apply 5
points of measurement scale, instead to apply 10 points of Likert scale. Most of them believe the result cannot be
affected even adapt short scale. In fact, the long scale (e.g: 10 points) is better than short scale (e.g: 5 points) in
determining of how much agree or disagree of respondent on particular questions.
For instance, if the enumerator ask the respondents to evaluate of the question consists of 5 points, the respondent
would rate the particular question in the range of 0% to 100%. If they decide to agree of that questions is 90%, the
enumerator would be claim that 90% was exactly represents of 5 points of the Likert scale which is the highest point in
scale of measurement in the questionnaire designed. This situation is contrary to the questionnaire that posits of 10
points for each items in which the enumerator would rate the question as 9 point that represents of 90% in scale of
measurement. Thereby, it can be inferred that the equidistance is not occurred in measurement scale when correlate
with short scale.
In fact, the scale of measurement is supposedly reflect the actual intention of respondent towards the question
submitted. If the 5 points is sustain in scale of measurement, the analysis provided will be not reliable even the
standard error produced is much lower than of 10 points. The short scale of measurement inhibit the actual scale of
18 Computational Methods in Social Sciences

respondent intention and thus pushing them to rate the question based on scale presented. Thus, this situation might
create a force measure in inferential statistics that provide a meaningless outcome induced there is no freedom for
respondents to make a choice of their intended.
In this respect, the expected value for mean and variance of IID random variables should be identical and
independent specifically as X1, X2,…, Xn . Means that, the scale used should be independent and identical for
constructing the linear formula as stipulated in z-test and t-test in order to prevent occurrence of force measure.
Moreover, the common method variance that tendency to harm the construct validity, measurement error and
covariance among latent variables resolved.

2.3. Sampling Technique

Sampling is the process of selecting a sample from a population. Since the information obtained from the sample is
used to generalize or to make a conclusion about the population, the sample must be selected in such a way that it will
accurately represent its population (Lau, Phang, & Zainudin, 2012).
Sampling technique can be classified broadly into two categories namely non-probability sampling technique and
the probability sampling technique. The non-probability sampling techniques encompasses of convenience sampling,
judgmental sampling, snowball sampling, and quota sampling. In a while, probability sampling encompasses of simple
random sampling, systematic sampling, stratified sampling, cluster sampling and multi-stage sampling. Both of these
sampling carry the different character of methodology in inferential statistics. In generals, non-probability sampling is
appropriate for nonparametric test whereas probability sampling is useful for parametric test.
However, the researchers fail to meet the requirement of parametric test especially during the data collection stage
that is associated with the sampling technique. At this stage, they collect the data or distribute the questionnaire based
on non-probability sampling and supposedly nonparametric test should be conducted in empirical research. Against on
that, they constantly attempt parametric analysis since the evidence clear that parametric analysis is better than non-
parametric analysis. Therefore, the finding produced is against their theoretical concept. In order to compare between 5
point and 10 points as suggested in this paper, the data was generated into two different of measurement scale that
consist same quantity of data sets. Both of these data sets will be tested at the same confirmatory model such that to
examine which kind of measurement scale was preferable for those model. In this stage, CB-SEM was ideal represents
of second generation modeling that is able to operate multiple construct measurement (refers Figure 2).

3. Finding

Figure 2: Comparison model between 5 points and 10 points

Zainudin Awang, Asyraf Afthanorhan, Mustafa Mamat 19

Figure 2 shows the result of construct measurement that consist of different measurement scales. Based on
that, the items retained in the model among them was not equivalent although the technical procedure to specify
the construct measurement was identical. Followed by those factor loadings, the next assessment such as
Composite Reliability (CR), Average Variance Extracted (AVE) and discriminant validity was performed.

3.1. The 5 points of measurement scales

Table 3: Factor loading of 5 points of measurement scale

Construct TTA TTA1 TTA2 TTA3 TR TTG LC TTE TE
TTA1 .88
TTA2 .89
TTA3 .92
B4.1 .75
B4.3 .84
B4.4 .81
B4.5 .81
C4.4 .80
D4.1 .80
D4.3 .70
D4.4 .80
D4.6 .81
A4.1 .74
A4.2 .85
A4.3 .58
C2.3 .55
C2.4 .75
A2.2 .60
A2.3 .58
A2.4 .83
A2.5 .84
B2.1 .78
B2.2 .75
B2.3 .90
B2.4 .67
E2.1 .66
E2.2 .70
E2.3 .56
D2.1 .69
D2.4 .66
CR .925 .915 .815 .772 .598 .810 .860 .677 .626
AVE .804 .643 .595 .536 .433 .523 .607 .413 .456

Table 4: Discriminant Validity

TTA TR TTG LC TTE TE
TTA .896
TR .48 .658
TTG .51 .70 .723
LC .61 .37 .89 .779
TTE .50 .75 .56 .51 .642
TE .47 .72 .80 .69 .75 .543
20 Computational Methods in Social Sciences

3.2. The 10 points of measurement scales

Table 5: Factor loadings of 10 points of measurement scales

Construct TTA TTA1 TTA2 TTA3 TR TTG LC TTE TE
TTA1 .98
TTA2 .99
TTA3 .97
B4.1 .77
B4.3 .82
B4.4 .84
B4.5 .78
C4.4 .84
D4.1 .85
D4.3 .84
D4.4 .66
D4.6 .78
A4.1 .74
A4.2 .72
A4.3 .80
C2.1 .65
C2.2 .76
C2.3 .72
C2.4 .65
A2.1 .77
A2.2 .82
A2.3 .86
A2.4 .82
A2.5 .86
B2.1 .78
B2.2 .71
B2.3 .92
B2.4 .83
E2.1 .75
E2.2 .79
E2.3 .86
E2.5 .87
D2.1 .80
D2.2 .93
D2.4 .90
CR .986 .923 .806 .798 .754 .915 .886 .890 .910
AVE .960 .668 .583 .569 .506 .683 .662 .671 .772

Table 6: Discriminant validity

TTA TR TTG LC TTE TE
TTA .997
TR .63 .711
TTG .56 .65 .826
LC .59 .49 .47 .813
TTE .79 .66 .66 .60 .819
TE .32 .43 .47 .32 .33 .878
Based on the results revealed, the construct measurement that consist of 10 points of measurement scales was
seemed more relevant under the confirmatory approach. Because the factor loadings appear in the model was
higher than 0.60 (Zainudin, 2015), value of CR was greater than 0.70 (Nunally & Bernstein, 1994; Afthanorhan,
Ahmad, & Mamat, 2014), value of AVE higher than 0.50 (Hair et al., 2009) and discriminant validity that
associated with latent variable correlation was satisfied (correlation must be less than the square root of AVE;
Fornell & Larcker, 1981). Therefore, the construct measurement of 10 points (refer Table 5 and Table 6) is able
to proceed to the next stage that is the relationships between latent variables can be confirmed. In contrast, the
results of construct measurement with 5 points of Likert scale (refer Table 3 and Table 4) was seemed fail to
satisfy the SEM requirement as this approach can be admissible when the construct measurement was satisfied in
terms of its reliability and validity. Other than that, the path analysis for estimating the relationships among latent
variables cannot be executed.
In SEM, the first priority is to ensure the construct involve in the model was reliable and valid. Thereby, this
method is frequently viewed as confirmatory tool (Bollen, 2014) that is the model must be have a strong theory.
The necessity of strong theory should be compatible with the data available to avoid more items deleted from the
Zainudin Awang, Asyraf Afthanorhan, Mustafa Mamat 21

model. In this case, we confirm that the 5 points of Likert scale was not meet with such requirement of
parametric based SEM.

4. Conclusion
The plethora of discussion and debate around measurement scale called for a fresh look at this scale as the use
of Likert scale become a common practice in social science and management research. As important aspect
endeavor, we provide an answer the question “What point of scale is preferable?” This answer is 10 points of
Likert scale as justified by our recent findings. It is particularly helpful for researchers who have educated in
measurement scale in the past, and who interest to add knowledge or understanding of the measurement scale.
With this, they know how to distinguish which one is called ordinal or interval and further they more convince to
choose the best scale for their research.
Finally, recent research confirms that 10 points of Likert scale serves a promising scale under parametric
based SEM. Both measurement and structural models can be assessed with 10 points of Likert scale that is
expected more success in determining the construct validity. We anticipate that once social science and
management’s researchers’ interest in developing questionnaire based of Likert scale becomes more pronounced,
10 points of Likert scale will face an additional substantial gain popularity and surely the danger of common
method variance in the model can be evaded.

5. References
[1] Afthanorhan, W. M. A. B. W., Ahmad, S., & Mamat, I. (2014). Pooled Confirmatory Factor Analysis (PCFA) using structural equation modeling
on volunteerism program: A step by step approach. International Journal of Asian Social Science, 4(5), 642-653.
[2] Bollen, K. A. (2014). Structural equations with latent variables. John Wiley & Sons.
[3] Edmondson, R.D. (2005), “Likert Scales: A History”, University Florida, USA, 127-133.
[4] Fornell, C., & Larcker, D. F. (1981). Structural equation models with unobservable variables and measurement error: Algebra and
statistics.Journal of marketing research, 382-388.
[5] Hair, J. F., Black, W. C., Babin, B. J., Anderson, R. E., & Tatham, R. L. (2009). Análise multivariada de dados. Bookman Editora.
[6] Henseler, J., Hubona, G., & Ray, P. A. (2016). Using PLS path modeling in new technology research: updated guidelines. Industrial
Management & Data Systems, 116(1), 2-20.
[7] Knapp, T. R. (1990). Treating ordinal scales as interval scales: an attempt to resolve the controversy. Nursing research, 39(2), 121-123.
[8] Korotayev, A. (2004). World religions and social evolution of the old world Oikumene civilizations: A cross-cultural perspective. E. Mellen Press.
[9] Kya, L. T., Ngor, P. Y., & Awang, Z. (2012). Statistics for UiTM.
[10] Likert, R. (1932). A technique for the measurement of attitudes. Archives of psychology.
[11] Marcus-Roberts, H. M., & Roberts, F. S. (1987). Meaningless statistics.Journal of Educational and Behavioral Statistics, 12(4), 383-394.
[12] Muthén, B. O. (1998). Mplus technical appendices. Los Angeles, CA: Muthén & Muthén, 2004.
[13] Nunally, J. C., & Bernstein, I. H. (1994). Psychonometric theory. New York.
[14] Velleman, P. F., & Wilkinson, L. (1993). Nominal, ordinal, interval, and ratio typologies are misleading. The American Statistician, 47(1), 65-72.
[15] Zainudin Awang. (2015). A Handbook on SEM. MPWS Publisher.
Reproduced with permission of the copyright owner. Further reproduction prohibited without
permission.