Article Type : Research Article
Authors : Mohammed SA, Workneh BD, Belayneh YM and Hailu AD
Keywords : Multilevel; Model; Assumption; In combination
Background:
Multilevel modelling is an approach that can be used to handle clustered or
grouped data using a statistical model that specifies an estimated relationship
between variables that have been observed at different levels of a hierarchical
structure. The presence of two random variables; the measurement level random
variable and the subject level random variable are the feature that
distinguishes the multilevel model from an ordinary regression model. The aim
of this review was to review types, assumptions and model building of
multilevel analysis.
Method:
Relevant kinds of literature were searched from Google Scholar, PubMed, Hinari,
Web of Science, Scopus, and Science Direct. A total of 426 kinds of literature
were searched. After the exclusion of redundant and irrelevant literature, 15
kinds of literature were reviewed.
Result:
The full multilevel regression model assumes that their hierarchal data set,
with one single dependent variable at all existing levels. The multilevel model
can be built by estimating regression lines separately for each unit or within
unit relationships for each unit and summarizes unit relationships between
group variables variance in intercepts and slopes predicted by between unit
variables. Multilevel models have similar assumptions with major general linear
models except some of the assumptions are modified for the hierarchical nature
of the design.
Conclusion:
Multilevel modelling is generalized linear modelling where regression
coefficients themselves given a model. The models of parameters vary at more
than one level that have a hierarchical structure in which the dependent
variable is measured at the lowest level and the independent variables are
measured at all available levels.
Social research regularly involves problems that
investigate the relationship between individuals and society where individuals
interact with the social groups, to which they belong, that individual persons
and the social groups are influenced by each other. The individuals and the
social groups are conceptualized as a hierarchical system of individuals nested
within groups, with individuals and groups defined at separate levels of this
hierarchical system. Naturally, such systems can be observed at different
hierarchical levels, and variables may be defined at each level [1]. This leads
to a statistical model that specifies and estimates relationships between
variables that have been observed at different levels of a hierarchical
structure referred to as multilevel research. Multilevel modelling is an
approach that can be used to handle clustered or grouped data [2].
Multilevel models (also known as hierarchical linear
models, nested models, mixed models, random coefficient, random-effects models,
random parameter models, or split-plot designs) are statistical models that
vary more than one level. The feature that distinguishes the multilevel model
from an ordinary regression model is the presence of two random variables; the
measurement level random variable and the subject level random variable.
Because multilevel models contain a mix of fixed and random effects, they are
sometimes known as mixed-effects models [2].
Research questions demand multilevel analysis when
health simultaneously affected at the level of individuals and at the level of
contexts. Multilevel relates to the levels of analysis in public health
research consists of individuals (at lower level) who are nested within spatial
units (at higher levels). The observations that are being analysed are
correlated or clustered along spatial, non-spatial, or/and temporal dimensions;
or the causal processes are thought to operate simultaneously at more than one
level; and/or there is an intrinsic interest in describing the variability and
heterogeneity in the population, over and above the focus on average
relationships [3]. The objective of this review was to systematically review
types, assumptions and model building of multilevel analysis.
Search
strategy, inclusion and exclusion criteria
The review was conducted by adapting a search strategy
in identified databases. Books that were peer-reviewed and non-reviewed with
the time period of 1991 to present and those written in English were included
in the review.
The search for literatures was conducted in three separate ways: searches in electronic databases on the internet, hand searches and iterative reviews of reference lists of papers. The databases searched were PubMed, Hinari and Goggle Scholar. The search was conducted using the following search terms: ‘Multilevel’, ‘Model’, ‘Assumption’, and in combination. All searches were performed from September 3-8, 2020.
Figure 1: Data
searching process.
When the electronic search using ‘Multilevel’, and
‘Model’ was done, as shown in (Figure 1) 426 studies were available from all
sources. Then, adding the search engines ‘Assumption’ 54 studies were availed.
From this study, 38 were excluded because of duplication and non-fulfilment of
inclusion criteria. Finally, a total of 15 separate studies were selected for
inclusion in the systematic review, out of these 4 is published articles in
peer-reviewed journals and 11 are books.
Assessment
of methodological quality
Methodological validity was checked prior to inclusion
of selected articles and during the review by undertaking critical appraisal
using preferred reporting items for systematic reviews and meta-analysis
(PRISMA) flow diagram and guidance set out by the centre for reviews and
dissemination.
There were four reviewers in this review and three
reviewers appraised the full text of each study independently. Any
discrepancies between the three reviewers were resolved through discussion
and/or involving a fourth reviewer as arbiter. Finally, fourth reviewers
validate the final selection of publications.
Type
and nature of variables
Variables can be defined at any level of the hierarchy
in the multi-level analysis. At each level in the hierarchy, we may have
several types of variables. In multilevel data, there is not one ‘proper’ level
at which the data should be analysed. Rather, all levels present in the data
are important in their own way. This becomes clear when we investigate cross-level
hypotheses, or multilevel problems [4].
The multilevel models assume hierarchical data, in
which the dependent variable is measured at the lowest level and the
independent variables are measured at all available levels [5]. Structural data
are hierarchical or nested data are likely to be correlated [6]. Clustered data
(nested data) are measurements are taken on subjects that share a common
category that leads to correlation. Clustered longitudinal data outcome is
measured repeatedly for the same subject over time, and subjects are clustered
within some unit. In a clustered data, each “level” represents a factor that
can be thought of as a random sample from a larger population, otherwise
classification [7].
The fixed effects are regression coefficients; in
which values of interest are all represented in the dataset while random
effects are variations of regression coefficients between levels (variance
components) ever-existing natural heterogeneity among subjects. Cross-level
interaction effects are fixed effects of the joint effect of variables at level
one in conjunction with variables at level two. Mixed-effects models combine
both factor(s) [8].
Types
of multilevel models
Random
intercepts model: In this model, the
intercepts are allowed to vary, so the dependent variable for each individual
observation will be predicted by the intercept that varies across groups [9].
The model also assumes that slopes are fixed (the same across different
contexts). In addition, this model provides information about intra class
correlations, which are helpful in determining whether multilevel models are
required in the first place [10].
Random
slopes model: A random slopes model is a model in
which slopes are allowed to vary, and therefore, the slopes are different across
groups. This model assumes that intercepts are fixed (the same across different
contexts) [10].
Random
intercepts and slopes model: In this model, both
intercepts and slopes are allowed to vary across groups, meaning that they are
different in different contexts [10].
Statistical
model analysis
The multilevel model can be conceptualized as a
two-stage equation.
Step 1: Estimates separate regression equations within
units. Relationships within units (intercept and slopes).
Step 2: Uses “summaries” of between unit relationships
as group variables.
Mathematically
Level 1: Regression lines estimated separately for
each unit or within unit relationships for each unit. Analyse the model with no
explanatory variables. The intercept only model is given by the model of
equations (1).
Y = ß 0j + ß 1j X ij
+ ? ij (1)
Where: ß 0j and ß 1j =
regression coefficients
Level 2: Variance in intercepts and slopes are
predicted by between unit variables. Level 2 models variance in level-1
parameters (intercepts and slopes) with between unit variables. Analyse the
model with all lower level explanatory variables.
ß 0j = ?00 + ?01
(Group j) + U0j (2)
ß 1j = ?10 + ?11
(Group j) + U1j (3)
Adding cross level interaction between explanatory
group level variables and those individual level explanatory variables leads a
full model formulated in equation (4). So, substitute equations (2) and (3) in
equation [1]:
Y ij = ?00 + ?01g
j + ?10Ii j + ?11g j Iij + U0j
+ U1j Iij + ?ij (4)
Where:
?00 is variances of group intercepts (over
all intercept)
?10 is regression coefficient (slop of
individual variable)
?11 is slop of interaction (how large Z
affect X)
?01 is variances of group slopes (group
regression coefficient)
U0j and U 1j are errors in the group-level
equations;
U0j is group level deviation of each
intercept from all intercept (?00?00)
The full multilevel regression model assumes that
their hierarchal data set, with one single dependent variable at all existing
levels. For example, Joop Hox’s multilevel analysis [4] assume data from J
classes, with a different number of pupils nj in each class. On the pupil
level, the outcome variable ‘popularity’ (Y), measured by a self-rating scale
that ranges from 0 (very unpopular) to 10 (very popular). They have two
explanatory variables on the pupil level: pupil gender (X1: 0 = boy, 1 = girl)
and pupil extraversion (X2, measured on a self-rating scale ranging from 1 to
10), and one class level explanatory variable teacher experience (Z: in years,
ranging from 2 to 25). Using variable labels the equation is:
Popularityij = ?0j + ?1j
genderij + ?2j extraversionij + eij (5)
In this regression equation, ?0j is the intercept; ?1j
is the regression coefficient (regression slope) for the dichotomous explanatory
variable gender, ?2j is the regression coefficient (slope) for the continuous
explanatory variable (extraversion), and the usual residual error term is eij.
The subscript j is for the classes (j = 1 . . . J) and the subscript i is for
individual pupils (i = 1 . . . nj).
The next step in the hierarchical regression model is
to explain the variation of the regression coefficients ?j introducing
explanatory variables at the class level:
?0j = ?00 + ?01Zj
+ u0j (6)
?1j = ?10 + ?11Zj
+ u1j
(7)
?2j = ?20 + ?21Zj
+ u2j (8)
The model with two pupil-level and one class-level
explanatory variable can be written as a single complex regression equation by
substituting equations (6), (7) and (8) into equation (5) gives:
Popularityij = ?00 +
?10 genderij + ?20 extraversionij +
?01 experiencej + ?11 genderij ×
experiencej + ?21 extraversionij × experiencej
+ u1j genderij + u2j extraversionij
+ u0j + eij (9)
Sample
size determination
It is generally accepted that increasing sample sizes
at all levels estimates and standard errors improve. To be statistically safe,
as “rule of thumb”, researchers should use ‘30/30’ rule, a sample of at least
30 groups with at least 30 individuals per group. On the other hand, the
numbers should be modified as if there is strong interest in cross-level
interactions, the number of groups should be larger, (a 50/20 rule-50 groups
with 20 individuals/ group); if there is stronger interest in the random part,
or in the variance and/ or covariance components, the number of involving groups
should be larger, leading to a 100/10 rule (100 groups with 10
individuals/group). One should take into account the costs attached to data
collection, so if the number of groups is increased, than the number of
individuals per group might decreases [5].
Assumptions
of multilevel modeling
Multilevel models have the same assumptions as other
major general linear models (e.g., Anova, regression), but some of the
assumptions are modified for the hierarchical nature of the design (i.e.,
nested data). Accordingly, checking and improving the specification of a
multilevel model in many cases can be carried out while staying within the
assumption of the multilevel model.
Linearity
The assumption of linearity states that there is a
rectilinear (straight-line, as opposed to non-linear or U-shaped) relationship
between variables [11]. However, the model can be extended to nonlinear
relationships [12]. A regression analysis expected to fit the best rectilinear
line that explains the most data given your set of parameters. Therefore, the
base models rely on the assumption that the data follow a straight line (though
the models can be expanded to handle curvilinear data). Graphically, by
plotting the model residuals (the difference between the observed value and the
model-estimated value) versus the predictor, linearity can be tested. If a
pattern emerges, a higher-order term may need to be included or you may need to
mathematically transform a predictor/response [13].
Normality
The assumption of normality states that the error
terms at every level of the model are normally distributed [11]. QQ plots which
are obtained in standard regression modeling in R can provide an estimation of
where the standardized residuals lie with respect to normal quintiles. Strong
deviation from the provided line indicates that the residuals themselves are
not normally distributed [13].
Homoscedasticity
The assumption of homoscedasticity or homogeneity of
variance assumes equality of population variances. However, different
variance-correlation matrix can be specified and the heterogeneity of variance
can itself be modelled [11]. In R, we extract the residuals from the model,
place them in our original table, take their absolute value, and then square
them (for a more robust analysis with respect to issues of normality. Finally, take
a look at the ANOVA of the between-subjects residuals [13].
Independence
of observations
Independence of observation is an assumption which
states that cases are random samples from the population and that scores on the
dependent variable are independent of each other [11].
Model
fitness
One way of assessing model fit is the chi-square
likelihood-ratio test, which assesses the difference between models. The
likelihood-ratio test can only be used when models are nested. It can be used
for examining what happens when effects in a model are allowed to vary, and
when testing a dummy-coded categorical variable as a single effect. However,
when testing non-nested models, comparisons between models can be made using
the Akaike information criterion or the Bayesian information criterion, among
others [10,14].
Statistical
tests and power
The types of statistical tests in multilevel models
depend on whether one is examining fixed effects or variance components. When
examining fixed effects, the tests should be compared with the standard error
of the fixed effect, which results in a Z-test. A t-test can also be computed.
When computing a t-test, it is important to consider degrees of freedom. For a
level one predictor, the degrees of freedom are based on the number of level
one predictor, the number of groups and the number of individual observations.
For a level two predictor, the degrees of freedom are based on the number of
level two predictors and the number of groups [10].
Statistical power for multilevel models differs
depending on whether it is level one or level two effects that are being
examined. Power for level one effect is dependent upon the number of individual
observations, whereas the power for level two effects is dependent upon the
number of groups [14].
Benefits
of multilevel analysis
The multilevel approach offers several advantages.
First, the result can generalize to a wider population. Second, fewer
parameters are needed. The dummy variables approach would require 25 additional
parameters. In the handling of more complex models and a limited amount of
data, a reduction in the number of parameters is important. Third, information
can be shared between groups. This is due to assuming that the random effects
resulted from a common distribution. As a result, the precision of predictions
for groups that have relatively little data improves. Finally, it can deal with
data in which the times of the measurements vary from subject to subject [2].
Limitation
of multilevel analysis
Analysing variables from different levels at one
single common level leads to two distinct types of problems. The first problem
is statistical. If data are aggregated, the result is that different data
values from many sub-units are combined into fewer values for fewer higher-level
units. As a result, statistical analysis loses power due to information loss.
On the other hand, for a larger number of sub-units, few data values from a
small number of super-units will ‘blown up’ into many more values, if data are
disaggregated. The second problem is conceptual. That is analysing the data at
one level and formulating conclusions at another level [4].
Acknowledgements
We are thankful to
Wollo University for providing the necessary facilities to conduct this review.
Declaration of Interest
The author has no
relevant affiliations or financial involvement with a financial interest in or financial with the subject matter or materials discussed in
the manuscript.
Data sharing is not
applicable to this article as no datasets were generated or analysed during the
current study.
All authors were
involved in searching for literatures and write up of the manuscript. All
authors read and approved the final draft of the manuscript.
There
is no conflict of interest.