Newsom

Psy 526/626 Multilevel Regression, Spring 2019 1

Distinguishing Between Random and Fixed:

Variables, Effects, and Coefficients 1

The terms “random” and “fixed” are used frequently in the multilevel modeling literature. The distinction is
a difficult one to begin with and becomes more confusing because the terms are used to refer to different
circumstances. Here are some summary comments that may help.

Random and Fixed Variables

A “fixed variable” is one that is assumed to be measured without error. It is also assumed that the values
of a fixed variable in one study are the same as the values of the fixed variable in another study.
“Random variables” are assumed to be values that are drawn from a larger population of values and thus
will represent them. You can think of the values of random variables as representing a random sample
of all possible values or instances of that variable. Thus, we expect to generalize the results obtained
with a random variable to all other possible instances of that value (e.g., a job candidate with a strong
résumé). Most of the time in ANOVA and regression analysis we assume the independent variables are
fixed.

Random and Fixed Effects

The terms “random” and “fixed” are used in the context of ANOVA and regression models and refer to a
certain type of statistical model. Almost always, researchers use fixed effects regression or ANOVA and
they are rarely faced with a situation involving random effects analyses. A fixed-effects ANOVA refers to
assumptions about the independent variable and that error distribution for the variable. An experimental
design is the easiest example for illustrating the principal. Usually, the researcher is interested in only
generalizing the results to experimental values used in the study. For instance, a drug study might use 0
mg, 5 mg, and 10 mg of an experimental drug. This is a circumstance when a fixed effects ANOVA
would be appropriate. In this example, the extrapolation is to other studies or treatments that might use
the same values of the drug (i.e., 0 mg, 5 mg, and 10 mg). However, if the researcher wants to make
inferences beyond the particular values of the independent variable used in the study, a random effects
model is used. A common example would be the use of public art works representing low, moderate,
and high abstractness (e.g., statue of a war hero vs. a pivoting geometric design). The researcher would
like to make inferences beyond just one art piece representing each category of abstractness, so the art
pieces are conceptualized as pieces randomly drawn from a larger universe of possible pieces that are
sampled from the domain for that level of abstractness. For example, one could imagine using several
instances of high abstract pieces that are randomly drawn from a larger population of high abstract
pieces and are thought to be only a few of the possible particular instances of high abstract art. Thus, the
inferences are made to a larger universe of art works with variations of abstractness within each
category. Such a generalization is more of an inferential leap, and, consequently, the random effects
model is less powerful because we are taking into account some additional expected random variation
on the independent variable. Random effects models are sometimes referred to as “Model II” or
“variance component models.” Analyses using both fixed and random effects are called “mixed models”
or "mixed effects models" which is one of the terms given to multilevel models.

Fixed and Random Coefficients in Multilevel Regression (MLR)

The random vs. fixed distinction for variables and effects is important in multilevel regression. In
multilevel regression models, both level-1 and level-2 predictors are assumed to be fixed. However, the
intercepts and slopes in the level-1 equations may be allowed to vary randomly across groups and are
therefore referred to as “random coefficients.” The variance of the intercepts is represented by τ20 =
var(U0j) and is nearly always estimated by default. The variance of the (first) predictor is represented by

1
My distinctions between random variables, effects, and coefficients was inspired by and relies heavily on Kreft & de Leeuw (1998).
Newsom
Psy 526/626 Multilevel Regression, Spring 2019 2

τ21 = var(U1j), and, depending on the circumstances, may not be allowed to vary for every predictor. 2 The
within-group variance, var(Rij) = σ2, is also referred to as “random.”
Because we may be interested in the variation of the coefficients across groups as a research
question, one can think about β0j and β1j as akin to the random variables I described above in the first
section. Instead of attempting to generalize beyond the particular values of the independent variable, we
are attempting to generalize beyond the particular groups in the study. For instance, we may have 100
companies, but we wish to generalize to a larger universe of companies when we examine the means
(intercepts) or the X-Y relationship (slopes). Output from software packages will usually have sections
labeled as fixed effects and random effects. The fixed effects are the coefficients (intercept, slope) as we
usually think about the. The random effects are the variances of the intercepts or slopes across groups.
In the HLM program, variances for the intercepts and slopes are estimated by default (U0j and U1j,
respectively). In SPSS Mixed and R (nlme or lme4), the user must specify which intercepts or slopes
should be estimated. If any variance, intercept or slope, is not specified their values are set to zero. By
setting variances to zero, we are testing a model in which we assume β0j and β1j do not vary randomly
across groups. Thus, the intercept or slope value is assumed to be constant or “nonvarying” across
groups. For example, fixed, nonvarying intercepts would imply the group average for the dependent
variable is assumed to be equal in each group—this implies no variance, so would be rarely assumed for
intercepts. Note that although researchers sometimes refer to this constraint as “fixing the intercepts” or
“fixing the slopes,” the term is somewhat loosely applied, because we are really assuming they are fixed
and nonvarying.

References
Kreft, I., & de Leeuw, J. (1998). Introducing multilevel modeling. London: Sage.
Raudenbush, S.W., & Bryk, A.S., (2002) Hierarchical linear models: Applications and data analysis methods. Thousand Oaks, CA: Sage.
Snijders, T.A.B., & Bosker, R.J. (2012). Multilevel analysis: An introduction to basic and advanced multilevel modeling (2nd Edition). London:
Sage.

2
My notation follows that of Snijders & Bosker (2012), but notation differs among texts. In Raudenbush & Bryk
(2002), for instance, the symbol for intercept variance is τ00, and the subscript refers to the diagonal element
corresponding to the row and column of the variance-covariance matrix for U0j, and the subscript for the slope
variance, τ11, refers to the diagonal element for the U1j column and row.
Newsom
Psy 526/626 Multilevel Regression, Spring 2019 3

Summary Table

Random vs. Definition Example Use in Multilevel

Fixed Regression
Variables Random variable: (1) is assumed to be measured Random variable: Predictor variables in
with measurement error. The scores are a function of a photographs MLR generally assumed
true score and random error; (2) the values come from representing to be fixed
and are intended to generalize to a much larger individuals with
population of possible values with a certain probability differing levels of
distribution (e.g., normal distribution); (3) the number of attractiveness
values in the study is small relative to the values of the manipulated in an
variable as it appears in the population it is drawn from. experiment, a subset
Fixed variable: (1) assumed to be measured without of census tracks
measurement error; (2) desired generalization to
population or other studies is to the same values; (3)
the variable used in the study contains all or most of Fixed variable:
the variable’s values in the population. gender, race, or
intervention vs.
It is important to distinguish between a variable that is control group.
varying and a variable that is random. A fixed variable
can have different values, it is not necessarily invariant
(equal) across groups.

Effects Random effect: (1) different statistical model of Random effect: Intercept only models in
regression or ANOVA model which assumes that an random effects MLR are equivalent to
independent variable is random; (2) generally used if ANOVA, random random effects ANOVA
the levels of the independent variable are thought to be effects regression and inclusion of one or
a small subset of the possible values which one wishes Fixed effect: fixed more level-1 predictors
to generalize to; (3) will probably produce larger effects ANOVA, makes the model
standard errors (less powerful). Fixed effect: (1) fixed effects equivalent to a random
statistical model typically used in regression and regression effects ANCOVA when
ANOVA assuming independent variable is fixed; (2) slopes do not vary
generalization of the results apply to similar values of across groups.
independent variable in the population or in other
studies; (3) will probably produce smaller standard
errors (more powerful).

Coefficients Random coefficient: term applies only to MLR Random Both used in MLR.
analyses in which intercepts, slopes, and variances can coefficient: the Slopes and intercept
be assumed to be random. MLR analyses most level-2 predictor, values can be
typically assume random coefficients. One can average income, is considered to be fixed or
conceptualize the coefficients obtained from the level-1 used to predict random, depending on
regressions as a type of random variable which comes school performance researchers'
from and generalizes to a distribution of possible in each school. assumptions and how
values. Groups are conceived of as a subset of the Intercept values for the model is specified.
possible groups. school performance The average intercept or
are assumed to be a slope is referred to as a
Fixed coefficient: a coefficient can be fixed to be non- sample of the "fixed effect." Variances
varying (invariant) across groups by setting the intercepts from a of the slopes and
between-group variance to zero. larger population of intercepts (if allowed to
schools. vary across groups) are
Random coefficients must be variable across groups. called “random
Conceptually, fixed coefficients may be invariant or Fixed coefficient: coefficients."
varying across groups. slopes or intercepts
constrained to be
equal over different
schools.