CoChat
Sign Up
Cited by 114Statistical Methods and Bayesian InferenceStatistics Education and MethodologiesEducational Research and Methods

The DOI matches its Crossref record; 25 of the 25 references it cites were found in a registry or an index.

Students Are Rarely Independent: When, Why, and How to Use Random Effects in Discipline-Based Education Research

Open access
Theobald, Elli·2018·CBE—Life Sciences Education
DOI 10.1187/cbe.17-12-0280PMC PMC6234830
Open this paper in CoChat — free accountOpen PDF

Abstract

Discipline-based education researchers have a natural laboratory—classrooms, programs, colleges, and universities. Studies that administer treatments to multiple sections, in multiple years, or at multiple institutions are particularly compelling for two reasons: first, the sample sizes increase, and second, the implementation of the treatments can be intentionally designed and carefully monitored, potentially negating the need for additional control variables. However, when studies are implemented in this way, the observations on students are not completely independent; rather, students are clustered in sections, terms, years, or other factors. Here, I demonstrate why this clustering can be problematic in regression analysis. Fortunately, nonindependence of sampling can often be accounted for with random effects in multilevel regression models. Using several examples, including an extended example with R code, this paper illustrates why and how to implement random effects in multilevel modeling. It also provides resources to promote implementation of analyses that control for the nonindependence inherent in many quasi-random sampling designs.

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1187/cbe.17-12-0280

Cites

25 of 25 references verified0 citation errors · 0 not checkable

Statistics

Figures

FIGURE 1. Workflow for multilevel modeling and a guide to using this paper. Each “decision point” in this figure represents a critical step in multilevel modeling and corresponds to a section in the paper and a figure and is illustrated with the extended example. R code for analyzing the data from the extended example can be found in Appendix 3 in the Supplemental Material (data are in Appendix 4). The code details calculating the ICC (helpful in determining whether random effects are needed), using model selection to select the appropriate random effect structure, and using model selection to select the appropriate fixed effects.
FIGURE 1. Workflow for multilevel modeling and a guide to using this paper. Each “decision point” in this figure represents a critical step in multilevel modeling and corresponds to a section in the paper and a figure and is illustrated with the extended example. R code for analyzing the data from the extended example can be found in Appendix 3 in the Supplemental Material (data are in Appendix 4). The code details calculating the ICC (helpful in determining whether random effects are needed), using model selection to select the appropriate random effect structure, and using model selection to select the appropriate fixed effects.CC BY-NC-SA · PubMed Central
FIGURE 2. Experimental design in the extended example. Squares represent class days (note that there were more class days in the course and the spacing between interventions is not to scale). Experimental treatment and control are color coded as yellow and blue, respectively. Unique topic is indicated with the letters “Q,” “R,” and “S.” The pre- and posttests (black triangles) on each day were identical. The fact that students took a posttest three times made the students “repeated measure” and warranted students being treated as random effects. Additionally, students were clustered in sections (and section was not synonymous with intervention), so section was tested as a possible random effect. The experimental design was unbalanced, such that section A received the control twice and the treatment once, whereas section B received the treatment twice and the control once.
FIGURE 2. Experimental design in the extended example. Squares represent class days (note that there were more class days in the course and the spacing between interventions is not to scale). Experimental treatment and control are color coded as yellow and blue, respectively. Unique topic is indicated with the letters “Q,” “R,” and “S.” The pre- and posttests (black triangles) on each day were identical. The fact that students took a posttest three times made the students “repeated measure” and warranted students being treated as random effects. Additionally, students were clustered in sections (and section was not synonymous with intervention), so section was tested as a possible random effect. The experimental design was unbalanced, such that section A received the control twice and the treatment once, whereas section B received the treatment twice and the control once.CC BY-NC-SA · PubMed Central
FIGURE 3. Random effects are important to include when modeling data from (A) nested design studies and (B) repeated-measures studies. In both nested designs and repeated measures, the outcome is quantified on the student level. (Note that this is not always the case in DEBR studies; e.g., a researcher may be interested in which courses have the highest proportion of women, in which case the outcome is likely measured on the section level.) In the nested design illustration (A), students are nested in sections, which are nested in courses, which are nested in years, which are nested in universities. Each of these nested levels, or clusters, is important to account for with random effects. Similarly, in the repeated-measures study (B), the outcome (students’ survey responses) is quantified on the student level, and students take the survey four times. In this case, it is important to account for the fact that students are repeated, so student 1 on survey 1 is not independent from student 1 on survey 2. Including a student random effect accounts for this nonindependence within a student.
FIGURE 3. Random effects are important to include when modeling data from (A) nested design studies and (B) repeated-measures studies. In both nested designs and repeated measures, the outcome is quantified on the student level. (Note that this is not always the case in DEBR studies; e.g., a researcher may be interested in which courses have the highest proportion of women, in which case the outcome is likely measured on the section level.) In the nested design illustration (A), students are nested in sections, which are nested in courses, which are nested in years, which are nested in universities. Each of these nested levels, or clusters, is important to account for with random effects. Similarly, in the repeated-measures study (B), the outcome (students’ survey responses) is quantified on the student level, and students take the survey four times. In this case, it is important to account for the fact that students are repeated, so student 1 on survey 1 is not independent from student 1 on survey 2. Including a student random effect accounts for this nonindependence within a student.CC BY-NC-SA · PubMed Central
FIGURE 4. Researchers can build intuition to determine whether variables should be included as fixed effects or random effects by asking three questions and following this decision tree.
FIGURE 4. Researchers can build intuition to determine whether variables should be included as fixed effects or random effects by asking three questions and following this decision tree.CC BY-NC-SA · PubMed Central
FIGURE 5. Depending on the structure of the data, a fixed effects–only model (A) may not be the best model. There are three main ways to specify random effects in multilevel models: (B) as random intercepts, in which the slope is the same for all groups, but the intercepts vary for each group; (C) as random slopes, in which the intercept is the same for all groups, but the slopes vary for each group; and (D) as random intercepts and random slopes, in which all groups are allowed to have different slopes and different intercepts. Illustrated here, students (points) are nested in six sections (illustrated by the six combinations of filled/unfilled symbols in different shapes). There are two treatments, illustrated by filled and unfilled symbols. The lines show how each approach models the relationship between PreScore and PostScore. In a fixed effects–only model (A), a single regression line is fit for the treatment (filled) and control (unfilled symbol), ignoring sections; in a random effects model (B–D), separate regression lines are fit for each section. In the random intercepts model (B), the intercepts are allowed to vary by section but not the slopes (thus, the lines are parallel); in the random slopes model (C), the slopes are allowed to vary for each section, but not the intercepts (thus, all the lines start at the same place); and in the random intercepts and random slopes model (D), both the intercepts and the slopes are allowed to vary for each section. The overall model fit in B–D is essentially the weighted average of the regression lines for each group and is not shown.
FIGURE 5. Depending on the structure of the data, a fixed effects–only model (A) may not be the best model. There are three main ways to specify random effects in multilevel models: (B) as random intercepts, in which the slope is the same for all groups, but the intercepts vary for each group; (C) as random slopes, in which the intercept is the same for all groups, but the slopes vary for each group; and (D) as random intercepts and random slopes, in which all groups are allowed to have different slopes and different intercepts. Illustrated here, students (points) are nested in six sections (illustrated by the six combinations of filled/unfilled symbols in different shapes). There are two treatments, illustrated by filled and unfilled symbols. The lines show how each approach models the relationship between PreScore and PostScore. In a fixed effects–only model (A), a single regression line is fit for the treatment (filled) and control (unfilled symbol), ignoring sections; in a random effects model (B–D), separate regression lines are fit for each section. In the random intercepts model (B), the intercepts are allowed to vary by section but not the slopes (thus, the lines are parallel); in the random slopes model (C), the slopes are allowed to vary for each section, but not the intercepts (thus, all the lines start at the same place); and in the random intercepts and random slopes model (D), both the intercepts and the slopes are allowed to vary for each section. The overall model fit in B–D is essentially the weighted average of the regression lines for each group and is not shown.CC BY-NC-SA · PubMed Central
FIGURE 6. Steps in random effect model selection, as recommended by Zuur et al. (2009) and Burnham and Anderson (2002). This example was implemented in R.
FIGURE 6. Steps in random effect model selection, as recommended by Zuur et al. (2009) and Burnham and Anderson (2002). This example was implemented in R.CC BY-NC-SA · PubMed Central
FIGURE 7. R output from code in Figure 3: (A) step 3 of model selection (comparing AIC values for each model that was fit); (B) model 1, which did not include any random effects; and (C) the best-fitting multilevel model with student as the only random effect. Note that treatment was retained in the final model, noting a treatment effect and supporting the hypothesis that the treatment is positively correlated with postscore.
FIGURE 7. R output from code in Figure 3: (A) step 3 of model selection (comparing AIC values for each model that was fit); (B) model 1, which did not include any random effects; and (C) the best-fitting multilevel model with student as the only random effect. Note that treatment was retained in the final model, noting a treatment effect and supporting the hypothesis that the treatment is positively correlated with postscore.CC BY-NC-SA · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 25 references

Research works printed but not linked: 2

  1. R Core Team, 2017, A language and environment for statistical computing
  2. Rosenthal R., 1975, The volunteer subject

Cited by

Sorted by citations

Related

See all

Metadata

Authors
Theobald, E.
Affiliations
University of Washington (US)
Corresponding author
Theobald, Elli
ORCID
Theobald, Elli
Venue
CBE—Life Sciences Education
Publisher
American Society for Cell Biology (ASCB)
Published
2018-09
Volume
17
Issue
3
Pages
rm2
ISSN
1931-7913 (ISSN-L)
DOI
10.1187/cbe.17-12-0280
PMC
PMC6234830
OpenAlex
W2888039885
Cited by
114
References
25
Citations by year
2026: 11 · 2025: 13 · 2024: 15 · 2023: 17 · 2022: 16 · 2021: 23 · 2020: 9 · 2019: 9 · 2018: 1
FWCI
8.71
Type
journal article
Language
en
Open access
Licence
cc-by-nc-sa
Locations
  • CBE—Life Sciences Education (publishedVersion, cc-by-nc-sa)
  • Europe PMC (PubMed Central) (publishedVersion, cc-by-nc-sa)
  • PubMed Central (publishedVersion, cc-by-nc-sa)
  • PubMed (publishedVersion)
Funders
University of Washington
Topics
  • Physical Sciences › Mathematics › Statistics and Probability › Statistical Methods and Bayesian Inference
  • Physical Sciences › Mathematics › Statistics and Probability › Statistics Education and Methodologies
  • Social Sciences › Education › Educational Research and Methods
Keywords
random effects, multilevel regression, student clustering, treatment implementation, discipline-based education research
MeSH terms
Humans, Random Allocation, Research Design, Science, Students, Cluster Analysis, Multilevel Analysis

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review