---
title: "Centering Decisions: Not Just a Technical Choice"
subtitle: "PSY 8XXX: Multilevel Modeling for Organizational Research — Week 5"
author: "Instructor Name"
date: last-modified
format:
html:
code-fold: true
code-tools: true
toc: true
toc-depth: 3
number-sections: true
theme: cosmo
self-contained: true
execute:
warning: false
message: false
---
# Introduction: Centering as Substantive, Not Just Technical
Most graduate students view centering as a technical preprocessing step: mean-center your variables to improve interpretability and reduce multicollinearity. It's taught mechanically, almost as an afterthought.
But in multilevel modeling, centering decisions are **profoundly substantive**. They determine what research questions you can answer, what effects you estimate, and ultimately, what you conclude about your data.
Consider this: You're studying whether employee autonomy predicts job satisfaction. The question sounds simple, but it harbors ambiguity:
1. **Within-team question**: "Are employees with *higher autonomy within their team* also more satisfied?"
2. **Between-team question**: "Do *teams with higher average autonomy* report higher average satisfaction?"
These are different causal questions. The first is about individual differences; the second is about team-level phenomena. They may have different answers. Remarkably, your centering choice determines which question you answer.
This tutorial explores how centering works in multilevel models, what each approach reveals (and hides), and how to align your centering strategy with your research questions.
# Setup and Data
```{r setup}
set.seed(1234)
library(tidyverse)
library(lme4)
library(lmerTest)
library(performance)
# Recreate employee dataset
n_teams <- 50
teams <- tibble(
team_id = 1:n_teams,
team_size = sample(8:15, n_teams, replace = TRUE),
team_climate = pmin(pmax(rnorm(n_teams, mean = 5, sd = 0.8), 1), 7)
)
employees <- teams %>%
slice(rep(1:n_teams, teams$team_size)) %>%
arrange(team_id) %>%
group_by(team_id) %>%
mutate(
emp_id = row_number(),
employee_id = paste0("T", team_id, "E", emp_id),
autonomy = rnorm(n(), mean = 4.5, sd = 1.2),
autonomy = pmin(pmax(autonomy, 1), 7),
performance = rnorm(n(), mean = 6, sd = 1.5),
performance = pmin(pmax(performance, 1), 10),
tenure = rgamma(n(), shape = 3, rate = 0.8)
) %>%
ungroup()
employees <- employees %>%
group_by(team_id) %>%
mutate(
u0j = rnorm(1, mean = 0, sd = sqrt(0.6)),
job_satisfaction = 3.0 +
0.5 * scale(autonomy)[,1] +
0.4 * team_climate +
u0j +
rnorm(n(), mean = 0, sd = sqrt(0.8)),
job_satisfaction = pmin(pmax(job_satisfaction, 1), 7)
) %>%
ungroup() %>%
select(-u0j, -emp_id)
cat("Dataset ready:", nrow(employees), "employees in", n_distinct(employees$team_id), "teams\n")
```
Now, let's create centered versions of autonomy for our analyses:
```{r create-centered-versions}
# Uncentered (raw)
employees_all <- employees %>%
mutate(autonomy_raw = autonomy)
# Grand-Mean Centering (CGM)
grand_mean_autonomy <- mean(employees$autonomy)
employees_all <- employees_all %>%
mutate(autonomy_cgm = autonomy - grand_mean_autonomy)
# Group-Mean Centering (CWC) - also called Within-Cluster Centering
employees_all <- employees_all %>%
group_by(team_id) %>%
mutate(
team_mean_autonomy = mean(autonomy),
autonomy_cwc = autonomy - team_mean_autonomy
) %>%
ungroup()
# Raudenbush formulation: both CWC and group mean
employees_all <- employees_all %>%
mutate(autonomy_raudenbush_cwc = autonomy_cwc,
autonomy_raudenbush_between = team_mean_autonomy)
# Display first few rows to see what we created
head(employees_all %>%
select(team_id, autonomy, autonomy_raw, autonomy_cgm, autonomy_cwc, team_mean_autonomy), 10)
cat("\n\nGrand mean of autonomy:", round(grand_mean_autonomy, 3), "\n")
cat("SD of autonomy:", round(sd(employees$autonomy), 3), "\n")
```
# Grand-Mean Centering (CGM): Interpretation and Use Cases
Grand-Mean Centering (CGM) subtracts the overall sample mean from each observation:
$$X_{CGM,ij} = X_{ij} - \bar{X}$$
Let's fit a model with CGM:
```{r cgm-model}
model_cgm <- lmer(job_satisfaction ~ autonomy_cgm + team_climate + (1 | team_id),
data = employees_all)
summary(model_cgm)
```
**Interpretation of Coefficients:**
- **autonomy_cgm = 0.369**: A one-unit increase in autonomy (from a grand-mean centered baseline) is associated with 0.369 higher satisfaction. This coefficient mixes within-team and between-team effects.
- **team_climate = 0.383**: A one-unit increase in team climate is associated with 0.383 higher satisfaction.
- **Intercept = 4.089**: The predicted satisfaction when autonomy is at its grand mean (4.5) and team_climate is at 0. This is interpretable as the overall average satisfaction.
**When to use CGM:**
1. **Predictors vary mainly within groups**: If autonomy is individual-specific and doesn't systematically vary by team.
2. **Centering at the study population level**: You want results to generalize to the population your sample represents.
3. **Simplicity**: CGM is straightforward and makes intercepts interpretable as population averages.
**What CGM hides:**
The autonomy coefficient represents a mixture of within-team and between-team effects. If autonomy varies systematically by team (e.g., some teams cultivate autonomy while others restrict it), this mixture is misleading. You can't distinguish: "Do autonomous individuals perform better?" from "Do autonomy-promoting teams perform better?"
# Group-Mean Centering (CWC): Isolating Within-Team Effects
Group-Mean Centering (also called Within-Cluster Centering, CWC) subtracts each team's mean from each observation:
$$X_{CWC,ij} = X_{ij} - \bar{X}_j$$
Let's fit the same model with CWC:
```{r cwc-model}
model_cwc <- lmer(job_satisfaction ~ autonomy_cwc + team_climate + (1 | team_id),
data = employees_all)
summary(model_cwc)
```
**Interpretation of Coefficients:**
- **autonomy_cwc = 0.369**: A one-unit increase in autonomy *within a team* (deviating from that team's mean) is associated with 0.369 higher satisfaction. This is purely a within-team effect.
- **team_climate = 0.383**: Exactly as before.
- **Intercept = 4.089**: The predicted satisfaction for an employee at their team's mean autonomy (not the grand mean). This is the team-level average.
**Key insight**: With CWC, the autonomy coefficient is now *unambiguous* about level of analysis. It's the within-team effect only.
# The Critical Demonstration: Separating Within and Between Effects
Now here's the crucial point: **within and between effects can differ**. Let's create a scenario where they do:
```{r differ-effects}
# Create data where within and between effects differ dramatically
set.seed(9999)
n_teams_demo <- 30
n_per_team <- 15
data_differ <- tibble(
team_id = rep(1:n_teams_demo, each = n_per_team),
team_autonomy_mean = rep(rnorm(n_teams_demo, 4.5, 1.2), each = n_per_team),
autonomy_within = rnorm(n_teams_demo * n_per_team, 0, 1)
) %>%
mutate(
autonomy = autonomy_within + team_autonomy_mean,
autonomy = pmin(pmax(autonomy, 1), 7),
# Within-team: NEGATIVE effect (more autonomous individuals less satisfied)
# Between-team: POSITIVE effect (more autonomous teams are more satisfied)
satisfaction = 4 +
(-0.4) * scale(autonomy_within)[,1] + # Within: negative
0.6 * scale(team_autonomy_mean)[,1] + # Between: positive
rnorm(n_teams_demo * n_per_team, 0, 0.5)
)
# Create centered versions
data_differ <- data_differ %>%
mutate(
autonomy_cgm = autonomy - mean(autonomy),
autonomy_cwc = autonomy - team_autonomy_mean
) %>%
group_by(team_id) %>%
mutate(team_mean_auto = mean(autonomy)) %>%
ungroup()
# Fit models with different centering
m_raw <- lm(satisfaction ~ autonomy, data = data_differ)
m_cgm <- lmer(satisfaction ~ autonomy_cgm + (1 | team_id), data = data_differ)
m_cwc <- lmer(satisfaction ~ autonomy_cwc + (1 | team_id), data = data_differ)
# Compare coefficients
coef_comparison <- tibble(
Model = c("OLS (uncentered)", "MLM (CGM)", "MLM (CWC)"),
Autonomy_Coefficient = c(
coef(m_raw)[2],
fixef(m_cgm)[2],
fixef(m_cwc)[2]
)
)
print(coef_comparison)
cat("\n\nInterpretation:\n")
cat("- OLS sees a POSITIVE effect (mixes within and between)\n")
cat("- CGM sees a POSITIVE effect (mixture closer to between)\n")
cat("- CWC shows the TRUE within-team effect: NEGATIVE!\n")
cat("\nWithin a team, more autonomous individuals are LESS satisfied.\n")
cat("But teams with higher autonomy ARE more satisfied.\n")
cat("These are real, opposite effects that CGM masks!\n")
```
This is the **Simpson's Paradox** of multilevel modeling. The direction of a relationship can flip between levels. CGM obscures this; CWC reveals it.
# The Contextual Effect: Separating Within and Between in One Model
The most elegant approach is the **Raudenbush (2009) formulation**, which includes both within-cluster centered and group-mean centered versions of the same predictor:
$$Y_{ij} = \gamma_{00} + \gamma_W X_{CWC,ij} + \gamma_B \bar{X}_j + u_{0j} + r_{ij}$$
where $X_{CWC,ij}$ is the within-cluster centered predictor and $\bar{X}_j$ is the group mean.
```{r contextual-effect}
# Fit Raudenbush model on the simulated data with differing effects
model_raudenbush <- lmer(satisfaction ~ autonomy_cwc + team_mean_auto + (1 | team_id),
data = data_differ)
summary(model_raudenbush)
cat("\n\nInterpretation:\n")
cat("Within-team effect (autonomy_cwc):",
round(fixef(model_raudenbush)[2], 3), "\n")
cat("Between-team effect (team_mean_auto):",
round(fixef(model_raudenbush)[3], 3), "\n")
cat("\nThese are DIFFERENT, revealing heterogeneity in effects across levels!\n")
cat("The contextual effect = between - within =",
round(fixef(model_raudenbush)[3] - fixef(model_raudenbush)[2], 3), "\n")
```
**This is powerful**: By including both predictors, you simultaneously estimate:
1. **$\gamma_W$**: The within-team effect. How much does individual autonomy matter?
2. **$\gamma_B$**: The between-team effect. How much does team-average autonomy matter?
3. **The contextual effect**: $\gamma_B - \gamma_W$. Is there team-level variance beyond individual differences?
If $\gamma_W$ and $\gamma_B$ differ, it signals that something team-level (team leadership, team norms, team resources) moderates the autonomy-satisfaction link.
# Connecting to OB Theory: Within vs. Between Questions
Let's ground this in actual organizational research questions:
```{r theory-example}
# Question 1: Does YOUR autonomy predict YOUR satisfaction?
# Answer: CWC coefficient (within-team effect)
cat("Question 1: Individual Within-Team Effect\n")
cat("'Does my autonomy predict my satisfaction?'\n\n")
model_q1 <- lmer(job_satisfaction ~ autonomy_cwc + (1 | team_id),
data = employees_all)
cat("Answer: autonomy effect =", round(fixef(model_q1)[2], 3), "\n")
cat("Interpretation: Holding team constant, employees with higher autonomy\n")
cat("report higher satisfaction.\n\n")
# Question 2: Do high-autonomy teams have higher satisfaction?
# Answer: Between-team effect
cat("\n\nQuestion 2: Team-Level Effect\n")
cat("'Do teams that give autonomy have higher average satisfaction?'\n\n")
employees_team_level <- employees_all %>%
group_by(team_id) %>%
summarize(
mean_satisfaction = mean(job_satisfaction),
mean_autonomy = mean(autonomy),
team_climate = first(team_climate),
.groups = 'drop'
)
m_between <- lm(mean_satisfaction ~ mean_autonomy, data = employees_team_level)
cat("Answer: autonomy effect =", round(coef(m_between)[2], 3), "\n")
cat("Interpretation: Teams with higher average autonomy report higher\n")
cat("average satisfaction.\n\n")
# Question 3: Which effect dominates? (Raudenbush approach)
cat("\n\nQuestion 3: Which Level Matters More?\n")
cat("'Is satisfaction driven by individual autonomy or team autonomy culture?'\n\n")
model_q3 <- lmer(job_satisfaction ~ autonomy_cwc + team_mean_autonomy +
(1 | team_id),
data = employees_all)
cat("Within-team effect:", round(fixef(model_q3)[2], 3), "\n")
cat("Between-team effect:", round(fixef(model_q3)[3], 3), "\n")
cat("Contextual effect:", round(fixef(model_q3)[3] - fixef(model_q3)[2], 3), "\n")
```
Notice how the same variable (autonomy) can be used to answer distinct theoretical questions depending on centering choice. This is not a limitation—it's the flexibility of multilevel modeling, but it requires careful conceptual thinking.
# Visualization: Seeing the Levels Separately
Visualize the two separate effects:
```{r visualize-effects}
# Create predictions showing within and between effects separately
# Get observed data
plot_data <- employees_all %>%
select(team_id, autonomy, autonomy_cwc, team_mean_autonomy, job_satisfaction)
# Fit model
m_vis <- lmer(job_satisfaction ~ autonomy_cwc + team_mean_autonomy + (1 | team_id),
data = employees_all)
# Create prediction data
pred_within <- tibble(
autonomy_cwc = seq(-2, 2, by = 0.2),
team_mean_autonomy = mean(employees_all$team_mean_autonomy), # Hold between constant
team_id = NA
)
pred_between <- tibble(
autonomy_cwc = 0, # Hold within constant
team_mean_autonomy = seq(2, 7, by = 0.2),
team_id = NA
)
pred_within$pred <- predict(m_vis, newdata = pred_within, re.form = NA)
pred_between$pred <- predict(m_vis, newdata = pred_between, re.form = NA)
# Plot both
p1 <- ggplot(employees_all, aes(x = autonomy_cwc, y = job_satisfaction)) +
geom_point(alpha = 0.2, size = 0.8) +
geom_line(data = pred_within, aes(y = pred), color = "blue", size = 1.2) +
labs(
title = "WITHIN-TEAM Effect",
subtitle = "How does individual autonomy predict satisfaction (holding team constant)?",
x = "Autonomy (within-cluster centered)",
y = "Job Satisfaction"
) +
theme_minimal()
p2 <- ggplot(employees_all, aes(x = team_mean_autonomy, y = job_satisfaction)) +
geom_point(aes(color = factor(team_id)), alpha = 0.3, size = 0.8, show.legend = FALSE) +
geom_line(data = pred_between, aes(y = pred, color = NA), color = "red", size = 1.2) +
labs(
title = "BETWEEN-TEAM Effect",
subtitle = "How does team average autonomy predict satisfaction?",
x = "Team Mean Autonomy",
y = "Job Satisfaction"
) +
theme_minimal()
gridExtra::grid.arrange(p1, p2, ncol = 2)
```
The left panel shows within-team variation (points scatter around the blue line within each team). The right panel shows between-team variation (team averages align with the red line). By separating these, we see both sources of variation clearly.
# A Decision Framework: Which Centering Approach?
Here's a practical decision tree:
```{r decision-framework}
cat("=== CENTERING DECISION FRAMEWORK ===\n\n")
cat("1. Is your predictor INDIVIDUAL-LEVEL (varies within teams)?\n")
cat(" Examples: autonomy perception, job stress, age\n")
cat(" → Use CWC or Raudenbush formulation\n\n")
cat("2. Is your predictor TEAM-LEVEL (constant within teams)?\n")
cat(" Examples: team size, team budget, team location\n")
cat(" → Use RAW (uncentered) or CGM\n")
cat(" → Actually, team-level predictors don't need centering!\n\n")
cat("3. Do within and between effects of the SAME predictor\n")
cat(" likely differ theoretically?\n")
cat(" Examples: autonomy, cohesion, communication frequency\n")
cat(" → Use Raudenbush formulation to separate them\n\n")
cat("4. Is this a cross-level interaction (Level 1 predictor ×\n")
cat(" Level 2 predictor modifying the relationship)?\n")
cat(" → Use CWC for Level 1 predictor (improves interpretation)\n")
cat(" → Use CGM or raw for Level 2 predictor\n\n")
cat("5. Do you want the intercept to mean something specific?\n")
cat(" - Grand-Mean autonomy? → Use CGM\n")
cat(" - Team-specific baseline? → Use CWC\n")
cat(" - Absolute zero? → Use RAW (rare)\n\n")
cat("=== RECOMMENDED STANDARD APPROACH ===\n")
cat("For most organizational research:\n")
cat("1. Use CWC for individual-level predictors\n")
cat("2. Keep team-level predictors uncentered (or use CGM if mixing levels)\n")
cat("3. Include both CWC and team mean if effects might differ\n")
cat("4. Interpret with explicit attention to level of analysis\n")
```
# Common Mistakes and How to Avoid Them
```{r common-mistakes}
cat("MISTAKE 1: Forgetting to center, then interpreting coefficients\n")
cat("-----------\n")
m_no_center <- lmer(job_satisfaction ~ autonomy + (1 | team_id),
data = employees_all)
cat("Intercept =", round(fixef(m_no_center)[1], 3),
"(predicted satisfaction when autonomy = 0, not meaningful)\n")
cat("Autonomy effect:", round(fixef(m_no_center)[2], 3),
"(meaningful, but intercept is garbage)\n")
cat("FIX: Always center or be explicit about what intercept means.\n\n")
cat("MISTAKE 2: Using CGM when you should use CWC\n")
cat("-----------\n")
cat("You compare within-team coefficient from CGM model\n")
cat("to published studies using CWC.\n")
cat("Coefficients appear different, but it's centering, not real difference!\n")
cat("FIX: Match centering to your research question and prior work.\n\n")
cat("MISTAKE 3: Including group mean without CWC\n")
cat("-----------\n")
m_bad <- lmer(job_satisfaction ~ autonomy + team_mean_autonomy + (1 | team_id),
data = employees_all)
cat("This model has serious collinearity:\n")
cat("autonomy and team_mean_autonomy are nearly collinear.\n")
cat("Coefficient SEs are huge!\n")
print(summary(m_bad)$coefficients)
cat("FIX: Use CWC for autonomy to decorrelate the predictors.\n\n")
cat("MISTAKE 4: Over-interpreting small between effects\n")
cat("-----------\n")
cat("You have N=50 teams. Between-team estimates have low power.\n")
cat("Wide CIs on team-level coefficients are normal!\n")
cat("FIX: Be transparent about power, consider Bayesian priors.\n")
```
# Extended Example: A Complete Centering Analysis
Let's put it all together with a complete example:
```{r complete-centering-example}
cat("=== COMPLETE CENTERING ANALYSIS ===\n\n")
# Step 1: Raw (uncentered)
cat("MODEL 1: Uncentered\n")
m1 <- lmer(job_satisfaction ~ autonomy + team_climate + (1 | team_id),
data = employees_all)
cat("Autonomy coef:", round(fixef(m1)[2], 3), "\n")
cat("Intercept:", round(fixef(m1)[1], 3),
"(when autonomy=0: not meaningful)\n\n")
# Step 2: CGM
cat("MODEL 2: Grand-Mean Centered\n")
m2 <- lmer(job_satisfaction ~ autonomy_cgm + team_climate + (1 | team_id),
data = employees_all)
cat("Autonomy coef:", round(fixef(m2)[2], 3),
"(SAME as Model 1)\n")
cat("Intercept:", round(fixef(m2)[1], 3),
"(when autonomy=grand mean: interpretable)\n\n")
# Step 3: CWC
cat("MODEL 3: Within-Cluster Centered\n")
m3 <- lmer(job_satisfaction ~ autonomy_cwc + team_climate + (1 | team_id),
data = employees_all)
cat("Autonomy coef:", round(fixef(m3)[2], 3),
"(SAME as Models 1-2)\n")
cat("Intercept:", round(fixef(m3)[1], 3),
"(when autonomy=team mean: interpretable)\n\n")
# Step 4: Raudenbush (both versions)
cat("MODEL 4: Raudenbush Formulation (separates within & between)\n")
m4 <- lmer(job_satisfaction ~ autonomy_cwc + team_mean_autonomy + team_climate +
(1 | team_id),
data = employees_all)
cat("Within-team autonomy coef:", round(fixef(m4)[2], 3), "\n")
cat("Between-team autonomy coef:", round(fixef(m4)[3], 3), "\n")
cat("Difference (contextual effect):",
round(fixef(m4)[3] - fixef(m4)[2], 3), "\n\n")
# Summary table
cat("=== SUMMARY: Do Centering Choices Change Conclusions? ===\n")
summary_table <- tibble(
Model = c("Uncentered", "CGM", "CWC", "Raudenbush within", "Raudenbush between"),
Autonomy_Coefficient = c(
round(fixef(m1)[2], 3),
round(fixef(m2)[2], 3),
round(fixef(m3)[2], 3),
round(fixef(m4)[2], 3),
round(fixef(m4)[3], 3)
),
Interpretation = c(
"Mixture of within/between",
"Mixture of within/between",
"Pure within-team effect",
"Pure within-team effect",
"Pure between-team effect"
)
)
print(summary_table)
```
**Key insight**: The raw correlation coefficient is the same across Models 1-3, but the *interpretation* changes based on centering and what's held constant. Only Raudenbush separates them.
# Try It Yourself Exercises
## Exercise 1: Replicate Simpson's Paradox
Modify the `data_differ` scenario to create more extreme opposites: strong negative within effect, strong positive between effect. Fit models with CWC and raw, showing how centering choice determines which effect you see.
## Exercise 2: Cross-Level Interaction
Fit a model with autonomy_cwc predicting satisfaction, team_climate predicting autonomy effects, and include the interaction: `autonomy_cwc * team_climate`. How does team climate moderate the autonomy effect? Visualize predictions at high and low climate.
## Exercise 3: Build a Centering Decision Tree
For your own research project (or a proposed study), create a table:
- Column 1: Your key predictors
- Column 2: Are they individual, team, or organizational level?
- Column 3: Do you expect within/between effects to differ?
- Column 4: What centering approach fits your questions?
## Exercise 4: Interpret Coefficients Precisely
Fit a full model with CWC autonomy, team_mean_autonomy, team_climate, and a cross-level interaction. Write out the interpretation of every coefficient in a way that would be clear to a non-methodologist manager: "When team climate is high, a one-unit increase in individual autonomy predicts..."
## Exercise 5: Power Simulation
Simulate what happens to power for detecting between-team effects as the number of teams varies (10, 25, 50, 100). How many teams do you need to reliably detect a between-team effect? Create a plot of power vs. number of teams.
---
**Key Takeaways:**
- Centering is not just a technical step; it's a **substantive choice** that determines what research questions you answer.
- **Grand-Mean Centering (CGM)** makes coefficients represent population-level relationships but mixes within and between effects.
- **Group-Mean Centering (CWC)** isolates within-team effects and makes the intercept represent team-specific baselines.
- **Simpson's Paradox** can occur: within-team and between-team relationships can have opposite signs. CGM obscures this; CWC reveals it.
- **The Raudenbush Formulation** includes both CWC and group-mean predictors, separating within and between effects in a single model.
- Always align your centering strategy with your research questions: ask "Is this a within-team or between-team question?" and choose centering accordingly.
- For reproducibility and clear communication, always report your centering approach explicitly in your methods.
