---
title: "REWB Models"
output: rmarkdown::html_vignette
vignette: >
%\VignetteIndexEntry{REWB Models}
%\VignetteEngine{knitr::rmarkdown}
%\VignetteEncoding{UTF-8}
---
```{r, include = FALSE}
knitr::opts_chunk$set(
collapse = TRUE,
comment = "#>"
)
# options(scipen = 10)
```
```{r setup, message = FALSE}
library(mlstats)
library(dplyr)
library(lme4)
library(lmerTest)
```
When observations are nested within groups (repeated measurements per person,
students within classrooms, etc.), the association between two variables has two
faces: how they co-vary *within* groups over time, and whether groups that score
higher on one variable also tend to score higher on the other. A naive linear
regression cannot distinguish these two effects, and conflating them can produce
severely misleading conclusions.
The **Random Effects Within-Between (REWB)** model (Bell et al., 2019) solves
this by including both components as separate predictors. This vignette shows
how to use `decompose_within_between()` to prepare data for REWB models and
how to fit and interpret those models.
## Example Data
We use `media_diary`, a simulated daily diary dataset included with **mlstats**
(100 participants over 14 days; *N* = 100 persons, *T* = 1,400 daily observations). See `?media_diary` for details.
```{r data}
data("media_diary")
```
The dataset was generated to illustrate two processes that can operate simultaneously — and in opposite directions — at within- and between-person levels:
1. **Between persons**: people who watch more entertainment media on average
tend to have lower average wellbeing — perhaps because chronic heavy media
use reflects lower trait self-control, which itself predicts lower wellbeing.
2. **Within persons**: on days when someone watches more than usual, their
wellbeing is slightly higher — consistent with short-term escapism or mood
repair through media use.
Because these processes were built into the simulation, they are present by design — not empirical discoveries. The purpose of the example is to show how REWB models recover effects that point in *opposite directions*, and what happens when they are conflated. A naive regression conflates them and produces a near-zero coefficient, making it appear that screen time has no relationship with wellbeing, when in fact it has two real and opposing effects (in the simulation).
## Decomposing Time-Varying Predictors
`decompose_within_between()` splits each specified variable into up to three
components:
- **`_grand_mean_centered`**: grand-mean-centered value
- **`_between_{group}`**: group mean (stable between-group component)
- **`_within_{group}`**: deviation from the group mean (within-group fluctuation)
The `vars` argument names the variables to decompose. `group` names the
grouping variable.
```{r decompose-basic}
media_diary |>
decompose_within_between(group = "person", vars = "screen_time") |>
select(starts_with("screen_time_"))
```
`screen_time_within_person` is the group-mean-centred score: how many more (or
fewer) minutes this person watched today compared to their own average.
`screen_time_between_person` is the person's average screen time, repeated for every
row belonging to that person. `screen_time_grand_mean_centered` is the grand-mean-centred
value, which shows each observation's deviation from the overall mean.
### Selecting Components
By default all three components are returned. Use the `components` argument to
select a subset. For REWB models, the within and between components are the
predictors you need.
```{r decompose-components}
media_diary |>
decompose_within_between(
group = "person",
vars = "screen_time",
components = c("within", "between")
) |>
select(starts_with("screen_time"))
```
Valid values for `components` are any non-empty subset of
`c("within", "between", "gmc")`.
### Customising Column Names
The `within_pattern`, `between_pattern`, and `gmc_pattern` arguments control
the naming of the new columns. Each pattern is a glue-style string where
`{col}` is replaced by the variable name and `{group}` is replaced by the
grouping variable name. For example, the default `{col}_within_{group}` produces `screen_time_within_person`. Here, we use `{col}_wg` and `{col}_bg` to produce shorter names:
```{r decompose-names}
media_diary |>
decompose_within_between(
group = "person",
vars = c("screen_time"),
components = c("within", "between"),
within_pattern = "{col}_wg",
between_pattern = "{col}_bg"
) |>
select(starts_with("screen_time"))
```
### Decomposing Multiple Variables at Once
Pass a character vector to `vars` to decompose several variables in a single
call. The same `components` and naming patterns apply to all variables:
```{r decompose-multi}
media_diary |>
decompose_within_between(
group = "person",
vars = c("screen_time", "stress"),
components = c("within", "between")
) |>
glimpse()
```
## Fitting the REWB Model
### Step 1 — Within and between effects
We start with a model that includes only the within- and between-person
components of `screen_time` and a random intercept for person. This is the
core REWB specification:
```{r rewb-model-base}
diary_decomp <- decompose_within_between(
data = media_diary,
group = "person",
vars = "screen_time",
components = c("within", "between"),
within_pattern = "{col}_within",
between_pattern = "{col}_between"
)
fit_rewb <- lmer(
wellbeing ~ screen_time_within + screen_time_between + (1 | person),
data = diary_decomp
)
summary(fit_rewb, correlation = FALSE)
```
**Interpreting the coefficients:**
In this simulated dataset, the within-person coefficient
(`r round(fixef(fit_rewb)["screen_time_within"], 4)`) is positive and highly
significant. To illustrate how such an effect would be interpreted: on days
when someone watches one minute more than their own average, their wellbeing is
`r round(fixef(fit_rewb)["screen_time_within"], 4)` points higher. For a person
watching 60 minutes more than usual, the expected gain would be
`r round(60 * fixef(fit_rewb)["screen_time_within"], 2)` wellbeing points.
The between-person coefficient (`r round(fixef(fit_rewb)["screen_time_between"], 4)`)
is negative and significant. Illustrating interpretation: people who watch one
minute more per day on average show
`r abs(round(fixef(fit_rewb)["screen_time_between"], 4))` lower wellbeing. For
someone who watches 60 minutes more per day on average than another person, the
expected wellbeing gap would be
`r abs(round(60 * fixef(fit_rewb)["screen_time_between"], 2))` points.
The two effects point in *opposite directions* — exactly the pattern built into
the simulation. A naive regression conflates them:
```{r naive-model}
fit_naive <- lm(wellbeing ~ screen_time, data = diary_decomp)
summary(fit_naive)
```
The naive coefficient is near zero because the positive within-person and
negative between-person effects cancel each other out — an entirely uninformative
result that hides two simulated effects pointing in opposite directions. This
illustrates why a naive regression can be misleading when within- and
between-group processes operate simultaneously.
### Step 2 — Accounting for confounding
`self_control` was identified above as a confounder of the *between-person*
effect: people with lower trait self-control may watch more media on average
*and* have lower wellbeing, making it appear as though heavy media use causes
worse wellbeing at the between-person level. Adding `self_control` as a
covariate lets us test whether the between-person association with screen time
persists after removing this alternative explanation.
```{r rewb-model-confounded}
fit_rewb_conf <- lmer(
wellbeing ~ screen_time_within + screen_time_between + self_control +
(1 | person),
data = diary_decomp
)
summary(fit_rewb_conf, correlation = FALSE)
```
**Interpreting the coefficients:**
In this simulated dataset, the within-person coefficient is unchanged
(`r round(fixef(fit_rewb_conf)["screen_time_within"], 4)`): `self_control` is a
stable trait measured once per person, so it carries no within-person variation
and cannot alter the within-person estimate. This is a general property of
between-person covariates in REWB models, not specific to these simulated data.
The between-person coefficient changes substantially — from
`r round(fixef(fit_rewb)["screen_time_between"], 4)` (significant) in the
unadjusted model to
`r round(fixef(fit_rewb_conf)["screen_time_between"], 4)` (*p* = .23,
non-significant) after adjusting for `self_control`. This illustrates confounding:
the simulation was designed so that the apparent between-person harm of screen
time is driven by self-control. In a real study, a similar pattern would suggest
that people with lower self-control watch more TV on average *and* have lower
wellbeing, and that the self-control deficit — not screen time — explains the
wellbeing gap.
## Adding More Predictors
When you have several time-varying predictors, decompose all of them at once:
```{r multi-predictor, eval = FALSE}
diary_decomp2 <- decompose_within_between(
data = media_diary,
group = "person",
vars = c("screen_time", "stress"),
components = c("within", "between"),
within_pattern = "{col}_within",
between_pattern = "{col}_between"
)
fit_multi <- lmer(
wellbeing ~ screen_time_within + screen_time_between +
stress_within + stress_between +
self_control + (1 | person),
data = diary_decomp2
)
```
Here `stress_within` captures whether more stressful days than usual predict
lower wellbeing on those days (within-person), while `stress_between` captures
whether chronically more stressed people have lower wellbeing overall
(between-person).
## Adding Random Slopes
The REWB model above assumes the within-person effect of screen time on wellbeing
is the same for all persons. You can allow this effect to vary by adding a
random slope:
```{r random-slopes, eval = FALSE}
fit_slopes <- lmer(
wellbeing ~ screen_time_within + screen_time_between + self_control +
(screen_time_within | person),
data = diary_decomp
)
```
A significant random slope variance indicates that the within-person association
between screen time and wellbeing differs across persons — for some, extra media
use lifts their mood more than for others.
## Further Reading
This vignette covers the data-preparation and basic modelling side of REWB
analysis. For thorough treatments of model specification, assumption checking,
and interpretation — including cross-level interactions — see Bell et al. (2019)
and Enders & Tofighi (2007). For descriptive statistics and correlation matrices
that can inform REWB model specification, see `vignette("multilevel-descriptives")`.
## References
Bell, A., Fairbrother, M., & Jones, K. (2019). Fixed and random effects models:
Making an informed choice. *Quality & Quantity, 53*(2), 1051–1074.
https://doi.org/10.1007/s11135-018-0802-x
Enders, C. K., & Tofighi, D. (2007). Centering predictor variables in
cross-sectional multilevel models: A new look at an old issue.
*Psychological Methods, 12*(2), 121–138.
https://doi.org/10.1037/1082-989X.12.2.121