Unit-Specific vs. Population-Average Models

One thing that I always felt uncomfortable in multilevel modeling (MLM) is the concept of a unit-specific (US)/subject-specific model vs. a population-average (PA) model. I’ve come across it several times, but for some reason I haven’t really made an effort to fully understand it. I happened to come across this paper by Harring and Blozis, which I read before, and think that why not try to really understand the relationship between the coefficient estimates in a US model and in a PA model in the context of generalized linear mixed-effect model (GLMM). So I have this learning note.

library(tidyverse)
library(modelsummary)
library(glmmTMB)

Warning in checkDepPackageVersion(dep_pkg = "TMB"): Package version inconsistency detected.
glmmTMB was built with TMB version 1.9.6
Current TMB version is 1.9.10
Please re-install glmmTMB from source or restore original 'TMB' package (see '?reinstalling' for more information)

library(geepack)

While MLM/GLMM is a US model, which models the associations between predictors and the outcome for each cluster, PA models are popular in some areas of research, with the popular method of the generalized estimating equation (GEE). Whereas the fixed effect coefficients in US are the same as the coefficients in PA in linear models, when it comes to generalized linear models with nonlinear link functions, the coefficients are not the same. This is because some of the generalized linear models typically assume constant variance on the latent continuous response variable. For example, in a single-level logistic model and a GEE model, the latent response $Y^{\ast }$ has a variance of ${\pi }^2/3$, but in a two-level model, the variance is ${\pi }^2/3+{\tau }_0^2$.¹ Because the coefficients are in the unit of the latent response, it means that the coefficients are on different units for US vs. PA. But how are they different? I will explore four link functions: identity, log, probit, and logit. But first, some notations.

Model Notations

While in actual modeling, the distributional assumptions of the response variables are important (e.g., normal, Poisson), the comparison of US vs. PA mainly concerns the mean of the outcome and the link function. For all models, the random effects are normally distributed.

Conditional (US) Model

$$\begin{array}{rl}\mathrm{E}(y_{ij}|u_j)& ={\mu }_{ij}\\ h({\mu }_{ij})& ={\mathbf{x}}_{ij}^{\mathrm{\top }}\mathit{\gamma }+{\mathbf{z}}_{ij}^{\mathrm{\top }}{\mathbf{u}}_j\end{array}$$ where $h(\cdot )$ is the link function, ${\mathbf{x}}_{ij}$ and ${\mathbf{z}}_{ij}$ are the fixed and random covariates for the $i$th person in the $j$th cluster. The distributional assumption is ${\mathbf{u}}_j\sim N_q(\mathbf{0},\mathbf{G})$

Marginal (PA) Model

Now one is modeling the marginal mean:

$$\begin{array}{rl}\mathrm{E}(y_{ij})& =\mathrm{E}[\mathrm{E}(y_{ij}|{\mu }_{ij})]={\mu }_{ij}^{\text{PA}}\\ h({\mu }_{ij}^{\text{PA}})& ={\mathbf{x}}_{ij}^{\mathrm{\top }}{\mathit{\gamma }}^{\text{PA}}\end{array}$$ The above two formulas can be used to find the transformation from the unit-specific coefficients, $\mathit{\gamma }$, to the population-average coefficients, ${\mathit{\gamma }}^{\text{PA}}$.

Identity Link

$$h({\mu }_{ij}^{\text{PA}})={\mu }_{ij}$$ From the US model $$\begin{array}{rl}\mathrm{E}(y_{ij})& =\mathrm{E}[\mathrm{E}(y_{ij}|u_j)]=\mathrm{E}[{\mathbf{x}}_{ij}^{\mathrm{\top }}\mathit{\gamma }+{\mathbf{z}}_{ij}^{\mathrm{\top }}{\mathbf{u}}_j]\\ & ={\mathbf{x}}_{ij}^{\mathrm{\top }}\mathit{\gamma }\end{array}$$ Compare to the PA model $$\mathrm{E}(y_{ij})={\mathbf{x}}_{ij}^{\mathrm{\top }}{\mathit{\gamma }}^{\text{PA}},$$ we have $\mathit{\gamma }={\mathit{\gamma }}^{\text{PA}}$

Plot

# Simulate predictor X ~ U(-2, 2), with ICC = 0
num_obs <- 2e4
num_subjects <- 200
x <- runif(num_obs, min = -1, max = 3)
# Subject IDs
subject_id <- rep(seq_len(num_subjects), each = num_obs / num_subjects)

# Fixed effects
gamma0 <- -1
gamma1 <- 1
# Random intercepts and 
tau2_u <- 0.25
# u <- rnorm(num_subjects, mean = 0, sd = sqrt(tau2_u))
u <- qnorm(seq(0.01, 0.99, length.out = num_subjects), sd = sqrt(tau2_u))
# Simulate mu_{ij}
mu <- gamma0 + gamma1 * x + u[subject_id]
# Plot
df <- tibble(y = mu, x = x, id = subject_id)
ggplot(df, aes(x = x, y = y)) + 
  geom_smooth(aes(group = id), 
              se = FALSE, 
              method = "lm",  
              size = 0.2) + 
  geom_smooth(se = FALSE, col = "red") + 
  stat_function(fun = function(x) gamma0 + gamma1 * x, 
                col = "blue", size = 1.4)

Warning: Using `size` aesthetic for lines was deprecated in ggplot2 3.4.0.
ℹ Please use `linewidth` instead.

`geom_smooth()` using formula = 'y ~ x'
`geom_smooth()` using method = 'gam' and formula = 'y ~ s(x, bs = "cs")'

The blue line represents the regression line when $u_j=0$, and the red line is the regression line for the population-average model. They are the same.

Compare NLME with GEE

# Unit-Specific
m_us <- glmmTMB(mu ~ x + (x | id), data = df, 
                family = gaussian(link = "identity"), 
                dispformula = ~ 0)
# Population-Average
m_pa <- geeglm(mu ~ x, 
               family = gaussian(link = "identity"),
               data = df, 
               id = id, 
               corstr = "exchangeable")
# The coefficients are the same
cbind(US = fixef(m_us)$cond, PA = coef(m_pa))

                    US PA
(Intercept) -0.9999857 -1
x            1.0000000  1

Log Link

The log link is commonly used in the Poisson model.

$$h({\mu }_{ij}^{\text{PA}})=\log ({\mu }_{ij})$$ From the US model $$\begin{array}{rl}\mathrm{E}(y_{ij})& =\mathrm{E}[\mathrm{E}(y_{ij}|u_j)]=\mathrm{E}[h^{-1}({\mathbf{x}}_{ij}^{\mathrm{\top }}\mathit{\gamma }+{\mathbf{z}}_{ij}^{\mathrm{\top }}{\mathbf{u}}_j)]\\ & =\mathrm{E}[\exp ({\mathbf{x}}_{ij}^{\mathrm{\top }}\mathit{\gamma }+{\mathbf{z}}_{ij}^{\mathrm{\top }}{\mathbf{u}}_j)]\\ & =\exp ({\mathbf{x}}_{ij}^{\mathrm{\top }}\mathit{\gamma })\mathrm{E}[\exp ({\mathbf{z}}_{ij}^{\mathrm{\top }}{\mathbf{u}}_j))]\\ & =\exp ({\mathbf{x}}_{ij}^{\mathrm{\top }}\mathit{\gamma })\exp [\mathbf{z}{\mathrm{\top }}_{ij}\mathbf{G}{\mathbf{z}}_{ij}/2]\\ & =\exp ({\mathbf{x}}_{ij}^{\mathrm{\top }}\mathit{\gamma }+\mathbf{z}{\mathrm{\top }}_{ij}\mathbf{G}{\mathbf{z}}_{ij}/2)\end{array}$$ Compare to the population-average model $$\mathrm{E}(y_{ij})=h^{-1}({\mathbf{x}}_{ij}^{\mathrm{\top }}{\mathit{\gamma }}^{\text{PA}})=\exp ({\mathbf{x}}_{ij}^{\mathrm{\top }}{\mathit{\gamma }}^{\text{PA}}),$$ we have ${\mathit{\gamma }}^{\text{PA}}=\mathit{\gamma }+[{\mathbf{z}}_{ij}^{\mathrm{\top }}\mathbf{G}{\mathbf{z}}_{ij}/2\mathbf{0}{]}^{\mathrm{\top }}$. So the intercept has an offset, while the other coefficients stay the same.

Plot

# Fixed effects
gamma0 <- 1
gamma1 <- 0.2
# Random intercepts and 
tau2_u <- 0.25
# u <- rnorm(num_subjects, mean = 0, sd = sqrt(tau2_u))
u <- qnorm(seq(0.01, 0.99, length.out = num_subjects), sd = sqrt(tau2_u))
# Simulate mu_{ij}
mu <- exp(gamma0 + gamma1 * x + u[subject_id])
# Plot
df <- tibble(y = mu, x = x, id = subject_id)
ggplot(df, aes(x = x, y = y)) + 
  geom_smooth(aes(group = id), 
              se = FALSE, 
              method = "glm",
              method.args = list(family = gaussian("log")), 
              size = 0.2) + 
  geom_smooth(se = FALSE, col = "red") + 
  stat_function(fun = function(x) exp(gamma0 + gamma1 * x), 
                col = "blue", size = 1.4)

`geom_smooth()` using formula = 'y ~ x'
`geom_smooth()` using method = 'gam' and formula = 'y ~ s(x, bs = "cs")'

The intercept for the red line has an offset of ${\tau }_0^2/2=0.125$.

Compare NLME with GEE

# Unit-Specific
m_us <- glmmTMB(mu ~ x + (x | id), data = df, 
                family = gaussian(link = "log"), 
                dispformula = ~ 0)
# Population-Average
m_pa <- geeglm(mu ~ x, 
               family = gaussian(link = "log"),
               data = df, 
               id = id, 
               corstr = "exchangeable")
# The coefficients are the same
cbind(US = fixef(m_us)$cond, PA = coef(m_pa))

             US        PA
(Intercept) 1.0 1.1178892
x           0.2 0.1981924

The offset is in the intercept.

Probit Link

The probit link, or the inverse normal cumulative density function, is commonly used in probit regression. $$h({\mu }_{ij}^{\text{PA}})=\mathrm{\Phi }({\mu }_{ij})$$ From the US model $$\begin{array}{rl}\mathrm{E}(y_{ij})& =\mathrm{E}[\mathrm{E}(y_{ij}|u_j)]=\mathrm{E}[h^{-1}({\mathbf{x}}_{ij}^{\mathrm{\top }}\mathit{\gamma }+{\mathbf{z}}_{ij}^{\mathrm{\top }}{\mathbf{u}}_j)]\\ & =\mathrm{E}[\mathrm{\Phi }({\mathbf{x}}_{ij}^{\mathrm{\top }}\mathit{\gamma }+{\mathbf{z}}_{ij}^{\mathrm{\top }}{\mathbf{u}}_j)]\\ & =P(Z\le {\mathbf{x}}_{ij}^{\mathrm{\top }}\mathit{\gamma }+{\mathbf{z}}_{ij}^{\mathrm{\top }}{\mathbf{u}}_j)\text{for }Z\sim N(0,1)\\ & =P(Z-{\mathbf{z}}_{ij}^{\mathrm{\top }}{\mathbf{u}}_j\le {\mathbf{x}}_{ij}^{\mathrm{\top }}\mathit{\gamma }),\end{array}$$ where $Z-{\mathbf{z}}_{ij}^{\mathrm{\top }}{\mathbf{u}}_j\sim N(0,\sqrt{1+{\mathbf{z}}_{ij}^{\mathrm{\top }}\mathbf{G}{\mathbf{z}}_{ij}})$. So $$\begin{array}{rl}\mathrm{E}(y_{ij})& =P(\frac{Z-{\mathbf{z}}_{ij}^{\mathrm{\top }}{\mathbf{u}}_j}{\sqrt{1+{\mathbf{z}}_{ij}^{\mathrm{\top }}\mathbf{G}{\mathbf{z}}_{ij}}}\le \frac{{\mathbf{x}}_{ij}^{\mathrm{\top }}\mathit{\gamma }}{\sqrt{1+{\mathbf{z}}_{ij}^{\mathrm{\top }}\mathbf{G}{\mathbf{z}}_{ij}}})\\ & =\mathrm{\Phi }[(1+{\mathbf{z}}_{ij}^{\mathrm{\top }}\mathbf{G}{\mathbf{z}}_{ij}{)}^{-1/2}{\mathbf{x}}_{ij}^{\mathrm{\top }}\mathit{\gamma }]\end{array}$$ So the PA coefficients shrinks by a factor of $(1+{\mathbf{z}}_{ij}^{\mathrm{\top }}\mathbf{G}{\mathbf{z}}_{ij}{)}^{-1/2}$.

Random intercepts only

Plot

# Fixed effects
gamma0 <- -0.5
gamma1 <- 1
# Random intercepts and 
tau2_u <- 0.3
# u <- rnorm(num_subjects, mean = 0, sd = sqrt(tau2_u))
u <- qnorm(seq(0.01, 0.99, length.out = num_subjects), sd = sqrt(tau2_u))
# Simulate mu_{ij}
mu <- pnorm(gamma0 + gamma1 * x + u[subject_id])
# Plot
df <- tibble(y = mu, x = x, id = subject_id)
ggplot(df, aes(x = x, y = y)) + 
  geom_smooth(aes(group = id), 
              se = FALSE, 
              method = "glm",
              method.args = list(family = gaussian("probit")), 
              size = 0.2) + 
  geom_smooth(se = FALSE, col = "red") + 
  stat_function(fun = function(x) pnorm(gamma0 + gamma1 * x), 
                col = "blue", size = 1.4)

`geom_smooth()` using formula = 'y ~ x'
`geom_smooth()` using method = 'gam' and formula = 'y ~ s(x, bs = "cs")'

The shrinkage factor is $(1+{\tau }_0^2{)}^{-1/2}=0.877058$.

Compare NLME with GEE

# Unit-Specific
m_us <- glmmTMB(mu ~ x + (x | id), data = df, 
                family = gaussian(link = "probit"), 
                dispformula = ~ 0)
# Population-Average
m_pa <- geeglm(mu ~ x, 
               family = gaussian(link = "probit"),
               data = df, 
               id = id, 
               corstr = "exchangeable")
# The coefficients are the same
cbind(US = fixef(m_us)$cond, PA = coef(m_pa))

                   US         PA
(Intercept) -0.499999 -0.4458653
x            1.000000  0.8858915

Smaller coefficients for both the intercept and $x$.

Random intercepts and slopes

Plot

# Add random slopes
tau2_u1 <- 0.15
# u1 <- rnorm(num_subjects, mean = 0, sd = sqrt(tau2_u1))
u1 <- qnorm(seq(0.01, 0.99, length.out = num_subjects), 
            sd = sqrt(tau2_u1))
# permutate the random effects to make u and u1 independent
u1 <- sample(u1)
# Simulate y
mu2 <- pnorm(gamma0 + (gamma1 + u1[subject_id]) * x + u[subject_id])
# Plot
df2 <- tibble(y = mu2, x = x, id = subject_id)
ggplot(df2, aes(x = x, y = y)) + 
  geom_smooth(aes(group = id), 
              se = FALSE, 
              method = "glm",
              method.args = list(family = gaussian("probit")), 
              size = 0.2) + 
  geom_smooth(se = FALSE, col = "red") + 
  stat_function(fun = function(x) pnorm(gamma0 + gamma1 * x), 
                col = "blue", size = 1.4)

`geom_smooth()` using formula = 'y ~ x'
`geom_smooth()` using method = 'gam' and formula = 'y ~ s(x, bs = "cs")'

The shrinking factor is $[1+({\tau }_0^2+2{\tau }_{01}\overline{x}+{\tau }_1^{2}var(x)){]}^{-1/2}$ = $[1+(0.3+0.15\times 4/3){]}^{-1/2}$ = 0.8164966.

Compare NLME with GEE

# Unit-Specific
m_us <- glmmTMB(mu2 ~ x + (x | id), data = df2, 
                family = gaussian(link = "probit"), 
                dispformula = ~ 0)
# Population-Average
m_pa <- geeglm(mu2 ~ x, 
               family = gaussian(link = "probit"),
               data = df2, 
               id = id, 
               corstr = "exchangeable")
# The coefficients are the same
cbind(US = fixef(m_us)$cond, PA = coef(m_pa))

                   US         PA
(Intercept) -0.499973 -0.3112705
x            1.000002  0.8302407

Smaller coefficients for both the intercept and $x$.

Logit Link

The logit link is commonly used in logistic regression. $$h({\mu }_{ij}^{\text{PA}})=\log \frac{{\mu }_{ij}}{1-{\mu }_{ij}}$$

From the US model $$\begin{array}{rl}\mathrm{E}(y_{ij})& =\mathrm{E}[\mathrm{E}(y_{ij}|u_j)]=\mathrm{E}[h^{-1}({\mathbf{x}}_{ij}^{\mathrm{\top }}\mathit{\gamma }+{\mathbf{z}}_{ij}^{\mathrm{\top }}{\mathbf{u}}_j)]\\ & =\mathrm{E}[\stackrel{-1}{\mathrm{logit}}({\mathbf{x}}_{ij}^{\mathrm{\top }}\mathit{\gamma }-{\mathbf{z}}_{ij}^{\mathrm{\top }}{\mathbf{u}}_j)\end{array}$$

The integral does not have a closed-form expression and cannot be expressed as a logistic function. However, one can approximates a normal cdf using a logistic function, and vice versa, and there are several ways to do it.² Zeger, Liang, and Albert (1988, p. 1054) provides one approximation that results in $$\mathrm{E}(y_{ij})\approx \stackrel{-1}{\mathrm{logit}}[a_l(\mathbf{G}){\mathbf{x}}_{ij}^{\mathrm{\top }}\mathit{\gamma }],$$ where $a_l(\mathbf{G})=(1+c^2{\mathbf{z}}_{ij}^{\mathrm{\top }}\mathbf{G}{\mathbf{z}}_{ij}{)}^{-1/2}$ and $c^2={\left(\frac{16}{15}\right)}^2\frac{3}{{\pi }^2}\approx 1/{1.7}^2=0.3460208$, which was used in Allison (2009). Some other authors, like Snijders and Bosker (2012), use a simpler approximation with $c^2=\frac{3}{{\pi }^2}\approx 0.3039636$.

Random intercepts only

Plot

# Fixed effects
gamma0 <- -1
gamma1 <- 2
# Random intercepts and subject IDs
tau2_u <- 1
# u <- rnorm(num_subjects, mean = 0, sd = sqrt(tau2_u))
u <- qnorm(seq(0.01, 0.99, length.out = num_subjects), sd = sqrt(tau2_u))
# Simulate y
mu <- plogis(gamma0 + gamma1 * x + u[subject_id])
# Plot
df <- tibble(y = mu, x = x, id = subject_id)
ggplot(df, aes(x = x, y = y)) + 
  geom_smooth(aes(group = id), 
              se = FALSE, 
              method = "glm", 
              method.args = list(family = gaussian("logit")), 
              size = 0.2) + 
  geom_smooth(se = FALSE, col = "red") + 
  stat_function(fun = function(x) plogis(gamma0 + gamma1 * x), 
                col = "blue", size = 1.4)

`geom_smooth()` using formula = 'y ~ x'
`geom_smooth()` using method = 'gam' and formula = 'y ~ s(x, bs = "cs")'

The shrinkage factor is $(1+{\tau }_0^2/{1.7}^2{)}^{-1/2}=0.8619342$.

Compare NLME with GEE

# Unit-Specific
m_us <- glmmTMB(mu ~ x + (x | id), data = df, 
                family = gaussian(link = "logit"), 
                dispformula = ~ 0)
# Population-Average
m_pa <- geeglm(mu ~ x, 
               family = gaussian(link = "logit"),
               data = df, 
               id = id, 
               corstr = "exchangeable")
# The coefficients are the same
cbind(US = fixef(m_us)$cond, PA = coef(m_pa))

                    US        PA
(Intercept) -0.9999994 -0.876522
x            2.0000000  1.711613

Smaller coefficients for both the intercept and $x$.

Random intercepts and slopes

# Add random slopes
tau2_u1 <- 0.5
# u1 <- rnorm(num_subjects, mean = 0, sd = sqrt(tau2_u1))
u1 <- qnorm(seq(0.01, 0.99, length.out = num_subjects), 
            sd = sqrt(tau2_u1))
# permutate the random effects to make u and u1 independent
u1 <- sample(u1)
# Simulate y
mu2 <- plogis(gamma0 + (gamma1 + u1[subject_id]) * x + u[subject_id])
# Plot
df2 <- tibble(y = mu2, x = x, id = subject_id)
ggplot(df2, aes(x = x, y = y)) + 
  geom_smooth(aes(group = id), 
              se = FALSE, 
              method = "glm", 
              method.args = list(family = gaussian("logit")), 
              size = 0.2) + 
  geom_smooth(se = FALSE, col = "red") + 
  stat_function(fun = function(x) plogis(gamma0 + gamma1 * x), 
                col = "blue", size = 1.4)

`geom_smooth()` using formula = 'y ~ x'
`geom_smooth()` using method = 'gam' and formula = 'y ~ s(x, bs = "cs")'

The shrinking factor is approximately $[1+({\tau }_0^2+{\tau }_1^{2}var(x))/{1.7}^2{]}^{-1/2}$ = $[1+(1+0.5\times 4/3)/{1.7}^2{]}^{-1/2}$ = 0.7963891.

Compare NLME with GEE

# Unit-Specific
m_us <- glmmTMB(mu2 ~ x + (x | id), data = df2, 
                family = gaussian(link = "logit"), 
                dispformula = ~ 0)
# Population-Average
m_pa <- geeglm(mu2 ~ x, 
               family = gaussian(link = "logit"),
               data = df2, 
               id = id, 
               corstr = "exchangeable")
# The coefficients are the same
cbind(US = fixef(m_us)$cond, PA = coef(m_pa))

                    US        PA
(Intercept) -0.9999936 -0.701903
x            1.9999867  1.585452

Footnotes

1. Snijders & Bosker (2012), chapter 17.

2. https://www.johndcook.com/blog/2010/05/18/normal-approximation-to-logistic/