Skip to contents

Parametric models

Source: vignettes/parametric_model.Rmd
parametric_model.Rmd

Frequentist methods

Polynomial models

Seroprevalence is modeled using the following serocatalytic model

\[ \pi(a) = 1 - \text{exp}({-\Sigma_{i=1}^k \beta_i a^i}) \]

Where:

  • \(a\) is the variable age

  • \(\pi\) is the seroprevalence of the population at age \(a\)

  • \(k\) is the degree of the polynomial

  • \(\beta_i\) are the model parameters

Which implies the force of infection is \(\lambda(a) = \Sigma_{i=1}^k \beta_i i a^{i-1}\)

This generalization encompasses several classical serocatalytic model including

Refer to Chapter 6.1.1 of the book by Hens et al. (2012) for a more detailed explanation of the methods.

Fitting data

We will use the Parvo B19 data from Finland 1997–1998 for this example.

data <- parvob19_fi_1997_1998[order(parvob19_fi_1997_1998$age), ]

To fit a polynomial model, use the polynomial_model() function.

# Fit a Muench model
muench <- polynomial_model(data, k = 1, status_col="seropositive")
summary(muench$info)
#> 
#> Call:
#> glm(formula = Age(k), family = binomial(link = link), data = df)
#> 
#> Coefficients:
#>      Estimate Std. Error z value Pr(>|z|)    
#> age -0.029088   0.001375  -21.15   <2e-16 ***
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#> 
#> (Dispersion parameter for binomial family taken to be 1)
#> 
#>     Null deviance:    Inf  on 1117  degrees of freedom
#> Residual deviance: 1366.9  on 1116  degrees of freedom
#> AIC: 1368.9
#> 
#> Number of Fisher Scoring iterations: 6
plot(muench)

The users can also choose to provide a range of values for \(k\) in which case the package will try to find the best \(k\) parameter determined by Loglikelihood Ratio test (LRT)

# Provide a range of values for k
best_param <- polynomial_model(data, k = 1:5, status_col = "seropositive")
plot(best_param)

# View the best model here which suggests k = 4 is the best parameter value
best_param
#> Polynomial model
#> 
#> Input type:  linelisting 
#> Degree (k):  4 
#> 
#> Call:  glm(formula = Age(k), family = binomial(link = link), data = df)
#> 
#> Coefficients:
#>        age    I(age^2)    I(age^3)    I(age^4)  
#> -3.381e-02  -9.950e-04   5.551e-05  -5.737e-07  
#> 
#> Degrees of Freedom: 1117 Total (i.e. Null);  1113 Residual
#> Null Deviance:       Inf 
#> Residual Deviance: 1336  AIC: 1344

Fractional polynomial model

Proposed model

Fractional polynomial model generalize conventional polynomial class of functions. In the context of binary responses, a fractional polynomial of degree \(m\) for the linear predictor is defined as followed

\[ \eta_m(a, \beta, p_1, p_2, ...,p_m) = \Sigma^m_{i=0} \beta_i H_i(a) \]

Where \(m\) is an integer, \(p_1 \le p_2 \le... \le p_m\) is a sequence of powers, and \(H_i(a)\) is a transformation given by

\[ H_i = \begin{cases} a^{p_i} & \text{ if } p_i \neq p_{i-1}, \\ H_{i-1}(a) \times log(a) & \text{ if } p_i = p_{i-1}, \end{cases} \]

Refer to Chapter 6.2 of the book by Hens et al. (2012) for a more detailed explanation of the methods.

Fitting data

Use fp_model() to fit a fractional polynomial model.

The parameter p specifies the powers of each polynomial term (length of p is thus the model’s degree)

hav <- hav_be_1993_1994
model <- fp_model(hav, p=c(1, 1.5), link="cloglog")
model
#> Fractional polynomial model 
#> 
#> Input type:  aggregated 
#> Powers:  1, 1.5 
#> 
#> Call:  glm(formula = as.formula(formulate(p)), family = binomial(link = link))
#> 
#> Coefficients:
#> (Intercept)     I(age^1)   I(age^1.5)  
#>    -4.56685      0.22340     -0.01775  
#> 
#> Degrees of Freedom: 85 Total (i.e. Null);  83 Residual
#> Null Deviance:       1320 
#> Residual Deviance: 85.81     AIC: 365.4
plot(model)

The users can also tell the package to perform parameter selection by providing p as a named list with 2 elements:

  • degree the maximum number of terms to search over
  • p_range the possible powers for each term
model <- fp_model(hav, 
                  p=list(
                    p_range=seq(-2,3,0.1),
                    degree=2
                  ), 
                  monotonic=FALSE,
                  link="cloglog")
plot(model)

# the best set of powers for this dataset is 1.5 and 1.6
model
#> Fractional polynomial model 
#> 
#> Input type:  aggregated 
#> Powers:  1.5, 1.6 
#> 
#> Call:  glm(formula = as.formula(formulate(curr_p)), family = binomial(link = link))
#> 
#> Coefficients:
#> (Intercept)   I(age^1.5)   I(age^1.6)  
#>    -3.61083      0.12443     -0.07656  
#> 
#> Degrees of Freedom: 85 Total (i.e. Null);  83 Residual
#> Null Deviance:       1320 
#> Residual Deviance: 81.6  AIC: 361.2

To restrict the parameters search such that the predictions are monotonic (thus ensuring the FOI to be \(\lambda \geq 0\)) set monotonic=TRUE

# ---- Best model with the monotonic constraint -----
model <- fp_model(hav, 
                  p=list(
                    p_range=seq(-2,3,0.1),
                    degree=2
                  ), 
                  monotonic=TRUE,
                  link="cloglog")
plot(model)

# the best set of powers with the monotonic constraint is 0.5 and 1.1
model
#> Fractional polynomial model 
#> 
#> Input type:  aggregated 
#> Powers:  0.5, 1.1 
#> 
#> Call:  glm(formula = as.formula(formulate(curr_p)), family = binomial(link = link))
#> 
#> Coefficients:
#> (Intercept)   I(age^0.5)   I(age^1.1)  
#>    -7.64170      1.67492     -0.05304  
#> 
#> Degrees of Freedom: 85 Total (i.e. Null);  83 Residual
#> Null Deviance:       1320 
#> Residual Deviance: 106   AIC: 385.5

Nonlinear models

Farrington model

Proposed model

For Farrington’s model, the force of infection was defined non-negative for all a \(\lambda(a) \geq 0\) and increases to a peak in a linear fashion followed by an exponential decrease

\[ \lambda(a) = (\alpha a - \gamma)e^{-\beta a} + \gamma \]

Where \(\gamma\) is called the long term residual for FOI, as \(a \rightarrow \infty\) , \(\lambda (a) \rightarrow \gamma\)

Integrating \(\lambda(a)\) would results in the following non-linear model for prevalence

\[ \pi (a) = 1 - e^{-\int_0^a \lambda(s) ds} \\ = 1 - exp\{ \frac{\alpha}{\beta}ae^{-\beta a} + \frac{1}{\beta}(\frac{\alpha}{\beta} - \gamma)(e^{-\beta a} - 1) -\gamma a \} \]

Refer to Chapter 6.1.2 of the book by Hens et al. (2012) for a more detailed explanation of the methods.

Fitting data

Use farrington_model() to fit a Farrington’s model.

farrington_md <- farrington_model(
   rubella_uk_1986_1987,
   start=list(alpha=0.07,beta=0.1,gamma=0.03)
   )
farrington_md
#> Farrington model 
#> 
#> Input type:  aggregated 
#> 
#> Call:
#> mle(minuslogl = farrington, start = start, fixed = fixed)
#> 
#> Coefficients:
#>      alpha       beta      gamma 
#> 0.07034904 0.20243950 0.03665599
plot(farrington_md)

Weibull model

Proposed model

For a Weibull model, the prevalence is given by

\[ \pi (d) = 1 - e^{ - \beta_0 d ^ {\beta_1}} \]

Where \(d\) is exposure time (difference between age of injection and age at test)

The model was reformulated as a GLM model with log - log link and linear predictor using log(d)

\[\eta(d) = log(\beta_0) + \beta_1 log(d)\]

Thus implies that the force of infection is a monotone function of the exposure time as followed

\[ \lambda(d) = \beta_0 \beta_1 d^{\beta_1 - 1} \]

Refer to Chapter 6.1.2 of the book by Hens et al. (2012) for a more detailed explanation of the methods.

Fitting data

Use weibull_model() to fit a Weibull model.

hcv <- hcv_be_2006[order(hcv_be_2006$dur), ]

wb_md <- hcv %>% weibull_model(t_lab = "dur", status_col="seropositive")
wb_md
#> Weibull model 
#> 
#> Input type:  linelisting 
#> b0=-0.276, b1=0.3807 
#> 
#> Call:  glm(formula = spos ~ log(t), family = binomial(link = "cloglog"))
#> 
#> Coefficients:
#> (Intercept)       log(t)  
#>     -0.2760       0.3807  
#> 
#> Degrees of Freedom: 420 Total (i.e. Null);  419 Residual
#> Null Deviance:       452.1 
#> Residual Deviance: 419.4     AIC: 423.4
plot(wb_md)

Bayesian methods

Proposed approach

Consider a model for prevalence that has a parametric form \(\pi(a_i, \alpha)\) where \(\alpha\) is a parameter vector

One can constraint the parameter space of the prior distribution \(P(\alpha)\) in order to achieve the desired monotonicity of the posterior distribution \(P(\pi_1, \pi_2, ..., \pi_m|y,n)\)

Where:

  • \(n = (n_1, n_2, ..., n_m)\) and \(n_i\) is the sample size at age \(a_i\)

  • \(y = (y_1, y_2, ..., y_m)\) and \(y_i\) is the number of infected individual from the \(n_i\) sampled subjects

Farrington

Proposed model

The model for prevalence is as followed

\[ \pi (a) = 1 - exp\{ \frac{\alpha_1}{\alpha_2}ae^{-\alpha_2 a} + \frac{1}{\alpha_2}(\frac{\alpha_1}{\alpha_2} - \alpha_3)(e^{-\alpha_2 a} - 1) -\alpha_3 a \} \]

For likelihood model, independent binomial distribution are assumed for the number of infected individuals at age \(a_i\)

\[ y_i \sim Bin(n_i, \pi_i), \text{ for } i = 1,2,3,...m \]

The constraint on the parameter space can be incorporated by assuming truncated normal distribution for the components of \(\alpha\), \(\alpha = (\alpha_1, \alpha_2, \alpha_3)\) in \(\pi_i = \pi(a_i,\alpha)\)

\[ \alpha_j \sim \text{truncated } \mathcal{N}(\mu_j, \tau_j), \text{ } j = 1,2,3 \]

The joint posterior distribution for \(\alpha\) can be derived by combining the likelihood and prior as followed

\[ P(\alpha|y) \propto \prod^m_{i=1} \text{Bin}(y_i|n_i, \pi(a_i, \alpha)) \prod^3_{i=1}-\frac{1}{\tau_j}\text{exp}(\frac{1}{2\tau^2_j} (\alpha_j - \mu_j)^2) \]

  • Where the flat hyperprior distribution is defined as followed:

    • \(\mu_j \sim \mathcal{N}(0, 10000)\)

    • \(\tau^{-2}_j \sim \Gamma(100,100)\)

The full conditional distribution of \(\alpha_i\) is thus \[ P(\alpha_i|\alpha_j,\alpha_k, k, j \neq i) \propto -\frac{1}{\tau_i}\text{exp}(\frac{1}{2\tau^2_i} (\alpha_i - \mu_i)^2) \prod^m_{i=1} \text{Bin}(y_i|n_i, \pi(a_i, \alpha)) \]

Refer to Chapter 10.3.1 of the book by Hens et al. (2012) for a more detailed explanation of the method.

Fitting data

To fit Farrington model, use hierarchical_bayesian_model() and define type = "far2" or type = "far3" where

  • type = "far2" refers to Farrington model with 2 parameters (\(\alpha_3 = 0\))

  • type = "far3" refers to Farrington model with 3 parameters (\(\alpha_3 > 0\))

df <- mumps_uk_1986_1987
model <- hierarchical_bayesian_model(df, type="far3")
#> 
#> SAMPLING FOR MODEL 'fra_3' NOW (CHAIN 1).
#> Chain 1: Rejecting initial value:
#> Chain 1:   Log probability evaluates to log(0), i.e. negative infinity.
#> Chain 1:   Stan can't start sampling from this initial value.
#> Chain 1: 
#> Chain 1: Gradient evaluation took 0.000165 seconds
#> Chain 1: 1000 transitions using 10 leapfrog steps per transition would take 1.65 seconds.
#> Chain 1: Adjust your expectations accordingly!
#> Chain 1: 
#> Chain 1: 
#> Chain 1: Iteration:    1 / 5000 [  0%]  (Warmup)
#> Chain 1: Iteration:  500 / 5000 [ 10%]  (Warmup)
#> Chain 1: Iteration: 1000 / 5000 [ 20%]  (Warmup)
#> Chain 1: Iteration: 1500 / 5000 [ 30%]  (Warmup)
#> Chain 1: Iteration: 1501 / 5000 [ 30%]  (Sampling)
#> Chain 1: Iteration: 2000 / 5000 [ 40%]  (Sampling)
#> Chain 1: Iteration: 2500 / 5000 [ 50%]  (Sampling)
#> Chain 1: Iteration: 3000 / 5000 [ 60%]  (Sampling)
#> Chain 1: Iteration: 3500 / 5000 [ 70%]  (Sampling)
#> Chain 1: Iteration: 4000 / 5000 [ 80%]  (Sampling)
#> Chain 1: Iteration: 4500 / 5000 [ 90%]  (Sampling)
#> Chain 1: Iteration: 5000 / 5000 [100%]  (Sampling)
#> Chain 1: 
#> Chain 1:  Elapsed Time: 17.494 seconds (Warm-up)
#> Chain 1:                100.35 seconds (Sampling)
#> Chain 1:                117.844 seconds (Total)
#> Chain 1:
model
#> Hierarchical Bayesian model 
#> 
#> Input type:  aggregated 
#> Model:  Farrington model with 3 parameters 
#> 
#> Fitted parameters:
#>  alpha1 = 0.1397 (95% CrI [0.129, 0.1521], sd = 0.005927)
#>  alpha2 = 0.199 (95% CrI [0.1848, 0.2181], sd = 0.008454)
#>  alpha3 = 0.009017 (95% CrI [0.000252, 0.02798], sd = 0.007503)
plot(model)

Log-logistic

Proposed approach

The model for seroprevalence is as followed

\[ \pi(a) = \frac{\beta a^\alpha}{1 + \beta a^\alpha}, \text{ } \alpha, \beta > 0 \]

The likelihood is specified to be the same as Farrington model (\(y_i \sim Bin(n_i, \pi_i)\)) with

\[ \text{logit}(\pi(a)) = \alpha_2 + \alpha_1\log(a) \]

  • Where \(\alpha_2 = \text{log}(\beta)\)

The prior model of \(\alpha_1\) is specified as \(\alpha_1 \sim \text{truncated } \mathcal{N}(\mu_1, \tau_1)\) with flat hyperprior as in Farrington model

\(\beta\) is constrained to be positive by specifying \(\alpha_2 \sim \mathcal{N}(\mu_2, \tau_2)\)

The full conditional distribution of \(\alpha_1\) is thus

\[ P(\alpha_1|\alpha_2) \propto -\frac{1}{\tau_1} \text{exp} (\frac{1}{2 \tau_1^2} (\alpha_1 - \mu_1)^2) \prod_{i=1}^m \text{Bin}(y_i|n_i,\pi(a_i, \alpha_1, \alpha_2) ) \]

And \(\alpha_2\) can be derived in the same way

Refer to Chapter 10.3.3 of the book by Hens et al. (2012) for a more detailed explanation of the method.

Fitting data

To fit Log-logistic model, use hierarchical_bayesian_model() and define type = "log_logistic"

df <- rubella_uk_1986_1987
model <- hierarchical_bayesian_model(df, type="log_logistic")
#> 
#> SAMPLING FOR MODEL 'log_logistic' NOW (CHAIN 1).
#> Chain 1: 
#> Chain 1: Gradient evaluation took 6.5e-05 seconds
#> Chain 1: 1000 transitions using 10 leapfrog steps per transition would take 0.65 seconds.
#> Chain 1: Adjust your expectations accordingly!
#> Chain 1: 
#> Chain 1: 
#> Chain 1: Iteration:    1 / 5000 [  0%]  (Warmup)
#> Chain 1: Iteration:  500 / 5000 [ 10%]  (Warmup)
#> Chain 1: Iteration: 1000 / 5000 [ 20%]  (Warmup)
#> Chain 1: Iteration: 1500 / 5000 [ 30%]  (Warmup)
#> Chain 1: Iteration: 1501 / 5000 [ 30%]  (Sampling)
#> Chain 1: Iteration: 2000 / 5000 [ 40%]  (Sampling)
#> Chain 1: Iteration: 2500 / 5000 [ 50%]  (Sampling)
#> Chain 1: Iteration: 3000 / 5000 [ 60%]  (Sampling)
#> Chain 1: Iteration: 3500 / 5000 [ 70%]  (Sampling)
#> Chain 1: Iteration: 4000 / 5000 [ 80%]  (Sampling)
#> Chain 1: Iteration: 4500 / 5000 [ 90%]  (Sampling)
#> Chain 1: Iteration: 5000 / 5000 [100%]  (Sampling)
#> Chain 1: 
#> Chain 1:  Elapsed Time: 4.337 seconds (Warm-up)
#> Chain 1:                6.096 seconds (Sampling)
#> Chain 1:                10.433 seconds (Total)
#> Chain 1:
model
#> Hierarchical Bayesian model 
#> 
#> Input type:  aggregated 
#> Model:  Log-logistic model 
#> 
#> Fitted parameters:
#>  alpha1 = 1.642 (95% CrI [1.539, 1.749], sd = 0.05377)
#>  alpha2 = -2.958 (95% CrI [-3.223, -2.722], sd = 0.1275)
plot(model)

Grenfell, B. T., and R. M. Anderson. 1985. “The Estimation of Age-Related Rates of Infection from Case Notifications and Serological Data.” The Journal of Hygiene 95 (2): 419–36. https://doi.org/10.1017/s0022172400062859.
Hens, Niel, Ziv Shkedy, Marc Aerts, Christel Faes, Pierre Van Damme, and Philippe Beutels. 2012. Modeling Infectious Disease Parameters Based on Serological and Social Contact Data: A Modern Statistical Perspective. Statistics for Biology and Health. Springer New York. https://doi.org/10.1007/978-1-4614-4072-7.
Muench, Hugo. 1934. “Derivation of Rates from Summation Data by the Catalytic Curve.” Journal of the American Statistical Association 29 (185): 25–38. https://doi.org/10.1080/01621459.1934.10502684.