hhh4: Fitting HHH Models with Random Effects and Neighbourhood Structure

Description

Fits an autoregressive Poisson or negative binomial model to a univariate or multivariate time series of counts. The characteristic feature of hhh4 models is the additive decomposition of the conditional mean into epidemic and endemic components (Held et al, 2005). Log-linear predictors of covariates and random intercepts are allowed in all components; see the Details below. A general introduction to the hhh4 modelling approach and its implementation is given in the vignette("hhh4"). Meyer et al (2017, Section 5, available as vignette("hhh4_spacetime")) describe hhh4 models for areal time series of infectious disease counts.

Usage

hhh4(stsObj,
     control = list(
         ar = list(f = ~ -1, offset = 1, lag = 1),
         ne = list(f = ~ -1, offset = 1, lag = 1,
                   weights = neighbourhood(stsObj) == 1,
                   scale = NULL, normalize = FALSE),
         end = list(f = ~ 1, offset = 1),
         family = c("Poisson", "NegBin1", "NegBinM"),
         subset = 2:nrow(stsObj),
         optimizer = list(stop = list(tol=1e-5, niter=100),
                          regression = list(method="nlminb"),
                          variance = list(method="nlminb")),
         verbose = FALSE,
         start = list(fixed=NULL, random=NULL, sd.corr=NULL),
         data = list(t = stsObj@epoch - min(stsObj@epoch)),
         keep.terms = FALSE
     ),
     check.analyticals = FALSE)

Value

hhh4 returns an object of class "hhh4", which is a list containing the following components:

coefficients: named vector with estimated (regression) parameters of the model
se: estimated standard errors (for regression parameters)
cov: covariance matrix (for regression parameters)
Sigma: estimated variance-covariance matrix of random effects
Sigma.orig: estimated variance parameters on internal scale used for optimization
Sigma.cov: inverse of marginal Fisher information (on internal scale), i.e., the asymptotic covariance matrix of Sigma.orig
call: the matched call
dim: vector with number of fixed and random effects in the model
loglikelihood: (penalized) loglikelihood evaluated at the MLE
margll: (approximate) log marginal likelihood should the model contain random effects
convergence: logical. Did optimizer converge?
fitted.values: fitted mean values $\mu_{i,t}$
control: control object of the fit
terms: the terms object used in the fit if keep.terms = TRUE and NULL otherwise
stsObj: the supplied stsObj
lags: named integer vector of length two containing the lags used for the epidemic components "ar" and "ne", respectively. The corresponding lag is NA if the component was not included in the model.
nObs: number of observations used for fitting the model
nTime: number of time points used for fitting the model
nUnit: number of units (e.g. areas) used for fitting the model
runtime: the proc.time-queried time taken to fit the model, i.e., a named numeric vector of length 5 of class "proc_time"

Arguments

stsObj

object of class "sts" containing the (multivariate) count data time series.

control

a list containing the model specification and control arguments:

ar

Model for the autoregressive component given as list with the following components:

f = ~ -1: a formula specifying $\log(\lambda_{it})$

offset = 1

optional multiplicative offset, either 1 or a matrix of the same dimension as observed(stsObj)

lag = 1

a positive integer meaning autoregression on $y_{i,t-lag}$

ne

Model for the neighbour-driven component given as list with the following components:

f = ~ -1: a formula specifying $\log(\phi_{it})$

offset = 1

optional multiplicative offset, either 1 or a matrix of the same dimension as observed(stsObj)

lag = 1

a non-negative integer meaning dependency on $y_{j,t-lag}$

weights = neighbourhood(stsObj) == 1

neighbourhood weights $w_{ji}$. The default corresponds to the original formulation by Held et al (2005), i.e., the spatio-temporal component incorporates an unweighted sum over the lagged cases of the first-order neighbours. See Paul et al (2008) and Meyer and Held (2014) for alternative specifications, e.g., W_powerlaw. Time-varying weights are possible by specifying an array of dim() c(nUnits, nUnits, nTime), where nUnits=ncol(stsObj) and nTime=nrow(stsObj).

scale = NULL

optional matrix of the same dimensions as weights (or a vector of length ncol(stsObj)) to scale the weights to scale * weights.

normalize = FALSE

logical indicating if the (scaled) weights should be normalized such that each row sums to 1.

end

Model for the endemic component given as list with the following components

f = ~ 1: a formula specifying $\log(\nu_{it})$

offset = 1

optional multiplicative offset $e_{it}$, either 1 or a matrix of the same dimension as observed(stsObj)

family

Distributional family -- either "Poisson", or the Negative Binomial distribution. For the latter, the overdispersion parameter can be assumed to be the same for all units ("NegBin1"), to vary freely over all units ("NegBinM"), or to be shared by some units (specified by a factor of length ncol(stsObj) such that its number of levels determines the number of overdispersion parameters). Note that "NegBinM" is equivalent to factor(colnames(stsObj), levels = colnames(stsObj)).

subset

Typically 2:nrow(obs) if model contains autoregression

optimizer

a list of three lists of control arguments.

The "stop" list specifies two criteria for the outer optimization of regression and variance parameters: the relative tolerance for parameter change using the criterion max(abs(x[i+1]-x[i])) / max(abs(x[i])), and the maximum number niter of outer iterations.

Control arguments for the single optimizers are specified in the lists named "regression" and "variance". method="nlminb" is the default optimizer for both (taking advantage of the analytical Fisher information matrices), however, the methods from optim may also be specified (as well as "nlm" but that one is not recommended here). Especially for the variance updates, Nelder-Mead optimization (method="Nelder-Mead") is an attractive alternative. All other elements of these two lists are passed as control arguments to the chosen method, e.g., if method="nlminb", adding iter.max=50 increases the maximum number of inner iterations from 20 (default) to 50. For method="Nelder-Mead", the respective argument is called maxit and defaults to 500.

verbose

non-negative integer (usually in the range 0:3) specifying the amount of tracing information to be output during optimization.

start

a list of initial parameter values replacing initial values set via fe and ri. Since surveillance 1.8-2, named vectors are matched against the coefficient names in the model (where unmatched start values are silently ignored), and need not be complete, e.g., start = list(fixed = c("-log(overdisp)" = 0.5)) (default: 2) for a family = "NegBin1" model. In contrast, an unnamed start vector must specify the full set of parameters as used by the model.

data

a named list of covariates that are to be included as fixed effects (see fe) in any of the 3 component formulae. By default, the time variable t is available and used for seasonal effects created by addSeason2formula. In general, covariates in this list can be either vectors of length nrow(stsObj) interpreted as time-varying but common across all units, or matrices of the same dimension as the disease counts observed(stsObj).

keep.terms

logical indicating if the terms object used in the fit is to be kept as part of the returned object. This is usually not necessary, since the terms object is reconstructed by the terms-method for class "hhh4" if necessary (based on stsObj and control, which are both part of the returned "hhh4" object).

The auxiliary function makeControl might be useful to create such a list of control parameters.

check.analyticals

logical (or a subset of c("numDeriv", "maxLik")), indicating if (how) the implemented analytical score vector and Fisher information matrix should be checked against numerical derivatives at the parameter starting values, using the packages numDeriv and/or maxLik. If activated, hhh4 will return a list containing the analytical and numerical derivatives for comparison (no ML estimation will be performed). This is mainly intended for internal use by the package developers.

Author

Michaela Paul, Sebastian Meyer, Leonhard Held

Details

An endemic-epidemic multivariate time-series model for infectious disease counts $Y_{it}$ from units $i=1,\dots,I$ during periods $t=1,\dots,T$ was proposed by Held et al (2005) and was later extended in a series of papers (Paul et al, 2008; Paul and Held, 2011; Held and Paul, 2012; Meyer and Held, 2014). In its most general formulation, this so-called hhh4 (or HHH or $H^3$ or triple-H) model assumes that, conditional on past observations, $Y_{it}$ has a Poisson or negative binomial distribution with mean $$\mu_{it} = \lambda_{it} y_{i,t-1} + \phi_{it} \sum_{j\neq i} w_{ji} y_{j,t-1} + e_{it} \nu_{it} $$ In the case of a negative binomial model, the conditional variance is $\mu_{it}(1+\psi_i\mu_{it})$ with overdispersion parameters $\psi_i > 0$ (possibly shared across different units, e.g., $\psi_i\equiv\psi$). Univariate time series of counts $Y_t$ are supported as well, in which case hhh4 can be regarded as an extension of glm.nb to account for autoregression. See the Examples below for a comparison of an endemic-only hhh4 model with a corresponding glm.nb.

The three unknown quantities of the mean $\mu_{it}$,

$\lambda_{it}$ in the autoregressive (ar) component,
$\phi_{it}$ in the neighbour-driven (ne) component, and
$\nu_{it}$ in the endemic (end) component,

are log-linear predictors incorporating time-/unit-specific covariates. They may also contain unit-specific random intercepts as proposed by Paul and Held (2011). The endemic mean is usually modelled proportional to a unit-specific offset $e_{it}$ (e.g., population numbers or fractions); it is possible to include such multiplicative offsets in the epidemic components as well. The $w_{ji}$ are transmission weights reflecting the flow of infections from unit $j$ to unit $i$. If weights vary over time (prespecified as a 3-dimensional array $(w_{jit})$), the ne sum in the mean uses $w_{jit} y_{j,t-1}$. In spatial hhh4 applications, the “units” refer to geographical regions and the weights could be derived from movement network data. Alternatively, the weights $w_{ji}$ can be estimated parametrically as a function of adjacency order (Meyer and Held, 2014), see W_powerlaw.

(Penalized) Likelihood inference for such hhh4 models has been established by Paul and Held (2011) with extensions for parametric neighbourhood weights by Meyer and Held (2014). Supplied with the analytical score function and Fisher information, the function hhh4 by default uses the quasi-Newton algorithm available through nlminb to maximize the log-likelihood. Convergence is usually fast even for a large number of parameters. If the model contains random effects, the penalized and marginal log-likelihoods are maximized alternately until convergence.

References

Held, L., Höhle, M. and Hofmann, M. (2005): A statistical framework for the analysis of multivariate infectious disease surveillance counts. Statistical Modelling, 5 (3), 187-199. tools:::Rd_expr_doi("10.1191/1471082X05st098oa")

Paul, M., Held, L. and Toschke, A. M. (2008): Multivariate modelling of infectious disease surveillance data. Statistics in Medicine, 27 (29), 6250-6267. tools:::Rd_expr_doi("10.1002/sim.4177")

Paul, M. and Held, L. (2011): Predictive assessment of a non-linear random effects model for multivariate time series of infectious disease counts. Statistics in Medicine, 30 (10), 1118-1136. tools:::Rd_expr_doi("10.1002/sim.4177")

Held, L. and Paul, M. (2012): Modeling seasonality in space-time infectious disease surveillance data. Biometrical Journal, 54 (6), 824-843. tools:::Rd_expr_doi("10.1002/bimj.201200037")

Meyer, S. and Held, L. (2014): Power-law models for infectious disease spread. The Annals of Applied Statistics, 8 (3), 1612-1639. tools:::Rd_expr_doi("10.1214/14-AOAS743")

Meyer, S., Held, L. and Höhle, M. (2017): Spatio-temporal analysis of epidemic phenomena using the R package surveillance. Journal of Statistical Software, 77 (11), 1-55. tools:::Rd_expr_doi("10.18637/jss.v077.i11")

Examples

Run this code

######################
## Univariate examples
######################

### weekly counts of salmonella agona cases, UK, 1990-1995

data("salmonella.agona")
## convert old "disProg" to new "sts" data class
salmonella <- disProg2sts(salmonella.agona)
salmonella
plot(salmonella)

## generate formula for an (endemic) time trend and seasonality
f.end <- addSeason2formula(f = ~1 + t, S = 1, period = 52)
f.end
## specify a simple autoregressive negative binomial model
model1 <- list(ar = list(f = ~1), end = list(f = f.end), family = "NegBin1")
## fit this model to the data
res <- hhh4(salmonella, model1)
## summarize the model fit
summary(res, idx2Exp=1, amplitudeShift=TRUE, maxEV=TRUE)
plot(res)
plot(res, type = "season", components = "end")


### weekly counts of meningococcal infections, Germany, 2001-2006

data("influMen")
fluMen <- disProg2sts(influMen)
meningo <- fluMen[, "meningococcus"]
meningo
plot(meningo)

## again a simple autoregressive NegBin model with endemic seasonality
meningoFit <- hhh4(stsObj = meningo, control = list(
    ar = list(f = ~1),
    end = list(f = addSeason2formula(f = ~1, S = 1, period = 52)),
    family = "NegBin1"
))

summary(meningoFit, idx2Exp=TRUE, amplitudeShift=TRUE, maxEV=TRUE)
plot(meningoFit)
plot(meningoFit, type = "season", components = "end")


########################
## Multivariate examples
########################

### bivariate analysis of influenza and meningococcal infections
### (see Paul et al, 2008)

plot(fluMen, same.scale = FALSE)
     
## Fit a negative binomial model with
## - autoregressive component: disease-specific intercepts
## - neighbour-driven component: only transmission from flu to men
## - endemic component: S=3 and S=1 sine/cosine pairs for flu and men, respectively
## - disease-specific overdispersion

WfluMen <- neighbourhood(fluMen)
WfluMen["meningococcus","influenza"] <- 0
WfluMen
f.end_fluMen <- addSeason2formula(f = ~ -1 + fe(1, which = c(TRUE, TRUE)),
                                  S = c(3, 1), period = 52)
f.end_fluMen
fluMenFit <- hhh4(fluMen, control = list(
    ar = list(f = ~ -1 + fe(1, unitSpecific = TRUE)),
    ne = list(f = ~ 1, weights = WfluMen),
    end = list(f = f.end_fluMen),
    family = "NegBinM"))
summary(fluMenFit, idx2Exp=1:3)
plot(fluMenFit, type = "season", components = "end", unit = 1)
plot(fluMenFit, type = "season", components = "end", unit = 2)
# \dontshow{
    ## regression test for amplitude/shift transformation of sine-cosine pairs
    ## coefficients were wrongly matched in surveillance < 1.18.0
    stopifnot(coef(fluMenFit, amplitudeShift = TRUE)["end.A(2 * pi * t/52).meningococcus"] == sqrt(sum(coef(fluMenFit)[paste0("end.", c("sin","cos"), "(2 * pi * t/52).meningococcus")]^2)))
# }


### weekly counts of measles, Weser-Ems region of Lower Saxony, Germany

data("measlesWeserEms")
measlesWeserEms
plot(measlesWeserEms)  # note the two districts with zero cases

## we could fit the same simple model as for the salmonella cases above
model1 <- list(
    ar = list(f = ~1),
    end = list(f = addSeason2formula(~1 + t, period = 52)),
    family = "NegBin1"
)
measlesFit <- hhh4(measlesWeserEms, model1)
summary(measlesFit, idx2Exp=TRUE, amplitudeShift=TRUE, maxEV=TRUE)

## but we should probably at least use a population offset in the endemic
## component to reflect heterogeneous incidence levels of the districts,
## and account for spatial dependence (here just using first-order adjacency)
measlesFit2 <- update(measlesFit,
    end = list(offset = population(measlesWeserEms)),
    ne = list(f = ~1, weights = neighbourhood(measlesWeserEms) == 1))
summary(measlesFit2, idx2Exp=TRUE, amplitudeShift=TRUE, maxEV=TRUE)
plot(measlesFit2, units = NULL, hide0s = TRUE)

## 'measlesFit2' corresponds to the 'measlesFit_basic' model in
## vignette("hhh4_spacetime"). See there for further analyses,
## including vaccination coverage as a covariate,
## spatial power-law weights, and random intercepts.


if (FALSE) {
### last but not least, a more sophisticated (and time-consuming)
### analysis of weekly counts of influenza from 140 districts in
### Southern Germany (originally analysed by Paul and Held, 2011,
### and revisited by Held and Paul, 2012, and Meyer and Held, 2014)

data("fluBYBW")
plot(fluBYBW, type = observed ~ time)
plot(fluBYBW, type = observed ~ unit,
     ## mean yearly incidence per 100.000 inhabitants (8 years)
     population = fluBYBW@map$X31_12_01 / 100000 * 8)

## For the full set of models for data("fluBYBW") as analysed by
## Paul and Held (2011), including predictive model assessement
## using proper scoring rules, see the (computer-intensive)
## demo("fluBYBW") script:
demoscript <- system.file(file.path("demo", "fluBYBW.R"),
                          package = "surveillance")
demoscript
#file.show(demoscript)

## Here we fit the improved power-law model of Meyer and Held (2014)
## - autoregressive component: random intercepts + S = 1 sine/cosine pair
## - neighbour-driven component: random intercepts + S = 1 sine/cosine pair
##   + population gravity with normalized power-law weights
## - endemic component: random intercepts + trend + S = 3 sine/cosine pairs
## - random intercepts are iid but correlated between components
f.S1 <- addSeason2formula(
    ~-1 + ri(type="iid", corr="all"),
    S = 1, period = 52)
f.end.S3 <- addSeason2formula(
    ~-1 + ri(type="iid", corr="all") + I((t-208)/100),
    S = 3, period = 52)

## for power-law weights, we need adjaceny orders, which can be
## computed from the binary adjacency indicator matrix
nbOrder1 <- neighbourhood(fluBYBW)
neighbourhood(fluBYBW) <- nbOrder(nbOrder1, 15)

## full model specification
fluModel <- list(
    ar = list(f = f.S1),
    ne = list(f = update.formula(f.S1, ~ . + log(pop)),
              weights = W_powerlaw(maxlag=max(neighbourhood(fluBYBW)),
                                   normalize = TRUE, log = TRUE)),
    end = list(f = f.end.S3, offset = population(fluBYBW)),
    family = "NegBin1", data = list(pop = population(fluBYBW)),
    optimizer = list(variance = list(method = "Nelder-Mead")),
    verbose = TRUE)

## CAVE: random effects considerably increase the runtime of model estimation
## (It is usually advantageous to first fit a model with simple intercepts
## to obtain reasonable start values for the other parameters.)
set.seed(1)  # because random intercepts are initialized randomly
fluFit <- hhh4(fluBYBW, fluModel)

summary(fluFit, idx2Exp = TRUE, amplitudeShift = TRUE)

plot(fluFit, type = "fitted", total = TRUE)

plot(fluFit, type = "season")
range(plot(fluFit, type = "maxEV"))

plot(fluFit, type = "maps", prop = TRUE)

gridExtra::grid.arrange(
    grobs = lapply(c("ar", "ne", "end"), function (comp)
        plot(fluFit, type = "ri", component = comp, main = comp,
             exp = TRUE, sub = "multiplicative effect")),
    nrow = 1, ncol = 3)

plot(fluFit, type = "neweights", xlab = "adjacency order")
}


########################################################################
## An endemic-only "hhh4" model can also be estimated using MASS::glm.nb
########################################################################

## weekly counts of measles, Weser-Ems region of Lower Saxony, Germany
data("measlesWeserEms")

## fit an endemic-only "hhh4" model
## with time covariates and a district-specific offset
hhh4fit <- hhh4(measlesWeserEms, control = list(
    end = list(f = addSeason2formula(~1 + t, period = measlesWeserEms@freq),
               offset = population(measlesWeserEms)),
    ar = list(f = ~-1), ne = list(f = ~-1), family = "NegBin1",
    subset = 1:nrow(measlesWeserEms)
))
summary(hhh4fit)

## fit the same model using MASS::glm.nb
measlesWeserEmsData <- as.data.frame(measlesWeserEms, tidy = TRUE)
measlesWeserEmsData$t <- c(hhh4fit$control$data$t)
glmnbfit <- MASS::glm.nb(
    update(formula(hhh4fit)$end, observed ~ . + offset(log(population))),
    data = measlesWeserEmsData
)
summary(glmnbfit)

## Note that the overdispersion parameter is parametrized inversely.
## The likelihood and point estimates are all the same.
## However, the variance estimates are different: in glm.nb, the parameters
## are estimated conditional on the overdispersion theta.

# \dontshow{
stopifnot(
    all.equal(logLik(hhh4fit), logLik(glmnbfit)),
    all.equal(1/coef(hhh4fit)[["overdisp"]], glmnbfit$theta, tolerance = 1e-6),
    all.equal(coef(hhh4fit)[1:4], coef(glmnbfit),
              tolerance = 1e-6, check.attributes = FALSE),
    all.equal(c(residuals(hhh4fit)), residuals(glmnbfit),
              tolerance = 1e-6, check.attributes = FALSE)
)
# }

Run the code above in your browser using DataLab