categoricalCUSUM: CUSUM detector for time-varying categorical time series

In surveillance: Temporal and Spatio-Temporal Modeling and Monitoring of Epidemic Phenomena

CUSUM detector for time-varying categorical time series

Description

Function to process sts object by binomial, beta-binomial or multinomial CUSUM as described by Höhle (2010). Logistic, multinomial logistic, proportional odds or Bradley-Terry regression models are used to specify in-control and out-of-control parameters. The implementation is illustrated in Salmon et al. (2016).

Usage

categoricalCUSUM(stsObj,control = list(range=NULL,h=5,pi0=NULL,
                 pi1=NULL, dfun=NULL, ret=c("cases","value")),...)

Arguments

stsObj	Object of class sts containing the number of counts in each of the k categories of the response variable. Time varying number of counts n_t is found in slot populationFrac.
control	Control object containing several items range Vector of length t_{max} with indices of the observed slot to monitor. h Threshold to use for the monitoring. Once the CUSUM statistics is larger or equal to h we have an alarm. pi0 (k-1) \times t_{max} in-control probability vector for all categories except the reference category. mu1 (k-1) \times t_{max} out-of-control probability vector for all categories except the reference category. dfun The probability mass function (PMF) or density used to compute the likelihood ratios of the CUSUM. In a negative binomial CUSUM this is dnbinom, in a binomial CUSUM dbinom and in a multinomial CUSUM dmultinom. The function must be able to handle the arguments y, size, mu and log. As a consequence, one in the case of, e.g, the beta-binomial distribution has to write a small wrapper function. ret Return the necessary proportion to sound an alarm in the slot upperbound or just the value of the CUSUM statistic. Thus, ret is one of the values in c("cases","value"). Note: For the binomial PMF it is possible to compute this value explicitly, which is much faster than the numeric search otherwise conducted. In case dfun just corresponds to dbinom just set the attribute isBinomialPMF for the dfun object.
...	Additional arguments to send to dfun.

Details

The function allows the monitoring of categorical time series as described by regression models for binomial, beta-binomial or multinomial data. The later includes e.g. multinomial logistic regression models, proportional odds models or Bradley-Terry models for paired comparisons. See the Höhle (2010) reference for further details about the methodology.

Once an alarm is found the CUSUM scheme is reset (to zero) and monitoring continues from there.

Value

An sts object with observed, alarm, etc. slots trimmed to the control$range indices.

Author(s)

M. Höhle

References

Höhle, M. (2010): Online Change-Point Detection in Categorical Time Series. In: T. Kneib and G. Tutz (Eds.), Statistical Modelling and Regression Structures, Physica-Verlag.

Salmon, M., Schumacher, D. and Höhle, M. (2016): Monitoring count time series in R: Aberration detection in public health surveillance. Journal of Statistical Software, 70 (10), 1-35. \Sexpr[results=rd]{tools:::Rd_expr_doi("10.18637/jss.v070.i10")}

See Also

LRCUSUM.runlength

Examples

## IGNORE_RDIFF_BEGIN
have_GAMLSS <- require("gamlss")
## IGNORE_RDIFF_END

if (have_GAMLSS) {
  ###########################################################################
  #Beta-binomial CUSUM for a small example containing the time-varying
  #number of positive test out of a time-varying number of total
  #test.
  #######################################

  #Load meat inspection data
  data("abattoir")

  #Use GAMLSS to fit beta-bin regression model
  phase1 <- 1:(2*52)
  phase2  <- (max(phase1)+1) : nrow(abattoir)

  #Fit beta-binomial model using GAMLSS
  abattoir.df <- as.data.frame(abattoir)

  #Replace the observed and epoch column names to something more convenient
  dict <- c("observed"="y", "epoch"="t", "population"="n")
  replace <- dict[colnames(abattoir.df)]
  colnames(abattoir.df)[!is.na(replace)] <- replace[!is.na(replace)]

  m.bbin <- gamlss( cbind(y,n-y) ~ 1 + t +
                      + sin(2*pi/52*t) + cos(2*pi/52*t) +
                      + sin(4*pi/52*t) + cos(4*pi/52*t), sigma.formula=~1,
                   family=BB(sigma.link="log"),
                   data=abattoir.df[phase1,c("n","y","t")])

  #CUSUM parameters
  R <- 2 #detect a doubling of the odds for a test being positive
  h <- 4 #threshold of the cusum

  #Compute in-control and out of control mean
  pi0 <- predict(m.bbin,newdata=abattoir.df[phase2,c("n","y","t")],type="response")
  pi1 <- plogis(qlogis(pi0)+log(R))
  #Create matrix with in control and out of control proportions.
  #Categories are D=1 and D=0, where the latter is the reference category
  pi0m <- rbind(pi0, 1-pi0)
  pi1m <- rbind(pi1, 1-pi1)


  ######################################################################
  # Use the multinomial surveillance function. To this end it is necessary
  # to create a new abattoir object containing counts and proportion for
  # each of the k=2 categories. For binomial data this appears a bit
  # redundant, but generalizes easier to k>2 categories.
  ######################################################################

  abattoir2 <- sts(epoch=1:nrow(abattoir), start=c(2006,1), freq=52,
    observed=cbind(abattoir@observed, abattoir@populationFrac-abattoir@observed),
    populationFrac=cbind(abattoir@populationFrac,abattoir@populationFrac),
    state=matrix(0,nrow=nrow(abattoir),ncol=2),
    multinomialTS=TRUE)

  ######################################################################
  #Function to use as dfun in the categoricalCUSUM
  #(just a wrapper to the dBB function). Note that from v 3.0-1 the
  #first argument of dBB changed its name from "y" to "x"!
  ######################################################################
  mydBB.cusum <- function(y, mu, sigma, size, log = FALSE) {
    return(dBB(y[1,], mu = mu[1,], sigma = sigma, bd = size, log = log))
  }


  #Create control object for multinom cusum and use the categoricalCUSUM
  #method
  control <- list(range=phase2,h=h,pi0=pi0m, pi1=pi1m, ret="cases",
		   dfun=mydBB.cusum)
  surv <- categoricalCUSUM(abattoir2, control=control,
			   sigma=exp(m.bbin$sigma.coef))

  #Show results
  plot(surv[,1],dx.upperbound=0)
  lines(pi0,col="green")
  lines(pi1,col="red")

  #Index of the alarm
  which.max(alarms(surv[,1]))
}

surveillance documentation built on Oct. 2, 2024, 1:08 a.m.