Biology Reference
In-Depth Information
coefficients),
142
and to allow k to depend on covariates.
143
Negative
binomial regression has now become a more commonly included
component of statistical software packages, and a number of studies
throughout the last decade have applied it to analyze egg count data on
A. lumbricoides
144,145
and other STH infections.
146
e
148
Zero-inflated Poisson
149
e
151
and negative binomial
152
regression tech-
niques were developed primarily for economic applications during the
mid-1980s to early 1990s.
153
More recently they have been adopted by the
parasitological research community, having been applied to spatial risk
models of Schistosoma mansoni infection
154
and S. mansoni
hookworm
co-infections;
155
the comparison of STH infections in ivermectin-na¨ve
and treated communities,
132
and the analysis of density- and female
weight-dependent fecundity in A. lumbricoides.
44
Aside from providing
a generally improved description of egg count data, zero-inflated models
permit modeling of the count and Bernoulli components. This has
been exploited to demonstrate that the preponderance of zeros in
A. lumbricoides egg counts depends on the female worm burden,
presumably arising from parasite density-dependent sensitivity of the
diagnostic test.
44
Modeling zeros in this way may also be useful for
random effects modeling, where diagnostic sensitivity may be heteroge-
neous among communities with different endemicities. A better under-
standing of the components of diagnostic performance of parasitological
assays will become increasingly important as control programs shift their
aim from morbidity control to elimination of infection.
e
Hierarchical, Mixed Effects Models
Hierarchical models, which are also referred to as mixed or random
effects models, are used to analyze non-independent, clustered data that
arise when observations are made from distinct or related units.
156,157
For
example, observations made on the same individual, either at the same or
at different points in time (longitudinal data) will generally be more
similar than observations made from different individuals. Similarly,
observations made from members of the same household, school or
community will generally be more similar than observations made from
different households, schools or communities. The former two examples
(i.e. individuals and households) are well illustrated by predisposition
and household clustering of A. lumbricoides infection, respectively.
106
In statistical parlance, these two phenomena simply represent the
clustering or correlation typically observed among repeated observations
made from on the same distinct units.
Interestingly, predisposition and household clustering have only
recently been explored (by analyzing data from Bangladesh) using
methods which exploit the natural three-level hierarchical structure of
worm counts measured repeatedly from individuals, before and after
