Kxy 067
Kxy 067
Kxy 067
499–517 DC
doi:10.1093/biostatistics/kxy067
Advance Access publication on December 26, 2018
S UMMARY
Microorganisms play critical roles in human health and disease. They live in diverse communities in
which they interact synergistically or antagonistically. Thus for estimating microbial associations with
clinical covariates, such as treatment effects, joint (multivariate) statistical models are preferred. Multi-
variate models allow one to estimate and exploit complex interdependencies among multiple taxa, yielding
more powerful tests of exposure or treatment effects than application of taxon-specific univariate anal-
yses. Analysis of microbial count data also requires special attention because data commonly exhibit
zero inflation, i.e., more zeros than expected from a standard count distribution. To meet these needs, we
developed a Bayesian variable selection model for multivariate count data with excess zeros that incorpo-
rates information on the covariance structure of the outcomes (counts for multiple taxa), while estimating
associations with the mean levels of these outcomes. Though there has been much work on zero-inflated
models for longitudinal data, little attention has been given to high-dimensional multivariate zero-inflated
data modeled via a general correlation structure. Through simulation, we compared performance of the
© The Author 2018. Published by Oxford University Press. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com.
500 K. H. LEE AND OTHERS
proposed method to that of existing univariate approaches, for both the binary (“excess zero”) and count
parts of the model. When outcomes were correlated the proposed variable selection method maintained type
I error while boosting the ability to identify true associations in the binary component of the model. For the
count part of the model, in some scenarios the univariate method had higher power than the multivariate
approach. This higher power was at a cost of a highly inflated false discovery rate not observed with
the proposed multivariate method. We applied the approach to oral microbiome data from the Pediatric
HIV/AIDS Cohort Oral Health Study and identified five (of 44) species associated with HIV infection.
Keywords: Bayesian variable selection; Markov chain Monte Carlo; Microbiome sequencing data; Multivariate
analysis; Zero-inflated models.
1. I NTRODUCTION
The human microbiome plays a critical role in maintaining health and causing both acute and chronic
disease. Microbes live in communities in which multiple species establish synergistic and antagonistic
relationships (Pflughoeft and Versalovic, 2012). These interactions allow some species to thrive and
keep others in check. The complex biological dependencies among taxa demand statistical methods that
account for and exploit this interdependence. There are valid and powerful methods for jointly analyzing
microbiome sequence data as predictors of health outcomes, but there are fewer methodological options for
analyzing microbiome community data as a set of joint endpoints. We specifically address three challenges
that commonly arise in analysis of microbiome sequencing data as responses (dependent variables): excess
zeros, interdependence of the endpoints, and the need for outcome-specific covariate selection.
First, in most human microbiome studies, a large proportion of microbial taxa is absent in the majority
of subjects, leading to many more zero counts for each taxon than expected on the basis of a Poisson,
negative binomial, or Dirichlet-multinomial distribution (e.g., see Supplementary Material A available at
Biostatistics online.) (Chen and Li, 2013). Application of a conventional linear model that uses untrans-
formed or logarithmic-transformed counts is inappropriate for zero-inflated data (Loeys and others, 2012;
Xu and others, 2015). An intuitive approach to analyzing zero-inflated count data is to view the data as
arising from an underlying zero-inflated distribution, which is a mixture of a point mass at zero and a
count distribution, such as Poisson (Lambert, 1992).
Second, as mentioned above, microbiome sequencing data are typically multivariate (joint response)
count data sampled from communities of interdependent species. Naïve application of a univariate, taxon-
by-taxon approach implicitly assumes that counts of each taxon are uncorrelated. Although one could
control for the type I error in this approach, this generally results in loss of power (Breiman and Friedman,
1997; La Rosa and others, 2012). Multivariate non-parametric methods that compare bacterial community
composition between two groups are available (Mantel, 1967; Mantel and Valand, 1970; Anderson, 2001);
these are generally less powerful than regression methods and often do not quantify the magnitude of
group differences. One approach for joint modeling of multivariate microbial sequence count data is
Dirichlet-multinomial regression (Holmes and others, 2012; La Rosa and others, 2012; Chen and Li,
2013; Wadsworth and others, 2017). However, the Dirichlet-multinomial model imposes restrictions
that may misrepresent features of multivariate taxa count data distributions. For example, despite that
relationships among microbial species can be either positively or negatively correlated, the dependence
between Dirichlet variates are always negatively correlated (Aitchison and Ho, 1989; Li, 2015).
Multivariate zero-inflated regression models can address both excess zeros as well as interdependent
responses. Such methods that have been developed to date have been scaled to model only a small
number of interdependent count endpoints, which include bivariate (Arab and others, 2012; Fox, 2013)
and trivariate (Li and others, 1999) zero-inflated Poisson models. In some cases, existing methods have
incorporated a restrictive covariance structure among outcomes, which may not always be appropriate.
Specific examples of such restrictions include zero-inflated models for longitudinal data only with variance
Bayesian variable selection for multivariate zero-inflated count data 501
components (with no covariance components) (Lee and others, 2006; Hall, 2000; Long and others, 2015),
models with dependence structures specific to spatial-temporal data (Earnest and others, 2007; Fernandes
and others, 2009; Wang and others, 2015a), and models including latent factors that can induce only
positive correlations among outcomes (Neelon and Chung, 2017).
A third impediment to developing and applying a multivariate analysis technique to microbiome data
is that due to having more than one endpoint, there is a large number of potential covariate-endpoint
associations to be modeled. It is well recognized that variable selection helps improve prediction accuracy
and reduce the cost of measurement and storage of future data. The need for variable selection techniques is
well appreciated for high-dimensional covariate data and may be less well known in the context of multiple
outcomes. Although there exist variable selection methods for multivariate normal (MVN) (Brown and
others, 1998; Lee and others, 2017) or multinomial responses (Wadsworth and others, 2017), we know
of no such technique applied to methods for multivariate zero-inflated outcomes.
We have developed multivariate zero-inflated regression models by relaxing requirements regarding the
covariance structure and incorporating a Bayesian variable selection approach. The proposed methods can
be used to identify zero-inflated count outcomes associated with a set of covariates while accounting for
the covariance structure of the outcomes. Since it is implausible that all outcomes are relevant to the same
subset of covariates, we enable the proposed model to perform outcome-specific variable selection, i.e., to
identify exposures or treatments associated with particular outcomes, in this case microbial taxa. Spike-
and-slab approaches have been widely used for Bayesian variable selection (George and McCulloch, 1993,
1997), including for multivariate linear regression problems (Brown and others, 1998; Lee and others,
2017). In this work, we extend the spike-and-slab approach to the context of multivariate zero-inflated data.
We use the newly developed model to analyze data from the Pediatric HIV/AIDS Cohort Oral Health
Study (PHACS). PHACS is an ongoing prospective cohort study at 15 US clinical sites, designed to assess
the health effects of HIV infection and antiretroviral therapy on youth perinatally infected with HIV
(PHIV) compared with exposed but uninfected (PHEU) youth (Alperen and others, 2014; Tassiopoulos
and others, 2016; Starr and others, 2018). The data analyzed were from a cross-sectional study focused on
oral health and the oral microbiome (Moscicki and others, 2016; Ryder and others, 2017). All participants
were exposed to HIV perinatally, the period when they became HIV infected if at all. Emerging from the
womb, it is likely that they began acquiring their oral microbiota at birth and should have had oral
microbiota similar to those of adults by 3 years of age (Mueller and others, 2015; Perez-Muñoz and
others, 2017). Thus, if there is a causal association, the oral bacterial sequences we measured at 10–22
years of age more likely resulted from perinatal HIV infection rather than the reverse. This is why we
treat taxa’s counts as endpoints and HIV as the exposure. The goals of this analysis are (i) to identify taxa
associated with HIV infection; and (ii) to estimate and test the association of HIV infection with counts
of the identified taxa.
The remainder of this article is organized as follows. Section 2 describes the proposed Bayesian
framework, including model formulation and specification of prior distributions. Section 3 describes
results from simulation studies conducted to compare the operating performance of the proposed variable
selection approach versus an existing univariate method. Section 4 describes results from the PHACS data
analysis. In Section 5, we further discuss the method and results.
2. M ETHODS
In this section, we describe the proposed multivariate zero-inflated model, present prior distributions for
model parameters and the variable selection strategy, discuss interpretation of the regression parameters,
and summarize the computational scheme and implementation (see Supplementary Material B available
at Biostatistics online. for a summary of model parameters and C for implementation details).
502 K. H. LEE AND OTHERS
where Uij is an unobservable indicator for the excess zeros for taxon j in subject i, and λij is the mean
of the Poisson distribution. The model implies that some zeros occur through a Poisson process whereas
others represent the impossibility for a given taxon to be observed in some subjects. In practice, regression
analysis based on the MZIP model proceeds by placing structure on λij and P[Uij = 1], specifically as a
function of the covariates and random effects. Toward this, let x i and z i be a vector of px and pz covariates
for the ith subject that will be considered in the model for λij and P[Uij = 1], respectively. With this
formulation it is not necessary for the presence and the count of a taxon to depend on the same set of
covariates.
For the count (Poisson) model part, we consider the following general modeling specification:
log(λij ) = β0j + x
i β j + log(ξi ) + Vij , (2.2)
where β 0 =(β01 ,…,β0q ) are the outcome-specific intercepts and β j =(βj1 ,…,βjpx ) are the outcome-specific
vectors of fixed-effect regression parameters. The random effects V i = (Vi1 , . . . , Viq ) characterize the
unobserved characteristics that are associated with the mean count for taxon j in subject i and account for
within-subject correlations. The term log(ξi ) is included as an offset variable for settings in which one is
interested in the incidence density λij /ξi . For application to genetic sequence counts, setting ξi to the total
number of sequencing reads accounts for individual variation in sequencing depth.
To account for the dependency structure in the binary part of the model, we adopt a multivariate probit
model (Ashford and Sowden, 1970). Letting Uij =I (wij ≥ 0), with indicator function I (·) , we consider a
MVN distribution for the latent variable wi =(wi1 ,…,wiq ) , with location vector α 0 + A z i and variance-
covariance matrix R. Here, α 0 =(α01 ,…,α0q ) are the intercepts and A is the pz × q coefficient matrix whose
columns are α j =(αj1 , …,αjpz ) , j=1,· · · ,q. Then the probability density function of Uij is given by
R is restricted to be a correlation matrix to ensure identifiability of all model parameters (Chib and
Greenberg, 1998; Liu, 2001). R measures the dependence between Uij and Uij by using the correlations
among the elements of the vector wi . Let Y , w, V , and φ denote the collections of Y i , wi , V i , and φi ,
respectively, across all subjects. Let B be the px × q coefficient matrix whose columnsβ j =(βj1 , …,βjpx ) ,
j=1,· · · ,q. Combining (2.1), (2.2), and (2.3), the augmented data likelihood function, as a function of the
unknown parameters ={α 0 , A, β 0 , B, V , R}, is:
n
q
exp yij log(λij ) − λij
L(Y , w|)
= I (wij ≥ 0) + I (wij < 0)I (yij = 0)
i=1 j=1
yij !
n
1 1
× (2π)−q/2 |R|− 2 exp − (wi − α 0 − A z i ) R−1 (wi − α 0 − A z i ) . (2.4)
i=1
2
Bayesian variable selection for multivariate zero-inflated count data 503
where (·) is the cumulative distribution function of the standard normal distribution. For models with
log-links and normally distributed random effects such as the one we propose, it is straightforward to
marginalize the conditional expectation over the random effects distribution (Long and others, 2015).
Under a model with x i = z i , the ratio of means for a one-unit increase in covariate k is given by
E[Yij |xik = x + 1, x i,(−k) ] E E[Yij |xik = x + 1, x i,(−k) , Vij ]
IDR(x i,(−k) ) = =
E[Yij |xik = x, x i,(−k) ] E E[Yij |xik = x, x i,(−k) , Vij ]
504 K. H. LEE AND OTHERS
βjk
α0j + x
i,(−k) α j,(−k) + (x + 1)αj,k
=e
, (2.6)
α0j + x i,(−k) α j,(−k) + xαj,k
where x i,(−k) and α j,(−k) are the vectors x i and α j , respectively, with the k-th element removed. Although we
obtain IDR(x i,(−k) ) by marginalizing E[Yij |Vij , Uij = 1] over the latent mixture distribution and the distri-
bution of random effect Vij , the quantity still depends on the values of other covariates x i,(−k) , which can be
addressed several ways (Preisser and others, 2012). For continuous covariates, one could obtain a covariate-
adjusted IDR(x̄ (−k) ) by either (i) inserting mean values of covariates x i,(−k) , or (ii) assuming specific covari-
ate distributions and marginalizing the quantity over these distributions. For discrete covariates, one could
(iii) empirically marginalize the IDR(x i,(−k) ) over the observed distribution of covariates, or (iv) present
multiple different values for the IDR(x i,(−k) ), one for each category defined by unique covariate profiles.
As has been well described, estimating the variance of measures such as the IDR(x i,(−k) ) in a frequentist
framework would require an extra statistical technique such as bootstrap resampling (Albert and others,
2014). An advantage of the Bayesian paradigm is that estimation of uncertainty for IDR(x i,(−k) ) follows
directly from the variance of its posterior distribution, estimated by evaluating its expression at each scan
of the Markov chain Monte Carlo (MCMC) scheme.
3. S IMULATION STUDIES
We evaluated the performance of the proposed method on simulated data. We generated data sets under
six scenarios with varying outcome correlation structures and association patterns between a covariate
and the vector of outcomes in the two model parts.
association between the covariate and outcomes in the two model parts by setting B = A =[0.05, 0.10,
0.15, 0.20, 0.25, −0.05, −0.10, −0.15, −0.20, −0.25, 0 10 ]. In Scenario IV, the covariate was associated
with outcomes in only one of the two model parts: B =[0.05, 0.10, 0.15, 0.20, 0.25, 0 15 ] and A =[05 , 0.05,
0.10, 0.15, 0.20, 0.25, 010 ]. We considered the null case in Scenario V by setting all elements of B and A
to zero. We set each variance-covariance matrix
V in the count part of the model and R in the binary part
of the model to a correlation matrix with an exchangeable structure with correlation c1 within the block
of the first ten outcomes, an exchangeable structure with correlation c2 within the block of the second 10
outcomes, and a common cross-block correlation of c3 for pairs of outcomes from different blocks. In
Scenarios I, IV, and V, the outcomes associated with the covariate were highly correlated and outcomes
unassociated with the covariate only moderately correlated, (c1 , c2 , c3 )=(0.70, 0.30, 0.20). In Scenario
II, the outcomes associated with the covariate were moderately correlated and the remaining outcomes
weakly correlated, (c1 , c2 , c3 )=(0.40, 0.05, 0.10). In Scenario III, the outcomes associated with the covariate
were weakly correlated and those unassociated with the covariate highly correlated, (c1 , c2 , c3 )=(0.20,
0.70, 0.30). In Scenario VI, each outcome is assumed to follow a univariate zero-inflated Poisson (UZIP)
distribution, indicating (c1 , c2 , c3 )=(0, 0, 0) (independence), and diag(
V )=diag(R)=0 (no overdispersion).
3.3. Results
We focus the presentation of results in this section on the MZIP model with ωβ,1 =ωα,1 =0.1. This is to
demonstrate the improvement gained by the proposed multivariate approach over an analogous univariate
506 K. H. LEE AND OTHERS
Table 1. Four operating characteristics∗ (%) and the number of outcomes selected to be associated with
the covariate (qsel ) for the univariate zero-inflated Poisson (UZIP)† and the proposed multivariate zero-
inflated Poisson (MZIP)‡ models across six simulation scenarios described in Section 3.1. The number of
outcomes generated (q) and the number of outcomes that are truly associated with the covariate (τ ) for
each scenario are presented under the corresponding scenario number
UZIP MZIP
(ωα = ωβ = 0.1) (ωα = ωβ = 0.5)
Scenario Binary Count Binary Count Binary Count
(q, τ ) Mean (SD) Mean (SD) Mean (SD) Mean (SD) Mean (SD) Mean (SD)
TPR 54.3 (9.7) 99.1 (2.9) 67.4 (8.4) 82.5 (5.5) 78.4 (8.1) 87.4 (5.5)
FPR 0.1 (1.0) 86.0 (11.1) 0.3 (1.8) 0.8 (2.8) 4.6 (6.5) 3.1 (5.5)
I PPV 99.8 (1.7) 53.7 (3.4) 99.6 (2.4) 99.1 (3.0) 95.0 (6.8) 96.9 (5.3)
(20, 10) NPV 68.9 (4.7) 94.4 (17.0) 75.6 (4.9) 85.2 (4.2) 81.9 (5.6) 88.7 (4.4)
qsel 5.4 (1.0) 18.5 (1.2) 6.8 (0.8) 8.3 (0.6) 8.3 (1.1) 9.1 (0.8)
TPR 53.7 (9.9) 99.3 (2.6) 60.8 (9.6) 75.2 (8.0) 73.0 (8.8) 84.1 (7.0)
FPR 0.4 (1.9) 87.1 (10.0) 0.5 (2.1) 0.6 (2.4) 4.0 (5.8) 2.7 (5.1)
II PPV 99.5 (2.8) 53.4 (2.9) 99.4 (2.8) 99.3 (2.8) 95.3 (6.7) 97.2 (5.4)
(20, 10) NPV 68.6 (4.6) 96.4 (14.2) 72.1 (4.9) 80.4 (5.3) 78.4 (5.6) 86.2 (5.5)
qsel 5.4 (1.0) 18.6 (1.1) 6.1 (1.0) 7.6 (0.8) 7.7 (1.1) 8.7 (0.8)
TPR 56.0 (10.2) 99.1 (2.8) 60.0 (10.0) 73.5 (8.3) 74.3 (9.2) 82.6 (7.9)
FPR 0.2 (1.5) 88.7 (11.1) 0.2 (1.5) 0.1 (1.0) 3.9 (5.8) 1.8 (4.1)
III PPV 99.6 (2.5) 53.0 (3.3) 99.6 (2.4) 99.9 (1.2) 95.2 (7.1) 98.1 (4.3)
(20, 10) NPV 69.8 (5.0) 92.3 (23.2) 71.7 (5.1) 79.4 (5.2) 79.2 (6.3) 85.4 (5.9)
qsel 5.6 (1.0) 18.8 (1.2) 6.0 (1.0) 7.4 (0.8) 7.8 (0.9) 8.4 (1.0)
TPR 56.8 (16.3) 98.8 (4.8) 63.0 (17.4) 82.6 (11.2) 79.0 (15.1) 88.6 (10.7)
FPR 0.1 (0.7) 86.8 (8.6) 0.4 (1.6) 0.3 (1.3) 3.3 (4.6) 2.5 (4.2)
IV PPV 99.8 (2.0) 27.6 (2.5) 98.5 (6.0) 99.3 (3.6) 90.3 (12.8) 93.3 (10.8)
(20, 5) NPV 87.6 (4.2) 95.9 (17.1) 89.2 (4.6) 94.6 (3.4) 93.4 (4.5) 96.3 (3.4)
qsel 2.9 (0.8) 18.0 (1.3) 3.2 (0.9) 4.2 (0.6) 4.5 (1.0) 4.8 (0.8)
TPR — — — — — — — — — — — —
FPR 0.1 (0.5) 86.8 (8.5) 0.1 (0.7) 0.1 (0.7) 1.9 (3.4) 1.1 (2.3)
V§ PPV — — — — — — — — — — — —
(20, 0) NPV 100.0 (0.0) 100.0 (0.0) 100.0 (0.0) 100.0 (0.0) 100.0 (0.0) 100.0 (0.0)
qsel 0.0 (0.1) 17.4 (1.7) 0.0 (0.1) 0.0 (0.1) 0.4 (0.7) 0.2 (0.5)
TPR 58.1 (10.7) 100.0 (0.0) 57.7 (10.8) 100.0 (0.0) 72.2 (9.6) 100.0 (0.0)
FPR 0.0 (0.0) 0.0 (0.0) 0.1 (1.2) 0.0 (0.0) 4.7 (6.3) 0.1 (1.1)
VI PPV 100.0 (0.0) 100.0 (0.0) 99.8 (1.9) 100.0 (0.0) 94.4 (7.3) 99.9 (1.0)
(20, 10) NPV 70.9 (5.5) 100.0 (0.0) 70.6 (5.4) 100.0 (0.0) 77.8 (6.1) 100.0 (0.0)
qsel 5.8 (1.1) 10.0 (0.0) 5.8 (1.1) 10.0 (0.0) 7.7 (1.1) 10.0 (0.1)
Note: Throughout the mean and standard deviation (SD) values are based on results from 100 simulated datasets.
∗ TPR=TP/τ , FPR=FP/(20 − τ ), PPV=TP/(TP+FP), NPV=(20 − τ − FP)/(20 − TP+FP), where TP is the number of outcomes
associated with the covariate and selected into the model, FP is the number of outcomes unrelated to the covariate and mistakenly
selected into the model.
† For variable selection in UZIP analyses, we used 95% confidence intervals with a false discovery rate controlling procedure.
‡ For variable selection in MZIP models, we applied a marginal posterior probability cutoff of 0.5 for both γ and δ. The hyperpa-
method, while implementing the fairest comparison to the univariate method. When the values of the
overall prior inclusion probabilities (ωβ,1 and ωα,1 ) increased from 0.1 to 0.5, the MZIP model tended to
select one to two more variables on average, yielding higher TPR and NPV but also a bit higher FPR and
lower PPV in both parts of the model (Table 1). When the outcome variables were uncorrelated (Scenario
VI), the variable selection capability for the MZIP model with ωβ,1 =ωα,1 =0.1 was almost the same as that
of UZIP model.
Across scenarios in which the outcomes were correlated (I–III), the binary part of the MZIP model was
more sensitive than in the UZIP approach (Table 1), for example, with TPR = 61% versus 54% in Scenario
II, respectively. The TPR in UZIP models was insensitive to the strength of correlation among outcomes,
whereas the MZIP TPR increased to 67% in Scenario I, in which there was a stronger correlation among
outcomes associated with the covariate. Both the UZIP and the MZIP methods successfully identified
the covariate associations for the four outcomes with the largest effect sizes (|αj,1 | ≥ 0.20; Table 2).
However, the MZIP model performed much better at detecting smaller-magnitude associations: when
|αj,1 | = 0.10 (outcomes 2 and 7), associations were correctly included in 40% and 20% of MZIP and
UZIP models, respectively; the corresponding inclusion rates were 95% and 60% when |αj,1 | = 0.15
(outcomes 3 and 8).
The multivariate approach yielded much more substantial improvement for the count part of model
when outcomes were correlated (Table 1). Even controlling the FDR, the univariate approach generally
exhibited inflated type I error, as high as 86% across Scenarios I–V. In contrast, across all scenarios
the MZIP model had a low probability of false discovery (FPR < 1%) while also exhibiting high TPR
that ranged from 73% to 83% in Scenarios I–IV. The relatively poor performance of the UZIP method
is not due to bias, since the estimated association for outcomes unassociated with the covariate (βj,1 ,
j=11,…,20) were very close to zero (Table 2). However, the medians of the asymptotic standard errors
(SE) for β̂j,1 were 20 times smaller than the empirical standard deviations of β̂j,1 (rang between 0.036 and
0.054) (Table 2), was also observed in Scenarios II–V (Supplementary Material D available at Biostatis-
tics online). Thus, the univariate approach appears to perform poorly in estimation of the SE for count
model parameters when outcomes are correlated. Consequently, the estimated confidence intervals are
too narrow.
In the null case (Scenario V), both approaches successfully excluded the covariate for all outcomes for
the binary part of the model even when outcomes were strongly correlated; the UZIP model exhibited a
high false discovery rate (87%) for the count part of the model.
We ran extensive additional simulations to explore other factors (detailed in Table E.1 and E.2 in
Supplementary Material E available at Biostatistics online), including a larger number of outcomes (q =
50), a lower signal density (4∼5%), a smaller sample size (n = 150) and negative correlations. Briefly, the
results were similar to those described above, with the MZIP performing much better for variable selection
and the UZIP exhibiting inflated type I error. We also compared the proposed MZIP to a univariate non-
parametric method, the Wilcoxon rank sum test. Although type I error was well controlled in the univariate
non-parametric method, substantially higher power was achieved by the MZIP.
To summarize, compared with the univariate approach, the proposed multivariate method improved
upon the UZIP’s performance for the binary part of the model by maintaining type I error while boosting
the ability to identify true associations under the simulated settings. For the count part of the model,
there were some scenarios in which the power of UZIP was higher than with the MZIP approach. This
higher power of UZIP was at a cost of a highly inflated false discovery rate, whereas the MZIP FPR was
< 1%. Performance of the MZIP model was enhanced by increasing the prior inclusion probabilities,
ωβ,k = ωα,l = 0.5. The TPR then exceeded 80% in all non-null scenarios and FPR remained < 4% across
all scenarios.
508 K. H. LEE AND OTHERS
Table 2. Estimated regression parameters and inclusion probabilities for the univariate zero-inflated
Poisson (UZIP)† ,∗ and the proposed multivariate zero-inflated Poisson (MZIP)‡ models for Scenario
I
UZIP MZIP†
Binary Count Binary Count
True αj,1 βj,1 αj,1 |δj,1 = 1 δj,1 βj,1 |γj,1 = 1 γj,1
j αj,1 , βj,1 Est (SE) mα,j Est (SE) mβ,j PM (SD) PM PM (SD) PM
1 0.05 0.05 (0.04) 0.02 0.042 (0.002) 0.95 0.06 (0.04) 0.04 0.03 (0.02) 0.03
2 0.10 0.11 (0.04) 0.14 0.088 (0.002) 1.00 0.11 (0.04) 0.43 0.09 (0.02) 1.00
3 0.15 0.15 (0.05) 0.63 0.142 (0.002) 1.00 0.15 (0.04) 0.95 0.15 (0.02) 1.00
4 0.20 0.20 (0.05) 0.91 0.191 (0.002) 1.00 0.20 (0.04) 1.00 0.19 (0.02) 1.00
5 0.25 0.25 (0.05) 0.99 0.233 (0.002) 1.00 0.24 (0.04) 1.00 0.24 (0.02) 1.00
6 −0.05 −0.04 (0.04) 0.01 −0.046 (0.002) 0.97 −0.06 (0.04) 0.04 −0.06 (0.02) 0.23
7 −0.10 −0.11 (0.04) 0.25 −0.104 (0.002) 0.99 −0.11 (0.04) 0.36 −0.11 (0.02) 1.00
8 −0.15 −0.14 (0.05) 0.55 −0.143 (0.002) 1.00 −0.15 (0.04) 0.95 −0.15 (0.02) 1.00
9 −0.20 −0.20 (0.05) 0.95 −0.194 (0.002) 1.00 −0.20 (0.04) 1.00 −0.20 (0.02) 1.00
10 −0.25 −0.25 (0.05) 0.99 −0.251 (0.002) 1.00 −0.24 (0.04) 1.00 −0.25 (0.02) 1.00
11 0.00 −0.01 (0.04) 0.00 0.001 (0.002) 0.82 −0.01 (0.04) 0.01 0.00 (0.01) 0.00
12 0.00 0.02 (0.04) 0.00 0.001 (0.002) 0.84 0.02 (0.04) 0.02 0.00 (0.01) 0.00
13 0.00 −0.01 (0.04) 0.00 −0.004 (0.002) 0.89 −0.00 (0.04) 0.01 −0.00 (0.01) 0.00
14 0.00 −0.00 (0.04) 0.00 −0.002 (0.002) 0.82 −0.00 (0.04) 0.01 −0.00 (0.01) 0.00
15 0.00 0.00 (0.04) 0.00 −0.000 (0.002) 0.93 −0.00 (0.04) 0.01 −0.00 (0.01) 0.01
16 0.00 −0.00 (0.04) 0.00 −0.004 (0.002) 0.80 −0.00 (0.04) 0.01 −0.00 (0.01) 0.00
17 0.00 0.00 (0.04) 0.00 −0.005 (0.002) 0.85 0.00 (0.04) 0.01 −0.00 (0.01) 0.00
18 0.00 0.01 (0.04) 0.00 0.007 (0.002) 0.96 0.00 (0.04) 0.01 −0.00 (0.01) 0.00
19 0.00 0.00 (0.04) 0.01 −0.003 (0.002) 0.85 −0.00 (0.04) 0.01 0.00 (0.01) 0.00
20 0.00 −0.01 (0.04) 0.00 0.002 (0.002) 0.83 −0.00 (0.04) 0.01 −0.00 (0.01) 0.00
Note: Throughout values are based on results from 100 simulated datasets.
∗ The medians of the maximum likelihood estimate (Est) and SE of α and β , the proportion that the covariate
j,1 j,1
is selected for the jth outcome (mα,j , mβ,j ) are calculated.
† The empirical standard deviations of Est(β ) range between 0.036 and 0.054. (not presented in the table)
j,1
‡ The medians of the posterior means (PM) and posterior standard deviation (SD) of α
j,1 and βj,1 (conditioning on δj,1 = 1 and
γj,1 = 1, respectively), the medians of the posterior means of δj,1 and γj,1 (marginal posterior probability of inclusion) are computed.
The hyper-parameters ωα = ωβ are set to 0.1.
4.2. Results
With a marginal posterior probability cutoff of 0.5, the MZIP method identified three species (Fusobac-
terium periodonticum, Lachnoanaerobaculum orale, and Veillonella genus NOI) for which counts were
associated with HIV infection among “susceptible” individuals, and it selected two species (Actinomyces
graevenitzii and Leptotrichia sp oral taxon 215) for which the probability of participants being susceptible
to these two species was associated with HIV infection (Figure 2). In contrast, the frequentist UZIP anal-
ysis with 95% confidence intervals identified 39 species whose levels were associated with HIV infection
in the susceptible population, including the three associations identified by MZIP. The UZIP analysis
identified no species for which the probability of being “susceptible” to that species was associated with
HIV infection. Based on the methods’ relative performance in the simulation study, it is plausible that
the UZIP’s lack of accounting for outcomes’ correlation patterns, which are complex (Figure 1a), grossly
inflated the type I error rate in the count model and may also have decreased sensitivity of the zero model.
Indeed, comparing the estimated associations between HIV infection and the five taxa selected based
on the MZIP model (Table 3), the uncertainty associated with the count part of UZIP model was much
smaller than that generated by the MZIP method. As with the simulation, this may have accounted for
the detection of 39 associations, many of which are presumably false positive associations. Because of
how we calculated the IDR estimate, it has an age-specific interpretation. For example, 16-year-old youth,
PHIV youth had 34% (95% credible interval 2%, 47%) and 48% (95% credible interval 14%, 76%) lower
abundance of F. periodonticum and L. orale species, respectively, compared with PHEU youth.
The proposed MZIP model captures within-subject dependence among multiple outcomes via two
correlation components, R and
V . The dependence patterns arising from the empirical correlations
appear to reflect, strongly, the correlation structure predicted by the binary component of the proposed
MZIP (Figure 1). This implies that in these data, the presence of taxa is more structured than their counts.
This result is not attributable to smoothing of the empirical correlation from having added one to every
count, because the results were not sensitive to changes of this value to 0.5, 0.1, and 0.01. The model also
provides an opportunity to quantify and compare the contribution of zero inflation versus other sources of
overdispersion to microbiome taxon abundance (see Supplementary Material F available at Biostatistics
online for posterior estimates of
Vj,j and α0,j ).
510 K. H. LEE AND OTHERS
(a) Observed
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Fig. 1. Observed and estimated correlations among counts of 44 microbial species in the Pediatrics HIV/AIDS Cohort
Study (PHACS): (a) empirical correlations calculated by cor(log(y + 1)); (b) Posterior median of R; (c) Posterior
median of RV , calculated based on posterior samples of
V from the proposed multivariate zero-inflated Poisson
(MZIP) model fit.
Bayesian variable selection for multivariate zero-inflated count data 511
Binary part
VeillonellaceaeG1 sp oral taxon 150 Count part
Veillonella sp oral taxon 780
Veillonella rogosae
Veillonella parvula
Veillonella genus NOI
Veillonella atypica
Streptococcus sobrinus
Streptococcus mutans
Stomatobaculum sp oral taxon 097
Stomatobaculum longum
Shuttleworthia satelles
Selenomonas sp oral taxon 136
Scardovia wiggsiae
Prevotella salivae
Prevotella oris
Prevotella melaninogenica
Prevotella histicola
Prevotella denticola
Oribacterium sinus
Oribacterium parvum
Megasphaera micronuciformis
Leptotrichia wadei
Leptotrichia sp oral taxon 215
Leptotrichia hongkongensis
Lactobacillus salivarius
Lactobacillus genus NOI
Lachnoanaerobaculum orale
Kingella oralis
Haemophilus sputorum
Haemophilus parainfluenzae
Haemophilus genus NOI
Granulicatella elegans
Gemella genus NOI
Fusobacterium periodonticum
Cryptobacterium curtum
Campylobacter concisus
Bifidobacterium dentium
Atopobium rimae
Atopobium parvulum
Alloprevotella sp oral taxon 473
Actinomyces sp oral taxon 448
Actinomyces lingnaeNVP
Actinomyces graevenitzii
Actinomyces genus NOI
Fig. 2. Analysis of Pediatric HIV/AIDS Cohort Study (PHACS) data: marginal posterior inclusion probabilities for
the HIV status covariate in relation to excess zeros and counts of 44 microbial species as estimated via the proposed
multivariate zero-inflated Poisson (MZIP) model. We adjust for participants age, a potential confounder, but do not
perform variable selection on it. “NOI” stands for “not otherwise identified.”
512 K. H. LEE AND OTHERS
Table 3. Estimated regression parameters and inclusion probabilities for the five species identified as
associated with HIV infection by using a multivariate zero-inflated Poisson (MZIP)∗ and the univariate
zero-inflated Poisson (UZIP)† models
Binary Count
IDR‡
αj,1 |δj,1 = 1 βj,1 |γj,1 = 1
IP IP
PM (95% CI) PM (95% CI) PM (95% CI)
Actinomyces graevenitzi −0.39 (−0.67 to −0.09) 0.54 0.25 (−0.22 to 0.72) 0.11 0.96 (0.57–1.08)
Fusobacterium periodonticum 0.16 (−0.60 to 0.48) 0.07 −0.43 (−0.66 to −0.14) 0.93 0.66 (0.53–0.98)
MZIP Lachnoanaerobaculum orale 0.10 (−0.17 to 0.33) 0.03 −0.67 (−1.15 to −0.25) 0.95 0.52 (0.24–0.86)
Leptotrichia sp oral taxon 215 −0.44 (−0.75 to −0.17) 0.63 −0.04 (−0.19 to 0.20) 0.02 0.86 (0.74–1.00)
Veillonella genus NOI§ −0.04 (−0.60 to 0.45) 0.05 −0.37 (−0.60 to −0.12) 0.88 0.71 (0.57–1.00)
αj,l βj,k
Est (95% CIF ) Est (95% CIF )
Actinomyces graevenitzi −0.27 (−0.89 to 0.35) −0.46 (−0.69 to −0.23)
Leptotrichia sp oral taxon 215 −0.41 (−1.04 to 0.21) −0.26 (−0.41 to−0.11)
∗ In the MZIP model, the posterior median (PM) and 95% credible interval (CI) of regression parameters and incidence density ratio
(IDR) are computed.
† In the UZIP model, the maximum likelihood estimates (Est) and 95% confidence intervals (CI ) of regression parameters are
F
computed.
‡ Adjusted for individuals with age of 16 years.
§ Not otherwise identified.
5. D ISCUSSION
We have described the development of a new Bayesian variable selection method that addresses challenges
arising in the analysis of microbiome sequencing data: excess zero counts and high-dimensional outcomes
with a complex association structure. Applying the proposed multivariate approach led to the identification
of two species for which the probability of being susceptible to those species was associated with HIV
infection; these associations did not meet FDR thresholds when the existing univariate approach was
applied. In addition, based on the estimated induced marginalized IDR under the proposed model, another
two species were less abundant in HIV-infected youth aged 16 years compared with PHEU youth of the
same age.
One might question how realistic are these analyses when they are adjusted for only one confounder,
age. Some reassurance might be provided from the observation that in univariate analyses that included
Bayesian variable selection for multivariate zero-inflated count data 513
additional confounders (e.g., sex, race, and dental visit in previous year as a marker of oral hygiene),
inference was not greatly altered compared with models including only HIV status and age. We are
continuing to study performance in a range of datasets, including more complete confounder adjustment.
We are also working to scale up the proposed method in the number of endpoints, as discussed further
below, and also in the number of covariates, both of which are required for integrated omics analyses.
The simulation study demonstrated superior performance of the proposed MZIP approach over the
existing UZIP method when outcomes were correlated. The sample size was small enough that asymp-
totic assumptions under the frequentist-based UZIP model did not hold. This affected estimation of the
asymptotic variance of the regression parameters for the count model, which was not a limitation for the
proposed multivariate Bayesian approach. This difference in performance is primarily because (i) for small
data settings, estimation is generally more stable with Bayesian approaches, which exploit information
from both the observed data likelihood and prior distribution; and (ii) the MZIP method uses information
not only on the mean model but also from the structure of covariance among outcomes.
We used a multivariate probit model for the binary part of the MZIP mixture model. An alternative
is to assume a multivariate logistic distribution for wi (O’Brien and Dunson, 2004), for which posterior
computation can be facilitated based on a data augmentation algorithm (Albert and Chib, 1993). However,
initial numerical studies using the latter approach resulted in prohibitively slow mixing of the MCMC
algorithms due to sparseness of data, even for data with q = 10 and assuming an unstructured covariance
pattern. This is because the multivariate logistic model specification requires the estimation of n more
latent parameters than does the multivariate probit model. Thus, the multivariate probit model in the MZIP
proved to be much more computationally tractable.
We have presented the model in its most general form that allows the importance of each covariate,
as well as the correlation structure among the multivariate outcomes, to vary across the binary and the
count components of the model. This gives the user maximal flexibility and provides evidence on how
a covariate is associated with each response, i.e., with more zeros or higher counts. The question arises
whether the complexity of the model is necessary or whether simpler models should suffice. Analysis of
the motivating data suggests that different correlation structures were needed in this case. It is difficult
to provide a general answer to this question until we have had more opportunity to apply it a range of
datasets and compare results with those obtained in simpler models. One would not be able to make this
comparison if the most general model is not available as a basis for comparison.Yet, simpler models might
well be useful in other datasets.
Thus, the software implementation of the proposed approach offers more parsimonious versions of the
model, simplified by imposing additional restrictions regarding the model parameters. For example, the
two model parts can be forced to have one common variable selection indicator by setting γj,k = δj,l . In
practice, such a restriction might facilitate implementation by providing a single vector of variable selection
indicators, i.e., one list as to which species are associated with each covariate. A different assumption is
that both model parts share the same covariance pattern (R = RV ), which will greatly reduce the number of
parameters to be estimated and thus the computational complexity in the MCMC algorithm, especially for
data with large q. In our initial analysis, fitting this restricted model to the PHACS data yielded unreliable
estimates of R, because the assumption that R = RV is violated in these data (Figure 1). Again, the restricted
model may serve well in other datasets. Therefore, we made available the algorithms to implement both
types of simpler MZIP models as options in the mBvs package.
There are several ways the proposed framework could be extended. First, marginalized zero-inflated
models have recently been developed so that inference can be made on the marginal mean of the sampled
population via a set of unified regression coefficients (Long and others, 2015; Tabb and others, 2016). The
unified regression coefficients have better interpretability, as the marginalized models do not require the
additional steps described in Section 2.3 to address the dependence of parameter values on x i,(−k) . In some
applications, it may be more appropriate to interpret the two sets of regression coefficients separately, yet
514 K. H. LEE AND OTHERS
there also may be other applications for which interpretability of regression parameters would be enhanced
by adopting a marginalized model within the proposed multivariate Bayesian variable selection method.
Marginalization may also provide more stable model fitting. Second, although we focused data analysis
on a preselected subset of species in the application, often microbiologists’ goal is to perform whole-
community oral microbiome analysis, which generally involves several hundred taxa. We are currently
working to address the computational issues arising from an even higher-dimensional parameter space.
Because the complexity mainly results from the flexible unstructured covariance model, we propose
to scale up the proposed MZIP method by adopting alternative correlation structures that can flexibly
accommodate potentially complicated patterns among hundreds of taxa. A final possibility is to study
the interaction between the marginal posterior probability cutoff and the prior inclusion probability in
controlling the FDR at the desired level under the proposed model.
In conclusion, the proposed framework gives researchers valid and powerful statistical tools to overcome
major methodological barriers in microbiome sequencing data analysis. Beyond the study of the human
microbiome, the methods, software, and guidance from simulation studies in this work will be useful in
any field requiring analysis of multivariate zero-inflated count data.
6. S OFTWARE
R-package mBvs contains codes to implement proposed Bayesian framework described in the article. The
package is currently available in CRAN (https://cran.r-project.org/web/packages/mBvs).
7. S UPPLEMENTARY MATERIAL
Supplementary material is available online at http://biostatistics.oxfordjournals.org.
A CKNOWLEDGMENTS
The authors thank the associate editor and four anonymous reviewers for their constructive comments
and suggestions. Portions of this research were conducted on the O2 High Performance Compute Cluster,
supported by the Research Computing Group, at Harvard Medical School. Conflict of Interest: None
declared.
F UNDING
National Institutes of Health (P01CA134294, P30ES000002, U01HD052104, U01HD052102,
R21DE026872, and R03DE027486).
REFERENCES
AAS, J. A., GRIFFEN, A. L., DARDIS, S. R., LEE, A. M., OLSEN, I., DEWHIRST, F. E., LEYS, E. J. AND PASTER, B. J.
(2008). Bacteria of dental caries in primary and permanent teeth in children and young adults. Journal of Clinical
Microbiology 46, 1407–1417.
AITCHISON, J. AND HO, C.H. (1989). The multivariate Poisson-log normal distribution. Biometrika 76, 643–653.
ALBERT, J. AND CHIB, S. (1993). Bayesian analysis of binary and polychotomous response data. Journal of the
American Statistical Association 88, 669–679.
ALBERT, J., WANG, W. AND NELSON, S. (2014). Estimating overall exposure effects for zero-inflated regression models
with application to dental caries. Statistical Methods in Medical Research 23, 257–278.
Bayesian variable selection for multivariate zero-inflated count data 515
ALPEREN, J., BRUMMEL, S., TASSIOPOULOS, K., MELLINS, C., KACANEK, D., SMITH, R., SEAGE, G. AND MOSCICKI, A.
(2014). Prevalence of and risk factors for substance use among perinatally human immunodeficiency virus–infected
and perinatally exposed but uninfected youth. Journal of Adolescent Health 54, 341–349.
ANDERSON, M. J. (2001). A new method for non-parametric multivariate analysis of variance. Austral Ecology 26,
32–46.
ARAB, A., HOLAN, S., WIKLE, C. AND WILDHABER, M. (2012). Semiparametric bivariate zero-inflated Poisson models
with application to studies of abundance for multiple species. Environmetrics 23, 183–196.
ASHFORD, J. AND SOWDEN, R. (1970). Multi-variate probit analysis. Biometrics 26, 535–546.
BENJAMINI, Y. AND HOCHBERG, Y. (1995). Controlling the false discovery rate: a practical and powerful approach to
multiple testing. Journal of the Royal Statistical Society: Series B 57, 289–300.
BOX, G. AND TIAO, G. (2011). Bayesian Inference in Statistical Analysis. Hoboken, NJ, USA: John Wiley & Sons.
BREIMAN, L. AND FRIEDMAN, J. (1997). Predicting multivariate responses in multiple linear regression. Journal of
the Royal Statistical Society: Series B 59, 3–54.
BROWN, P., VANNUCCI, M. AND FEARN, T. (1998). Multivariate Bayesian variable selection and prediction. Journal
of the Royal Statistical Society: Series B 60, 627–641.
CAPORASO, J., LAUBER, C., WALTERS, W., BERG-LYONS, D., LOZUPONE, C., TURNBAUGH, P., FIERER, N. AND KNIGHT,
R. (2011). Global patterns of 16S rRNA diversity at a depth of millions of sequences per sample. Proceedings of
the National Academy of Sciences of the United States of America 108, 4516–4522.
CHEN, J. AND LI, H. (2013). Variable selection for sparse Dirichlet-multinomial regression with an application to
microbiome data analysis. The Annals of Applied Statistics 7, 418–442.
CHIB, S. AND GREENBERG, E. (1998). Analysis of multivariate probit models. Biometrika 85, 347–361.
DEWHIRST, F. E., CHEN, T., IZARD, J., PASTER, B. J., TANNER, A.C.R., YU, W.-H., LAKSHMANAN, A. AND WADE,
W. G. (2010). The human oral microbiome. Journal of Bacteriology 192, 5002–5017.
EARNEST, A., MORGAN, G., MENGERSEN, K., RYAN, L., SUMMERHAYES, R. AND BEARD, J. (2007). Evaluating the
effect of neighbourhood weight matrices on smoothing properties of conditional autoregressive (car) models.
International Journal of Health Geographics 6, 54.
FERNANDES, M. V. M., SCHMIDT, A. M. AND MIGON, H. S. (2009). Modelling zero-inflated spatio-temporal processes.
Statistical Modelling 9, 3–25.
FOX, J. (2013). Multivariate zero-inflated modeling with latent predictors: modeling feedback behavior. Computational
Statistics & Data Analysis 68, 361–374.
GEORGE, E. I. AND MCCULLOCH, R. E. (1993). Variable selection via Gibbs sampling. Journal of the American
Statistical Association 88, 881–889.
GEORGE, E.I. AND MCCULLOCH, R.E. (1997). Approaches for Bayesian variable selection. Statistica Sinica 7,
339–374.
GOMES, B., BERBER, V., KOKARAS, A., CHEN, T. AND PASTER, B. (2015). Microbiomes of endodontic-periodontal
lesions before and after chemomechanical preparation. Journal of Endodontics 41, 1975–1984.
HALL, D. B. (2000). Zero-inflated Poisson and binomial regression with random effects: a case study. Biometrics 56,
1030–1039.
HOLMES, I., HARRIS, K. AND QUINCE, C. (2012). Dirichlet multinomial mixtures: generative models for microbial
metagenomics. PLoS One 7, e30126.
LA ROSA, P. S., BROOKS, J. P., DEYCH, E., BOONE, E. L., EDWARDS, D. J., WANG, Q., SODERGREN, E., WEINSTOCK, G.
AND SHANNON, W. D. (2012). Hypothesis testing and power calculations for taxonomic-based human microbiome
data. PLoS One 7, e52078.
516 K. H. LEE AND OTHERS
LAMBERT, D. (1992). Zero-inflated Poisson regression, with an application to defects in manufacturing. Technomet-
rics 34, 1–14.
LEE, A. H., WANG, K., SCOTT, J. A., YAU, K. K. W. AND MCLACHLAN, G. J. (2006). Multi-level zero-inflated Poisson
regression modelling of correlated count data with excess zeros. Statistical Methods in Medical Research 15,
47–61.
LEE, K. H., TADESSE, M. G., BACCARELLI, A. A., SCHWARTZ, J. AND COULL, B. A. (2017). Multivariate Bayesian
variable selection exploiting dependence structure among outcomes: application to air pollution effects on DNA
methylation. Biometrics 73, 232–241.
LI, C., LU, J., PARK, J., KIM, K., BRINKLEY, P. A. AND PETERSON, J. P. (1999). Multivariate zero-inflated Poisson
models and their applications. Technometrics 41, 29–38.
LI, H. (2015). Microbiome, metagenomics, and high-dimensional compositional data analysis. Annual Review of
Statistics and Its Application 2, 73–94.
LIU, C. (2001). Bayesian analysis of multivariate probit models—discussion on the art of data augmentation. Journal
of Computational and Graphical Statistics 10, 75–81.
LIU, C. AND SUN, D. X. (2000). Analysis of interval-censored data from fractionated experiments using covariance
adjustment. Technometrics 42, 353–365.
LOEYS, T., MOERKERKE, B., DE SMET, O. AND BUYSSE, A. (2012). The analysis of zero-inflated count data: Beyond
zero-inflated Poisson regression. British Journal of Mathematical and Statistical Psychology 65, 163–180.
LONG, D. L., PREISSER, J. S., HERRING,A. H. AND GOLIN, C. E. (2015).A marginalized zero-inflated Poisson regression
model with random effects. Journal of the Royal Statistical Society: Series C 64, 815–830.
MANTEL, N. (1967). The detection of disease clustering and a generalized regression approach. Cancer Research 27
(2 Part 1), 209–220.
MANTEL, N. AND VALAND, R. S. (1970). A technique of nonparametric multivariate analysis. Biometrics 26,
547–558.
MOSCICKI, A.-B., YAO, T.-J., RYDER, M. I., RUSSELL, J. S., DOMINY, S. S., PATEL, K., MCKENNA, M., VAN DYKE,
R. B., SEAGE III, G. R., HAZRA, R. and others. (2016). The burden of oral disease among perinatally HIV-infected
and HIV-exposed uninfected youth. PLoS One 11, e0156459.
MUELLER, N. T., BAKACS, E., COMBELLICK, J., GRIGORYAN, Z. AND DOMINGUEZ-BELLO, M. G. (2015). The infant
microbiome development: mom matters. Trends in Molecular Medicine 21, 109–117.
NEELON, B. AND CHUNG, D. (2017). The LZIP: a Bayesian latent factor model for correlated zero-inflated counts.
Biometrics 73, 185–196.
O’BRIEN, S. M. AND DUNSON, D. B. (2004). Bayesian multivariate logistic regression. Biometrics 60, 739–746.
PEREZ-MUÑOZ, M. E., ARRIETA, MARIE-CLAIRE, RAMER-TAIT, AMANDA E. AND WALTER, JENS. (2017). A critical
assessment of the “sterile womb” and “in utero colonization” hypotheses: implications for research on the pioneer
infant microbiome. Microbiome 5, 48.
PFLUGHOEFT, K. J. AND VERSALOVIC, J. (2012). Human microbiome in health and disease. Annual Review of
Pathology: Mechanisms of Disease 7, 99–122.
PREISSER, J. S., STAMM, J. W., LONG, D. L. AND KINCADE, M. E. (2012). Review and recommendations for zero-
inflated count regression modeling of dental caries indices in epidemiological studies. Caries Research 46,
413–423.
RYDER, M. I, YAO, T.-J., RUSSELL, J. S., MOSCICKI, A.-B. AND SHIBOSKI, C. H. (2017). Prevalence of periodontal
diseases in a multicenter cohort of perinatally HIV-infected and HIV-exposed and uninfected youth. Journal of
Clinical Periodontology 44, 2–12.
Bayesian variable selection for multivariate zero-inflated count data 517
STARR, J. R., HUANG, Y., LEE, K. H., MURPHY, C. M., MOSCICKI, A.-B., SHIBOSKI, C. H., RYDER, M. I., YAO, T.-J.,
FALLER, L. L., VAN DYKE, R. B. and others. (2018). Oral microbiota in youth with perinatally acquired HIV
infection. Microbiome 6, 100.
TABB, L. P., TCHETGEN, E. J., WELLENIUS, G. A. AND COULL, B. A. (2016). Marginalized zero-altered models for
longitudinal count data. Statistics in Biosciences 8, 181–203.
TANNER, M. A. AND WONG, W. H. (1987). The calculation of posterior distributions by data augmentation. Journal
of the American Statistical Association 82, 528–540.
TASSIOPOULOS, K., PATEL, K., ALPEREN, J., KACANEK, D., ELLIS, A., BERMAN, C., ALLISON, S. M., HAZRA, R., BARR,
E., CANTOS, K. and others. (2016). Following young people with perinatal HIV infection from adolescence into
adulthood: the protocol for PHACS AMP Up, a prospective cohort study. BMJ Open 6, e011396.
WADSWORTH, W. D., ARGIENTO, R., GUINDANI, M., GALLOWAY-PENA, J., SHELBURNE, S. A. AND VANNUCCI, M.
(2017). An integrative Bayesian Dirichlet-multinomial regression model for the analysis of taxonomic abundances
in microbiome data. BMC Bioinformatics 18, 94.
WANG, X., CHEN, M., KUO, R. C. AND DEY, D. K. (2015a). Bayesian spatial-temporal modeling of ecological
zero-inflated count data. Statistica Sinica 25, 189–204.
WANG, Z., MA, S. AND WANG, C. (2015b). Variable selection for zero-inflated and overdispersed data with application
to health care demand in Germany. Biometrical Journal 57, 867–884.
XU, L., PATERSON, A. D., TURPIN, W. AND XU, W. (2015). Assessment and selection of competing models for
zero-inflated microbiome data. PLoS One 10, e0129606.
ZEILEIS, A., KLEIBER, C. AND JACKMAN, S. (2008). Regression models for count data in R. Journal of Statistical
Software 27, 1–25.
[Received October 27, 2017; revised September 6, 2018; accepted for publication September 9, 2018]