Modelling the co-occurrence of Streptococcus pneumoniae with other bacterial and viral pathogens in the upper respiratory tract.

Otitis media (OM) is a major burden for all children, particularly for Australian Aboriginal children. Streptococcus pneumoniae, Moraxella catarrhalis, Haemophilus influenzae and viruses (including rhinovirus and adenovirus) are associated with OM. We investigated nasopharyngeal microbial interactions in 435 samples collected from 79 Aboriginal and 570 samples from 88 non-Aboriginal children in Western Australia. We describe a multivariate random effects model appropriate for analysis of longitudinal data, which enables the identification of two independent levels of correlation between pairs of pathogens. At the microbe level, rhinovirus infection was positively correlated with carriage of S. pneumoniae, H. influenzae and M. catarrhalis, and adenovirus with M. catarrhalis. Generally, there were positive associations between bacterial pathogens at both the host and microbe level. Positive viral-bacterial associations at the microbe level support previous findings indicating that viral infection can predispose an individual to bacterial carriage. Viral vaccines may assist in reducing the burden of bacterial disease.

Introduction

Upper respiratory tract (URT) carriage of respiratory bacterial pathogens is associated with otitis media (OM), which is a major burden of disease in young children, particularly in Australian Aboriginal children [1], [2], [3], [4]. Colonisation and survival of an individual microbe in the human nasopharynx are determined by numerous factors, some of which are host-specific, such as environmental exposure to infectious agents and immune function, and some microbial, including interactions between microbial species. It is the extent and mechanisms by which microbes interact and the resulting potential to cause disease that are of particular interest. Several studies have suggested that viral infections of the URT predispose to bacterial respiratory infection and that viruses in the URT enhance bacterial adherence to respiratory tract cells [5], [6], [7], [8]. Pathogens can also alter the effects of co-infecting pathogens by destroying or damaging a physical barrier, decreasing mucociliary activity, altering immune cell function or up-regulating gene expression [9]. Data are limited on the co-occurrence of bacteria in the URT [10], [11], [12] and on associations between colonising pathogens in the absence of symptoms.

In a cohort study of young children in Western Australia (WA), we noted that age-specific patterns of carriage of Streptococcus pneumoniae, Moraxella catarrhalis and Haemophilus influenzae were very similar [13], suggesting that the same host factors might influence carriage of all three pathogens. In addition, there may be positive interactions between pathogens in the ecological niche of the URT. Bacterial interference may also occur: there are reports of negative associations between carriage of serotypes included in a 7-valent pneumococcal conjugate vaccine (7vPCV) and Staphylococcus aureus [11], [12], [14].

A major limitation of cross-sectional analyses of longitudinal data is that they do not account for the possibility that samples displaying multiple carriage at different time points may come from the same individuals. Cross-sectional analyses are therefore unable to distinguish between host-level (between subjects) associations and associations at the microbe-level (within subjects), both of which will result in co-occurrence of pathogens within samples.

A positive association between pathogens at the host level suggests that individuals with high susceptibility to infection with one pathogen, as evidenced by higher than average incidence over time, are also susceptible to infection with another pathogen under consideration. Environmental factors and immune function may play a role [15], [16], [17], [18], [19].

A positive association at the microbe level is observed when co-occurrence of two pathogens within a particular host occurs more frequently than would be expected given the independent susceptibilities of the host to these pathogens. In the case of nasopharyngeal carriage, this might be the result of a mechanism whereby the presence of one pathogen creates conditions conducive to adherence of another. Conversely, a negative correlation would suggest that the two pathogens of interest are competing for colonisation sites in the nasopharynx.

Statistical modelling of longitudinal data, consisting of repeated sampling of individuals over time, can be used to differentiate between host- and microbe-level correlations and thus assist in understanding the ecology of the URT and identifying appropriate interventions. We have investigated patterns of co-occurrence of pathogens identified in the nasopharynx of Aboriginal and non-Aboriginal children living in a semi-arid area of WA. We describe a multivariate random effects model appropriate for analysis of longitudinal data, which enables the identification of two independent levels of association between pairs of pathogens.

Methods

Study design

This study was conducted in the Kalgoorlie–Boulder area of the Eastern Goldfields located 600 km east of Perth, WA. We used data from the Kalgoorlie Otitis Media Research Project, which aims to investigate causal pathways to OM and the epidemiology of URT micro-organism carriage. The project design has been described elsewhere [13]. Briefly, 100 Aboriginal and 180 non-Aboriginal children were enrolled at birth and followed regularly to age 2 years. Nasopharyngeal aspirates (NPA) were collected at ages 1–3, 6–8 and 4, 6, 12, 18 and 24 months. Primary bacterial isolation of S. pneumoniae, H. influenzae, M. catarrhalis and S. aureus was performed using selective media [13], [20]. To identify viruses, nucleic acid extraction for nucleic acid amplification was carried out using QiaAmp Viral RNA kit (QiaGen corp) in accordance to the manufacturer's protocol. In-house nested or semi-nested PCR amplification was performed for rhinoviruses, adenoviruses, respiratory syncytial virus, human metapneumovirus, influenza viruses A and B, coronaviruses and parainfluenza virus types 1–3. However, only adenovirus and rhinovirus were detected in sufficient numbers to be included in this data analysis.

Statistical analysis

Data from children who contributed four or more samples were included in the analysis. The presence or absence of S. pneumoniae, H. influenzae, M. catarrhalis, S. aureus, adenovirus and rhinovirus was recorded for every sample. A hierarchical logistic model was constructed with joint binary responses for each of the pathogens. This provided estimates of multivariate normally distributed random effects (under a logit link) at the host level and also residuals at the microbe level which have a multivariate Bernoulli distribution. Modelling of multivariate longitudinal data is frequently performed but in most cases the correlations between dependent variables are regarded as nuisance parameters which must be accounted for in order to estimate fixed effects accurately [21], [22]. However, in this study, they are the main focus of interest. One previous study has investigated separate levels of multivariate association in longitudinal binary data with a similar approach to that used here [23]. Estimated covariances of the random effects and their standard errors provide an indication of significance of the association between each pair of pathogens, at both the host and microbe levels. The models incorporated fixed effects to adjust for a quadratic dependence of log odds of carriage on age and also for seasonal effects. Patterns of co-occurrence were analysed separately in Aboriginal and non-Aboriginal children. The analysis was conducted using the MLwiN software package [24] with a penalised quasi-likelihood (PQL) solution method. A detailed model description is presented in Appendix B.

Individual children provided between four and seven samples. Random effects modelling require that data are missing at random (MAR) in order to avoid any bias of parameter estimates [25]. Since data were collected from healthy children during routine follow-up visits, there should be little correspondence between a child's carriage status and attendance for routine follow-up, and therefore the MAR assumption is unlikely to be violated.

Results

A total of 435 samples from 79 Aboriginal children (median 6, mean 5.5/child) and 570 samples from 88 non-Aboriginal children (median 7, mean 6.5/child) who had at least four specimens examined for all relevant pathogens were included in this analysis. The frequencies of positive samples for each pathogen under investigation are shown in Table 1 . The patterns of co-occurrence of S. pneumoniae with M. catarrhalis, S. aureus, rhinovirus and adenovirus are shown in Fig. 1 , stratified by Aboriginality and age group. In Aboriginal and non-Aboriginal children of all ages, S. pneumoniae was isolated more frequently when M. catarrhalis was also isolated than when M. catarrhalis was not isolated. A similar pattern was observed for S. pneumoniae and H. influenzae, and also for S. pneumoniae and rhinovirus except in children aged <1 month. On cross-sectional examination, co-occurrence of S. pneumoniae with adenovirus and with S. aureus varied with age (Fig. 1(b and d)).

Table 1

Numbers (%) of S. pneumoniae, H. influenzae, M. catarrhalis, S. aureus, adenovirus and rhinovirus identified in NPAs collected from Aboriginal and non-Aboriginal children

	Aboriginal	Non-Aboriginal
Number of samples	435	570
S. pneumoniae	219 (50.3%)	158 (27.7%)
H. influenzae	182 (41.8%)	73 (12.8%)
M. catarrhalis	225 (51.7%)	171 (30.0%)
S. aureus	88 (20.2%)	147 (25.8%)
Adenovirus	37 (8.5%)	20 (3.5%)
Rhinovirus	102 (23.4%)	94 (16.5%)

Fig. 1

Proportion of NPAs from which S. pneumoniae was isolated in presence or absence of: (a) rhinovirus, (b) adenovirus, (c) M. catarrhalis and (d) S. aureus by age and Aboriginality.

Associations between pairs of bacteria at host and microbe levels obtained from longitudinal models are shown in Table 2 . There were significant positive host-level associations between S. pneumoniae, M. catarrhalis and H. influenzae in both Aboriginal and non-Aboriginal children. Only the association between S. pneumoniae and H. influenzae in non-Aboriginal children failed to reach significance. We also observed significant positive associations at the microbe level between S. pneumoniae, M. catarrhalis and H. influenzae in both Aboriginal and non-Aboriginal children; only the association between S. pneumoniae and H. influenzae in Aboriginal children failed to reach significance. There were no significant bacterial associations involving S. aureus with the exception of a negative association with M. catarrhalis at the microbe level in non-Aboriginal children.

Table 2

Associations between pairs of bacteria in Aboriginal and non-Aboriginal children

	Host-level covariance		Microbe-level covariance
	Aboriginal	Non-Aboriginal	Aboriginal	Non-Aboriginal
S. pneumoniae × H. influenzae	1.26 (0.72,1.81)	0.32 (−0.10,0.75)	0.09 (−0.01,0.19)	0.20 (0.12,0.28)
S. pneumoniae × M. catarrhalis	1.34 (0.72,1.96)	0.71 (0.31,1.11)	0.27 (0.18,0.36)	0.25 (0.17,0.33)
H. influenzae × M. catarrhalis	1.14 (0.57,1.71)	0.56 (0.11,1.00)	0.19 (0.09,0.29)	0.26 (0.18,0.34)
S. pneumoniae × S. aureus	−0.26 (−0.76,0.24)	−0.40 (−0.88,0.08)	0.04 (−0.07,0.14)	0.05 (−0.04,0.14)
H. influenzae × S. aureus	0.25 (−0.24,0.74)	−0.24 (−0.79,0.31)	−0.04 (−0.15,0.06)	−0.04 (−0.12,0.05)
M. catarrhalis × S. aureus	0.07 (−0.47,0.61)	−0.04 (−0.51,0.42)	−0.05 (−0.15,0.06)	−0.10 (−0.19,−0.01)

Associations are statistically significant (p < 0.05) where the confidence interval for the covariance does not include zero.

Ninety-five percent confidence limits are shown in parentheses; significant values appear in bold.

Associations for each bacterium/virus pair at host and microbe level are shown in Table 3 . Generally, there were no significant associations between the viruses under investigation and bacteria at the host level. At the microbe level, we observed significant positive associations between the presence of rhinovirus and the three major respiratory bacteria in all children. There was no consistent pattern of co-occurrence between adenovirus and bacteria but a significant positive association of adenovirus with M. catarrhalis in all children and with H. influenzae in Aboriginal children.

Table 3

Associations between bacterium/virus pairs in Aboriginal and non-Aboriginal children

	Host-level covariance		Microbe-level covariance
	Aboriginal	Non-Aboriginal	Aboriginal	Non-Aboriginal
S. pneumoniae × Rhinovirus	0.44 (0.05,0.83)	−0.01 (−0.32,0.31)	0.14(0.04,0.23)	0.14 (0.05,0.23)
H. influenzae × Rhinovirus	0.39 (0.00,0.77)	−0.11 (−0.47,0.26)	0.25 (0.16,0.35)	0.12 (0.04,0.21)
M. catarrhalis × Rhinovirus	0.36 (−0.07,0.78)	0.12 (−0.19,0.43)	0.21 (0.12,0.31)	0.14 (0.05,0.22)
S. aureus × Rhinovirus	0.34 (−0.07,0.74)	0.20 (−0.21,0.61)	0.01 (−0.10,0.11)	0.12 (0.03,0.20)
S. pneumoniae × Adenovirus	0.61 (0.00,1.22)	−0.33 (−0.97,0.32)	−0.05 (−0.15,0.05)	0.21 (0.13,0.29)
H. influenzae × Adenovirus	0.41 (−0.18,1.00)	0.52 (−0.23,1.26)	0.15 (0.05,0.25)	0.07 (−0.02,0.16)
M. catarrhalis × Adenovirus	0.33 (−0.33,0.99)	0.00 (−0.65,0.64)	0.11 (0.01,0.20)	0.33 (0.25,0.40)
S. aureus × Adenovirus	0.12 (−0.51,0.76)	0.56 (−0.28,1.39)	−0.05 (−0.15,0.05)	0.08 (−0.01,0.17)
Adenovirus × Rhinovirus	0.48 (−0.01,0.96)	−0.02 (−0.59,0.55)	0.09 (−0.01,0.19)	0.27 (0.19,0.35)

Associations are statistically significant (p < 0.05) where the confidence interval for the covariance does not include zero.

Ninety-five percent confidence limits are shown in parentheses; significant values appear in bold.

The covariances computed by our models as indicators of association may perhaps be more easily understood in terms of more frequently used measures such as the odds ratio (OR) [26]. A microbe-level covariance of zero for two pathogens describes a situation where the probability of presence of one pathogen is unaffected by whether the other pathogen is also present, i.e. an OR of 1. Similarly, a positive covariance will always equate to an OR of greater than 1. However, there is no simple correspondence between non-zero covariances and ORs when the prevalence of carriage or infection changes over time and individuals have differing susceptibilities. Nevertheless, some indications can be provided: the most significant microbe-level covariances found in this study were in the region of 0.25 which would correspond to an OR of 2.8 if both pathogens were to have a constant prevalence of 50% in all individuals. This rises to an OR of 5.9 for less frequently occurring pathogens with a prevalence of 10%.

Discussion

Statistical modelling of longitudinal data has the potential to distinguish two independent levels of association between pairs of pathogens, each of which has a distinct biological interpretation. To our knowledge this approach has not been used before in the area of infectious diseases. A similar method of analysis has, however, been reported for longitudinal data consisting of the incidence of different symptoms in a palliative care trial [23]. This study used an identical modelling approach to that used in the current study and identified associations between symptoms both between and within subjects. The statistical properties of the models were also investigated.

Through statistical modelling we found positive host-level associations between carriage of the three bacterial pathogens causing OM, which is consistent with previous studies showing that environmental factors such as crowding, presence of siblings in the household, day care attendance or exposure to environmental tobacco smoke are associated with high rates of bacterial carriage [15], [16], [17], [18], [27]. Concurrent multiple bacterial infections may be more pathogenic than single infections. It has been suggested that the more persistent severe OM experienced by Aboriginal children compared with non-Aboriginal children may be related to simultaneous prolonged carriage of multiple OM pathogens in the nasopharynx of Aboriginal children [2]. Hence the host-level social and environmental factors giving a high force of infection must be addressed to reduce the concurrent carriage of pathogens in the URT and the resulting severe outcomes of OM in Aboriginal children. Environmental factors are also likely to contribute to the positive host-level association observed between rhinovirus and S. pneumoniae in Aboriginal children. At the host level, mucosal immune function may play an important role in preventing colonisation by potential pathogens [28], [29]. Salivary IgA to OM pathogens has been measured as part of the Kalgoorlie OM Research Project and we will investigate associations between immune status and URT carriage.

We also found positive associations at the microbe level between carriage of all the bacterial OM pathogens, suggesting that these pathogens have an affinity to colonise simultaneously. Our findings are supported by an in vitro study by Budhani and Struthers who showed that S. pneumoniae was protected from benzylpenicillin and amoxycillin when grown in the presence of β-lactamase-producing M. catarrhalis but susceptible to these antibiotics when grown with non-β-lactamase-producing moraxellae [30]. Thus, β-lactamase-producing M. catarrhalis may act as an indirect pathogen. On the other hand, micro-organisms may compete for successful colonisation in the URT [31] and we found negative associations at the microbe level between S. aureus and both M. catarrhalis and H. influenzae. We did not find significant negative associations between all serotypes of S. pneumoniae and S. aureus. However, significant negative associations have been reported between S. aureus and the serotypes included in 7vPCV [11], [12], [14] and the potential for serotype and species replacement after immunisation with 7vPCV exists [32]. Density of colonisation may also play a role in bacterial interactions and we will examine this further.

The positive viral–bacterial associations observed at the microbe level, particularly those involving rhinovirus, do not necessarily indicate any causal relationship between viral infection and subsequent bacterial colonisation or vice versa. However, our findings support previous studies, which postulate that an existing respiratory viral infection can predispose an individual to bacterial colonisation of the URT by creating conditions for increased adherence [7], [8], [33], [34]. Thus, vaccines targeting URT viruses might have a role in reducing S. pneumoniae-related disease. Conversely, a trial in South Africa reported that 7vPCV provided some protection against viral pneumonia, suggesting that S. pneumoniae plays an important role in the development of viral pneumonia [35].

We have presented here our initial investigation of host- and microbe-level associations between potential pathogens in the URT. Future analyses of data collected as part of the Kalgoorlie OM Research Project will consider density of bacterial colonisation and specific subtypes of bacteria and we will investigate the impact of social and environmental factors on associations between pathogens. Statistical modelling of longitudinal data is a powerful tool to examine the ecology of the URT and will assist in determining appropriate interventions aimed at reducing the burden of URT microbial carriage and infections.