The natural history of Chlamydia trachomatis infection in women: a multi parameter evidence synthesis

The early history of the programme

The NCSP was set up in 2003 and was implemented in some way in every region of England by 2008. It is an opportunistic approach involving a combination of health and non-health screening venues, targeted at sexually active young people aged 15–24 years. 3 The objectives of the programme include both individual and population benefit.

1. Reduce CT incidence (population benefit). Detection and treatment of CT directly lowers CT prevalence and this will reduce the incidence of new cases in future.

2. Treat CT and prevent sequelae (individual benefit). Detection and early treatment of CT should prevent progression to sequelae.

3. Treat sexual partners of index cases.

The website (www.chlamydiascreening.nhs.uk/) provides further information on how the programme operates.

The precise evidence base for the NCSP has been controversial both before4 and since its introduction. 5 In the UK, new screening programmes are normally examined by the National Screening Committee (NSC), based on a specifically commissioned systematic review and cost-effectiveness analysis (CEA) of candidate proposals for screening. The committee applies a set of criteria (see www.screening.nhs.uk/criteria) derived from the Wilson and Junger criteria for screening,6 and considers the infrastructural implications of the screening and diagnostic tests, training in counselling and other services. In the case of CT, the decision to go ahead with a screening programme was originally based on the report from the Chief Medical Officer’s Expert Advisory Group on Chlamydia trachomatis. 7 This group considered that the proposed programme met the necessary criteria, although others5 have found otherwise. Although the report recommended screening in sexually active women aged < 25 years, women aged 25 years or over with a new partner in the previous 12 months, and high-risk men, there was a recognition that data on the costs, feasibility and acceptability of the programme were lacking, and these were to be addressed by pilot studies to be initiated the following year, in 1999. An early dynamic modelling study undertaken during this period by the Department of Health8 may also have influenced the decision. It had concluded that screening could be cost neutral within 4 years.

In the event, following supportive experience during the pilot studies, CT screening formed an important plank in the English government’s 2001 sexual health strategy9 and has continued to do so ever since. The NSC reviewed the population-based screening programme in 2006 but did not recommend it. 10 The NSC may revisit this, or it may review opportunistic screening as currently practiced in England, in the future. Sheringham et al. 11 recently provided an insightful analysis of the development of policy leading to the NCSP.

Evaluating the benefits of Chlamydia trachomatis screening

As with any screening programme, evaluation of its effects can be undertaken in two ways: one method is experimental studies in which individuals are randomised to either screening and early treatment or to no screening, with outcomes assessed at appropriate later time points. At the time when the NCSP was set up, two randomised controlled trials (RCTs) had been conducted,12,13 one of which, a large-scale study13 on women considered to be at higher risk in the USA, appeared to show that the frequency of pelvic inflammatory disease (PID) could be approximately halved in the screened group. This trial13 has been criticised by a number of commentators,14–16 although the US Task Force for Prevention did not see fit to adjust the estimates of benefit17 and we return to consider this more closely in later chapters. A second study, by Ostergaard et al. ,12 was relatively small. Both of these trials12,13 studied an intervention and target population somewhat different from that proposed and implemented in England. Further RCTs have been published since the NCSP was introduced.

By 2009 there had been more reviews of the trial evidence than there had been trials. 18 Once again, although there was good evidence that screening could be effective in lowering the incidence of PID in high-risk populations, and also in high school students, some felt that there was a lack of evidence that screening would be successful in the general population. 18 Since then, the Prevention of Pelvic Infection (POPI) trial19 has studied whether or not screening can lower the PID rate in women in higher education. The findings of this trial19 were again positive. However, the results were considered to be less than definitive: the trial was somewhat underpowered because of difficulties with recruitment and lower-than-expected rates of PID. 20,21

There have also been trials that have sought to estimate the effect of screening on population prevalence, rather than its effect on morbidity in the recruited population. Studies published to date have not delivered convincing results. 18 A stepped-wedge pilot in the Netherlands has reported no substantial fall in positivity among those screened over three screening rounds. The intended primary outcome of impact on population prevalence was aborted as a result of difficulties obtaining this measure. 22 A Danish trial23 randomised women to a single postal invitation to be screened compared with no intervention, and followed up to PID, EP and infertility outcomes over a 9-year period. In view of the uptake of testing in both arms as part of routine care, and the low uptake of the offer of testing in the treatment arm, the result is not surprising. 24 In addition, the length of follow-up is likely to lead to considerable dilution of the relative risk (see Chapter 3, Population Excess Fraction). Results from the large-scale ACCEPt trial (Australian Chlamydia Control Effectiveness Pilot)25 in Australia are still awaited.

A second approach to evaluation is by systematic review of the evidence and CEA. CEA applied to screening decisions can be informed directly by randomised studies, but is more often based on putting together the wide range of evidence on the population epidemiology of the condition, its clinical natural history with and without treatment, and – in the case of infectious disease programmes – models of the dynamic effects of interventions on future incidence. With sexually transmitted diseases (STDs), dynamic models must also make a range of assumptions about sexual partnership formation rates, duration of partnerships, their concurrency, and the impact of screening on partner recall and subsequent sexual behaviour, and so on. The advantage of evaluation by modelling, compared with RCTs, is that it is possible to consider the longer-term outcomes, beyond the trial, such as EP and tubal factor infertility (TFI). One can also study the expected dynamic effects of reducing prevalence. In 2007, two major CEAs of CT screening in the UK context were published. 1,2,15

The results from the two CEAs were very different: one2 concluded that screening was highly unlikely to be cost-effective (ClaSS, Chlamydia Screening Studies project), whereas the other concluded that screening was likely to be cost-effective, and most clearly so in men and women aged < 20 years if the risk of PID attributable to chlamydia was 10% or more. The fact that the two models produced different results is, in itself, not at all surprising. They were not examining exactly the same screening protocol, different structural assumptions were made about the natural history and different parameter values were assumed, particularly touching on transmissibility and sexual behaviour. 26,27

However, what is surprising, particularly given the wide acceptance of systematic review as the methodology for analysing problems of this type, is that the disagreement between the two UK cost-effectiveness analyses1,2 was not only about the parameter values, or even about how parameter values should be estimated from the available data, but also went as far as a conflict about what kinds of data should be used to inform parameters. Later in this chapter we illustrate the extent of the controversy concerning just one single parameter.

Statistical combination of evidence

To carry out this programme, we have relied on Multi-parameter Evidence Synthesis (MPES). This is an approach to model building and evidence synthesis42,43 that has properties that will prove valuable in the present context. It recognises that epidemiological studies do not necessarily generate direct evidence on the parameters of interest, but instead can inform functions of parameters. It has the technical means to combine information on parameters and functions of parameters within a single coherent structure; it offers standard statistical approaches to checking model fit, and specifically for determining whether or not different sources of evidence are in conflict regarding the information they contribute to any specified parameter; all data sources are formally incorporated via their likelihood, and parameter uncertainty, sampling uncertainty, and variation owing to heterogeneity is readily propagated through the evidence network.

Multi-parameter Evidence Synthesis is an essentially Bayesian methodology and it has been implemented using the popular Bayesian Markov chain Monte Carlo (MCMC) package WinBUGS version 1.4.3 (Medical Research Council Biostatistics Unit, Cambridge, UK). 44 But we emphasise that the use of Bayesian MCMC is not a necessary feature of our approach to evidence synthesis. Bayesian MCMC in WinBUGS makes it relatively easy for investigators to fit complex models, and the adoption of simulation-based Bayesian computation also greatly facilitates uncertainty propagation. However, similar results could be achieved, in principle, though perhaps with some difficulty, using Frequentist methods.

The idea of statistical combination of information on different functions of parameters appears to have been independently rediscovered a number of times in the last 80 years, with applications in physics, environmental risk control, deterministic dynamic models of population change, and in a range of computational frameworks, in both Frequentist and Bayesian forms. 45–48 The term ‘multi-parameter synthesis’ was coined by Victor Hassleblad,49 one of the originators of the Confidence Profile Method,50 a programme of research in clinical and descriptive epidemiology that was the immediate precursor to MPES and which made particular use of bias adjustment.

Besides review papers,42,43,51 there are a number of applications of MPES to infectious disease epidemiology, often oriented to screening problems. 52–56 A relatively simple application, and perhaps the best introduction, setting out all of the main properties of MPES, is a stylised application to prenatal human immunodeficiency virus (HIV) testing. 57

Multi-parameter Evidence Synthesis can be viewed as a purely technical extension of the single parameter meta-analysis, as described in numerous textbooks and tutorial papers,58–60 to multi-parameter problems. But it is probably more accurate to see it as a philosophically different approach to evidence and evidence synthesis.

Coherence of estimates

An important property of this approach to synthesis is its internal coherence. The generation of epidemiological evidence tends to be fragmented: separate studies are made of CT prevalence, CT incidence and CT duration, and separate estimates are published. But, in truth, these three quantities are known to be related: prevalence = incidence × duration. Estimates of any one of these parameters need to be consistent with estimates of the other two and with whatever data inform any of them. In the same way, studies of CT incidence, progression rates to PID, EP and TFI should produce results that are mathematically related to the results from retrospective studies of the role of CT in PID, EP and TFI, and to routine statistics. MPES is simply a methodology in which these relationships can be explored in a formal way, while taking account of the full extent of uncertainty due to sampling variation, additional sources of variation and any concerns that might come under the heading of ‘data quality’.

There is, in fact, nothing very new or original about putting different pieces of information together to see if they ‘add up’. Many chlamydiologists have carried out ‘back of the envelope’ calculations with exactly this objective in mind, to see whether information on incidence and prevalence, put together with information on progression to PID, EP or TFI could be consistent with routine data or not. Or asking whether routine data, put together with information on the proportion of PID, EP or TFI attributable to CT could be consistent with CT incidence and prevalence information or not. Indeed, many calculations of this type have been published. 5,39,40,61,62 The programme of work described in this monograph uses MPES to examine exactly the same ideas, but within the framework of a formal statistical method, and perhaps in a more complete and systematic way.

Role of interpretation

Note, however, that MPES is not simply a ‘synthesis exercise’. The synthesis is carried out under a very specific model of how the different data sources were generated. There is, therefore, as in any scientific exercise, an element of interpretation, and therefore of judgement. This arises in several ways. First, goodness-of-fit tests may not be able to distinguish between alternative models, forcing investigators to use their judgement. But a far more serious element of subjectivity lies in the fact that our overall judgement about the mechanisms that generate a set of data may depend greatly on our interpretation of just one or two studies. It is notoriously difficult, for example, to know if confounding variables or effect modifiers are present, or how great their impact is. Although conflicts between evidence sources may alert us to the presence of confounders and effect modifiers, the use of MPES does not remove the problems they create. We have therefore been explicit about where there is a subjective element to model choice, and particularly where there is an element of data-driven (post hoc) re-interpretation of evidence. In these instances, we follow the usual practice of describing the results as ‘hypothesis generating’, and we prioritise research recommendations in this light.

Summary

1. At the time the NCSP was introduced, the evidence base for the cost-effectiveness of CT screening against its aims of reducing infection rates and reducing sequelae was mixed and questionable.

2. Two major UK studies1,2 of the cost-effectiveness of CT screening sourced critical parameter estimates from completely different types of study, indicating a lack of consensus about what kinds of evidence should be used to inform natural history parameters.

3. Our analytic strategy is to look at all sources of evidence: retrospective, prospective and routine data.

4. MPES is used to assess the consistency of different types of evidence, and to develop an internally coherent model of the natural history and population epidemiology of CT.

5. The use of MPES does not remove the subjective elements in model choice and evidence interpretation.

Chlamydia trachomatis infection

Chlamydia trachomatis is an intracellular bacterial pathogen and a major cause of genital and eye disease. It comprises three human biovars. One causes trachoma, the most common form of blindness worldwide, and particularly prevalent in Africa. A second is the cause of lymphogranuloma venereum, a sexually transmitted infection (STI), mainly of gay men, in the developed world. The third biovar, consisting of serovars D to K, is the topic of this monograph. It is a major cause of PID, EP and TFI in women, epididymitis in men, and neonatal conjunctivitis and neonatal pneumonia in newborns. 3,28,63,64

Chlamydia trachomatis is transmitted by oral, vaginal or anal sex, and can also be transmitted from mother to newborn during a vaginal delivery. 3,63,64 Diagnosis is generally based on nucleic acid amplification tests (NAATs) carried out on cervical or vaginal swabs in women, urethral swabs in men or on urine samples. 3,63,64 NAAT has largely replaced culture or enzyme immunoassays for diagnostic testing. 3,63,64 However, culture played a key role in earlier epidemiological studies, and the lower sensitivity of culture, 60–80%,64,65 needs to be taken into account when interpreting the earlier research results.

The infection can be treated effectively by antibiotics, such as azithromycin or doxycycline, but it is often asymptomatic, especially in women. 3,63,64 Re-infection by the same partner is an important issue in prevention: partner notification rates can be improved and risks of re-infection reduced by providing index cases with prescriptions intended for their partners, or with the medication itself. 66–71

Chlamydia trachomatis infection

Approximately 75% of incident infections in women are asymptomatic. 73 Although studies of the duration of asymptomatic infection have produced a wide range of estimates – between 1 and 18 months (see Chapter 4) – it is evident that the infection can clear spontaneously in the absence of treatment as a consequence of the adaptive and innate immune responses. 38,74

The risk of transmission of infection from one episode of sexual intercourse is estimated to be between 10% and 20%. 75,76 Given the high sensitivity of current NAATs for detecting CT,8–10 and the fact that non-host DNA and sperm can be recovered from the female genital tract up to 7 days following sexual intercourse, this strongly implies that at least some women who test CT positive (CT+) are probably only passively infected. 77,78 This is consistent with the observations of Joyner et al. 79 and Geisler et al. 80 that a proportion of women testing CT+ without treatment become CT NAAT-negative within 2 weeks.

Prevalence of CT can be studied by population surveys, using NAATs on genitourinary specimens. As with most STIs, incidence and prevalence are highest among young people, and decline steeply with age. 81 Infection is more common among the more sexually active individuals, with the greatest number of sexual partners. 8–10,81

Definition of pelvic inflammatory disease

Pelvic inflammatory disease comprises a spectrum of upper genital tract inflammatory disorders among women, which includes any combination of endometritis, salpingitis, tubo-ovarian abscess and pelvic peritonitis. 82 The principal clinical and economic significance of CT lies in the fact that it is a causal agent of PID and the further damage to reproductive health that follows from salpingitis. 33,83,84 Confusion can arise with the term PID. Historically, it was often used interchangeably with the term ‘acute salpingitis’ as detected at laparoscopy, which, historically, was the gold standard for diagnosing PID. 85,86 More recently endometritis, which is associated with salpingitis but can be detected by endometrial sampling, has been equated with PID because of the unacceptability of performing laparoscopy on large numbers of women with minimal symptoms. 85–87

Pelvic inflammatory disease is not easy to diagnose and the criteria for a clinical diagnosis of PID have changed over time. 33 In the original Lund studies33 the minimum criteria for diagnosis were: (1) lower abdominal or pelvic pain of < 3 weeks; (2) a purulent vaginal discharge diagnosed on microscopy or painful intermenstrual bleeding; and (3) increased motion tenderness of the uterus and adnexa on bimanual examination. In the UK national PID guideline 2011,83 recent onset of lower abdominal pain in association with local tenderness on bimanual examination is now considered sufficient to establish a diagnosis and initiate treatment. This is because it is now recognised that many women with salpingitis have subtle or mild symptoms. 63,83 Even when present, the clinical symptoms and signs of PID lack sensitivity and specificity for detecting women with salpingitis. 83 Because of the difficulty of diagnosis and the potential for damage to the reproductive health of women (even by apparently mild PID) and the benefit of early antimicrobial therapy, health-care providers are now advised to maintain a low threshold for the diagnosis of PID. 83

Laparoscopic diagnosis of PID, which identifies women with salpingitis, is no longer undertaken in women with a clinical diagnosis of PID. However, clinical information can be used to classify PID as ‘possible’, ‘probable’ and ‘definite’ PID, based on Hager’s criteria. 88,89 Table 1 details the clinical definitions adopted by Taylor-Robinson et al. ,88 based on a modified version of the Hager criteria, and which were used to estimate the proportion of women with salpingitis following a clinical diagnosis of PID. This classification is often used in clinical trials such as POPI, and in studies of patient data such as the GPRD database, although in recent work using GPRD ‘probable’ PID includes chronic disease as well. 19,88,90 As the probability that patients diagnosed with PID based on clinical presentation have salpingitis decreases, the likelihood of other diagnoses, such as endometriosis, irritable bowel syndrome and functional pain, increases.

For the purposes of this report we define clinically diagnosed PID as women with ‘probable’ and ‘definite’ diagnostic criteria. Women with ‘possible’ diagnostic criteria may be treated for PID depending on to where they present, which is probably less likely in primary care compared with departments of genitourinary medicine (GUM) as there is greater expertise in managing PID in the latter setting. Diagnostic thresholds have changed over time as the role of PID in reproductive damage has been recognised. Accordingly, the proportion of women with PID regarded as having ‘probable/definite’ or ‘possible’ PID diagnostic criteria is likely to have increased over time, with fewer women remaining symptomatic and undiagnosed. 19,90 This probably began in the mid to late 1990s in the UK.

Infections associated with pelvic inflammatory disease

Pelvic inflammatory disease is not exclusively caused by CT, with about 30–40% of PID cases being CT+37,82 (see Chapter 7). A range of STIs have been implicated, particularly gonorrhoea and, more recently, Mycoplasma genitalium. 63,83,91 It is also likely that PID can be caused by micro-organisms associated with bacterial vaginosis (BV), which is commonly present in women with PID. 63,83,92 These micro-organisms include Gardnerella vaginalis and anaerobes (including Prevotella, Atopobium and Leptotrichia),83,93 which are probably sexually transmissible but would not be classed as STIs. 92,93 Anaerobes are isolated more often from women with severe PID. As a consequence, broad-spectrum antibiotic therapy is recommended for PID to cover Neisseria gonorrhoeae, CT, and a variety of aerobic and anaerobic bacteria commonly isolated from the upper genital tract in women with PID. 83

Bacterial vaginosis is a common condition affecting women of reproductive age. 94,95 It is a condition characterised by vaginal flora imbalance, in which the normal plentiful lactobacillus is scarce and other anaerobic bacteria abundant. 94–96 It is present in between 12% and 25% of women. 94,95 Women are usually asymptomatic and if symptoms are present these consist of vaginal discharge with or without an odour. 63 Only some BV-associated bacteria have been associated with PID, and it remains to be established whether or not these bacteria are indeed causally implicated in reproductive damage. 63

Undiagnosed pelvic inflammatory disease and asymptomatic pelvic inflammatory disease

Over 50% of PID episodes in the UK are treated in primary care, but it is also treated in STI clinics and in hospital. Data from the GPRD, Hospital Episode Statistics (HES), and KC-60 returns from STI clinics for 2002 are shown in Table 2, which is based on the ‘definite/probable PID’ definition. 90 We have assumed that women admitted for inpatient treatment of PID in the HES data, and women diagnosed at departments of GUM by clinicians trained in diagnosing PID, are ‘probable’ or ‘definite’ cases. An unknown proportion of diagnosed PID episodes meeting these definitions will be recorded under more than one of these locations, although the study by Nicholson et al. 97 using the GPRD database suggests that approximately one-third of cases of probable or definite PID diagnosed in primary care have been treated elsewhere.

These figures on diagnosed PID incidence certainly underestimate the true incidence of PID episodes. It is widely accepted that a relatively high proportion of PID is undiagnosed. This is inferred from historical studies of TFI. PID is considered to be a necessary precursor of TFI,98 and yet a high proportion of women with TFI have no history of clinical PID. 99–102 Wolner-Hanssen103 found that only 34% of TFI cases reported a previous diagnosis of PID. However, only 11% reported never having had clinical symptoms. 103 The implication is that a large proportion of the PID that causes TFI is undiagnosed, but it is relatively unusual for PID which causes reproductive damage to be completely asymptomatic. We therefore distinguish diagnosed PID, by which we mean PID meeting the ‘probable/definite’ criterion, undiagnosed symptomatic PID and asymptomatic or silent PID (see Chapter 7).

Therefore, even if we maintain a single diagnostic criteria, such as ‘probable/definite’, across all studies, and these definitions were consistently applied – which we know is not the case – it would still be the case that the proportions of PID that would be ascertained would depend on the study design. For example, the routine PID statistics in Table 2 depend on patient self-referral, whereas in a prospective study of women attending a STI clinic, or in the context of a screening trial, a higher proportion of PID meeting the same definition may come to light (see Chapter 7).

Current guidance on pelvic inflammatory disease management

Recognising the difficulty of achieving a definitive diagnosis, the British Association for Sexual Health and HIV (BASHH) guidelines83 emphasise a low threshold for empiric treatment, to lower the risk of subsequent EP and infertility. The same advice is seen in primary care guidelines and the National Institute for Health and Care Excellence (NICE) Clinical Knowledge Summaries. 104,105 However, the latter cites multiple partners, a recent new partner, young age and previous history of PID or a STI as risk factors for PID. The NICE Clinical Knowledge Summary104 explicitly states that ‘PID is almost always due to a sexually transmitted disease’. This contrasts somewhat with BASHH guidelines,83 which acknowledge that only 25% of cases of can be accounted for by CT or N. gonorrhoeae. US guidelines highlight young age as a risk factor but avoid citing any behavioural risk factors. 63 All guidance advocates early use of broad-spectrum antibiotics.

Age profiles of Chlamydia trachomatis and pelvic inflammatory disease

In this section we briefly compare the age profiles of CT and PID, to draw attention to an important, but generally overlooked, issue in the natural history of CT. Estimates of CT prevalence are approximately fourfold higher in women aged 16–24 years than women aged 25–44 years. 106 The age profile of incident PID is very different. HES show an increasing trend in PID incidence rates with age (see Table 2). Evidence from the GPRD shows incidence rates that are higher in the age group 20–24 years, but the difference in PID incidence rates across ages is far less than the difference in chlamydia prevalence. The age profile in PID presenting in STI clinics is more similar to the age profile of incident CT, in that there is a relatively high proportion of young women. However, of the three sources, these make up the smallest proportion of diagnosed cases.

There are three plausible contributory reasons for these discrepancies in age profile. First, the PID diagnosis rates may increase with age. There is some evidence that the severity of PID might increase with the number of previous PID episodes. The proportion of women who become infertile because of tubal scarring from adhesions, and the proportion of pregnancies that are ectopic, appears to increases multiplicatively with number of previous PID episodes,33 which is likely to be positively correlated with age, and it is likely that more severe PID has a higher chance of being diagnosed and hospitalised. This hypothesis is supported by the dramatic increase with age of hospital-diagnosed PID.

A second possibility is that the probability that a case of CT causes an episode of PID may be higher in older women, perhaps owing to more previous episodes of CT, other STIs and PID. 107

Finally, the proportion of PID that is caused by CT may decline with age (see Chapter 7), possibly because of differences in the age distributions of the other risk factors for PID. Facultative and anaerobic bacteria are frequently isolated from women with PID; although often present in the normal vaginal flora, they occur more often and in increased concentrations, in women with BV,82 which has been associated with PID. Interestingly, isolation of endogenous bacteria from the upper genital tract appears more common in older women and may be associated with severe suppurative disease, including tubo-ovarian abscess and recurrent PID. 82,108 Although BV does not appear to be associated with age, it is associated with increasing lifetime number of sexual partners. 94,109

These three different explanations would have quite different public health implications. Analyses in subsequent chapters throw some light on the issues, but we do not believe it can be resolved on the basis of existing evidence, and this is addressed in the research recommendations in Chapter 12.

Salpingitis

Much of our knowledge of the impact of PID on reproductive health is based on the Swedish Lund study,33,34,84,110 which recruited PID cases admitted to hospital between 1960 and 1984, and followed them forward to observe reproductive outcomes over a median 8-year period. The study was based on laparoscopic examination. Women with macroscopic inflammation of the fallopian tubes, ‘salpingitis’, were distinguished from PID cases with no macroscopic inflammation at laparoscopy. The more severe the macroscopic salpingitis, the greater the risk of adverse reproductive outcomes. Remarkably, the incidence of TFI and EP in the control group, those admitted to hospital with PID but with no salpingitis on laparoscopy, appeared to be no different from the general population, even though a lower genital tract infection was identified in the majority of cases. 33 It is likely that many of these women would have had endometritis or microscopic inflammation of the fallopian tubes but some will have had other pathology or causes for their pelvic pain. 83,111,112 Incidence of EP and TFI among women with salpingitis depended on age, the number of clinical PID episodes – and thus presumably on the number of salpingitis episodes – and on the severity of the salpingitis on admission (see Chapters 9 and 10).

These results raise the question: what proportion of clinically diagnosed PID is accompanied by an underlying macroscopic salpingitis? Unfortunately for our purposes, laparoscopy is no longer a routine procedure in examination of PID, so it is not possible to match up the definitions of PID and salpingitis employed in the Lund study33,34,84,110 to current practice or to recent research studies. However, we can deduce from the fact that they were hospitalised that the cases of PID recruited into the Lund study33,34,84,110 would be at the more severe end of the spectrum. What is known is that the proportion of clinical PID cases with salpingitis fell during the course of the Lund study33,34,84,110 from an initial 80% down towards 60% among those recruited at the end of the study. 110 In the analyses we present in Chapters 10 and 11, we have assumed that the proportion of PID cases with salpingitis has continued to fall.

Finally, there is little quantitative information on whether the organism causing PID impacts on the probability or extent of reproductive damage. In the Lund study33,34,84,110 it was observed that the severity of CT salpingitis at laparoscopy was generally greater than would be expected from the relatively benign clinical picture. 101 For PID with salpingitis (including that caused by CT), delay of treatment for more than 3 days from onset of symptoms increases the likelihood of tubal damage. 84 Therefore, if the clinical symptoms of CT-related PID appear to be less severe than gonococcal PID, it is possible that women with chlamydia PID may present later than women with gonococcal PID,101,113–115 and as a result suffer more reproductive damage. The same comments apply to women with asymptomatic PID.

Ectopic pregnancy

In a normal pregnancy the fertilised egg implants in the uterus, where it divides and develops into an embryo and, eventually, a fetus. An ectopic pregnancy is one in which the egg implants outside the uterus, most often in the Fallopian tubes. EP is an important cause of maternal morbidity and, occasionally, mortality. 116 It usually presents with acute symptoms including pelvic pain and vaginal bleeding, although increasingly EPs are diagnosed before the onset of symptoms, allowing early, conservative treatment. Prevention of maternal morbidity and mortality relies on early diagnosis and appropriate management, which is almost invariably in a specialist setting. 116

Among the common causes of EP are fallopian tube damage because of salpingitis, surgery and smoking. 116 In a normal pregnancy, cilia in the Fallopian tubes carry the fertilised egg towards the uterus. If the cilia are damaged, the egg may become lodged in the tube. PID is an inflammatory process that can result in the development of scar tissue in the tubes. If both tubes become entirely blocked then fertilisation cannot take place at all. This constitutes a single common pathway leading from salpingitis to both EP and to TFI (which will be exploited in the modelling in Chapters 9 and 10). 98,116

Data on overall EP rates in England are based on HES (HES code O00). 117 If these are compared with conception rates (maternities plus abortions),118 it is evident that the EP rate, defined as the number of EPs divided by the number of conceptions, has remained remarkably constant between 2000 and 2009 (Figure 3), at around 1% of conceptions. Similar rates have been reported in France, although there was evidence that a fall in intrauterine device (IUD)-related EPs had masked a rise in EPs caused by salpingitis. 119 EP rates reported in the USA have been somewhat lower, but, again, have remained stable over recent years. 120

FIGURE 3.

Proportion of conceptions that result in EP, by year and age group.

Although the overall conception rate falls steeply from the age of 30 years onwards, the proportion of conceptions that are EPs rises with age (see Figure 3). The same phenomenon has been observed many times. 120–122 This points to the role of cumulative exposure to the causal factors, whether these are smoking or tubal damage, which can only increase with age.

The fact that EP risk links to cumulative exposure to risk factors, including salpingitis, rather than to, say, incident exposure, is critical to understanding the role of PID. 123 Egger et al. 121 report an ecological analysis of EP rates and CT prevalence by age and location in Uppsala county, Sweden. They establish that the correlation between CT prevalence and EP rates is strongest in 20- to 24-year-olds, and that it becomes increasingly weak in the 25- to 29-year-old, 30- to 34-year-old, and finally the 35- to 39-year-old groups. At the same time the regression of EP rate against CT prevalence becomes increasingly steep, which would not be expected if variation in EP rates was driven by variation in CT incidence. However, these phenomena can be readily explained if cumulative salpingitis incidence is the variable determining EP rates. Most CT occurs in the younger age groups, but current incidence and prevalence become increasingly poor predictors of cumulative exposure to CT as age increases. On the other hand, the regression against recent incidence becomes increasingly steep with age – although with increasing error about the regression – because cumulative incidence is higher per unit of current prevalence. These ecological observations provide further support for the conclusions we draw from Figure 3 in the previous paragraph.

Another important feature of the relation between salpingitis and EP is the strong ‘dose–response’ relationship reported in prospective studies following women with salpingitis. The Lund studies,33,110,124 in particular, followed women for an average of approximately 8 years from hospitalisation for PID. The proportion of women whose first subsequent pregnancy was an EP increased systematically with severity as determined by laparoscopy, and also with the number of previous salpingitis episodes. 124 Among the women who subsequently conceived, those with two, or three or more previous episodes of salpingitis had a non-linear increase in risk of EP (see Chapter 9). A similar pattern of findings was reported for TFI. 33 A record linkage study in the USA following high-risk women with a diagnosed CT infection over an average 6-year period again reported a similar increase in risk of EP, but this time in relation to the number of previous CT infections rather than the number of PID episodes. 107 These studies constitute some of the strongest evidence for a causal link between salpingitis (PID) and EP. Of particular significance is the fact that the dose–response relation is exactly what would be expected on the basis of cumulative exposure to risk factors. In addition, the similarity of the ‘dose–response’ effect attaching to CT and salpingitis suggests that the risk of EP may be approximately the same whatever the cause of the salpingitis, although the agreement could be coincidental if the aetiology changes with age in a way that approximately cancels out any differences.

Tubal factor infertility

Women are usually considered ‘infertile’ if they have tried to have a baby for ≥ 1 year and have been unsuccessful. Defined in this way, infertility is clearly not an absolute condition, as only 85% of couples succeed in conceiving within 1 year, whereas 95% may succeed within 3 or 4 years. 125 This immediately reveals a major difficulty in the study of infertility, as apparently minor changes in definition readily generate threefold changes in the number considered infertile.

Infertility is a complex construct: the infertile population includes both couples who are sterile, and couples who are subfertile. The former cannot conceive, but the latter can, if given enough time. Population surveys of infertility in the UK126–128 distinguish between primary unresolved failure to conceive, between 2% and 3.5%, representing women who have not conceived at the end of their reproductive life, and secondary infertility, which includes women who have conceived at least once, but who subsequently experience difficulty in conceiving. Secondary infertility must be included in our study because it is quite possible for salpingitis to be acquired by women who have already had children.

Surveys reporting secondary infertility have reported the numbers who have experienced at least two years of being unable to conceive. 126–128 According to one study,126 the secondary infertility group included women who had already had one or two, and in one case eight, previous children. The 2–3.5% estimates of secondary infertility from these surveys are therefore probably an overestimate, but the extent of the overestimate depends on the distribution of duration of infertility.

A further problem is the difficulty of estimating the proportion of women who are infertile at any particular age. Not only is it necessary to ensure that their infertility has lasted sufficiently long to meet a specified definition, but infertility can only manifest itself, and thus be diagnosed, in women who are trying to become pregnant, although the underlying condition may have been present for a long time. This contrasts with the study of EP where reliable age-specific incidence data are available from routine hospital statistics.

To avoid these problems in this monograph we have followed a strategy that has been adopted in a number of epidemiological studies of infertility in the UK,126–129 which is to focus attention on the proportion of the population who are infertile, which specifically means unable to conceive, at the end of their reproductive life, which we take to be the age of 44 years (see Chapter 10).

Tubal factor infertility is one of several causes of female infertility (Figure 4). 130 According to a 2006 survey by the Human Fertilisation and Embryology Authority (HFEA), 22% of those being treated by in vitro fertilisation (IVF) have TFI (Table 3), but a further 30% of infertility in women undergoing treatment is unexplained. 131

FIGURE 4.

Types of infertility treated 2000–7. Source HFEA.

The Lund studies124 established the key relationship between PID and TFI, which parallels the relationship between EP and TFI. A higher rate of TFI was observed only in PID cases with confirmed salpingitis on laparoscopy, and the risk increased with the number of subsequent PID episodes. Critically, none of the 601 control group women with PID who were not confirmed as having salpingitis went on to have TFI. A difference between EP and TFI, however, is that TFI can occur only following an infective process,35 namely salpingitis; EP on the other hand has a multifactorial aetiology.

There is an extensive literature on the role of CT in TFI. Much of it consists of retrospective studies demonstrating a higher prevalence of serum antibody to CT in TFI cases compared with non-TFI controls. Because of the poor sensitivity and specificity of the assays, the imprecise characterisation of their diagnostic accuracy, and the variation in cut-offs, these studies have not generated reliable quantitative estimates of the role of CT in TFI. However, a number of these retrospective studies have reported marked differences in the titre distributions in antibody-positive TFI cases and controls (see Chapter 11). The higher prevalence of high titres in TFI cases constitutes powerful evidence of the causal role of CT.

Chronic pelvic pain

Chronic pelvic pain is a well-recognised complication of PID, which has a significant morbidity. 33,132 It is defined as non-menstrual pain of ≥ 6 months’ duration that is severe enough to cause functional disability or require medical or surgical treatment. Chronic pelvic pain is common in women, with an estimated prevalence of 3.8% in women aged 15–73 years in the USA. Often the aetiology is not clear. There are many physical disorders of the reproductive tract, gastrointestinal system, urological, musculoskeletal system and psychoneurological system that may be associated with chronic pelvic pain in women,132 and these may often co-exist. Although adhesions resulting from pelvic inflammatory disease are diagnosed in about 25% of women undergoing investigation, it is not clear whether these are always the cause of the pain. 132 The prevalence of adhesions in women with chronic pelvic pain will vary depending on the population studied. Most information is obtained from women attending secondary and tertiary care, who are likely to experience more severe forms and may not be representative of women with chronic pelvic pain in the community. An accurate diagnosis of pelvic adhesions is only possible following laparoscopy and, even then, it is not possible to ascertain the cause, as the infectious agent will have usually resolved prior to investigation.

We have therefore chosen not to investigate chronic pelvic pain because of the paucity of data on the role of PID, and the lack of evidence on the contribution of CT.

Premature delivery, neonatal pneumonia and conjunctivitis

There is conflicting evidence that CT may result in miscarriage and/or premature labour. One hypothesis is that adverse pregnancy outcomes may only occur following primary infection. No studies have investigated this since 1990. In view of the limited and conflicting data on adverse pregnancy outcomes, we decided not to include this in our work plan.

Chlamydia trachomatis is transmitted to approximately half of children of mothers infected with CT and about 10–50% of these develop neonatal conjunctivitis and/or pneumonia. This was recently reviewed by Thorne133 for the NSC. Data in the UK are based on those presenting to hospital and thus biased towards severe and prolonged cases. Neonatal infection with chlamydia is usually asymptomatic, and the most common manifestations (conjunctivitis and pneumonia) are non-specific. Furthermore, chlamydial disease is usually of a mild to moderate nature and easily treated or is often self-limiting. For these reasons, perinatal and neonatal outcomes have not been included in this study.

Bayesian Markov chain Monte Carlo in WinBUGS

We have adopted Bayesian methods throughout this monograph. 150 Bayesian inference seeks to find the probability distribution of parameters (the posterior distribution) given the data (likelihood) and a prior distribution representing beliefs before the data were available. Classical inference (maximum likelihood) considers the parameter values at which the probability of obtaining the data is greatest. The Bayesian viewpoint is generally considered to be the more ‘natural’, in that people want to know what they can say about parameters given the data, not what they can say about the data given specific parameter values. However, it has been considered to have two disadvantages: the ‘subjectivity’ of the prior distribution and the computational difficulty of carrying out the high-dimensional integration required. We have avoided the subjectivity issue by using vague priors wherever possible (but see below). Advances in statistical theory151 and computing power, and the development of highly flexible programming packages, such as WinBUGS,44 have not only solved the computational problems, but also have made Bayesian methods the common choice for complex synthesis. 42

The general principles of Bayesian MCMC are set out in standard texts,152 and the procedures for Bayesian MCMC using WinBUGS are to be found in the WinBUGs on-line user manual. 44 The WinBUGS user is required to specify a model in algebraic terms. The model consists of (1) specification of prior distribution for ‘basic‘50 parameters; (2) definitions of ‘functional’ parameters in terms of basic parameters, effectively the ‘model’; and (3) specification of data likelihoods. Each item of data informs either a basic or a functional parameter.

Given some starting values for the basic parameters, parameter distributions converge to the joint posterior distribution. The WinBUGS 1.4.3 software makes use of a number of different MCMC algorithms. The most commonly used is Gibbs sampling,153,154 in which samples are drawn from the conditional distribution of each parameter in turn, conditioning on the other parameters and on the data. Eventually, the marginal distributions of the parameters converge to their true marginal posterior distribution. In essence, all the parameters are uncertain: the marginal distribution expresses the uncertainty in each parameter, averaging over the distributions of all the other parameters.

Users find out about the posterior distributions of parameters of interest by sampling from them. Posterior means or medians are commonly cited, along with 95% CrIs or posterior standard deviations (SDs). CrIs have the interpretation that many people attribute to classical confidence intervals (CIs): namely that they express the probability that the parameter lies within that range, given the model, priors and data. Posterior SDs have the same interpretation as standard errors of estimates in classical statistics.

Because of the need for the process to converge, MCMC users must pay careful attention to a number of technical issues. First, diagnostics must be used to assess whether or not the process has converged. The Brooks–Gelman statistics155 are commonly used as these are available in WinBUGS. The initial samples – the ‘burn-in’ – are discarded. Second, a large number of posterior samples may be needed to obtain accurate summary statistics. It is usually recommended that the process is run until the Monte Carlo error in the mean is < 5% of the posterior SD. It is also good practice to base posterior inference on more than one MCMC chain, because occasionally chains converge to different posteriors. Our practice here is to use two chains, to check they converge to the same point, to use a conservatively large burn-in period, and also conservatively large numbers of posterior samples. In all the analyses in the monograph convergence has been achieved within about 25,000 samples, and we discard the first 50,000 as ‘burn-in’. Posterior inference is based on 200,000 samples, 100,000 from each of two chains.

Sensitivity analyses

A further issue arising in Bayesian analysis is sensitivity to prior distributions. 150 Our general approach has been to introduce all information through likelihoods, and to put vague priors on all the basic parameters. When data are relatively sparse, or when the number of parameters to be estimated is large in relation to the amount of data informing them, results can be sensitive to the distributional form of the priors. Sensitivity analyses are therefore included wherever this is the case. The main findings from sensitivity analyses and their implications for the conclusions are discussed in the main text, but the full details are provided in the appendices.

In some cases it is preferable to fix parameters at a range of set values, to see the impact on the final results, rather than give the parameters a distribution. This, again, may generate a series of sensitivity analyses that are reported in full only in the appendices.

Model diagnostics and model selection

Our methods for assessing goodness of fit are based on Spiegelhalter et al. 156 The key statistic is the posterior mean residual deviance D¯, corrected for the saturated model. This is the familiar residual deviance statistic from classical statistics, except that it is computed on each MCMC cycle rather than at the maximum likelihood values of the parameters. As a result a distribution of samples is generated and the mean is taken as a measure of goodness of fit. In a good-fitting model, the summed residual deviances should approximate to the number of data points. 156,157 When comparing models with different numbers of parameters a common approach in classical statistics is the Akaike Information criterion,158 in which the models are ‘penalised’ by adding the number of parameters to the deviance statistic. We adopt this approach where appropriate using the Bayesian Deviance Information Criterion (DIC):156,159 DIC=D¯+pD, where p_D represents the effective number of parameters. This statistic is calculated by WinBUGS and is available through the DIC tool. In models lacking any hierarchical elements, p_D is simply the number of free parameters in the model. With hierarchical models it reflects the degree of shrinkage. We make sparing use of hierarchical models in this monograph because of difficulties of interpretation (see Chapter 4), and they play no role in our final estimates.

D¯ and DIC can be examined as measures of global model fit, or the contributions of individual data points to D¯ and DIC can be examined in order to locate more precisely the particular data points that are in conflict with the other data under the model. It should be emphasised that these statistics are sensitive to prior distributions. This is one of the reasons why we tend to put vague priors on all parameters and introduce all of the information through data likelihoods. In this way, we can use residual deviance to assess conflict between different data sources.

There are exceptions to this rule, however, where we introduce evidence through informative priors. This is done where the parameter’s role is local to a source of data, or when it is not central to the issues under study. In these cases, we may also use the WinBUGS ‘Cut Function’. 44 This is a special function in WinBUGS that allows users to impose a prior distribution on a set of parameters that is not updated by the likelihood (the posterior is equal to the prior). The flow of information from elsewhere back to those parameters is blocked. This is done in cases when we wish to impose a belief (probability distribution) on a set of parameters, which we do not want to be altered by the data, or more pertinently, when we want the key parameters to be estimated conditional on a set of specific assumptions over which we have complete control.

An example that illustrates the reasoning behind this is the duration of symptomatic CT infection in women. It was difficult to source evidence for this parameter from formally collected data, and we turned, instead, to expert clinical opinion. The duration of symptomatic infection is the sum of the ‘incubation time’ between infection and the appearance of symptoms; the time from symptoms until treatment is obtained; and the time from treatment to clearance of infection. It was considered by clinicians that a uniform distribution between 4 and 8 weeks would represent clinically informed belief about this parameter, and this was the chosen prior. This information was combined with information on incidence, prevalence, duration of asymptomatic infection, and the proportion of infections that are symptomatic, all of which had been based on formal consideration of likelihoods of data sourced in a reproducible way. If all the evidence was combined without the ‘cut’ function, the incidence and prevalence information would have changed our view of the duration of symptomatic infection.

In our view, this was not appropriate. We therefore used the ‘cut’ function to prevent this flow of information between sets of parameters. This allows us to estimate the other parameters, conditional on the assumption that this duration parameter is 4–8 weeks. This is clear and reproducible. If we built a full probability model with feedback, we would have to report that the parameter estimates were conditional on having a prior of 4–8 weeks on the duration parameter, but leading to a posterior distribution for duration that was somewhat different. We do, of course, assess the sensitivity of the results on incidence, prevalence, and other parameters to this assumption.

WinBUGS Development Interface and WinBUGS Differential Interface

WinBUGS itself has a limited repertoire of mathematical functions and statistical distributions. However, WBDev (WinBUGS Development Interface)160 allows users to specify their own mathematical subroutines in the Component Pascal programming language. Although the functions must be evaluated on every MCMC iteration, the programming is ‘hard wired’ in a way that allows for very efficient computation. WBDev is used in to solve the discrete time Markov chain for repeat PID in Chapter 8.

A second WinBUGS ‘add-on’, WBDiff (WinBUGS Differential Interface)161 allows users to specify systems of first-order ordinary differential equations, which are solved on each MCMC iteration using the Runge–Kutta method. WBDiff is used to solve the equations relating transition rates to transition probabilities in the continuous time Markov model fitted in Chapter 6.

When these add-ons are used, the WBDev and WBDiff code are provided in the appendices.

Directed acyclic graphs (influence diagrams)

The directed acyclic graph (DAG), or influence diagram, is a way of portraying the relationships between the data and the model parameters. The DAG is a formal entity in Bayesian statistics, and, in the WinBUGS language,44 it is in fact possible to specify a model as a DAG rather than write WinBUGS code. In this monograph, following the previous work in Bayesian evidence synthesis,42,50,52 we use schematic DAGs to show how all of the data sources ‘fit together’ around the synthesis model.

The schematic DAG in Figure 5 illustrates how they are used and how they should be interpreted. The figure shows four numbered data items in rectangles, and five model parameters α, β, γ, δ, θ in ellipses. The direction of ‘influence’ is shown in the direction of the arrows. The model is said to generate the data, so that the data are always pointed to. (In cases where information is introduced via informative priors, rather than as data, the information is shown in a rectangle, with an arrow pointing to the parameter.) Three of the five parameters, α, β, δ have no ‘parents’, that is, no arrows point to them, only from them. These are called ‘basic’ parameters,50 and they are the parameters that must be assigned prior distributions. The remaining two parameters γ, θ are ‘functional’ or ‘logical’ parameters, and they can always be defined in terms of the basic parameters. The functional relationships between parameters are shown as equations in the ellipses, and these equations would be exactly those in the WinBUGs code.

FIGURE 5.

Directed acyclic graph. (a) Two sources of evidence on a parameter; and (b) one source is blocked from influencing the parameter.

Note that data can inform either basic or functional parameters. Also, the same model can be parameterised in many different ways, with different parameters defined as basic or functional. As long as the data are not too sparse in the relation to the number of parameters, this should not have any material impact on estimates.

Methods to examine evidence conflict

There are effectively two evidence groups in Figure 5a, each representing a different source of information on the parameter θ. One source of information is from Data item 4, the other is a combination of information from Data items 1, 2 and 3. Figure 5a represents a DAG in which both sources of information are combined to give a pooled estimate of θ. In this situation the inclusion of Data 4 not only informs θ, but it also feeds back to impact on the estimates of all the other parameters. We can examine conflict between the two groups of evidence by simply blocking Data item 4, to obtain an estimate of θ in its absence (see Figure 5b). Similarly, we can block out the other data to see what the estimate of θ would be if informed from Data item 4 alone. This technique is a version of ‘node splitting’. 162,163

A very useful way of assessing conflict between different sources of evidence visually is to plot the posterior densities. In this way, the posterior distributions of the parameter based on separate sources of evidence can be compared. The posterior density when the different sources are pooled can also be plotted on the same graph. This approach has been used by Dias et al. 164 in the context of splitting nodes in mixed treatment comparison evidence networks. A further refinement is to generate ‘significance tests’ based on Bayesian p-values: these test the hypothesis that one source of evidence generates a posterior distribution that is different from the other source.

Evidence identification

Our target parameter is the mean duration of an episode of untreated, asymptomatic CT in women. Studies in any country on any population were included, as long as they reported sufficient information to determine: numbers at risk of clearing infection, numbers who cleared infection, and total time at risk. We define an episode of CT as the time from infection to the time that the infection has cleared, in the absence of treatment. Studies of CT duration have the inherent limitation that patients may clear infection and be re-infected during the follow-up period. For this reason we consider same-partner re-infections, which microbiological evidence suggest comprise the great majority of re-infections169 to be part of a continuous episode. Symptomatic infection is defined as infection that is diagnosed and treated.

We identified studies through three recent reviews,74,139,142 reference searches of published economic and population transmission models,1,2,8,29,140,141,170–174 and consultation with an expert panel (see Acknowledgements). A 2012 search identifying papers that had cited the three reviews revealed no further studies of duration of asymptomatic infection in women. The recent review paper by Geisler et al. 74 was part of an authoritative review of CT epidemiology set up by an expert group of the CDC. 38 Two studies175,176 were excluded because patients were treated for CT or other infection soon before recruitment175 or during follow-up. 176 Three were excluded because they provided insufficient information. 177–179 Results from the 11 studies73,79,165,169,180–186 included are shown below (see Table 4).

Initial consideration of the data: study design

The studies differed in their design with some being set in STD clinics, and others based on women identified as infected by population screening (Table 4). Two were set in antenatal clinics, where women were screened for CT at a prenatal visit. In most studies, patients were followed forward from an initial examination, when samples were collected, until they returned for a second visit, when CT status was assessed again; patients attended for the second test after different intervals but each patient was retested only once. In other studies,169,180 the same patients were tested repeatedly over several follow-up periods: those at risk in subsequent periods had already ‘survived’ the previous period without clearing infection.

A preliminary analysis is presented in Table 4, showing the numbers clearing infection during each follow-up period i of study j with mean follow-up time t_ij. (Where this was not reported, we estimated it from the data at hand: details in Appendix 1. ) We assume that in each study the clearance of infection in each woman is the result of an independent Bernoulli trial, and that the number clearing infection in each period is therefore binomially distributed with parameter θ(t_ij), and a constant clearance rate λ_ij. As the observations are interval censored, the exact clearance times are unobserved, but we can obtain an estimate of the clearance rate λ_i and CIs through the relation θ(t_ij) = 1– exp(–λ_ijt_ij).This is the complementary log–log (cloglog) link commonly assumed for binomial data generated in this way. 187 The mean CT duration Δ_ij in that segment is estimated by Δ_ij = 1/λ_ij.

The crude estimates of mean duration vary by over 40-fold, varying between studies and also within. Although it is possible to fit either FE or RE models to these data, neither would capture the strong association between clearance rate, length of follow-up and study design. The studies based in STD clinics tend to have much shorter follow-up periods, and have a higher clearance rate than the longer-term studies, which tend to be screening studies. In the STD studies, women were recruited only if they were both asymptomatic at first attendance and had no indicators, such as a sexual partner with CT, that would trigger immediate treatment. Our interpretation is that women who attend STD clinics in the absence of any symptoms are most likely to be seeking to assess their infection status following a possible recent sexual exposure to CT. Those found to be infected in this setting might therefore have their first sample collected within a few days of the date of infection, and thus represent incident cases of infection.

In screening studies, observations are left truncated as well as interval censored: those found to be infected have already ‘survived’ clearance for an unknown period prior to screening. Screening studies therefore reflect prevalent infection. This, together with the apparently slower clearance rate seen in screening studies, provides a prima facie case for believing that CT infections may clear at different rates: screening studies may generate apparently longer clearance rates because they selectively oversample the longer-lasting infections, as short-duration infections are more likely to clear before screening, a phenomenon known as length-biased sampling. 188

The high rates of clearance reported at the shortest follow-up times are consistent with observations made in Chapter 2 on the likelihood of ‘passive infection’ (see also Joyner et al. 79), according to which CT is recoverable from the female genital tract for several days after sexual intercourse, but without leading to an active infection.

This provides a basis, independent of the evidence in Table 4, for a two-class model comprising passive and true infections. However, examination of Table 4 suggests that, even within the screening studies, longer follow-up times seem to be associated with lower average clearance rates. We have therefore also considered a three-class model, comprising passive infection, and fast- and slow-clearing infections. These alternative conceptions of the natural history of transmission, infection and clearance can be captured by mixture-of-exponentials models. We return to a consideration of the a priori scientific plausibility of both the two- and three-class models in the discussion, where the role of alternative survival models is also considered.

We begin by describing the standard FE and RE models for these data. We then describe the parameters of the mixture of exponentials models and develop expressions for their functional relationship to the parameters in the likelihood.

Statistical models for studies of the duration of Chlamydia trachomatis infection

Mixture of exponentials model

In both FE and RE models described above, within every study, each woman has the same clearance rate. We now develop the three-class mixture model. This allows women within every study to have one of three different rates. It is the same three rates that are present in every study. We then introduce the two-class model as a special case. As we now show, there are three types of data: each provides information on a different function of a common set of parameters. For each type of data we define the likelihood and explain the relation between the parameters in the likelihood and the common set of rate and probability parameters in the model.

We assume that incident infections are in one of three categories (‘passive’, fast-clearing and slow-clearing). The proportion of incident infections in each of the three classes is p₁, p₂, p₃; p₃ = 1 – p₁ – p₂, and the clearance rates are λ₁, λ₂, λ₃, with each rate having a fixed effect across studies. It can readily be established that the two-class model is a special case of the three-class model, in which p₃ = 0, and that the FE model is a one-class model in which p₂ = 0 and p₃ = 0. We henceforth refer to the latter as the one-class FE model. The schematic DAG (Figure 6) clarifies the relationships between the shared rate and proportion parameters, and the likelihoods generated by the sampling and recruitments schemes represented by the three study designs, under the mixture-of-exponentials models.

FIGURE 6.

Directed acyclic graph of the CT duration synthesis. Subscripts: i, observation; j, study; k, class; n, number followed up; p_k, proportion of infections in class k; r, number of patients clearing infection; t_ij, duration of follow-up period; T_ij, duration of previous segments; w_k, expected proportion of infections in class k; z_ijk, number of infections in class k at risk of clearing in time interval ij; θ(), probability of clearance; λ_k, clearance rate of class k infections.

Sexually transmitted disease clinic studies

We interpret clinic studies as studies of incident infection, as explained above, and the infections observed are therefore a mixture of the three types of infections, the clearance rates of which are exponentially distributed. In any given sample, the denominator number of infections ‘at risk’ of clearing is an unobserved mixture of infection types. The proportions recruited from each class k will vary by chance around the population mean proportions p₁, p₂, p₃. For each observation i in study j, the number of infections z_i,j,k in class k at risk of clearing will be drawn from a multinomial distribution, with probabilities p₁, p₂, p₃ and denominator n_ij: z_{i,j,k = 1,2,3} ∼ Multinomial (p₁, p₂, p₃, n_ij).

The expected clearance probability in follow-up period t_ij is therefore a weighted average of the clearance probabilities for the three classes with the z_ijk as weights:

⁽⁴⁾ θ ( t i j ) = z i , j , 1 n i j . ( 1 − exp ⁡ ( − λ 1 . t i j ) ) + z i , j , 2 n i j . ( 1 − exp ⁡ ( − λ 2 . t i j ) ) + z i , j , 3 n i j . ( 1 − exp ⁡ ( − λ 3 . t i j ) )

Screening studies: follow-up from the initial screen

The expected proportion of incident infections of each type is p₁, p₂, p₃. But screening studies must over-sample longer lasting infections. The prevalence π_k of infection of type k can be derived from the relationship: prevalence = incidence × duration:

⁽⁵⁾ π k = η . p k λ k

where η is the incidence of CT. Thus the expected proportion w_k of women in each group k in a prevalent population is:

⁽⁶⁾ w k = ( p k λ k ) / ( p 1 λ 1 + p 2 λ 2 + p 3 λ 3 )

Again, to take account of the chance variation between the proportions in the general population and the proportions sampled for each observed group, we take the data collected in each observation as a multinomial sample, now with parameters w_k. For the first observation (i = 1) in study j we may now use the same expression (see equation 4) for the probability of clearance, with weights:

⁽⁷⁾ Z 1 , j , k = 1 , 2 , 3 ∼ M u l t i n o m i a l ( w 1 , w 2 , w 3 , n i j )

Screening studies: subsequent follow-up periods

In the Molano et al. 169 and Morre et al. 180 studies, after the first observation period, patients who are ‘at risk’ of clearance in the subsequent follow-up segments should be considered to have ‘survived’ both the original period between infection and screening, and the previous follow-up segments, of duration say T_ij, since screening. This is shown in Table 4. For example in the Molano et al. 169 study, 2 out of 6 cleared CT during a 1-year period of observation, having already ‘survived’ a further T_ij = 3 years since their recruitment to the study. This leads to another, rather complex, sampling structure. The proportion z_i,j,k in each category k for each observation i in study j is determined by two factors. First, a single multinomial sample at the start of the first time period (the time of screening) determines the numbers w_k in each group, exactly as in equation 6. Second, in subsequent time periods there is no further multinomial sampling. Instead, each original sample in category k must be depleted by a further exp(–λ_k T_ij), so that, through substitution into equation 5, and then equation 4, the observed number of women clearing CT in each time period T_ij is equal to:

⁽⁸⁾ θ i j ( t i j , T i j ) = ( ∑ k = 1 3 z 1 , j , k n 1 , j exp ⁡ ( − λ k T i j ) ∑ k = 1 3 z 1 , j , k n 1 , j exp ⁡ ( − λ k T i j ) ( 1 − exp ⁡ ( − λ k t i j ) ) )

Studies based in antenatal clinics

Two studies181,182 were set in antenatal clinics. One clinic181 recruited pregnant women between 8 and 23 weeks of gestation, and the other clinic182 between 36 and 40 weeks of gestation. These studies, in principle, could be considered as having the same recruitment process as a screening study. However, it would also be plausible to consider at least a proportion of those recruited to have been infected at the same time that they conceived, on average, say, 16 weeks or 38 weeks prior to the initial sample. We therefore explored including the antenatal studies under the extreme assumptions that (1) they are left-truncated screening studies, and (2) infection was acquired at the point of conception.

Sensitivity and specificity of diagnostic tests for Chlamydia trachomatis used in the studies

Poor test sensitivity may overestimate the clearance rate, as patients who have not cleared infection may be thought to have done so. Most of the studies use NAATs, which we assumed to be 100% sensitive and 100% specific. Sensitivity is defined as the probability of detecting a CT infection, given that it is present. For example, if sensitivity of culture is ψ then studies of clearance rate that test for CT using culture estimate θ_ij(t_ij)/ψ, rather than θ_i(t_i). Geisler et al. 73 used both NAAT and culture tests, and reported the sensitivity of culture given a previous culture positive result for that infection as 78 out of 86 (90%), and we use this as an estimate of ψ.

Note that we distinguish between test sensitivity in the initial testing to determine infection status, and sensitivity in subsequent tests to establish whether or not the infection has cleared. A failure to detect infection initially will mean that the individual is not recruited into the study, and we assume that this has no bearing on the estimates of clearance rates. Although we can estimate the sensitivity of culture in women who are initially culture positive, it is difficult to estimate sensitivity of NAAT in women who are initially NAAT positive, and we make the simplifying assumption that this is 100%. An alternative, but equivalent, assumption would be that if the bacterial load is so low that it cannot be detected by NAAT, then it does not constitute a clinically relevant infection.

Prior distributions

For the RE model, we place a flat normal prior on the log mean clearance rate and a Uniform(0,5) prior on the log SD, and the effect of this prior is assessed through sensitivity analysis. Because the shortest follow-up period is 1 week, the Table 4 data do not allow us to adequately identify λ₁, the rate of clearance of passive infection, in the three-class model. To improve the identifiability, we set λ₁ = 120 per year, implying that passive infection lasts, on average, only 3 days. We confirmed subsequently that the precise value chosen has negligible impact on the other parameters. We assigned priors of the following form to the clearance rates: log(λ₂) ∼ Uniform(A,B), and log(λ₃) ∼ Uniform(C,λ₂). We chose the values A = –1, B = 4.8 and C = –2.62, corresponding to durations between 3 days and 2.7 years for the duration of a fast-clearing infection and between 1/λ₂ and 13.7 years for slow clearing. As a sensitivity analysis for the three-class model, we varied the values of A, B and C, and also considered placing Uniform priors on clearance rate, and on duration (1/clearance rate), as alternatives. We also estimated a two-rate model, in which p₃ = 0, λ₃ = 0, and log(λ₃) ∼ Exponential(0.001). A prior distribution for the sensitivity of diagnostic tests of ψ ∼ Beta(78,8) was based on the results from Geisler et al. 73 The mixing proportions p₁, p₂, p₃ were given Beta(1,1) and Dirichlet(1,1,1) priors in the two- and three-rate mixture models, respectively.

Estimation of Chlamydia trachomatis infection duration from each model

The target parameter is the average duration of asymptomatic infection. For the one-class FE model, this is simply 1/λ₁. For the RE model, at least two estimates can be generated, depending on how the model is interpreted (see Discussion, below). One would focus on the RE mean, which would generate an estimate of duration. In this case we monitor a node 1/exp(µ) and report its summary statistics. Another interpretation might be that the rate obtained in any future situation is effectively another sample from the RE distribution whose parameters have been estimated from the data. The estimate produced by this ‘predictive’ interpretation is readily found in the MCMC framework, by monitoring the nodes α_new ∼ N(µ,σ²) and 1/exp(α_new), and reporting summary statistics for the distribution of the latter.

In the two-class mixture model, a proportion p₁ of CT+ patients do not have a true infection, so the estimate of CT duration is 1/λ₂. Under the three-class mixture model, the estimate of CT duration is a weighted average of the durations of the fast- and slow-clearing infections:

The WinBUGS code for the two-class model is presented in Appendix 2.

Model fit statistics

There were a total of 28 data points from 11 studies informing CT duration, constituting the independent observations for the evidence synthesis. The models that include the antenatal data assuming it represents prevalent infection fit better than when these data are assumed to represent infection that is incident at the time of conception (results not shown). The model fit statistics (Table 5) show, however, that when the antenatal data (assumed to be prevalent cases) is included none of the models fits well. The residual deviance D¯ is expected to approximate the number of data points (28), but the best-fitting model, the RE model, has a D¯ of 36. The FE and the latent class models fit even worse. If the antenatal studies are excluded, the two- and three-class models now fit the data well. The RE model is an acceptable fit, but the DIC, a measure that sums D¯ and the number of effective parameters p_D, strongly favours the mixture models, as they have fewer effective parameters as well as having a better fit. Further results are reported for only the reduced data set, excluding the antenatal studies, to which we return in the discussion.

Parameter estimates

The rate estimates from each model are shown in Table 6. The rate estimate generated by the one-class FE model (1.1 per year) is lower than the RE model (1.7 per year) because the former is more strongly influenced by the largest studies, which happen to be the screening studies with their longer follow-up and slower clearance, whereas the latter allows greater influence of the more numerous high-clearance-rate studies. A prediction for a new study has an expected rate of 18.5. This is due to the skew in the posterior distribution (median 1.5). The degree of between-studies heterogeneity according to the RE model, as indicated by the between-studies SD, is extraordinary. The posterior mean of the between studies SD is 1.33 on the log scale, indicating that a study on the 95th percentile of clearance rate would have a clearance rate almost 200 times higher than the fifth percentile. The mean clearance rates estimated from the exponential mixture models were 0.74 per year for the two-class model, and 9.5 and 0.64 per year for fast and slow clearers, respectively, in the three-class model. They identify 19–23% of infections as ‘passive’, clearing exceptionally rapidly. Note that the three-class mixture model is barely identified by the data, with the proportions in each class and the class-specific clearance rates having extremely wide CrIs. This is largely a result of the sparseness of data at long follow-up times.

The estimates of the target parameter, mean duration of CT (Table 7), indicate broad agreement between the two- and three-class models (1.36 years, 1.24 years, respectively), and considerably lower estimates from the one-class FE and RE models. The posterior medians of the mean duration in the RE model, whether based on the mean or on the predictive distribution, are barely half those from the mixture models. The RE distribution is so skewed, however, and its variance so large, that the mean of the predictive distribution is 23 years.

The ‘observed’ durations in each observation from Table 4 are compared with those predicted by the two-class model taking into account study design and duration of follow-up, and with duration plotted on a log scale (Figure 7). Note that for the longer-duration screening studies the predicted duration stabilises to a single value, which is, in fact, the expected duration of true CT infection. This is because in a screening study the chances of recruiting a woman with passive infection are very low indeed. This is also evident from equation 6. It is interesting that the results from the longer-duration screening studies systematically deviate from the two-rate model in showing that observed duration increases with length of follow-up. This is evidence that favours a three-rate model but, because the data are so sparse, the global statistical fit does not support the three-rate against the two-rate model.

FIGURE 7.

Forest plot of the ‘observed’ log duration (black) and predicted (green) based on the two-rate model, mean and 95% CrI, for each study observation (in the same order as Table 4). 68 Adapted from: Mixture-of-exponentials models to explain heterogeneity in studies of the duration of Chlamydia trachomatis infection. Stat Med 2013;32:1547–60. Results of the RE meta-analysis are presented at the base.

Sensitivity analysis

Sensitivity to priors was also explored. When the RE variance prior was changed from Uniform(0,5) to Uniform(0,3), this changed the posterior expectation of the RE mean by less than 1%. Similarly, the posterior between-studies SD was changed by only 1%. This suggests the prior is having only a small effect on the between-study variance and we conclude that the Table 4 data are adequate to estimate the RE model.

We examined a two-class model in which λ₁ was not restricted to a fixed value: it did not fit better and has the same residual deviance as the model restricted to λ₁ = 120. It converged to a very high rate of clearance of passive infection, with p₁ estimated at 0.23 (0.16,0.32), very close to the estimate from the restricted model. This finding supports our use of the value of 120 to improve model stability. All reasonable priors for λ₂ gave essentially identical answers. Sensitivity analysis of the three-class models showed that the rate and probability parameters were sensitive to priors, which might be expected from the fact that the model is only barely identifiable. The estimated duration of CT, the target parameter, was, however, relatively insensitive, with values ranging from 1.17 to 1.24 years.

Assumptions

1. Studies of duration have been either based in STI clinics, or they have been studies following individuals whose infection was detected by screening. We have assumed that these two designs recruit incident and prevalent infections respectively.

Summary of findings

1. The data are not compatible with a single-class (FE) model, and the degree of variability implied by the RE model is not scientifically plausible.

2. The exponential mixture models, whether two- or three-class, fit the data best, and accord with our understanding on the biological mechanisms of transmission and infection. The data does not distinguish between two- and three-class models.

3. The expected duration of asymptomatic infection in women, based on this data alone, is 1.36 years (95% CI 1.13 to 1.63 years), based on the two-class model.

Evidence sources

Statistical models

Regression estimates of infection and re-infection rates by age and setting

We model the infection rates as a function of a baseline infection rate λ_1,1,1 multiplied by the between-setting HRs ρ_s and the between-age group HRs γ_a. The age- and setting-specific re-infection rates λ_a,s,2 equal the respective infection rate multiplied by a setting-specific re-infection HR η_s (equation 11). Other regression models are considered in Appendix 3:

⁽¹¹⁾ λ a , s , 1 = γ a ρ s λ 1 , 1 , 1 and λ a , s , 2 = η s λ a , s , 1

The infection rates λ_a,s,i in equation 11 are informed by the data in Table 9, which shows the number of initially uninfected women in each age within each setting who were found to be infected after a 6-month follow-up period. However, the mathematical relationship between the infection rates in each group and the proportions infected κ(t)_a,s,i at the end of a period of time length t is complex. The formula shown in the DAG (Figure 8) allows for the fact that in the LaMontagne et al. data192 it is possible for a woman to clear infection spontaneously or through treatment, and then re-acquire infection within the 6-month follow-up. It is necessary, therefore, to take account of the clearance rates of symptomatic and asymptomatic infection, and the proportion of incident infections that become symptomatic (see Appendix 4).

FIGURE 8.

Directed acyclic graph of the evidence network. 193 Reproduced from: Price M, Ades A, De Angelis D, Welton NJ, Macleod J, Turner K, et al. Incidence of Chlamydia trachomatis infection in England: two methods of estimation. Epidemiol Infect 2014;142:562–7. This work is under the Commons Attribution-NonCommercial license 3.0: http://creativecommons.org/licenses/by-nc/3.0/. The basic parameters are: λ_1,1,1, the infection rate in age group 1, setting 1; γ_a, the HR for infection in age group a relative to group 1 (age 16–17 years); ρ_s, the HR for infection in setting s relative to setting 1 (GP setting); η_s, the setting-specific re-infection–infection rate ratio; p_a,GP, the proportion of patients at recruitment in the GP attenders in age group a in the LaMontagne study192 that were in the re-infection group reweighted to account for differential recruitment; Δ_A and Δ_S, the durations of asymptomatic and symptomatic infection; φ the proportion of incident infections in which symptoms develop. The black bar indicates where the network can be cut to obtain separate estimates of incidence (see text).

Key assumptions

1. The LaMontagne study192 of incidence in FP, STI and GP groups is nationally representative.

2. The incidence and prevalence estimates are based on several data sources, but are relevant to the years 2000–5.

Summary of findings

1. The available data on incidence in the UK agree with predictions based on duration (see Chapter 4) and UK prevalence.

2. The mean duration of CT infection in women (including symptomatic and asymptomatic infections) is 1.03 years (95% CrI 0.82 to 1.25 years). The proportion of incident infections that are symptomatic is 0.23 (95% CrI 0.16 to 0.32).

3. CT incidence in the general population aged 16–24 years is 5.04 per 100 person-years (95% CrI 3.52 to 7.09) and in women aged 16–44 years it is 2.07 per 100 person-years (95% CrI 1.50 to 2.81).

4. CT prevalence in the general population aged 16–24 years is 5.15% (95% CrI 3.76% to 6.93%), and in women aged 16–44 years it is 2.11% (95% CrI 1.63% to 2.71%).

Evidence identification

We sought prospective studies of women with CT including observational studies with or without a control group, and RCTs of screening interventions. We identified studies from recent review papers,20,61,98,144 reference searches of published economic and population transmission models,1,2,8,29,140,141,170–174 and as part of a wider review of the literature on the natural history of CT. Included among these was a recent systematic review instituted at an Expert Advisory Meeting of the Centers for Disease Control and Prevention. 98 We exclude the study by Bachmann et al. 202 because the follow-up time was not reported; the study by Stamm et al. 203 because the patients were co-infected with gonorrhoea; and the study by Westergaard et al. 204 because the patients were women who had just undergone an abortion, and this is a very unrepresentative population. The study by Morre et al. 180 is excluded because of uncertainty about the adequacy of outcome assessment. 40 We identified eight studies12,13,19,73,165,184,199,200 (Table 15) that followed up women with CT to assess the proportion who developed PID in the absence of, or prior to, treatment. Note that four of the studies73,165,184,200 shown in Table 15 are excluded from the modelling because of a lack of a control group.

A model for progression to pelvic inflammatory disease in the controlled prospective studies

In our model for prospective studies (Figure 11), women who are CT+ and CT– at the start of the study begin in states 1 and 2, respectively. Women in state 1 may clear infection and progress to the CT– state, state 2, at rate λ^C, or they may develop PID (state 3) at a study-specific rate θsCT+ for study s. Women in state 2 may become infected with CT, which occurs at rate λsl or may develop PID at rate θsCT− if they acquire, or already had, a non-CT infection that carries a PID risk. We assume that the prevalence and incidence rates for other STIs are the same in each arm. The difference δ between θsCT+ and θsCT− represents the rate of acquiring PID in women with a current CT infection that can be causally attributed to CT. The Figure 11 model is represented by the Markov transition rate matrix (G) shown in equation 18:

FIGURE 11.

Multi-state transition model for disease progression in prospective studies. From Price et al. 201

Only comparative studies can contribute information on δ but we cannot derive estimates of δ from comparative studies without information on λ^C and λsl. We assume that the CT clearance rate for asymptomatic infection, λ^C, is constant across studies, at 0.74 (95% CrI 0.61 to 0.89) per year, based on the Chapter 4 evidence synthesis of studies on CT duration. Similar estimates have been reported elsewhere. 166 Note that the information on incidence and prevalence can back-propagate through the evidence network and impact on the posterior distribution of the duration of asymptomatic infection, although the latter was not materially changed.

For CT incidence λsl we assume an infection rate of 5% per year in women who are CT– at the outset, and a re-infection rate of 15% per year in women who are CT+ and tested and treated. Given a CT average duration of 1–1.5 years, a 5% incidence reflects a CT prevalence of 5–7.5%, which accords with the baseline CT prevalence in the three trials. Re-infection rates are higher by approximately a factor of 3. 192 We assume that all women in the Scholes13 and Ostergaard12 (95% CT–) studies are at risk of infection, whereas all women in the Rahm165 (50% CT+) and POPI19 (100% CT+) studies are assumed to be at risk of re-infection. Detailed sensitivity analyses (see Appendix 6) show that conclusions are relatively insensitive to the values assumed.

Some of the screening studies record the proportion of women who were tested and treated for CT during follow-up, thus shortening their time at risk of CT-related PID. This was 43% for the POPI trial,19 32% for Ostergaard et al. 12 and 8.1% for Rees. 199 The Scholes et al. paper13 implies that some women were tested but does not report how many: we have assumed a proportion of 32%, following Ostergaard et al. ,12 but the impact of a lower rate is explored in sensitivity analyses (see Appendix 6).

We consider two models. In the first, the causal rate of PID due to CT (δ) is constant throughout the duration of infection. In the second, the causal rate in the period immediately following CT infection (δ₁) is different from the causal rate during the remainder of the infection (δ₂). We fit models with the rate δ₁ lasting for periods of 30 days, 60 days and 90 days post infection.

Estimation

Details of the estimation methods are given in Appendix 7. In brief, each study arm provided data with a binomial likelihood, which estimates a probability parameter, say ps,ti→j, representing the study-specific transition probability that a person in state i occupies state j, t years later. These probabilities are entries in the transition probability matrix P_t. If the transition rates are constant over the period t then the transition probabilities, on which we have data (see Table 15), are functionally related to the transition rates in equation 18 by Kolmogorov’s Forward Equations (equation 19):209

Estimation is carried out using a Bayesian approach, where the posterior distribution is sampled through Markov chain Monte Carlo (MCMC) simulation. The rate parameters in equation 18 are given vague priors, the values of which are updated by the data. Numerical solutions to the Forward Equations are found in each MCMC cycle, using the WinBUGS44 add-on WBDiff,161 which uses the Runge–Kutta method. 210,211

Only the CT+ groups from the POPI trial19 were included in the analysis because the treatment effect estimate is thereby based on a randomised comparison. The proportion with CT in the early treated arm are considered, in effect, as being CT– at the start of the observation period. Those in the deferred treatment group are, of course CT+. However, in the Scholes13 and Ostergaard12 trials, although the women in the screened arm are again taken to be CT– at the start of follow-up, those in the unscreened arm are a mixture of (untreated) CT+ and CT–. The study-specific CT prevalences in the screened arms are estimators of these mixing proportions.

Further details on how the observed data informs the rate parameters in the matrix G (equation 18) are given in Appendix 7. The joint posterior distribution of the rate parameters is used to obtain posterior distributions for the two parameters of interest: the proportion of CT infections that result in PID and the proportion of CT-caused PID preventable by annual screening.

Estimation of the proportion of pelvic inflammatory disease episodes prevented by screening

Even with 100% coverage, an annual screening programme for CT would not prevent all cases of PID. We estimate the proportion of episodes of CT-related PID that could be prevented in women who become infected with CT. The expressions for this quantity are derived and presented in Appendix 7. This applies to the benefits to CT+ women who are screened and treated, and is not designed to measure the full effect of screening as it does not take into account reduction in PID due to reduced CT incidence.

Plan of analysis

The data analysis is carried out in several steps in order to show the influence of the different types of data on the estimates of interest. First, we examine each of the comparative studies separately. We then synthesise information from all three RCTs, and then from the four comparative studies. Studies are combined on the basis that the progression rates in CT– women vary between studies in an arbitrary way, but the difference δ between the progression rates from CT– and CT+ has the same fixed effect across studies. WinBUGS code is provided in Appendix 8.

Assumptions

1. Studies based on screening asymptomatic individuals recruit prevalent infection, whereas clinical-based studies recruit recently incident cases.

2. The definitions of PID used in the POPI study19 correspond closely with the definition of probable/definite PID used in Chapter 2.

Summary of findings

1. The data do not distinguish between a model that assumes a constant hazard of PID following CT infection, and a model in which the rate is higher over the initial 1–3 months.

2. The results of the POPI trial19 are consistent with the results of other trials and controlled prospective studies although there is a high degree of uncertainty.

3. The CT-related risk of PID following CT infection is 14.8% (95% CrI 4.8% to 24.8%), based on the trial evidence on PID that would be ascertained in a prospective study.

4. The total CT-related risk of PID following CT, including undiagnosed PID, is 17.1% (95% CrI 5.6% to 28.9%).

5. 61% (95% CrI 55% to 67%) of CT-related PID could be prevented by screening at annual intervals. The estimate is slightly lower if CT-to-PID progression rates are higher in the first 1–3 months.

Estimates of the incidence of all-cause pelvic inflammatory disease

This section describes two independent sources of PID incidence estimates, and a third estimate based on pooling them.

Synthesis of pelvic inflammatory disease incidence estimates

The two estimates of PID incidence were compared. A DAG (Figure 14) shows how the two estimates were combined. The WinBUGS code and data sets are set out in Appendix 10.

FIGURE 14.

Directed acyclic graph showing combination of two sources of evidence on PID incidence.

Population Excess Fraction: how much PID is caused by chlamydia?

In this section we examine estimates of the PEF based on microbiological studies, estimates based on retrospective case–control studies, and estimates derived from the ratio of CT-related PID incidence and all-cause PID incidence, and finally estimates based on randomised trials. DAGs for each of the estimates are shown in Figure 15.

FIGURE 15.

Directed acyclic graph showing methods of estimating Population Excess Fraction. Symbols: PEF-1: π_a, CT prevalence in age band a; OR from case–control study. PEF-2 OR* adjusted OR; ψ^Erf correct for under-ascertainment of CT infection based on Erfurt study. 225 PEF-3: λaCT incidence of CT; λaALL PID incidence of all-cause PID; θaCT→PID CT-to-PID progression risk. PEF-4: P^trt probability of PID in POPI19 treatment arm; p^Ctrl probability of PID in POPI19 deferred treatment (control) arm; p*Ctrl adjusted p^Ctrl; p^IT probability independent testing in CT+ controls.

Consistency of evidence on pelvic inflammatory disease incidence

Table 20 provides a series of age-specific estimates for the incidence of (all cause) PID, from the different data sources. The ‘direct’ estimate from the POPI trial,19 2.0% per year – when adjusted for under-ascertainment due to asymptomatic presentation – represents a 2.4% per year incidence. This is consistent with the estimate derived from routine data, 3.0% per year for that age range, following adjustment for the underdiagnosis inherent in routine PID statistics. Figure 16 shows the posterior distributions of PID incidence from the two sources, and the pooled estimate, and establishes the consistency of the estimates. The consistency of registry data with the POPI study19 can be studied only for the 16- to 24-year-old age group.

FIGURE 16.

Evidence consistency plot for the incidence of all-cause PID in England in women aged 16–24 years.

Note that when combining the direct and indirect evidence sources, the incidence information back-propagates to ‘update’ the estimate of the proportions of PID that is undiagnosed. This was 31% in the original study but is estimated to be 36% following synthesis with the POPI19 data.

The final pooled PID incidence estimates are 2.5% per year (95% CrI 1.8% to 3.4%) in women aged 16–24 years, and 1.6% (95% CrI 1.1% to 2.2%) in women aged 25–44 years. Taking into account the age structure of the female population, we can estimate that 62.9% (95% CrI 57.8% to 67.4%) of PID episodes occur in women over 24 years.

Estimates of the Population Excess Fraction

The synthesis of the three retrospective studies31,37,230 of CT in women with PID and control subjects (see Table 19) produces a posterior mean OR of 9.2 (95% CrI 4.4 to 18.1). The model fitted well (residual deviance of 5, compared with six data points). The OR adjusted by the under-ascertainment factor from the Erfurt study225 is 17.1 (95% CrI 7.9 to 34).

The several estimates of the PEF are shown in Table 21. The adjustment for under-detection of CT infection in the case–control studies, based on the Erfurt study,225 almost doubles the overall PEF (women aged 16–44 years) from 14.5% to 24.6%. The dramatic fall in PEF with age is no more than a reflection of our assumption that the OR is not related to age, whereas the incidence and prevalence of PID varies, as is evident from the routine data (see Table 18).

The PEF-3 estimates (column 3) based on the ratio of CT-related PID to all-cause PID also show a marked falling off with age. It is reassuring that the adjusted estimates based on case–control studies are close to estimates derived in a very different way from the progression rate, CT incidence and all-cause PID incidence.

The direct estimate of the PEF based on the POPI trial,19 PEF-4, is 35.7%, but if this is corrected for independent testing in the deferred treatment (control) arm, this becomes 49.4%. The CrIs on both estimates are wide, and span zero because the difference in the crude PID rates between the two trial arms itself is too uncertain to rule out no difference with 95% confidence.

Limitations of the analysis

Impact of age on the relationship between Chlamydia trachomatis and pelvic inflammatory disease

The PEF estimates show a clearly declining trend with age. In the retrospective estimates, this follows as a consequence of assuming a constant OR and applying this to a prevalence that declines steeply with age. In the case of PEF-3, the decline with age is partly due to the decline in CT incidence with age (exactly mirroring the decline in prevalence with age), the age profile of PID in the routine data, and the assumption that the proportion of PID that is undiagnosed is constant over age.

Although the decrease in PEF with age is to some extent a result of our assumption that neither the probability that PID is diagnosed nor the risk of PID following CT are age-dependent, the degree of variation with age, a factor of 4–6 between age 16–19 years and age 35–44 years, is so extreme that we would have to propose quite extreme trends in one or both of those quantities to reverse it. The proposal that PEF decreases with age is, apparently, not one that has been made before, but it is consistent with the observation that estimates based on the trial evidence tend to be higher because the trials are heavily weighted towards younger women.

Age dependency in PEF, if confirmed (see Chapter 2 for discussion of alternative explanations), would have significant impact on the public health importance of CT, as the majority of EP and TFI occurs in older women. Although it is certainly possible that CT infections in younger women have a key role in reproductive health problems that emerge many years later, these results focus attention on the distinctly different age profiles of CT and PID. There are insufficient data in either the prospective or retrospective studies to address this question more fully. Further work to examine these issues directly is proposed in our research recommendations.

Post-hoc element in choice of estimates

We record here that, when the work in this chapter was first carried out, we included only the unadjusted (PEF-1), and we pooled that information with the evidence sources contributing to estimate PEF-3. At that time we believed that the estimates based on retrospective studies must be upper bounds, and we did not examine the trial-based estimates. After we had completed the work on TFI (see Chapters 10 and 11), we became aware that the PEF-1 estimates were too low to be consistent with the results of those chapters. We then reconsidered the relationship between CT and PID, and re-examined the trial-based estimates of PEF, particularly in the light of the comments in Gottlieb et al. 226 At the same time, results from the recent Erfurt study225 made it clear that the retrospective estimates were not necessarily upper bounds. There is, therefore, a post-hoc element to the reasoning in this chapter. The ‘hypothesis-generating’ aspect to our findings on PEF are reflected in the research recommendations. Nevertheless, that fact that two estimates of PID incidence are consistent with each other, and with an account of the PEF that is consistent with other independent sources of data, to some extent mitigates the post-hoc process in our investigation.

Coherent estimates of CT and PID incidence, CT-to-PID risk, and Population Excess Fraction

There is an option to strengthen our estimate of the PEF by pooling PEF-3 with either the adjusted retrospective evidence PEF-2, or the evidence directly derived from the POPI trial,19 PEF-4 or PEF-5, or both. We have chosen not to do this, on the basis that the PEF-2 estimate is vulnerable to many biases, and relates to a different time and place, as discussed above. The estimate of PEF from the POPI trial19 is not statistically independent of the estimate of the CT-to-PID risk estimate, and is also subject to uncertainties described above. However, we regard the similarity of these independently derived estimates as an important validation.

In taking forward the PEF-3 estimate γaCT←PID, which is calculated from the CT incidence estimates in Chapter 5, the CT-to-PID progression risk in Chapter 6, and all-cause PID incidence in this chapter, we have generated a fully coherent set of estimates for CT incidence, prevalence, duration and the proportion of infection that is symptomatic; PID incidence, the proportion of PID that is symptomatic and the proportion that is diagnosed, the CT-to-PID progression risk, and the PEF. The summary posterior means and CrIs can be found in Tables 13 (full synthesis model), 17 (trials only), 20 (synthesis model) and 21 (column 3).

These estimates are carried forward to the analyses in the subsequent chapters.

Summary of assumptions

1. The probability that PID is diagnosed is independent of age.

2. The risk of PID due to CT is independent of age.

3. The probability of diagnosis is the same in CT- and non-CT-related PID.

4. In prospective studies, including the screening trials, all PID – except for asymptomatic PID – is diagnosed.

5. In routinely collected data on PID, a lower proportion of all PID is diagnosed than in prospective studies.

Summary of findings

1. Two estimates of PID incidence in the UK were consistent: one derived from the control arm of the POPI trial,19 and the other from routine statistics combined with evidence on the proportion of PID that is diagnosed.

2. A pooled estimate of PID incidence in age 16- to 24-year-olds is 2.5% per year (95% CrI 1.8 to 3.4), and 1.8% per year (95% CrI 1.3% to 2.5%) in 16- to 44-year-olds.

3. 36% (95% CrI 26% to 48%) of PID that causes reproductive damage is diagnosed; 12.6% (95% CrI 4 to 25) is asymptomatic.

4. Estimates of the PEF were compared: (1) an estimate based on retrospective studies adjusted for under-ascertainment; (2) an estimate based on CT incidence, CT-to-PID progression rate, and PID incidence; and (3) estimates based on the POPI19 trial. There was no evidence of inconsistency although Crls were wide.

5. We estimate that 62.9% of PID episodes (95% CrI 57.8% to 67.4%) occur in women aged over 24 years.

6. The PEF declines sharply with age.

7. Our estimate of PEF, derived from CT and PID incidence data and the CT-to-PID progression risk is 19.7% (95% CrI 5.9% to 38.1%) in women aged 16–44 years, and 35.3% (95% CrI 10.5% to 68.5%) in women aged 16–24 years.

The analyses thus far in the report establish a set of coherent estimates for the UK of CT incidence, CT prevalence, CT duration, probability that an incident CT infection is symptomatic, PID incidence, proportion of PID that is diagnosed, proportion of PID that is symptomatic, the CT-to-PID progression, and the PEF of PID due to CT.

Model for repeat pelvic inflammatory disease episodes

Estimates of age-specific incidence of CT-related PID, all-cause PID, and therefore non-CT-related PID are available from earlier chapters. What we need, however, is a way to estimate what proportion of these PID episodes are occurring in women who have already experienced a previous episode. In a homogeneous population, PID incidence would be the same in women who had previously had PID and those who had not. However, in the Lund study33 it was observed that the majority of women who experience a second PID did so within 2 years. We therefore assume that women acquire PID at a rate λa,1PID in age group a, and that this rate increases to λa,2PID for the 2 years following a PID. To determine the ratio of λa,2PID to λa,1PID we use information from the LaMontagne study192 of CT infection and re-infection rates for up to a 18-month period. This was described in Chapter 5. Specifically, we use the ratio on the GP group, which was 7.08 (3.97 to 11.6) from Table 13, and assume that this same ratio applies to PID incidence so that: λa,2PID=ηGP. λa,1PID.

The GP group was chosen, rather than the FP or STI groups, as it was the most likely to reflect re-infection rates in the general population. Using data on recurrent CT makes the further strong assumption that the pattern of infection and re-infection in CT is the same as the pattern of PID and repeat PID, for PID from any cause. Apart from the observation by Westrom33 that repeat PIDs usually occurred within 2 years of follow-up, this assumption, and its many implications, do not have strong a priori validity. Our results are, however, consistent with the Lund study33 observations on the incidence of repeat PID, and with other data, as we show below.

We have posterior estimates of η^GP from Chapter 4 and λaALL PID from Chapter 7, and our task is to calculate λa,1PID and λa,2PID. The parameter λaALL PID denotes the incidence of all PID, whether first, second, or subsequent, and from any cause. There are several ways to parameterise the problem but the simplest is as follows:

The equations, together with estimates of λaALL PID and η^GP, uniquely determine ρ, which is a calibration constant ensuring that when we sum first PIDs and subsequent PIDs we end up with the correct total PID incidence. It is interpreted as the ratio of PID incidence in women who have not had a PID in the last 2 years to the incidence of PID in all women. ρ is calculated through a trial and error process. The model was run from ages 16–44 years for different values of ρ until a value was found where the estimated incidence of any PID agreed with our estimates of PID incidence from Chapter 7 (see Table 20).

Markov model

Figure 17 shows a Markov model of PID development. It is assumed that women start in state 1 at age 15 years and progress through the model until age 44 years. It is a discrete time model with 1-year cycles, programmed in WBDEV and WinBUGS161,236 (see Appendix 11). Expressions for the transition probabilities are shown in Table 22, with the parameters defined as in equations 30 and 31. Note that the model shows progression to first salpingitis and subsequent PID episodes. Exactly the same structure is used to assess numbers of PID episodes, or number of salpingitis as described later in the section.

FIGURE 17.

Markov model of salpingitis and subsequent PID incidence. Individuals start in state 1 (no previous salpingitis) and progress to state 2 (salpingitis in the last 0–1 year). Once in state 2, women can (a) fail to have a second PID and enter state 3 (salpingitis 1-2 years ago) or (b) have a second PID episode within a 2-year period and enter state 5 (salpingitis and 1 PID 0–1 year ago). From state 3 it is also possible to progress to state 5 at the faster ‘recurrence’ rate or to progress to state 4 (salpingitis 2+ years ago) by failing to experience a further PID episode within 2 years.

To generate estimates of the proportion of women who have ever had one, two or three or more PIDs, we assume that further PIDs occur within a 2-year time period. The logic of the model is that individuals start in state 1 (no previous salpingitis), and progress to state 2 (salpingitis 0–1 years) at the constant incidence rate λa,1Salp. The probability of a transition to state 2 with a 1-year period is therefore 1−e−λa,1Salp. Once in state 2, women can (1) fail to have a subsequent PID episode and enter state 3 (salpingitis 1–2 years ago), or (2) have a subsequent PID episode and enter state 5 (salpingitis and 1 PID 0–1 years ago) at a rate 1−e−λa,2PID. From state 3 it is also possible to progress to state 5 at the faster ‘recurrence’ rate, or to progress to state 4 (salpingitis 2+ years ago) by failing to experience a PID episode. The subsequent transitions follow the same logic.

Given values of λa,1PID, λa,2PID and λa,1Salp, the model is run for a cohort of women aged 16–44 years on a 1-year cycle, generating proportions of the female population φk,a in each state k, at each age a. From these state probabilities the proportions of the population who have experienced no salpingitis, or salpingitis with zero, one or two or more subsequent PID episodes can readily be found (see below).

We also validate our results against results from the Lund data set. 33 The study reports data on the distribution of numbers of PID episodes in women who have had salpingitis, for a mean follow-up period of approximately 8 years, for women aged < 25 years, and for women aged ≥ 25 years. The Markov model was run separately for women starting at each of the 22 ages, 16–37 years, in each case starting in state 2. An 8-year time-horizon was used in order to compare results with those observed in the Lund study. 33 The average predicted number of women with a single subsequent PID and women with two or more subsequent PIDs were obtained by averaging across the age ranges 16–24 years and 25–37 years, respectively. Results were obtained for all PID and diagnosed PID only, and these are shown alongside the observed results in the Lund study33 (Table 23).

Model inputs

Besides ϕ^salp|PID and ρ (see above), the key parameters are:

1. the proportion of all PID that is due to CT, the PEF, for each age group, γaCT←PID

2. the incidence of all-cause PID, by age group, λaALL PID

3. the CT re-infection to infection ratio rate for the GP setting, η^GP

4. the proportion of PID that is diagnosed in the general population, ψ^Diag.

These parameters are all correlated, as they are estimated simultaneously from a common combined data set in Chapter 7. After suitable transformations (log transformations for λaPID and η^GP; a logit transformation for γaCT←PID; ψ^Diag on the natural scale) the joint uncertainty in these parameters is described by a multivariate normal distribution.

Model outputs

Our primary output from the chapter is πm,aS+PID, the proportion of women age a who have had an episode of salpingitis and zero, one, or two subsequent episodes of PID (indexed by m = 0,1,2 respectively). At age 15 years, all women are assumed to have never had PID so that φ1,15=1, φ2,15⋯φ8,15=0 .  If φk,a is the proportion of women at age a in state k, then:

where:

Pj',j,a is the transition probability from state j’ to state j where a indexes from ages 16 to 44. The model was run first for all-cause PID, and then separately for diagnosed all-cause PID. Although the true progression through the model is the same, only a proportion ψ^Diag of PIDs are observed. The probability of diagnosis for each PID is assumed to be independent of the number of previous PIDs, and it is assumed that no women have more than three PIDs, so the results will be a slight underestimate. As such, the proportion of women observed to have had zero, one, two or three or more PIDs, respectively, πn,aobsPID, is:

Summary of assumptions

1. The distribution of zero, one and two or more subsequent PID episodes in women with salpingitis in the UK approximates that found in the Lund study.

2. The probability that PID is diagnosed does not depend on the number of previous episodes, nor on whether previous episodes would have confirmed as salpingitis on laparoscopy.

3. It is assumed that the re-infection rate for CT applies equally to all other causes of PID and salpingitis.

Summary of findings

1. 33.6% of women age 35–44 years have experienced at least one episode of PID (diagnosed or not).

2. 16.1% of women age 35–44 years have experienced at least one episode of salpingitis (diagnosed or not).

Prospective evidence: ectopic pregnancy following pelvic inflammatory disease

We identified studies from a recent systematic review of the sequelae of CT,98 from the knowledge of collaborators, and as part of a wider search of the literature on CT. We exclude a study by Buchan et al. 239 because it does not report data on number of pregnancies (see below). Table 27 summarises the results of the six prospective studies33,228,240–243 following women forward from PID to a next pregnancy. These studies have very different follow-up times. However, we can control for this approximately by estimating the proportion of pregnancies that are ectopic, φEP|preg. Study-specific crude estimates are presented in column 7. Apart from the small study by Heinonen et al. ,240 in which EP rates are unaccountably low, estimates from studies following women with laparoscopically confirmed salpingitis are reasonably homogeneous, at ≈7–10%. The study by Ness et al. 228 estimates a much lower proportion (≈1%). It follows women with mild to moderate PID, not laparoscopically confirmed salpingitis, and the most serious 3% of cases are excluded. The results from the prospective studies are otherwise broadly similar.

Our quantitative analysis of the prospective risk of EP following salpingitis is based solely on the Lund study,124 for four reasons: it is by far the largest study; it is the only study to include a control group; uniquely it presents EP risk by age, severity of salpingitis in women with PID, and numbers of episodes of PIDs; and also it delivers a related set of estimates of the risk of TFI following hospital diagnosed PID with salpingitis. This last feature is especially significant, as it provides an opportunity to check the consistency between observed and predicted EP and TFI frequency in the UK on the same basis.

Application of the Lund results to the epidemiology of pelvic inflammatory disease in the UK

As described in Chapter 8, we assume that the proportion of diagnosed ‘probable/definite’ PID that would be accompanied by salpingitis if examined on laparoscopy is 0.429 (95% CrI 0.27 to 0.63). We further assume that this proportion is the same regardless of PID severity, clinical presentation, place of diagnosis or whether it is diagnosed or not. Note that this implies the existence of undiagnosed but symptomatic salpingitis, and asymptomatic or ‘silent’ salpingitis, as discussed by Wolner-Hanssen. 103 We assume that only PID that would be confirmed under laparoscopy (salpingitis) can cause EP or TFI. This is supported by observations in the control group in the Lund study33 (women with hospital-diagnosed PID but no visual evidence of laparoscopy) that 0 out of 601 women developed TFI. The EP rate in this group was slightly higher than the general population but there is considerable uncertainty (only six cases).

Outline of prospective analysis

The prospective analysis of EP proceeds in the following four steps:

1. A logistic regression model is fitted to the data available from the Lund study (Table 28),124 to estimate the risks that the pregnancy will be ectopic η_m,s,a in women whose index salpingitis of severity s at age a was followed by m = 0,1,2+ subsequent episodes of PID. The non-salpingitis control group informs an age-specific baseline θControl,aEP|preg, the probability that a pregnancy is ectopic in age group a in women with no previous salpingitis.

2. Because the Lund analysis relates only to hospital-diagnosed PID, we explore a range of five scenarios that relate the EP risks θm,d,aEP|Salp,preg following d = hospital diagnosed PID (HD), PID diagnosed outside of hospital (nHD), or undiagnosed PID (UD), to the EP risks at different severities s adjusted by the proportion of PID estimated to be salpingitis. These five models can be thought of as ‘mappings’ between the η_m,s,a and the θm,d,aEP|Salp,preg parameters.

3. The proportion of PID at age a in the categories HD, nHD, and UD, κ_d,a, is derived from the UK routine data on PID and information on the proportion of PID that is diagnosed, ψ^Diag. This is multiplied into the proportion of population prevalence of salpingitis in age band a, followed by m = 0, 1, 2+ subsequent PID episodes πm,aS+PID from Chapter 8, to form πm,d,aS+PID, which has the same interpretation but breaks the population into nine groups by age, diagnosis setting, and number of subsequent PIDs.

4. Estimates of the population proportion of conceptions that are ectopic because of salpingitis are formed by weighting the estimates θm,d,aEP|Salp,preg, generated by the five ‘mapping’ models (see below), by the population proportions πm,d,aS+PID.

The full prospective synthesis is shown in the form of a DAG in Figure 22.

FIGURE 22.

Directed acyclic graph of prospective analysis of EP.

Distribution of pelvic inflammatory disease diagnostic status in the English population

In order to apply the prospective risks informed by the Lund studies to predict EP outcomes in England, it is necessary to have estimates of the risk factor distributions, by age. Estimates of the proportion of women in age band a who have had an episode of salpingitis that has been followed by m = 0,1,2+ episodes of PID, πm,aS+PID, where obtained in Chapter 8. However, we need to consider the proportions κ_d,a of women in age band a who were hospital diagnosed (d = HD), diagnosed elsewhere (d = nHD) or undiagnosed (d = UD). To do this, we make several fundamental simplifying assumptions, in the absence of clear information to the contrary, and in order to make the analytic task tractable. First, that the probability a PID is diagnosed in hospital is independent of the number of previous PIDs a woman has had; second that the increase in odds of EP given one or two or more as compared to zero, subsequent PID episodes conditional on having an initial PID that is confirmed as salpingitis is independent of whether or where the initial PID was diagnosed, and clinical severity of the initial salpingitis.

Routine ectopic pregnancy data

The premise behind the use of routine data on EP is that it should be possible to compare the observed risk of EP in pregnant woman of age a, based on routine HES data for 2002 (Table 29), taking account that not all EP is caused by salpingitis, to what would be predicted from the Lund study.