A critique of cohort studies examining the role of human papillomavirus infection in cervical neoplasia
Odds ratios, which range from 1.2 to 327, have been reported from cohort studies which have examined the association between human papillomavirus (HPV) infection and cervical neoplasia1–16. Although these studies differ with respect to definitions of exposure and disease, they suggest a heterogeneity of effect which is unlikely to have a purely biological explanation. Therefore, it is necessary to consider the extent to which the variation in these estimates can be attributed to idiosyncrasies in study design and data analysis. Towards this end, we first consider the importance of cohort studies in defining the role of HPV in the aetiology of cervical neoplasia, we next summarise the essential elements of their design and analysis, and then use examples to illustrate in some detail the methodological challenges posed by the study of this particular exposure–disease relationship. Finally, and because they address an issue of relevance to the cervical screening programme, we scrutinise those studies which have explored the relationship between the risk of disease and duration of infection. We hope that this critique, which is not intended to be another systematic review, will result in a greater shared understanding between clinicians, virologists and epidemiologists of the methodological issues underpinning the design and analysis of cohort studies in this setting.
Identification of cohort studies
This critique includes 34 cohort studies which were identified from the PubMed database and from the citations of research reports relating to HPV1–34. There have been two reviews which have listed cohort studies examining the association between HPV infection and cervical neoplasia: one was undertaken by the International Agency for Research on Cancer [IARC] (1995)35, the other was a systematic review of the role of HPV testing in the Cervical Screening Programme36, commissioned by the NHS Research and Development Health Technology Assessment Programme (1999). Some reports identified in these reviews as cohort studies have been excluded from this critique, either because of an ambiguity about the study design, or because there is some doubt as to whether the investigators had intended any formal comparison of exposed and non-exposed subjects, or because more than one report on the same study was available37–45. However, this critique includes an additional 11 studies, two of which were available only in abstract at the time of the systematic review.
Importance of cohort studies
The epidemiological evidence that lends support to a role for HPV in the aetiology of cervical neoplasia is largely based on studies with a case–control design; these have consistently revealed a strong association between cervical neoplasia and the detection of HPV DNA in samples of exfoliated cervical cells taken at, or subsequent to, diagnosis of disease35. However, with case–control studies there is always ambiguity about the temporal relation between exposure and outcome which, for example, does not allow us to eliminate the possibility that the presence of disease could itself increase the ability to detect HPV. Only cohort studies can provide compelling evidence of temporality, the sine qua non for causality, and this study design has been used to examine the role of HPV at different points in the natural history of cervical neoplasia. Some studies have focussed on the risk of disease acquisition in women who were free of disease at study entry, others on the risk of disease progression in women who were already to some extent diseased. In many of these studies, there has also been an interest in the relative importance of infection with low and high risk HPV types and the duration of exposure necessary for disease to occur.
Essential elements of cohort study design
The essence of a cohort study is that non-exposed and exposed subjects, or subjects with varying degrees of exposure, are observed over time for the outcome of interest. The strength of this design is that the method of subject selection should leave little doubt about the temporal relationship between the exposure and the disease under study.
Prospective and retrospective designs
Cohort studies may be prospective or retrospective. For both, exposure status is, or can be, defined before outcome status is known. In a prospective study, observations are made after the point in time at which the study begins; samples are then collected for cytological, histological and virological examination. In a retrospective study, the material necessary to determine exposure and disease status will already have been collected; exposure and disease status are then assigned following examination of this archival material. This strategy has been used successfully in Scandinavian countries, where linkage of mature population-based screening registers and cancer registration systems facilitate the identification of populations suitable, for example, for the study of the association between HPV exposure and invasive cancer, an objective likely to fall outside of the scope of any prospective enquiry11.
The nested case–control design
The primary drawback of cohort studies is the large sample size required, often leading to high data collection costs. For example, the testing of a large number of samples for the presence of HPV DNA is an expensive and time-consuming exercise. One solution is to collect and store samples from all subjects within the cohort at baseline. Women who subsequently develop disease (the cases) are then matched with women who have remained free of disease at least until the time of diagnosis of their matched case (the controls). This approach, which retains many of the advantages of the cohort design, is potentially more efficient, because the analysis is restricted to subjects who develop the outcome during follow up and only a sample of those who do not. Typically, cases and controls are matched for age and duration of follow up. When exposure status is assigned using archival material, cases and controls should also be matched on the date of the baseline sample, because the ability to detect HPV DNA may diminish with increasing storage time; this matching criterion has been used in some studies6,11,19 but not in others5. By definition, a control must be free of disease when their matched case is diagnosed, and therefore should have had a normal smear at a time corresponding to the date of diagnosis of the case; again, this requirement has been met in some studies3,6,12 but not in others11,19; the failure to verify absence of disease in ‘controls’ is perhaps most remarkable when the study end-point is high grade CIN, an asymptomatic condition.
General analytical considerations
The primary objective of a cohort study is to compare the frequency of disease occurrence in exposed and non-exposed subjects. The basic analysis involves estimation of the rate of occurrence of a specified outcome among exposed and non-exposed subjects, and the provision of a measure of association which quantifies the absolute or relative difference in risk between these groups; the latter is usually expressed as a relative risk, hazards ratio or an odds ratio. Unlike cross sectional and case–control studies, the cohort design explicitly incorporates the passage of time: time on follow up; time to an event (e.g. the first detection of cytological abnormality or the diagnosis of high grade CIN); time since first exposure; and duration of exposure.
Occasionally, all women in the cohort are followed for the same fixed period after study entry, before outcome status is ascertained27; in a nested case–control study, the study design should ensure that women who develop the outcome of interest have the same period of observation as those who do not. Usually, however, subjects enter the study at different times and remain under observation for different periods without necessarily reaching the study end-point. In these circumstances, methods suitable for the analysis of survival data should be used when estimating and comparing rates; this requirement is frequently ignored. Some investigators have used methods which are valid only when all subjects have the same period of observation, even when this was demonstrably not the case18,20–24,26,28,29–32,34. It is for these reasons that we believe it is unwise to attempt to independently estimate measures of association when these have not been explicitly calculated by the investigators; the earlier systematic review appears to have erred in this respect36. Table 1 lists only those studies for which a measure of association was explicitly recorded by the investigators; the measure of association cited is that which most strongly links some aspect of HPV exposure and cervical neoplasia.
Table 1. Risk of cervical neoplasia associated with HPV infectiona.
|Nobbenhuis et al.1||Persistent HRc HPV||CIN3||327 (42 to 2468)|
|Rozendaal et al.2||HR HPV||CIN3||116 (13 to 990)|
|Liaw et al.3||HPV 16||HSIL||63.9 (16.4 to 248.1)|
|Hopman et al.4||Persistent HR HPV||2 × 3A or 1 × 3B+||28.2 (3.72 to 215.2)|
|Zielinski et al.5||HR HPV||Cancer||28 (11 to 72)|
|Wallin et al.6||HPV||Cancer||16.4 (4.4 to 75.1)|
|Koutsky et al.7||HPV 16, 18||CIN2–3||11 (4.6 to 26)|
|Moscicki et al.8||HPV infection for one to two years||LSIL||10.3 (5.6 to 18.7)|
|Woodman et al.9||HPV 16||CIN2–3||8.5 (3.7 to 19.2)|
|Remmink et al.10||HR HPV||Progressive CINd||6.5 (1.5 to 28.4)|
|Ylitalo et al.11||HPV 16||Carcinoma in situ||>5 (na)|
|Coker et al.12||HR HPV||HSIL||3.8 (1.5 to 9.6)|
|Kataja et al.13||HPV 16||Progressive CINe||3.2 (1.4 to 7.3)|
|Ho et al.14||HPV||SIL||3 (2 to 7)|
|Woodman et al.15||HPV 16, 18||Progressive CINe||2.3 (1.2 to 4.3)|
|Liu et al.16||HPV 16||Progressive dysplasiaf||1.19 (1.03 to 1.4)|
A fundamental requirement for valid epidemiological inference is that exposed and unexposed groups are comparable with respect to any factor that might explain differences in outcome, other than the exposure under consideration. In cohort studies, matching at the start of follow up or stratification in the analysis can be used to achieve a valid comparison. For example, in one study which included women with different degrees of histological abnormality, the prevalence of HPV infection in each of these subgroups varied, as did their underlying risk of disease progression; therefore the comparison of the risk of progression may have been distorted by the failure to stratify exposed and unexposed subjects by baseline disease status25.
A critical assumption underpinning all cohort studies is that subjects are free of the outcome of interest at study entry. When the intention is to recruit women who are cytologically free of disease, some investigators review all baseline cytological material and exclude from further analysis those found to have abnormal smears3. Others report the results of such a review but nevertheless retain all subjects in the cohort. For example, Rozendaal et al.2 report that “women with normal Pap smears containing high-risk HPV genotypes were 116 times (95% CI 13 to 990) more at risk of developing CIN III, in contrast to women without high-risk HPV”; the confidence intervals are wide because the estimate is based on only seven cases of CIN3, of whom only one tested negative for high risk HPV types. The characterisation of these women as cytologically normal is unfortunate, because at baseline 21% of the study population had a smear reported as PAP 2 (very mild squamous dyskariosis, including atypical squamous cells of unknown significance). Furthermore, cytological review revealed that in fact two of the seven women subsequently found to have CIN3 had moderate dyskariosis in their baseline smear, and in two more a normal baseline smear was re-graded as PAP 2. Wallin et al.6 argue that the exclusion from their cohort of women found to have cytological abnormality on review “would have made the result difficult to interpret, since it would have redefined the study population by using additional tests that are related to both the exposure and the outcome”. Zielinski et al.5 report measures of association calculated both before and after the exclusion of women found to have cytological abnormality following review of baseline smears, which were initially reported as normal. When the primary interest is in the natural history of the exposure–disease relationship, then the inclusion in the analysis of subjects who are not free of disease at recruitment would seem inappropriate. However, when there is pragmatic interest in the utility of the additional information, which might be provided by a HPV test in the context of a screening programme, where inevitably some women will have false negative smears, then the inclusion of these subjects might be considered legitimate.
In those studies focussing on the risk of disease progression in women who are already diseased, admission to the cohort is based on the results of cytological or colposcopic examination, or both1,11,17–19,28,33. In others, admission only follows histological examination of a colposcopically directed punch biopsy, an investigation which offers greater diagnostic certainty that the end-point of interest has not already been attained prior to study entry13,15,19–22,24–26,28–31. For example, Nobbenhuis et al.1 report that “women with persistent infection with high risk human papillomavirus from baseline till the last visit were at higher risk of end histology CIN3 than those who had negative results throughout the study (odds ratio 327 [95% CI 42 to 2468])”; the confidence intervals are wide because only one unexposed woman progressed to CIN3. However, the use of this end-point presents a difficulty in that women were not histologically sampled at baseline. Of the 353 women in the cohort, 133 were colposcopically adjudged to have at least CIN2 at study entry. Despite repeated colposcopic examinations, only 33 of the 103 women who reached the study end-point (CIN3) were considered to have clinically progressed during follow up; the remainder were only detected when all women with abnormal cytology were biopsied at the end of the study period. The inability of the investigators to colposcopically predict progression to CIN3 substantially undermines the specificity of their initial colposcopic diagnosis, and therefore we cannot exclude the possibility that a substantial number of cohort members may already have had CIN3 at study entry. Were this the case, this study could not be considered a cohort study in any meaningful sense, at least when this end-point is used.
Measurement of exposure status
In cohort studies, disease status is observed on at least two occasions, but HPV status is sometimes only measured at baseline12,19,25. Frequently, however, repeated measurements of exposure status are also made after study entry and prior to diagnosis. Studies with this design are potentially more informative as they are able to illustrate if and how a woman's HPV status changes during follow up: a woman who is HPV negative at study entry may become HPV positive during follow up and prior to diagnosis. Therefore, during her period of observation, she is non-exposed for some of the time and exposed for the remainder. Some investigators choose to ignore the information provided by repeated measurements of exposure status and report outcome in relation to baseline HPV status alone14. Others treat women who become HPV positive after study entry as exposed for the entire duration of the follow up period10. Both of these analytical approaches will under-estimate the risk of outcome associated with HPV exposure. Very few investigators use analytical methods that distinguish the time during which a woman is unexposed and that during which she is exposed7,9,16.
Ascertainment of disease status during follow up
Given the episodic nature of observations on disease status, the imperfections of our detection systems and the often transient nature of epithelial abnormalities of the cervix, it is unlikely that the outcome of interest will always be identified, should it occur. However, from an epidemiological viewpoint, the critical consideration is often not whether all outcomes of interest have been identified but whether the failure to ascertain outcome is related to exposure (differential misclassification) or not (non-differential misclassification).
We first examine the impact of a non-differential failure to ascertain all cases of disease. Consider a ‘perfect’ study, one in which all cases of disease are ascertained. In such a study, were 20 cases of high grade CIN to occur in 100 women exposed to HPV, and 10 cases in 1000 non-exposed women, the relative frequency with which disease occurs in exposed compared with non-exposed subjects, the relative risk, would be 20 [(20/100) ÷ (10/1000)]. In a less than perfect study which is subject to non-differential ascertainment of disease where, for example, we fail to ascertain 20% of all cases of high grade CIN, incidence rates will be lower to the same extent in both exposed (16/100) and non-exposed women (8/1000); but the relative risk will be exactly the same [(16/100) ÷ (8/1000) = 20]. This conclusion is only valid when the diagnosis of high grade CIN has 100% specificity (i.e. when all putative cases of high grade CIN are in fact high grade CIN); this can usually be assured by pathological review. Where there is both imperfect specificity and sensitivity, or when there is also the potential for exposure misclassification, the situation is more complex. These issues have been extensively reviewed elsewhere46.
The preceding example is helpful when addressing concerns that histological sampling may distort the natural history of the disease. This would be an important consideration if it were true, and if our primary objective was a comparison of the absolute rates of disease progression in exposed and non-exposed subjects. However, we are usually more interested in the relative rates of disease progression. As we have no a priori reason to believe that a colposcopically directed punch biopsy is more likely to disrupt the natural history of a HPV-associated abnormality than that of an abnormality not associated with HPV, then, as we have just demonstrated, the effect of any potential non-differential disruption of the natural history of the disease is to under-estimate the rates of disease progression in both exposed and non-exposed women, but not the relative frequency with which progression occurs in these groups. Similar considerations apply to any colposcopically directed punch biopsies removed during follow up. In any event, concerns over the distortion of the natural history of CIN caused by histological sampling may be overstated because a randomised controlled trial has shown that a colposcopically directed punch biopsy has minimal impact on the natural history of cervical abnormalities47.
Differential ascertainment of disease status
Risk estimates will be biased if knowledge of exposure status influences the ascertainment of disease status. In some studies, HPV status is not determined until after the end of clinical follow up, while in others, it is made explicit that clinical staff are blind to a subject's HPV status. It is considered good practice for cytological and histological material to be reviewed to confirm the final diagnosis and the absence of disease at baseline; however, this may inadvertently result in differential misclassification if, for example, material from exposed subjects is more intensively reviewed than that from non-exposed subjects. In three studies, the intensity and duration of follow up varied between exposed and non-exposed subjects, by design; this inevitably results in differential ascertainment, as one group of subjects will have greater opportunity for detection of outcome than the other4,30,34.
Risk of cervical neoplasia in relation to duration of HPV infection
In an attempt to explain how HPV infection can simultaneously be an extremely common sexually transmitted infection and the cause of cervical cancer, it has been suggested that the development of disease only follows a prolonged infection during which the same high risk HPV type can be continuously detected. Three strategies have been used to test this hypothesis: the first compares the risk of disease in women considered to have infections of short duration with that in women with infections of long duration; the second compares the duration of infections in women in whom disease progressed with that in women in whom it did not; and the third estimates how soon after the first detection of HPV the risk of disease is greatest. Before considering the outcome of inquiries based on these strategies, we make some further general observations on the measurement of the HPV exposure–disease relationship.
Measurement of the exposure–disease relationship
Exposure and disease status are usually defined by examining material from the same cytological sample. Both HPV infection and CIN are asymptomatic conditions and their presence or absence can only be determined by examinations made at intervals. The times of onset of both exposure and disease can never be precisely established; we can only infer that onset occurred at some point between the last negative observation and the first positive observation. The longer the interval between observations, the more uncertain the time of onset. In practice, the time of onset of an exposure or disease is usually assumed to have been at the midpoint of the interval. If the objective of a study is to estimate the time to the first occurrence of cytological abnormality, then the accuracy of this estimate will be dependent on the sensitivity of the screening test and the interval between observations. If, however, the objective is to describe the time to onset of high grade CIN, then the situation is more complex. Progression to high grade CIN will only be suspected if periodic observations are made using cytological and colposcopic examination, and can only be confirmed or refuted by histological examination. Inevitably, there is a ‘latent period’ between the onset of morphological changes and their histological identification, the duration of which is dependent not only on the interval between smear tests and the sensitivity of cytological examination, but also the threshold for referral for colposcopic assessment and the ability of the colposcopist to recognise and remove for histological assessment the area of most severe abnormality. The latent period will inevitably be prolonged when, as is sometimes the case, a biopsy is postponed until the end of a fixed period of follow up or until some cytological or colposcopic threshold has been reached; this is because it is possible for progression to high grade CIN to occur at the same time as cytological examination reveals only minor cellular changes, and because, as we have already discussed, colposcopic examination can be an unreliable predictor of disease progression1. Observations on exposure status made after the onset of the latent period are clearly irrelevant with respect to the duration of exposure necessary for that outcome to occur. A continuing HPV infection may be necessary to maintain a CIN lesion once it has occurred, or a continuing CIN lesion may itself facilitate the continuing detectability of HPV; however, this is a different issue and one which is likely to prove more difficult to disentangle.
Risk of disease associated with ‘persistent’ or ‘transient’ infection
In an attempt to test the hypothesis that a prolonged type-specific infection is a prerequisite for disease, the occurrence of disease in women who test positive for HPV on two or more occasions (persistent infection) has been compared with that in women who test positive only once (transient infection). In 3 of the 10 studies which took this approach, the investigators used the sample taken at diagnosis to provide the second of the two consecutive positive samples necessary to define an infection as persistent3,6,34. This is a serious source of error because observations on exposure status made at or after the time of diagnosis are uninformative with respect to estimates of the duration of exposure necessary for that disease to occur; an outcome cannot be attributed to a given level of exposure until that period of exposure has been completed. An analogy may be helpful. If we wish to estimate the risk of lung cancer after five years of smoking, then only lung cancers occurring after this time contribute to the analysis; cancers occurring before five years have elapsed must be excluded. A similar conceptual error appears to have been made in two other studies, in which time to an event was measured from the beginning of the posited exposure period, rather than from the end, and inspection of the survival curves presented suggests that some of the observations necessary to define an infection as persistent were only made after the disease process had begun4,33. In three further studies, it is unclear as to whether the second and/or subsequent HPV positive samples preceded, or coincided with, that taken at diagnosis7,22,30. Only Ylitalo et al.11 appear to have recognised this problem and excluded from their comparisons smears taken during the year before diagnosis; however, this provides only limited reassurance because no information is provided on the presence or absence of cytological abnormality in the second of these two samples.
In addition to the possibility of exposure misclassification, there are a number of other conceptual problems with the analytical approach just described. In the studies reviewed, the interval between two positive tests was so short (median four months) that these infections cannot be considered prolonged in any meaningful sense. A more fundamental problem relates to inferences drawn from observations made at indeterminate points during the natural history of the infection: the distinction between a persistent and transient infection is arbitrary to the extent that it is dependent on the timing of the samples in relation to the natural history of the infection, and on the interval between them. For example, when a woman tests positive in her first sample, it is impossible to say how long she was infected before study entry. One study, which included women with incident HPV infections, was potentially more informative, because it at least provided a reference point from which the duration of infection could be measured14. However, the investigators' lower limit for persistent infection (six months) was less than the median duration of HPV in cohort members, thus classifying more than 50% as having a persistent infection.
Although some investigators report the distribution of exposure times necessary for an event to occur, the practice of characterising exposure levels according to the number of positive tests, and the failure to recognise exposure misclassification, have often drawn them into conflicting interpretations of the same data. For example, Koutsky et al.7 report that “the time from the detection of HPV to the development of cervical intraepithelial neoplasia grade 2 or 3 is remarkably short”, but they also conclude that “each additional positive HPV DNA test was associated with an increased risk of cervical intraepithelial neoplasia grade 2 or 3”; Liaw et al.3 conclude that “… the duration between initial HPV infection and detection of new SIL can often be even shorter than the approximately 2-year median time to diagnosis of our case patients”, but they also comment that “a persistent or repeated infection with a cancer-associated type of HPV strongly increases the risk of high grade SIL”. Wallin et al.6 report that “… a shorter time between the base-line Pap smear and the diagnosis of cancer [was] associated with increased detectability of DNA”, but also comment that “the odds ratio for cervical cancer associated with type-specific persistence of HPV was 58.7 (95% CI 10.2 to ∞)”.
Duration of preceding infection in women with, and without, progressive disease
The impact of duration of infection has also been explored by comparing the exposure history of women who were already to some extent diseased at study entry and in whom disease progressed, with that of women in whom it did not. Paraskevaidis et al.32 found no difference in the proportion of women with progressive and non-progressive lesions who had been “permanently HPV DNA positive”; the duration of follow up in the two groups is not reported and therefore no inferences can be drawn. Remmink et al.10 report that “HPV DNA was continuously present from the start of the study” in all women with progressive CIN; how often this occurred in women who had not reached the study end-point is not reported. We have already referred to a later report from this study in which the authors conclude that “persistent infection with high risk human papillomavirus is necessary for development and maintenance of CIN 3”1. This study has been influential in so far as it has been extensively cited by those who advocate the inclusion of HPV testing in the cervical screening programme. However, it is not clear how the authors reached their conclusion because they acknowledge that “since we did not take biopsy samples during follow-up, the time to first occurrence of CIN 3 was not known. We therefore omitted time from the analysis…”. In the absence of any measurement of time to an event, it is not possible to comment on the duration of the infection necessary for that event to occur.
Risk of disease in relation to time from first detection of HPV
Moscicki et al.8 report that in a cohort of young women, most of whom were HPV positive at study entry, the risk of LSIL increased rapidly after first detection of HPV, reaching a maximum after one to two years, before declining. Coker et al.12 also report that “high-risk HPV positivity [at baseline] was significantly associated with SIL development only within the first year of follow-up”. However, the incubation period for HPV-associated abnormality, that is the time from first detection of HPV to the first detection of cytological abnormality, can only be reliably measured if: the cohort includes women who are cytologically normal and HPV negative at study entry; repeated measures of HPV status are made at relatively short intervals; and the length of the latent period is minimised by immediately referring for colposcopic assessment all women found to have any grade of cytological abnormality. When Woodman et al.9 used this design to study the natural history of early cervical neoplasia in a cohort of young women, they found that the risk of high grade CIN was maximum 6–12 months after first detection of HPV 16, declining rapidly thereafter. This study also revealed that the incubation period for the discovery of cytological abnormality was frequently shorter than the interval between tests, with the result that the first detection of exposure and disease was often contemporaneous. In these circumstances, it is impossible to determine whether the onset of exposure precedes the onset of disease and therefore these cases are non-informative with respect to the temporal relationship between HPV detection and the occurrence of cytological abnormality. This analysis revealed a further difficulty: a precise measurement of the risk of high grade CIN associated with a given type-specific infection is dependent on the identification of the relevant exposure; this identification is problematic when, as was found in this study, HPV infection can be episodic, more than one HPV type is associated with high grade CIN and more than one HPV type is detected before diagnosis.
Recommendations for study design
With many variations possible according to study population, exposure and outcome, the adequacy of a particular cohort study design can only be judged in the light of the hypothesis or hypotheses being investigated. However, the inadequacies of design and analysis highlighted in this critique can be avoided by considering the following broad guidelines, not all of which are specific to the investigation of the role of HPV infection in cervical neoplasia, but instead reflect basic principles of cohort study design48:
- •All women must be free of the outcome of interest at the start of follow up.
- •Observations on exposure and outcome status should be made as frequently as possible.
- •Intensity and duration of follow up must be independent of exposure status.
- •Exposed and unexposed groups must be comparable with respect to any factor that might explain differences in outcome, other than the exposure under consideration; any differences must be accounted for in the analysis.
- •Characterise exposure levels according to duration of exposure rather than number of positive tests.
- •Exclude observations on exposure status made at or after the time of diagnosis from estimates of the duration of exposure necessary for that disease to occur.
- •Use analytical methods that distinguish the time during which a woman is unexposed from that during which she is exposed.
- •Use methods suitable for the analysis of survival data when subjects enter the study at different times or remain under observation for different periods.
Nested case–control studies should only be conducted within a well designed and conducted cohort study. Further:
- •When exposure status is assigned using archival material, match cases and controls on the date of the baseline sample.
- •Controls must be known to be free of the outcome of interest when their matched case is diagnosed.
- •Women who develop the outcome of interest must have the same period of observation as those who do not.
Although a number of cohort studies have now shown that HPV can be detected before the onset of epithelial abnormality, and one study has demonstrated that it is likely that the first episode of HPV precedes the first detection of abnormality, it is disappointing that only a few studies, those which made repeated measurements of HPV status, had the potential to fully explore the complexity of the exposure–disease relationship. Many of the studies reviewed had one or more substantial methodological flaws which may provide a partial explanation for the heterogeneity of effect observed in Table 1. Those studies, which have concluded that a ‘persistent’ infection is a prerequisite for the development of high grade CIN, are conceptually and methodologically flawed. The converse is more likely to be true: epithelial abnormalities of the cervix frequently appear shortly after the first detection of HPV. This is not to deny that HPV can establish a persistent viral infection or that a persistent viral infection is necessary for the development of invasive cancer. However, as to whether such persistent infections are characterised by the continuing detection of HPV or by a state of viral latency during which the virus remains undetectable for prolonged periods, again appearing shortly before the detection of cellular abnormality, is clearly a matter of some importance to the cervical screening programme, but one which remains to be established49. A woman cannot be defined as having a persistent infection in any meaningful virological sense merely because she tests positive for HPV on two occasions, some months apart, and therefore she should not, on the basis of this evidence, be considered to have a higher risk of cervical cancer.