Effects of Referral Bias on Estimates of Anal Intraepithelial Neoplasia Progression and Regression Rates in a 3-State Markov Model.

The study aim is to compare anal intraepithelial neoplasia (AIN) progression and regression rates in a cytology inception cohort to estimates based on the subcohort referred for ≥1 high-resolution anoscopies (HRAs).A cytology-based retrospective cohort was assembled including the anal cytology histories and invasive anal cancer (IAC) outcomes of all HIV-infected adults under care between 2001 and 2012. A 3-state Markov model (<HSIL↔HSIL→IAC) was estimated separately for all patients and for the subcohort undergoing ≥ 1 HRAs with biopsy. Cytology was adjusted for misclassification. State transition rates (per person-year) and covariate hazard ratios were estimated using the R package msm.Of 2804 eligible patients in the inception cohort, 629 (22%) were in the HRA subcohort and 2175 (78%) in the non-HRA subcohort. Patients in the HRA subcohort were more likely to have baseline CD4<350, viral load >400, and to have HSIL at baseline and thereafter. They also had more anal cytology examinations (median 6 vs 3) and longer follow-up (median 5.5 vs 3.6 years). State transition rates were overestimated in the HRA subcohort relative to inception cohort, but the degree of discordance varied by transition: for <HSIL to HSIL (0.44 vs 0.04); for HSIL to <HSIL (0.56 vs 0.17); and for HSIL to IAC (0.014 vs 0.011). Beneficial covariate effects on the <HSIL to HSIL transition were concordant (P < 0.05) for time-updated HIV viral load, CD4 count, and antiretroviral therapy. The observed effects of HRA-triage bias may be relevant to estimates of AIN state transitions from other cohorts subject to referral bias.

INTRODUCTION

Among the 10 classic criteria of Wilson and Jungner to justify implementing a screening program is an adequate understanding of the natural history of the index disease.[1] However, a number of biases may distort accurate estimation of key transition rates along the natural history pathway.[2] Referral bias, a form of selection bias, may occur when natural history is modeled using a referred study cohort with a prognostic factor distribution that differs from that of the source population at risk of disease. Modeling of natural history and the impact of prognostic factors is therefore ideally based on the experience of inception cohorts so as to minimize referral bias. Screening programs for anal intraepithelial neoplasia (AIN) typically limit referral for high-resolution anoscopy (HRA) to those with abnormal anal cytology results (high-grade squamous intraepithelial lesion [HSIL], low-grade SIL [LSIL], atypical cells of uncertain significance [ASCUS] or atypical cells, cannot rule out high grade [ASC-H]) or other clinical abnormalities.[3] The aim of this study was to compare estimates of AIN progression and regression rates in a cytology inception cohort to estimates based on the subset of patients referred for ≥ 1 HRA with biopsy procedures. We propose the term HRA-triage bias to designate the form of referral bias demonstrated in the following analysis.

METHODS

A clinical care cytology-based retrospective cohort was assembled including the anal cytology histories and invasive anal cancer (IAC) outcomes of all HIV-infected adults under care at the UCSD Owen Clinic between 2001 and 2012. Eligibility criteria, screening program characteristics, and study measure definitions were as previously reported.[4] A 3-state Markov model (5] We present absolute and relative differences in state transition rates between the inception cohort (RIC) and HRA referral cohort (RHRA) where the relative difference is defined as (RHRA– RIC)/RIC. The study was approved by the UCSD Human Research Protection Program (Project 071931).

RESULTS

Of 2804 eligible patients in the inception cohort, 629 (22%) were in the HRA subcohort and 2175 (78%) in the non-HRA subcohort. Patient characteristics are presented in Table 1. Patients in the HRA subcohort were: more likely to be male (91.6% vs 88.2%, P = 0.02), white (64.7% vs 60.8%, P = 0.01), men having sex with men (MSM) (82.7% vs 76.3%, P = 0.002), to have baseline CD4<350 cells/mm3 (50.6% vs 42.4%, P < 0.0001) and HIV plasma viral load >400 copies/mL (58.7% vs 49.6%, P < 0.0001), and more likely to have HSIL at baseline (32.0% vs 5.3%, P < 0.0001) and thereafter (98.9% vs 22.6%, P < 0.0001). They also had more anal cytology examinations (median 6 vs 3, P < 0.0001) and longer follow-up (median 5.5 vs 3.6 years, P < 0.0001) than the non-HRA subcohort. The groups did not differ by age at entry (median 40.2, interquartile range 34.1 – 46.4), use of antiretroviral therapy at entry (75%), or smoking (29.9%). Of the 23 confirmed incident IAC cases, all occurred in the HRA subcohort. One or more infrared coagulation (IRC) treatments of HSIL lesions were documented in 26% of HRA subcohort patients.

TABLE 1

Patient Characteristics by Study Group

Table 2 presents estimates of state transition rates (per person-year) for the inception cohort and for the HRA subcohort. State transition rates were overestimated in the HRA subcohort relative to the inception cohort, but the degree of discordance varied by transition: for

TABLE 2

Estimates of State Transition Rates (per Person-Year) Adjusted for Cytology Misclassification Assumptions, by Study Group (All Patient Inception Cohort vs HRA Subcohort)

Covariate effects on state transition rates, estimated separately for the inception cohort and for the HRA subcohort are presented in Table 3. Beneficial covariate effects on the

TABLE 3

Estimated Unadjusted Hazard Ratios (95% CI) of Time-Updated Covariates, by State-Transition and by Study Group

DISCUSSION

An inception cohort has been defined as “a group of individuals identified and assembled for subsequent study at an early and uniform point in the course of the specified health condition.”[6] Failure to assemble an inception cohort can have unpredictable and often important effects on estimates in natural history studies.[7] In this analysis, we found that Markov model state transition rates were overestimated if cohort membership was conditioned on receipt of one or more HRA procedures as compared to unconditional estimates from the cytology inception cohort. Relative to inception cohort estimates, the effect of HRA referral bias was greatest for the

What factors may account for the differential effects of conditioning analytic cohort membership on receipt of HRA? We discuss first the observed differential inflation of state transition rates and then effects on covariate effect estimation. Because the development of HSIL cytology was the primary criterion for referral to HRA, it is unsurprising that the transition rate from

With regard to the impact of HRA subcohort selection on covariate effect estimation, it is noteworthy that there was concordance between both analytic cohorts in the significant protective effects of antiretroviral therapy, HIV viral load suppression, and CD4 + lymphocyte count on the

Our explanatory speculations highlight the previously noted unpredictability of the effects of failure to assemble an inception cohort owing to limited ability to account for discordant distributions of both measured and unmeasured prognostic factors in the context of referral bias.[7] Although we have identified no other publications examining the effects of referral bias on estimation of AIN clinical evolution, we do call attention to work examining the effect of referral bias on estimation of progression to cirrhosis in hepatitis C infected analytic cohorts. Fu et al found in a simulation study that the estimated 20-year probability of progression to cirrhosis in patients referred to a liver clinic was 4-fold higher (20%) than the estimate for community-based samples (5%). The authors concluded that “When attempting to establish the natural history of new diseases with long incubation periods, researchers should be on the lookout for potential biases that result from the way patients are referred into clinical cohorts.”[8]

Although not the primary focus of our analysis, it is notable that all 23 IAC cases were documented in the HRA referral subcohort. We believe that there are several potential explanations for this observation. First, because HSIL was the primary criterion for triage to HRA, the HRA subcohort was greatly enriched with patients at highest risk of progression to IAC. Second, we did not implement IRC treatment of HSIL lesions until 2007 (year 7 of a 12-year study period). Third, although we found in our primary analysis that IRC increased the downgrading of HSIL to 4] Finally, even among patients undergoing HRA surveillance during the IRC treatment era (2007 and thereafter), assuring regularity of followup examinations was a challenge because of limited HRA operator availability and patient adherence to followup recommendations.

Our study results are subject to several limitations. (1) Unknown or imprecisely measured factors may have contributed to prognostic differences between the inception and HRA subcohorts. (2) For both analytic cohorts, there may be residual state misclassification after correcting for the fallibility of cytology using HRA-directed biopsy as the reference standard; however, HRA-directed biopsy is itself a fallible reference standard.[9] (3) The 10] (4) Because of the composition of our analytic cohorts, model inferences are most robust for HIV-infected MSM. (5) Power to detect covariate effects on the transition from HSIL to IAC is limited by the small number of IAC endpoints.

CONCLUSIONS

Modeling AIN state transition rates in a noninception cohort defined by differential HRA referral resulted in substantial overestimation of the