# Jamie Cohen

Research Scientist, Institute for Disease Modeling

Dissertation Title:  "Applications of Mathematical Modeling to Evaluate Cervical Cancer Prevention in an HIV-Endemic Setting"Mathematical models that simulate the burden of disease and project health and economic outcomes under various scenarios can help policymakers decide how to optimally allocate resources. This dissertation uses simulation modeling to evaluate cervical cancer prevention policy and describe its effects in South Africa, an HIV-endemic setting where the risk of cervical cancer is heterogeneous.

In the first chapter, I evaluated the cost-effectiveness of cervical cancer screening in unvaccinated women. I built a microsimulation model of HIV infection, HPV infection and cervical disease, which captured the impact of HIV infection on HPV and cervical natural history. Disease dynamics were represented by transitions between mutually exclusive health states, including oncogenic and non-oncogenic HPV infection, grade of pre-cancer, and stage of cancer. I calibrated the model to South African epidemiologic data and compared current screening guidelines to 55 alternative strategies that varied the age to start screening and screen frequency for women based on their HIV status and history of HIV testing. Costs included cancer and HIV screening, diagnosis and treatment. Health outcomes included cancer cases and disability-adjusted life years (DALYs) averted. I conducted a cost-effectiveness analysis to determine the optimal screening strategy at different willingness to pay (WTP) values.

I found it was always optimal to screen HIV-uninfected and women of unknown HIV status starting at a younger age and more frequently than current guidelines recommend. For women with a diagnosed HIV infection, optimal screen frequency depended upon WTP; at a WTP of $1,300, it was optimal to reduce screen frequency compared to current guidelines and at a WTP of$5,200 it was optimal to increase screen frequency relative to current guidelines. These findings were robust to variations in cost and improvements in screen coverage and efficacy.

In the second chapter, I quantified the impact of HPV vaccination on the cost-effectiveness of cervical cancer screening and determined whether the impact was moderated by differential vaccine protection in women living with HIV. Human papillomavirus (HPV) vaccination may offer an opportunity to reduce the frequency of cervical cancer screening.

I refined a microsimulation model of HIV infection, HPV infection and cervical carcinogenesis. I assumed all women in the model received a completed course of the HPV vaccine at age 9. I modeled infection with HPV genotypes 16, 18, 31, 33, 45, 52, 58, other oncogenic genotypes, and all non-oncogenic genotypes. I calibrated the model to South African epidemiologic data and compared current screening guidelines (which are agnostic to vaccination status) to 15 alternative strategies that varied screen start age and screen frequency for “low-risk” women and screen frequency for “high-risk” women, and considered various rates of vaccine waning in immunocompromised women. Costs included cancer screening, diagnosis and treatment. Health outcomes included cancer cases and disability-adjusted life years (DALYs) averted. I conducted a cost-effectiveness analysis to determine the optimal screening strategy at different willingness to pay (WTP) values and rates of vaccine waning.

I found that at a willingness to pay of $5,200 per DALY averted, the upper end of an empirical WTP range for South Africa, it would always be cost-effective to increase screening relative to current guidelines for bivalent vaccinated women. The optimal strategy at this WTP was consistent with the optimal strategy in Chapter 1 in unvaccinated women, suggesting that screening guidelines need not be differentiated by vaccination status. At the lower end of the WTP range ($1,300 per DALY averted), it would be cost-effective to increase the screen start age to 40 years old for low-risk women. These results were robust to changes in rate of vaccinate decline in immunocompromised women, vaccine efficacy in women living with HIV, and screen coverage and compliance. When we considered the nonavalent vaccine and when CIN and HPV treatment was perfectly effective, the optimal screening strategy was less frequent across the WTP range.

This analysis suggested that it could be cost-effective to relax screening frequency in vaccinated women, depending upon WTP and vaccine type used. They also provided confidence that any differences in vaccine efficacy in women living with HIV will not drive major differences in screening.

In the third chapter, I explored the impact of model structural uncertainty on our epidemiologic inference and policy results. Decisions about model structure will undoubtedly have consequences for epidemiologic outcomes, estimates of cost-effectiveness, and policy conclusions. Yet structural uncertainty is frequently ignored, even though it may have a much greater impact on model results and conclusions than parameter uncertainty. While it is common to acknowledge potential limitations of model structure and identify assumptions that have been made, there is no clear guidance on methods to explicitly evaluate structural uncertainties. And while modelers are guided by the principle of making a model only as complex as necessary, there is little consensus on what qualifies as necessary.

In this chapter, I compared several alternative model structures that capture the process of natural immunity and meaning of HPV re-detection and quantified the impact associated with these structural decisions both in terms of our prediction accuracy and policy implications. I found that all five model structures fit the calibration targets well, with only small variations in performance. The fitted models resulted in significant variation in key model parameters, such as the level and duration of natural immunity, and rates of progression between HPV infection, lesion and invasive cervical cancer. Allowing for infections to become latent and re-activate impacted the age distribution of causal HPV infections and the subsequent health impact and cost-effectiveness of vaccination strategies that vary the end age of vaccination. Model structures that do not allow for latency predicted a four-year older average age of causal HPV infection compared to models that accounted for latency. Structural decisions regarding who acquires natural immunity did not produce much difference in other model natural history outcomes nor cost-effectiveness of vaccination policy.

These results imply that the specific structural uncertainties I explored are meaningful for the way we have, and potentially should, model HPV. Specifically, models that ignore the possibility of HPV latency and re-activation may over-estimate the benefit of vaccinating up to older ages. They also demonstrate that decisions regarding who acquires natural immunity and at what level are less influential, so long as natural immunity exists in the model. While this analysis was specific to HPV modeling decisions, it serves as an example of how structural decisions matter for modeling in general.