Contribution of H. pylori and Smoking Trends to US Incidence of Intestinal-Type Noncardia Gastric Adenocarcinoma: A Microsimulation Model

Jennifer Yeh and colleagues examine the contribution of IHelicobacter pyloriI and smoking trends to the incidence of past and future intestinal-type noncardia gastric adenocarcinoma. Please see later in the article for the Editors' Summary


Intestinal-type noncardia gastric adenocarcinoma (NCGA) microsimulation model
The population-based gastric cancer model simulates the natural history of intestinaltype NCGA. In a Monte Carlo simulation, individuals transition among health states one at a time and the detailed information for each individual is continuously tracked, allowing the natural history and course of disease to be conditional on that individual's risk factor profile.
Specifically, events are simulated for a sequence of individuals using random numbers based on event probabilities (e.g., the probability of progressing from intestinal metaplasia to dysplasia).
An individual's risk factors may be allowed to change each year to reflect changes in status.
Based on epidemiologic data [1], we assumed that precancerous lesions were already present in a subset of 20-years olds, with a proportion being greater among those infected with H. pylori. Because smoking initiation largely occurs between the ages of 15 and 30, we assumed smoking did not increase the presence of precancerous lesions at age 20. Individuals are simulated until they die or reach the final calendar year of analysis.

Smoking-specific background mortality
To reflect the higher risk of background mortality among smokers, we calculated smoking intensity-specific rates using age-specific birth cohort rates from the Berkeley Mortality Database (http://www.demog.berkeley.edu/~bmd/) and U.S. Social Security Administration [2], relative risk (RR) estimates (range = 1.1-3.8) based on the American Cancer Society Cancer Prevention Study II (CPS-II) [3] and the prevalence of smokers by intensity type at each year. We assumed that 1) for smokers ages 55 and younger, heavy smokers faced a 1.3 and 1.2 higher risk of background mortality compared to low and moderate smokers, respectively, 2) for individuals who quit smoking, background mortality risk remained elevated to intensity-specific levels for 5 years, and then declined afterwards to a constant relative risk for the remainder of his/her lifetime (RR = 1.5) [4], and 3) relative risks were constant across birth cohorts.

Natural history progression rates (identified via model calibration)
To infer unobservable natural history progression rates among precancerous lesions to invasive cancer, we used a likelihood-based calibration approach, previously described [5], to estimate natural history parameters and ensure model predictions are consistent with epidemiologic data. In short, based on the published literature, we established a priori plausible ranges for all model input parameters, including natural history parameters and relative risks associated with H. pylori and smoking [6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23] (see Table S1). We then specified calibration targets from epidemiologic studies and Surveillance, Epidemiology and End Results (SEER): age-specific intestinal metaplasia prevalence in 1990 (10-year age groups between ages 30 and 89), age-specific intestinal-type NCGA incidence for each To explore the parameter space defined by the plausible ranges for our model parameters, we used a simulating annealing algorithm, a computationally feasible and timeefficient directed parameter search for microsimulation models [29]. Goodness-of-fit was defined based on the sum of log likelihoods for each outcome (assuming binomial distributions for all parameters). In addition, to identify parameter sets that fit both to our calibration targets and exhibited a decline in age-standardized intestinal-type NCGA incidence over time as observed in SEER, we included a slope variable to our goodness-of-fit, equal to the difference squared between the simulated and SEER slope between 1978-79 and 2007-08. We assumed that baseline disease progression rates remained constant across birth cohorts, although among birth cohorts, overall rates would vary due to birth-cohort specific risk factor patterns.
To reflect the uncertainty in disease natural history on our modeled outcomes, we conducted 500 independent simulating annealing searches to identify multiple parameter sets that fit equally well to the calibration targets. Each simulating annealing search entailed 500 directed searches, each simulating 15.5 million individuals (7.5% of the population size for each birth cohort).
Among the 500 searches, we then identified 119 parameter sets with statistically similar fit to the best-fitting parameter set (α= 0.05). Among the 500 searches, we then identified 119 parameter sets with statistically similar fit to the best-fitting parameter set (α= 0.05). From these "good-fitting" parameter sets, we randomly selected a subset of 50 to reflect uncertainty in disease natural history and serve as natural history parameters for our scenario analyses (see Table 1 of the main text). For all modeled outcomes, we report the expected value (mean) and range (minimum, maximum) among the 50 parameter sets. As depicted in Figures S1-S3, the majority of modeled outcomes fell within the 95% confidence intervals of the calibration targets on intestinal metaplasia prevalence, cancer incidence and cancer stage distribution, though there was variability among the parameter sets.
In addition, the relative risks of intestinal-type NCGA associated with H. pylori (4.7 vs.

ADDITIONAL RESULTS: AGE-SPECIFIC INTESTINAL-TYPE NCGA INCIDENCE
Modeled age-specific intestinal-type NCGA incidence suggest that rates have declined and are projected to continue to decline due to observed risk factor trends (see Figure S5 for base case scenario). Under the 'no tobacco control' scenario, as depicted in Figure S6, agespecific rates were also estimated to decline between 1978 and 2040. Gastritis to atrophy* 0-0.10 assumption *Constant exponential rate (r) of decline per birth cohort as described in the following equation: (1-r)^t, where t = year of birth -1901.