Partitioning of Time Trends in Prevalence and Mortality of Bladder Cancer in the U.S.

Abstract

Purpose

To evaluate the relative contributions of incidence, stage-specific relative survival, and stage ascertainment to changes in Bladder Cancer (BC) prevalence and incidence-based mortality.

Methods

Partitioning of prevalence and incidence-based mortality trends into their epidemiological components.

Results

Bladder cancer prevalence estimated from our model increased but at monotonically decreasing rates until 2007 after which it decreased. The main forces underlying observed trends in BC prevalence were relative BC survival, which improved throughout period, and BC incidence, which increased at a decreasing rate until 2005 and declined thereafter. Mortality of persons ever diagnosed with BC increased at an increasing rate until 1997, increased at a decreasing rate 1997– 2005 and decreased thereafter. The primary forces accounting for mortality trends were changes in mortality in the general population, which improved at an increasing rate during most of 1992–2010, the most important factor, and changes in incidence. Stage ascertainment did not improve during 1992–2010.

Conclusions

While mortality rates improved, these gains largely reflected improvements in U.S. population survival rather than from improvements in BC-specific outcomes.

Introduction

Bladder cancer (BC) is a common and costly cancer for the health care system. Specifically bladder cancer is the 4^th most common cancer among men in the U.S. and the 9^th most prevalent cancer globally^1–5. Understanding the dynamics of current and historical epidemiological trends in BC and their epidemiologic causes provides important insight into reducing the health and human burden associated with this condition. This is because time trends in measures such as disease prevalence and especially incidence-based mortality reflect important information on periods of success and failure in disease care and prevention. Such analysis is made possible by the availability of historical longitudinal data from the Surveillance Epidemiology and End Results (SEER) Registries, which provide sufficient statistical power and detail to accurately document the trends of BC prevalence and incidence-based mortality as well as trends in their epidemiological causes in the United States.

BC is a heterogeneous disease with regard to patient and tumor characteristics. Even though the SEER provides sufficient statistical power to account for these differences by stratifying available cases into well-defined homogeneous groups, the time trend in associated prevalence/mortality is itself heterogeneous due to differences in the relative impact of the causal components that contribute to changes in these measures over time. The partitioning analysis^6–9 used in this study created reliable quantitative estimates of the relative contributions of a number of such components derivable from SEER data to the overall trends in BC prevalence/mortality. Changes in the direction and the magnitude of the contribution of a component detected allowed us to identify weak and modifiable links in BC diagnosis or treatment.

Methods

Data Sources

Data from SEER covering the years 1973–2015 were used in this study¹⁰. Estimates were for the 1992–2010 period. Data prior to 1992 were used as a look-back period to calibrate the model. The last year for which a full 5-year survival estimate could be calculated was 2010.

Statistical analysis

This study used a recently developed partitioning approach^6–9 based on an explicit representation of prevalence and mortality with no simplifying assumptions. The full mathematical derivation of the partitioning approach and its application to lung cancer is contained in Akushevich et al.⁷. In brief, the method predicts prevalence and mortality and decomposes (or partitions) their time trends into their constituent components by calculating the relative impact each component has on the overall trend as well as intertemporal changes in the strength and direction of these impacts. Specific outcome measures are age-specific and age-adjusted prevalence and incidence-based mortality and the constituent components are incidence, relative survival, frequency of cancer stage at diagnoses, and mortality for the general population (for mortality only).

The model of cancer prevalence is based on the idea that the probability of being prevalent at a given age requires being incident prior to and survival post that age. Similarly, for incidence-based mortality the probability of dying in an age interval requires death in this interval and being incident at an earlier age. The mortality model includes two components: M = R_p + R_is. Non-disease-specific mortality (R_p) represents the mortality in the subset of the population prevalent with BC that is not directly or indirectly caused by BC. Disease associated mortality (R_is) represents the additional mortality due to the presence of BC (R_is = 0 when relative survival equals 1⁷).

The time trends in age-adjusted prevalence and mortality are defined as their derivatives with respect to time y. Explicit calculation of time trends in age-adjusted prevalence results in: P′(y)/P(y) = T_inc(y) + T_π(y) + T_S(y), representing changes in incidence, cancer stage at time of diagnosis (localized regional and distant as defined by SEER), and relative survival respectively. The time trend in incidence-based mortality is initially partitioned into five components: M′(y)/M(y) = T̂_p(y) + T̂_μ(y) + T̂_inc(y) + T̂_π(y) + T̂_S(y), where T̂_P(y), is the change in disease prevalence, T̂_μ(y) is the change in mortality in the general population and the effects of the terms T̂_inc(y), T̂_π(y), and T̂_S(y) describe the additional effects of the prevalence components. Of these T̂_P(y) and T̂_μ(y) are derived from R_p while T̂_inc(y), T̂_π(y), and T̂_S(y) are derived from R_is. However, since the expression for T̂_P(y) (eq. (7) of ref.⁷) contains three terms associated with these partitioning components, they can be joined with T̂_inc(y), T̂_π(y), and T̂_S(y) resulting in a more parsimonious representation of the trend in incidence based mortality. The downside of such consolidation is the loss of attribution of the effects to R_p and R_is.

Explicit expressions for partitioning components are given in eqs (5,7–9) of ref.⁷. Each component is to be interpreted as the rate of change at any point in time––increasing if > 0, decreasing if < 0––with the magnitude of the effect indicating the speed of the change. The sum of the contributions adds to +100% if the decomposed measure is increasing and −100% if the decomposed measure is decreasing at a given point in time.

To assess goodness of fit of our model, time patterns of prevalence and mortality calculated using the partitioning approach are shown in comparison to corresponding estimates provided by the SEER*Stat tool that produces similar statistics. When possible, SEER*Stat estimates were generated using 2 alternative lookback periods: 18 years, the longest period available for all years of BC initial diagnosis; and up to 23 years, the longest period available in the data, though not for all years of BC initial diagnosis.

Adding look-back years increased prevalence by identifying additional still surviving persons diagnosed with BC over the lookback period (Figure 1). Consequently, our model, which used a B-spline-based extrapolation approach to estimate the epidemiological measures required for the prevalence calculation outside of the bounds of the lookback periods, yielded the highest prevalence levels. The agreement between empirical data and our model is good: the shape of the model mirrors empirical patterns for 18 years, with differences in magnitude reflecting the effect of accumulated prevalence from earlier years as expected.

Figure 1. Estimates of bladder cancer prevalence.

Age-adjusted prevalence using our model (solid black line), age-adjusted prevalence using SEER*Stat using an 18-year look-back period (solid black (red) line with filled circles), age-adjusted prevalence using SEER*Stat using an up to 23-year look-back period (solid black (blue) line with filled squares).

Estimates from our model for BC incidence-based mortality shows good agreement with rates derived from SEER*Stat (Figure 2): Mortality patterns from our model and those from SEER*Stat (solid squares) are similar. Our estimates of mortality are higher than those from SEER*Stat in part because our estimates of BC prevalence are higher. As with prevalence, BC mortality increased gradually for most of the observational period, reaching a peak in 2007, followed by a modest decline thereafter.

Figure 2. Estimates of bladder cancer incidence-based mortality.

Total mortality using our model (solid black line), non-disease-specific mortality R_p using our model (dotted line), disease associated mortality R_is using our model (dashed line), total mortality using SEER*Stat (black (red) filled squares).

Results

Decomposition of changes in bladder cancer prevalence

Bladder cancer prevalence estimated from our model increased but at monotonically decreasing rates until 2007 after which it decreased (Figure 3). Partitioning the rate of change in BC prevalence showed that increases between 1992–2003 (rates of change > 0 throughout this subperiod) reflected increased incidence and improved relative BC survival.

Figure 3. Partitioning of bladder cancer age-adjusted prevalence.

See color or black and white figure.

Relative survival increased BC prevalence at a constantly decreasing rate. Relative survival improved over 1992–2010 overall, but the strength of this effect declined over time as accumulated persons with a BC diagnosis who survived in the early study years died at a later time (e.g. accumulation of survivors). Incidence decreased during the study period with a sharp rise in the rate of decrease after 2003. By 2007 the strength of the decrease in incidence overpowered the effect of improved survival leading to the overall decrease in BC prevalence. Increased incidence accounted for almost a third (32%) of the prevalence increase during 1992–2003, and improved relative survival accounted for almost all of the rest (65%). The contribution of earlier BC ascertainment was minimal (3%). Evaluated at 1995, the contributions were 33%, 63%, and 3%, respectively. At 2000, the contribution of relative survival was slightly higher (66%) and increased incidence slightly lower (31%). In 2010, the decline in prevalence largely reflected a decline in incidence (−131%), with improved relative survival contributing 38% and worsening stage ascertainment, contributing −7% with individual contributions summing to −100% versus 100% at 1995 and 2000 when BC prevalence was increasing.

Decomposition of changes in bladder cancer incidence-based mortality: five component model

Bladder cancer incidence-based mortality increased at an increasing rate until 1997, increased but at a decreasing rate from 1997 to 2005 and decreased thereafter (Figure 4). The rate of change in incidence-based mortality evaluated at 2000, when the mortality trend was increasing showed that the contribution of the prevalence trend was 153% (part of the R_p component of mortality). Other factors contributed more modestly: −53% from the mortality in the general population (R_p); 33% from incidence (R_is); −32% from relative survival (R_is), and 0% from stage ascertainment (R_is). Increased prevalence was the major driver of the increasing death rate. After accounting for immediately prior prevalence, the flow of additional newly diagnosed cases added to the prevalence pool and thereby the number of deaths. By contrast, the decrease in general mortality lowered the rate of change in mortality of BC patients below what it otherwise would have been. At 2010, when overall BC mortality was decreasing, these contributions were: 25% from prevalence, −124% from the mortality in the general population, 22% from incidence, 19% from relative survival, and 2% from stage ascertainment. The last three factors also contributed to the mortality trend through the trend in prevalence (not explicitly shown in Figure 4).

Figure 4. Partitioning of bladder cancer age-adjusted incidence-based mortality – 5 components.

See color or black and white figure.

Decomposition of changes in bladder cancer incidence-based mortality: four component model

Combining effects of trends in general morality, relative survival, and stage ascertainment operating through trends in prevalence with the effects of trends in these factors operating in addition to prevalence yielded a 4-factor specification (Figure 5). This consolidation changed the component contributions to 96% from incidence, 4% from stage ascertainment, 53% from relative survival, and −53% from the mortality in the general populations at 2000. At 2010, the component contributions were −22% from incidence, 3% from stage ascertainment, 44% from relative survival, and −124% from general mortality.

Figure 5. Partitioning of bladder cancer age-adjusted incidence-based mortality – 4 components.

See color or black and white figure.

The improvement in incidence-based mortality was largely driven by reduced general mortality. Although boosting mortality in both Figures 4 and 5, the contribution of relative survival was over twice as high in Figure 4, which implies that somewhat less than half of the contribution of relative survival of persons ever diagnosed with BC operated through the prevalence trend in 2010. The contribution of incidence is much higher in Figure 5 than in Figure 4; a large portion of the effect of incidence on incidence-based mortality operated through prevalence. Some of the remaining contribution of relative survival may have been from deaths that might have occurred earlier if there had not been improvements in BC management.

Discussion

In this study, we decomposed historical trends in bladder cancer prevalence and incidence-based mortality into their epidemiological causal components to identify sources of change in the epidemiology of this cancer and identify potential actionable targets for future private and public policy health interventions.

By the end of the study period, both BC prevalence and incidence-based mortality were declining. The turning point from increasing to decreasing occurred in 2005 for incidence-based mortality and 2007 for prevalence. Most of the improvement in incidence-based mortality was attributable to reductions in mortality due to causes other than BC. Decreased BC incidence was the major determinant of recent declines in BC prevalence. Neither improvements in general mortality nor declining BC incidence could be attributed to improved BC treatment.

Life expectancy increased substantially in the U.S. Much of the increase is attributable to reductions in cardiovascular mortality^11,12. For the U.S. elderly, the age strata at which most BC is initially diagnosed, life expectancy increased between 1990 and 2010 by 1.9 years at age 65 and by 1.2 years at age 75¹³. These trends benefited persons with and without a BC diagnosis and accounted for most of the life expectancy gains experienced by BC-diagnosed persons during 1992–2010.

Bladder cancer incidence reductions mainly reflected reduced exposure to BC risk factors, most importantly to smoking^1,14–17. One study estimated that smoking accounted for 50% of BC tumors¹⁷. Based on a literature review, Burger et al., concluded that about 20% of BC incidence is related to carcinogenic exposures in the workplace¹. However, contributions of trends in workplace and other environmental exposures such as arsenic in the drinking water¹⁸ are much more difficult to quantify than are declines in smoking rates¹⁹.

Earlier stage ascertainment of BC might increase BC incidence, at least initially, however we documented little change in stage ascertainment during 1992–2010, at least for the stages broadly defined by SEER.

A study of lung cancer (LC) employed the same methodology as this study over a comparable observational period⁷. In contrast to BC, a strong adverse role for the stage ascertainment component was identified for LC. Although survival improved for both LC and BC, the primary reason for improvement in LC was disease-associated mortality (R_is) – that is, improvements in the treatment of LC, while for BC most mortality gains were due to improvements in non-BC-specific mortality (R_p). Finally, changes in incidence-based mortality among patients with LC were almost totally determined by changes in incidence, while for BC changes in incidence-based mortality were shared more equally among the sources considered. The stark differences in sources of changes in prevalence and incidence mortality trends for LC and BC point to the importance of conducting separate partitioning by cancer type.

Finally, we identified that the accumulation of survivors can, over time, confound observed survival trends. This can be observed by comparing the effects of the relative survival components in Figures 3, 4, and 5. The contribution is beneficial in prevalence, beneficial in 5-component mortality, and harmful in 4-component mortality. Since the 4-component plot for partitioning of mortality is obtained from the 5-component mortality partitioning by decomposing prevalence and combining respective contributions (i.e., survival from 5-component mortality with the contributions of survival from prevalence), occurrence of the harmful effect of relative survival can be explained by accumulation of survivors in previous years and their higher mortality (because of higher prevalence of survivors) in last years.

We acknowledge several study limitations. The calculations were based on empiric parametric models for age-dependence of incidence and fractions of stage at diagnosis as well as age- and time-after-diagnosis dependence for survival. Although we carefully constructed such models based on their empiric properties or widely accepted modeling approaches, statistical uncertainties still remain. The data contained an insufficient number of female cases to perform trend decomposition stratified by gender. The overwhelming part of the combined trend consisted of male trends. SEER did not have access to individual-level data on common BC-risk factors such as smoking. Therefore, trends in these measures were not integrated into the decomposition analysis. Finally, improvements in relative BC survival might have been greater if the data had included most recent years and data from the immediate future as the U.S. Food and Drug Administration recently approved 5 immune-checkpoint inhibitors for locally advanced and metastatic BC²⁰. However, technological progress in treatment of non-muscle invasive BC has been quite limted²¹ while muscle-invasive BC treatment is prone to delays²².

Conclusion

Long term trends in prevalence and incidence-based mortality of BC are affected by the accumulation of survivors resulting from reductions in mortality in early years. Bladder cancer prevalence has declined in recent years, largely reflecting a decline in incidence, most likely attributable to a reduction in smoking. While incidence-based mortality rates of persons ever diagnosed with bladder cancer have improved, these gains largely reflect improvements in the health of the general US population during 1992–2010 rather than improvements in bladder cancer-specific outcomes or in stage at which the cancer is first ascertained.

List of Abbreviations

BC

Bladder Cancer

SEER

Surveillance Epidemiology and End Results

LC

Lung Cancer