Social Network Effects of Nonlifesaving Early-Stage Breast Cancer Detection on Mammography Rates

Abstract

Objectives. We estimated the effect of anecdotes of early-stage, screen-detected cancer for which screening was not lifesaving on the demand for mammography.

Methods. We constructed an agent-based model of mammography decisions, in which 10 000 agents that represent women aged 40 to 100 years were linked together on a social network, which was parameterized with a survey of 716 women conducted through the RAND American Life Panel. Our model represents a population in equilibrium, with demographics reflecting the current US population based on the most recent available census data.

Results. The aggregate effect of women learning about 1 category of cancers—those that would be detected but would not be lethal in the absence of screening—was a 13.8 percentage point increase in annual screening rates.

Conclusions. Anecdotes of detection of early-stage cancers relayed through social networks may substantially increase demand for a screening test even when the detection through screening was nonlifesaving.

Women often overestimate the mortality reduction from mammograms,^1–3 and anecdotes of women with early-stage breast cancer being diagnosed through mammography can serve as particularly powerful motivators in encouraging other women to have screening mammograms.^4–6 Such diagnoses, which are often shared through discussions between friends, family, coworkers, and other acquaintances, are also often viewed by patients as lifesaving.⁷ These discussions can be viewed as the transmission of information through a social network. Typically, social networks are defined as sets of individuals with links between pairs of individuals. In this article, we define 2 individuals as linked if they discuss their health history and outcomes (such as breast cancer diagnosis and treatment) with each other.

Although women often view the detection of early-stage breast cancer through mammography as lifesaving, a recent study estimated that more than three quarters of women with screen-detected cancer have not actually had their lives saved by early detection.⁸ Rather, these women have cancers that (1) would never have been detected clinically if not detected through screening (thus, they were overdiagnosed), (2) eventually would have been detected clinically but would not be lethal, or (3) eventually would be lethal despite being detected by screening. Physicians and epidemiologists have long understood that high rates of early-stage diagnosis through screening do not necessarily indicate that the screening reduces mortality. Furthermore, when issuing cancer screening recommendations, professional organizations generally use mortality reduction as a primary measure of the screening’s efficacy.^9–12

Understanding how nonlifesaving early detection of breast cancer through screening drives patient demand for screening has several important implications. First, professional organizations currently disagree about when and how frequently women should be screened for breast cancer.^9,10 Those who believe that some women are screened too early and too often may find demand for mammograms driven by such nonlifesaving diagnoses problematic. Our analysis estimated how much demand for mammograms may be driven by women’s incorrect assumptions about whether breast cancer screening was lifesaving for individuals in their social networks. More generally, the fact that patients often use frequency of early detection as a proxy for the screening efficacy⁷ means that patients with limited time and resources might prioritize a screening test that is relatively ineffective at reducing mortality over a test that more successfully reduces mortality if the former has higher early detection rates than the latter. We examined the potential patient-driven demand for screening that is motivated by nonlifesaving screen-detected cancers discussed in a social network.

It is challenging to quantify this demand through observation or experiment because it is often not possible to determine at the individual level whether early detection through screening affected mortality. If a woman with early-stage, screen-detected breast cancer does not die of breast cancer, it is usually not possible to know with certainty if she would have died from breast cancer had she not been screened. If a woman with early-stage, screen-detected breast cancer ultimately dies of breast cancer, we know that screening was not lifesaving, but there may be many years between her initial diagnosis and death from breast cancer. Therefore, we used an agent-based simulation model to estimate how nonlifesaving early-stage breast cancer detection relayed through a social network might influence mammography rates. A major advantage of using a simulation model to study this problem is that, within the model, we can distinguish between agents (individual women) whose early detection of breast cancer through screening was lifesaving and those whose early detection was not lifesaving.

In our analysis, we explored how population-level mammography rates might change if individuals were able to account for the fact that not all early-stage, screen-detected cancers are lifesaving. We used the results of a customized survey conducted through RAND’s American Life Panel to inform and parameterize our behavioral model. Many individuals assume that detection of early-stage breast cancer through mammography is lifesaving,⁷ and we assumed that if those diagnoses were known to be nonlifesaving, they would not serve as such powerful motivators in encouraging others to screen. Mammography is lifesaving only when a cancer that would have been lethal when detected clinically is not lethal when detected by mammography. We focused our analysis on categories of cancers whose detection by screening was not lifesaving. First, we simulated changing the incidence of cancers that would not have been detected in the absence of mammography (i.e., were overdiagnosed). Second, we simulated hypothetical interventions that allowed individuals in our model to know that mammography was nonlifesaving for (1) cancers that would be detected but would not have been lethal in the absence of mammography (never-lethal cancers) and (2) cancers that were ultimately lethal despite being screen detected (always-lethal cancers).

METHODS

To inform and parameterize our model of mammography decision-making, we conducted a survey of 716 women through RAND’s American Life Panel.¹³ Because the focus of this article is on results from our simulation model, we briefly describe the survey and how it informed our model here and provide additional detail in sections 4 and 5 of the Appendix (available as a supplement to the online version of this article at http://www.ajph.org). We designed this survey to allow us to

1. determine which mechanisms dominate the breast cancer screening decision-making process (section 5.2 of the Appendix),

2. estimate the effects of these dominant mechanisms on future screening behavior (section 5.2 of the Appendix),

3. estimate a memory parameter for our model (section 5.1 of the Appendix),

4. estimate the average number of connections for individuals in the network (section 3.2 of the Appendix), and

5. estimate age-mixing patterns in the network (section 5.4 of the Appendix).

Simulation Model

We constructed a discrete-time agent-based model in which 10 000 agents that represented women aged 40 to 100 years were linked together on a social network. In our model, each time step represented 1 year. Each year, agents (1) decide whether to be screened for breast cancer, (2) can be diagnosed with breast cancer through screening or clinically, (3) may die from breast cancer or other causes, and (4) can inform other agents about their breast cancer diagnosis and can be informed about the breast cancer diagnoses and deaths of other agents.

Our methodology for modeling breast cancer diagnoses and mortality was based on the SPECTRUM breast cancer model,¹⁴ which is used to study population-level effects of breast cancer screening and treatment. This model uses stages (in situ, local, regional, or distant) to describe how advanced a cancer is, which was the same measure we used in our survey, rather than incorporating other measures such as tumor size^15–18 or nodal status.¹⁹ Our model differed from the SPECTRUM model in 4 main ways. First, SPECTRUM explicitly models estrogen receptor status and adjuvant therapy with multivalent chemotherapy or tamoxifen. Because our goal was not to understand the effect of adjuvant therapy on mortality, we did not model it explicitly. Second, because we were interested in mammography rates as a model outcome, we explicitly modeled agents’ screening decisions, whereas SPECTRUM uses historical data to estimate screening rates. Third, we included overdiagnosis in our model, but SPECTRUM does not. Finally, as presented in Mandelblatt et al.,¹⁴ SPECTRUM models a historical period (in this case, 1975–2000); by contrast, our model represented a population in equilibrium, with demographics reflecting the current US population based on the most recent available census data. (We describe the biological model in detail in section 3.2.1 of the Appendix.)

We modeled each agent’s propensity to be screened, which is the logit of her yearly probability of being screened. This allowed us to parameterize our model by using the results of a logistic regression. As noted earlier, one of our goals in designing the survey was to determine which factors were most important in influencing individuals’ decisions to be screened. We used logistic regression analysis of data collected in our survey to test whether the following individual-level factors, identified in earlier studies, were associated with the likelihood of future screening in our study population: physician recommendation,^20–22 insurance status,^21,23–25 past screening behavior,²² age,^21,26 race,^26,27 education,²⁶ and income.^26,27 Furthermore, we tested whether learning about an early- or late-stage breast cancer diagnosis in a woman who had or had not been regularly screened within the past year affected a respondent’s future likelihood of being screened. (This regression analysis is described in section 5.2 of the Appendix.)

We found from the results of our survey that the following individual-level factors affected a woman’s likelihood of being screened: (1) whether the woman had insurance that partially or fully paid for mammograms, (2) whether the woman’s health care provider recommended screening mammograms, (3) the woman’s age, and (4) the woman’s past frequency of being screened. We therefore included these 4 individual-level factors in our simulation model. We also found that on learning first hand that a neighbor in the woman’s social network was diagnosed with early-stage breast cancer through screening, the woman became more likely to be screened; we included this effect in our simulation as well. (More detail on how we defined reports of cancer diagnoses from our survey as early- or late-stage is presented in section 4.3.1 of the Appendix.)

For each individual, i, we defined a baseline propensity to be screened, which depended on age, insurance status, and whether the individual currently had a provider who recommended breast cancer screening. We assumed that being screened or learning about a case of early-stage, screen-detected breast cancer temporarily made an individual more likely to be screened and that the propensity to be screened tended to decay back to the baseline propensity at a rate , where is a memory parameter estimated from our survey (section 5.1 of the Appendix). Similarly, not being screened temporarily made an individual less likely to be screened in the future. We assumed that the propensity at iteration n+1 evolved as follows:

where is the change in propensity to be screened for individual at iteration In any iteration, can take on one of 4 possible values, which depend on , a change in propensity that results from getting a mammogram in iteration n; , a change in propensity that results from not having a mammogram in iteration n; and , a change in propensity that results from learning of at least 1 person with an early-stage, screen-detected breast cancer in iteration n. The possible values for are shown in Table 1.

TABLE 1—

Values for Based on Individual i’s Experience in Iteration n^a

Screening Status	Learned of an Early-Stage, Screen-Detected Breast Cancer in Iteration n	Did Not Learn of an Early-Stage, Screen-Detected Breast Cancer in Iteration n
Screened in iteration n
Not screened in iteration n

^aValues for (the change at iteration n in propensity for individual i to be screened) depend on whether the individual was screened in iteration n and whether the individual learned about an early-stage, screen-detected breast cancer in iteration n.

Specific Scenarios Considered

In this article, we examine the effects of each of 3 categories of nonlifesaving early-stage, screen-detected cancers on population-level mammography rates: (1) cancers that would never have been detected clinically (i.e., were overdiagnosed), (2) cancers that would have been detected but would not have been lethal in the absence of mammography, and (3) cancers that were ultimately lethal despite being screen detected. Throughout the article, we refer to these respective categories of cancers as overdiagnosed, never-lethal, and always-lethal.

Because overdiagnosis of a cancer never benefits the patient, whereas diagnoses through screening of always-lethal and never-lethal cancers can benefit the patient when his or her life is prolonged or when less aggressive treatment can be used compared with a later clinical diagnosis, we modeled overdiagnosis differently from other nonlifesaving diagnoses. In the first set of scenarios, we varied the overdiagnosis rate. In the second and third sets of scenarios, we did not change the number of breast cancer diagnoses. Rather, we modeled an intervention designed to inform agents that detection of certain cancers through mammography was not lifesaving. To do this, we varied the amount by which an agent’s propensity to be screened increased when she learned about the diagnosis of never-lethal or always-lethal cancers through her social network.

Normally, in our model, whenever an agent learned of an early-stage, screen-detected breast cancer in her network, her propensity to be screened increased by Recall that in the real world, when a woman is screened and diagnosed with early-stage breast cancer, it is not possible to predict what her outcome would be if the cancer were diagnosed clinically. Therefore, it is not possible to identify a woman’s cancer as never-lethal or always-lethal. However, in our model, we simulated both a screen-detected outcome and a clinically detected outcome for each woman diagnosed with cancer, which allowed us to identify never-lethal and always-lethal cancers. We assumed that if women knew that certain diagnoses were nonlifesaving, these diagnoses would serve as less powerful motivators compared with when women assume these diagnoses are lifesaving.

In the second set of scenarios, we simulated a hypothetical intervention that informed women that the never-lethal diagnoses in their network were nonlifesaving. To do this, we decreased (the change in screening propensity that results from learning of at least 1 person with an early-stage, screen-detected breast cancer) when an agent learned of a diagnosis we identified as never-lethal. In the third set of scenarios, we simulated a hypothetical intervention that informed women that the always-lethal diagnoses in their network were nonlifesaving. We expected that the changes in population-level screening rates would be highly sensitive to ; therefore, we conducted a sensitivity analysis in which we ran the model using values of that were 1 SE above or below the mean (base case) estimate for estimated from our network survey data. From those data, we estimated a mean value of Δ_e = 1.06 (mean + SE = 0.54; mean − SE = 1.58), which corresponds to an odds ratio (OR) of 2.89 (± SE = 1.72, 4.85). We ran our sensitivity analysis with a lower bound estimate Δ_e = 0.54 and an upper bound estimate Δ_e = 1.58.

We performed 5 simulations of 200 iterations for each scenario. Simulations reached equilibrium after approximately 50 iterations; therefore, we excluded the first 50 iterations of each simulation and averaged the mammography rates found in iterations 51 through 200 from all 5 simulations to find the average estimated mammography rate for each scenario.

RESULTS

We first examined overdiagnosed breast cancers, which are nonlifesaving, early-stage cancers that never would have been detected clinically if not detected through mammography. Most overdiagnosis rate estimates lie between 1% and 10%,¹⁰ and some studies estimate the overdiagnosis rate to be as high as 30%¹⁰ or higher.²⁸ We varied the overdiagnosis rate from 0% to 30% to examine the possible effects of overdiagnosis on population-level mammography rates. Figure 1 shows these results.

FIGURE 1—

Upper bound, base case, and lower bound estimates for the average population-level yearly mammography rates as a function of the overdiagnosis rate based on simulation of 10 000 women’s screening decisions.

Note. Error bars show the SD of the mammography rates estimated across all iterations for each scenario.

Effect of Overdiagnosis on Mammography Rates

We found that an overdiagnosis rate of 30% would be expected to increase yearly mammography rates by approximately 4.7 percentage points compared with those we would expect without overdiagnosis if we assume our base case of Δ_e = 1.06 (OR = 2.89). This effect was approximately linear over the range we considered such that we would expect each 10 percentage point increase in the overdiagnosis rate to increase population-level mammography rates by approximately 1.6 percentage points. Our sensitivity analysis showed that the effect likely falls between 0.92 (assuming our lower bound on Δ_e) and 1.70 (assuming our upper bound on Δ_e) per 10 percentage points of the overdiagnosis rate.

Effect of Never-Lethal Cancers on Mammography Rates

We modeled an intervention that informed agents in our model that mammography was nonlifesaving for the never-lethal cancers in the social network.

We modeled this intervention by reducing Δ_e only for those cancers that would not be lethal if detected clinically. Figure 2 shows how reducing Δ_e for these cancers would affect population-level mammography rates. In our base case, we found that overall mammography rates would decrease by 13.8 percentage points per year if we were able to reduce the effect of learning about cancers through a social network from the status quo (0% reduction) to no effect at all (100% reduction). On the basis of our sensitivity analysis, we estimated that the reduction is most likely between 6.7 percentage points (assuming our lower bound on ) and 17.8 percentage points (assuming our upper bound on ).

FIGURE 2—

Upper bound, base case, and lower bound estimates for the average population-level yearly mammography rates as a function of the percent reduction in for never-lethal cancers based on simulation of 10 000 women’s screening decisions.

Note. Error bars show the SD of the mammography rates estimated across all iterations for each scenario.

Effect of Always-Lethal Cancers on Mammography Rates

Finally, we modeled an intervention that informed agents in our model that mammography was nonlifesaving for the always-lethal cancers in the social network. We modeled the intervention by reducing only for those cancers that are ultimately lethal regardless of whether they are detected clinically or through screening. Figure 3 shows how reducing for these cancers would affect population-level mammography rates. In our base case, we found that overall mammography rates would decrease by 2.9 percentage points per year if we were able to reduce the effect of learning about these cancers through a social network from the status quo (0% reduction) to no effect at all (100% reduction). From our sensitivity analysis, we found that the reduction is most likely between 1.5 (assuming our lower bound on ) percentage points and 3.2 percentage points (assuming our upper bound on ).

FIGURE 3—

Upper bound, base case, and lower bound estimates for the average population-level yearly mammography rates as a function of the percent reduction in for always-lethal cancers based on simulation of 10 000 women’s screening decisions.

Note. Error bars show the SD of the mammography rates estimated across all iterations for each scenario.

DISCUSSION

We constructed an agent-based simulation model to estimate the effect of nonlifesaving, early-stage, screen-detected breast cancers on screening rates in a social network. These cancers represent instances in which mammography does not reduce mortality. We assumed that individuals who learn about early-stage, screen-detected breast cancers from their social network neighbors are more likely to be screened in future iterations and based the magnitude of this effect on empirical survey results. We found that learning through a social network about 2 types of cancer—overdiagnosed cancers and early-stage, screen-detected cancers that later become lethal—had only modest effects on overall screening rates: about 1.6 points per 10% overdiagnosis and 2.9 percentage points, respectively. By contrast, we found that learning through a social network about a third type of cancer—early stage, screen-detected cancers that would not be lethal if detected clinically—did substantially affect mammography rates, by approximately 13.8 percentage points. These never-lethal cancers had a large effect on screening rates because they constituted most—about 55%—early-stage, screen-detected cancers in our model, reflecting that many early-stage, screen-detected cancers would be treatable if detected clinically. By contrast, overdiagnosed cancers and always-lethal cancers represented about 13% (assuming 10% overdiagnosis) and 9% of early-stage, screen-detected cancers in our model. Screening was lifesaving in the remaining 23% of early-stage detections in our model.

Given the lack of consensus on mammography screening guidelines, the implications of our analysis for breast cancer screening vary depending on one’s view of such screening. For those who believe that women are screened too early and too often, our analysis suggests that targeted interventions designed to inform patients about the true likelihood that detection of early-stage breast cancer is lifesaving could reduce overscreening. However, interventions may not be necessary if overscreening is deemed not to be a problem.

Our analysis has implications beyond breast cancer screening. It suggests that when there is a disconnect between social network information used by individuals to make preventive health decisions and population-level information used by professional organizations to make recommendations, there is the potential for a wide disconnect between demand for the intervention driven from the “bottom up” by patients and demand driven from the “top down” by recommendations. It is not difficult to imagine 2 screening tests: one that produces a moderate reduction in mortality with a moderate number of “early detections” and another that produces no mortality reduction but a large number of “early detections.” Professional organizations would clearly prefer the former, but our results suggested that patient demand for the test could be substantially greater for the latter. Given patients’ limited time and resources, patients may neglect truly beneficial interventions and favor others of questionable utility.

Limitations

Our analysis had several limitations in both the survey used to parameterize the model and the model itself. The survey we used to parameterize the model was cross-sectional; therefore, we measured intention to screen in the future rather than actual future screening behavior. We conducted our survey at the individual level; therefore, we collected information about respondents’ perceptions of the breast cancer histories of those in their network rather than both these perceptions and respondents’ neighbors’ reports of their true health histories. An important extension of this work would be to collect both true health histories and perceptions of the health histories of others in the social network. This would allow us to model misinformation.

In addition, we did not use an empirical network structure, so our model may lack certain nuances of the true social network structure. In particular, we assumed an “undirected” network so that for any pair of neighbors, either neighbor would learn if the other was diagnosed with early-stage, screen-detected breast cancer. We also gave all ties equal weight. In reality, social networks are directional and weighted.²⁹ Despite this limitation, we were able to capture several important aspects of women’s social networks, including the mean number of contacts and age mixing.

Finally, our model was designed to study mammography rates only in equilibrium; we did not predict how mammography rates might evolve over time. Our model could be extended in future work to examine dynamic effects. This would be important for predicting how screening behavior might change shortly after an intervention.

Conclusions

Despite these limitations, we believe that this study helped address some important issues in cancer screening and represents important methodological advances. We developed a simulation model of breast cancer screening and diagnosis that allowed us to endogenously estimate breast cancer screening rates with an empirically based model of women’s breast cancer screening decisions. This represents an important extension of existing agent-based simulations of breast cancer screening and outcomes. In general, such simulation models are important tools in estimating the potential effects of screening methods or treatments.^30–36

We have shown that screening rates may be significantly affected by outcomes in the social network. Therefore, if a novel intervention significantly altered the rate of an outcome or people’s awareness or perception of an outcome (such as the rate of early-stage, screen-detected cancers), we would expect screening rates to be significantly affected, and historical screening rates may be a poor indicator of screening rates after the introduction of the intervention. Furthermore, the methodology we have developed could be applied to screening tests for other cancers or, even more broadly, to models of other preventive health interventions.