The Stability of Influenza Vaccination Behavior Over Time: A Longitudinal Analysis of Individuals Across 8 Years

Abstract

Background

Seasonal influenza vaccination is an important behavior with significant individual and public health consequences, yet fewer than half of individuals in the USA are vaccinated annually. To promote vaccination adherence, it is important to understand the factors that affect vaccination behavior.

Purpose

In this research, we focused on one such factor, an individual’s vaccination history. We gathered longitudinal data to track and understand the relationship between an individual’s vaccination history and their current behaviors.

Methods

U.S. adults completed multiple surveys over an 8 year period, which asked about whether they had received the influenza vaccination during the previous flu season. We analyzed the data to determine the strength of the relationship between vaccination decisions across single-year and multiyear intervals. Additionally, we fitted two mathematical models to the data to determine whether individuals were better characterized as having a stable propensity to vaccinate or a stable propensity to repeat their previous decisions.

Results

Individuals exhibited highly consistent behavior across adjacent years, yet, across the complete extent of the longitudinal study, they were far more likely to repeat the earlier decision to vaccinate. Surprisingly, the results of the mathematical model suggest that individuals are better characterized as having a stable propensity to repeat their previous decisions rather than a stable propensity to vaccinate per se. Although most individuals had an extremely strong tendency to repeat the previous decision, some had a far weaker propensity to do so.

Conclusions

This suggests that interventions intended to increase vaccination uptake might be most impactful for those individuals with only a weak tendency to vaccinate or not to vaccinate.

Individuals exhibit highly consistent influenza vaccination behavior across adjacent years, yet, across longer time horizons, they are far more likely to repeat decisions to vaccinate.

Seasonal influenza accounts for over $15 billion annually in morbidity and mortality in the USA [1]. Influenza vaccination is the most effective means of preventing the illness. During the 2016–2017 season, influenza vaccination prevented an estimated 5.3 million influenza illnesses, 2.6 million influenza-associated medical visits, and 85,000 influenza-associated hospitalizations [2]. Thus, influenza vaccination is an effective way to avert the illness and to avert its most severe manifestations if contracted [3].

Because of its effectiveness, influenza vaccination is recommended for all individuals aged 6 months or older in the USA, yet fewer than half are vaccinated each year [4]. Most other countries do not have a universal influenza vaccine recommendation. In comparing vaccination rates among those aged 65 years and older, the USA has a comparatively high vaccination rate at 69% [5]. While this statistic is somewhat reassuring, it means that more than 30% of high-risk adults in the USA remain unvaccinated.

Many interventions to increase vaccination uptake have been studied. These include enhancing access to vaccines (e.g., through home visits or by reducing costs); increasing community demand (e.g., through incentive rewards or recall and reminder systems); and through provider- or system-based interventions (e.g., through provider reminders) [6]. Interventions have had positive effects as partially evident in increased vaccination coverage in the USA over the past decade [5], yet current vaccination rates leave a wide margin for increased uptake.

One of the strongest predictors of present influenza vaccination behavior is past decisions. In this paper, we adopt a longitudinal perspective to examine the stability of individuals’ vaccination decisions across multiple seasons. This allows us to study the association between past and present vaccination decisions over longer periods than previously possible, to examine whether individuals differ in the strength of their propensity to repeat past behaviors, and to identify events that may cause an individual to change their vaccination decisions from one year to the next.

Stability of Health Behaviors Over Time

In the case of influenza vaccination, as with many other decisions, past behavior predicts future behavior [7, 8]. For influenza vaccination, this means that people who vaccinate are likely to do so again and that people who do not vaccinate are unlikely to do so in the future. This is supported by empirical studies, which show the strong influence of past behavior on perceived risk of illness [9], intention to vaccinate [10], and actual vaccination decisions [11–13]. Myriad other factors influence the decision to vaccinate, including knowledge about influenza and the vaccination, perceived risk and protection, past experiences, provider recommendation, social influences, and other sociocognitive variables [9, 14–16]. To the extent that these factors are stable or change only gradually, individuals who vaccinated in the past will tend to vaccinate again.

A tacit assumption in much of the influenza vaccination literature is that vaccination decisions are deliberate rather than habitual. Given the limited (i.e., annual) opportunity to practice vaccination, it seems likely that influenza vaccination does not become habitual but rather requires a deliberate decision. And, whereas past behavior predicts future behavior for both habitual and deliberate processes, in the case of deliberate processes, the relationship is mediated by intention [7, 8]. Notwithstanding this distinction, past behavior can serve as a proxy variable for the constellation of factors—including intention—that contribute to behavior and are relatively stable across time. Additionally, past behavior may directly contribute to intention by affecting attitudes, perceived norms of the behavior, and perceived control over the behavior, thereby strengthening the tendency to perpetuate the behavior.

Many social cognition theories emphasize a relationship between intention and behavior [17–19], including in the case of vaccination [20, 21]. That said, they still allow for behavior to deviate from intentions, for instance, if intentions are unstable or if unforeseen barriers arise after the intention has been formed. DiBonaventura and Chapman explored these issues for vaccination behavior [20]. During each year of a 4 year longitudinal study, university faculty and students reported whether they had received the influenza vaccination in the previous year and whether they intended to vaccinate in the coming year. Individuals differed in their intentional stability. About two-thirds reported the same vaccination intention across all 4 years, and about one-third reported different vaccination intentions across at least one pair of adjacent years. The intention–behavior relationship in the final year of the study was strongest for individuals with the greatest intentional stability over the preceding 3 years. This suggests that individuals differ in the stability of their intentions and that intentional stability mediates the relationship between intention and behavior. Additionally, during one year of the study, flu shot availability was delayed, imposing an unforeseen barrier. The intention–behavior relationship was weakest during that year, suggesting a moderating effect of the unforeseen barrier. As discussed below, these theories provide multiple predictions for longer-term patterns of behavior.

Empirical Studies of Seasonal Influenza Vaccination Stability

Existing studies have consistently shown the association between past vaccination decisions and current decisions [11–13]. One limitation of these studies is that they typically employ cross-sectional designs or, at best, track respondents across just one influenza season. Another limitation is that they typically use samples that are not nationally representative. These studies have contributed significant understanding of the factors associated with vaccination and how population-level vaccination rates change, but they cannot address how the multiple-season dynamics play out at the individual level. Other studies have examined the stability of individuals’ intentions across multiple years [20] but not the stability of their decisions.

Understanding the dynamics of vaccination behavior across longer periods of time has the potential to inform interventions. For example, if there are asymmetries in the stability of vaccination versus nonvaccination, this could have implications for differential interventions to increase influenza vaccination uptake. Understanding these dynamics also has the potential to advance scientific inquiry. For example, the current study was conducted in the context of a larger interdisciplinary project using nationally representative behavioral surveys to inform an agent-based simulation of the dynamics of individual vaccination decisions across many years. Providing empirical evidence of the stability of vaccination within individuals over time will increase the realism of population simulations and the fidelity of their results.

Overview of Current Study

In this study, we examined the stability of individuals’ self-reported vaccination behavior across many years using a nationally representative sample. The focus is on behavior itself, measured multiple times over 8 years using self-reports within surveys conducted on the same longitudinal panel. Year-to-year stability in vaccination behavior could potentially conform to two different behavioral models: people could have stable propensities to vaccinate or, alternatively, they could have stable propensities to repeat past behaviors.

Figure 1 depicts the two accounts graphically. According to the first account, people have a stable propensity to vaccinate [20]. Deviations from past decisions reflect some combination of weak intentions or unforeseen barriers toward vaccinating. The decision to vaccinate each year can be seen as a (biased) coin flip. The association between behavior across years is correlational—a coin with a 90% chance of heads will tend to produce heads across adjacent flips. According to the second account, the decision to vaccinate is a decision about whether or not to repeat the previous year’s behavior. The (typically) strong disposition to repeat the behavior could arise from the effortful processing needed to deviate from established tendencies [7, 8] or from the self-reinforcing nature of those tendencies. For example, an individual who remains healthy after vaccinating may take that as evidence that vaccination is effective, whereas an individual who remains healthy after not vaccinating may take that as evidence that vaccination is unnecessary. By this account, the association between behaviors across years is nearer to causal—the situation is tantamount to biasing the outcome of a coin flip based on the previous outcome.

Fig. 1.

Graphical depiction of vaccinate model (left) and repeat model (right). Shading of circles in vaccinate model reflects propensity to vaccinate, and weight of arrows in repeat model indicate probability of changing behavior across seasons.

These two simple models relate to theories about the stability of intentions and behavior and the relationship between them [7, 8, 17–20]. Both models predict that people will tend to repeat the last year’s decision, but they make different predictions for years after people deviate from their predominant tendencies. According to the first model, people will return to established tendencies after deviating from them. At least for people who predominantly vaccinate, the intention is stable but unforeseen barriers may sometimes prevent them from doing so. According to the second model, people will perpetuate new behaviors after switching to them [7, 8]. This is more consistent with the idea that the intention to vaccinate has changed in those individuals, producing the behavioral shift.

An even more nuanced question is whether people differ in the likelihood of changing their behavior. Several theories that postulate a relationship between intentions and health behaviors also postulate that intentions (and behaviors by extension) may fluctuate over time and across individuals [17–20]. If individuals do indeed differ in their intentional stability, and if intentions influence health behaviors, then it stands to reason that individuals will also vary in their behavioral stability. The implication is that vaccination interventions may be more impactful in increasing vaccination rates for certain individuals (i.e., those who have at least vaccinated periodically in the past), which is important to inform the design of such interventions.

In this paper, we capitalize on a unique data set to longitudinally examine individuals’ vaccination behavior. We use a series of influenza-related surveys of U.S. adults administered at six time points, spanning 8 years, all using the RAND American Life Panel (ALP). The surveys cut across multiple studies and, hence, differ in content and procedure, but all asked respondents whether they received the influenza vaccination during the previous flu season. The longitudinal nature of the repeated measures paired with reasonably large sample sizes provides a rare opportunity to examine the dynamics of individuals’ vaccination behavior over an extended timeframe and, in doing so, to address three questions:

1. How consistent are individuals’ vaccination decisions across multiple years?

2. Do individuals have a stable propensity to vaccinate or, alternatively, a stable propensity to repeat past behaviors?

3. Does the propensity to repeat past behaviors vary across individuals?

These questions can only be addressed by studying the vaccination decisions of the same individuals across multiple years. The survey results, along with the computational modeling outcomes, provide answers to these questions. The implications of these findings bear on both how to think about patients’ behavioral characteristics and how to design effective, evidence-based interventions to increase influenza vaccination uptake.

Methods

From May 2010 to June 2017, six online surveys were fielded to random samples from a panel of U.S. adults aged 18 years and older participating in the ALP, a nationally representative panel of U.S. adults. The study protocols were approved by RAND’s Human Subjects Protection Committee, and analyses were conducted in 2017–2018. Surveys contained from 5 to 58 questions with most about vaccination experiences and expectations and influenza experiences and expectations. All surveys included minor variants of a question about the previous season’s vaccination behavior (Supplementary Table S1). The surveys were conducted as part of multiple distinct studies and, hence, the remaining items in each survey were tailored to the specific research questions that the survey addressed. For example, one survey also asked about vaccination and influenza experiences of the respondent’s family and acquaintances in order to examine social network effects on personnel health behavior. In that survey, questions about vaccination and flu experiences of family and acquaintances followed individuals’ reports about their own experiences to avoid biasing their responses. For more information about how ALP surveys are designed and fielded, see Pollard and Baird [22].

Sample

Overall, 6,234 respondents completed 15,679 surveys, timed to correspond to the end of the respective influenza season. As part of these surveys, respondents reported on influenza vaccination during the previous season, the primary outcome measure in our analyses. They also reported on whether they contracted influenza during the previous season, a secondary variable in some of our analyses. A challenge of this sort of secondary analysis is that the data were not specifically collected for our purposes. Hence, the surveys were completed by different but overlapping subsets of respondents. A strength of the ALP, however, is the richness of the available data. In three instances, we were able to use responses to later surveys (e.g., in the following Fall) to retrospectively account for missing observations. The completion rate aggregated across surveys was 83.5% (calculated as the number of respondents who completed the surveys divided by the sizes of the selected samples). Vaccination data were not gathered during 2013 and 2014, leaving data from six of the eight intervening influenza seasons. Supplementary Table S2 provides the number of respondents and demographic comparisons across all survey samples.

Analysis

We computed unweighted seasonal vaccination percentages and weighted percentages using raking weights for each wave to account for sampling and nonresponse. Raking weights were benchmarked against the Current Population Survey using gender, age, race/ethnicity, education, and the number of household members. To examine the stability of individuals’ behavior across time, we computed transition probabilities (i.e., the probability of vaccinating conditioned on the previous season’s decision), and we compared probabilities using chi-square tests. To explore the heterogeneity of individuals in the sample, we computed the proportions of individuals vaccinating from zero to six times during six seasons, and we then computed the proportions of individuals repeating the previous season’s behavior from zero to five times. Given that these were secondary analyses of existing data sets, we did not perform a power analysis to determine suitable sample sizes a priori. However, we performed power analyses in a post hoc manner. Given the sample sizes, the power to detect small effects (w = 0.1) with 95% confidence exceeded 85% for all reported analyses.

The statistical analyses revealed substantial heterogeneity in vaccination propensity and stability within the sample. To systematically explore individual differences, we fitted two mathematical models to the data. The first model treated individuals as coming from one of n subgroups, each with a different probability of vaccinating. The second model treated individuals as coming from one of n subgroups, but with each having a different probability of repeating the previous season’s behavior. For the respective models, we estimated the vaccination and repeat probabilities that maximized the likelihood of the data [23]. The technical details of the models are described in the Results section and in the Supplementary Material.

Results

Respondent Demographics

Of the respondents, mean age in 2017 (the final year included in our analyses) was 53.8 years, 41% were male, 88% were white, and 41% had a bachelor’s degree or higher. A total of 3,955 respondents completed surveys for two or more seasons, and 214 completed surveys for all six seasons during the 8 year period. The numbers of respondents completing from one to six surveys were: 2,279, 992, 1,623, 367, 759, and 214.

The fact that relatively few respondents completed all surveys is due to the secondary nature of the analysis—different studies sampled different groups of respondents—and is not a reflection of wave-on-wave attrition. Completion rates were quite high for all waves (Supplementary Table S1). This presented a trade-off between the number of respondents and the number of waves to retain for analysis. Retaining responses only from individuals who completed all six surveys would have resulted in discarding most of the data, much of it from individuals with responses across multiple seasons. To manage this trade-off, we present empirical and modeling results from the complete sample and from the subset of individuals who completed all six surveys. Supplementary Table S3 fully presents the number of respondents completing surveys from different combinations of influenza seasons.

Overall Vaccination Rates

The overall vaccination rate across all respondents and seasons was 46%. Vaccination rates increased across survey seasons (Table 1), paralleling national increases in vaccination during that time (40.4% in 2009–2010 to 46.8% in 2016–2017). One concern is that completing surveys may have influenced respondents’ future vaccination decisions. A second concern is that respondents’ vaccination decisions may have influenced further participation in surveys. We present evidence against these two accounts in the Supplementary Material, along with an analysis of the impact of demographic factors on vaccination decisions.

Table 1.

Sample size and self-reported vaccination rates for each surveyed influenza season

Influenza season ending	Sample size	Unweighted percentage vaccinate	Weighted percentage vaccinate
2010a	2,756	41.6	39.6
2011b	3,244	41.0	40.4
2012	457	44.9	41.9
2015	5,184	48.7	44.4
2016c	2,193	48.3	43.7
2017	1,845	52.7	45.8

^aIncludes 104 respondents who provided 2010 vaccination data retrospectively in Fall 2011.

^bIncludes 37 respondents who provided 2011 vaccination data retrospectively in Fall 2011.

^cIncludes 1,186 respondents who provided 2016 vaccination data retrospectively in Fall 2016.

Stability of Behavior Across Time

We began by analyzing data from the subsets of respondents who completed pairs of surveys across adjacent years. Table 2 presents the observed likelihood of vaccinating in each influenza season conditional on behavior in the previous surveyed season, along with the number of respondents completing each pair of surveys. Individuals’ behaviors were stable, though imperfectly so. The overall likelihood of vaccinating after having done so in the previous season was 0.85 and the overall likelihood of vaccinating after having not done so was 0.15. These values were symmetrical (the likelihood of not vaccinating conditioned on previous nonvaccination was 1.00 − 0.15 = 0.85), even though they did not have to be. In other words, respondents were equally likely to repeat the decision to vaccinate as to repeat the decision not to (χ ²₍₁₎ = 0.13, ns). The stability of vaccination behavior across time did not vary by individuals’ race, gender, or educational level.

Table 2.

Numbers of respondents completing successive surveys and likelihood of vaccinating in current season conditioned on behavior previous season

			Probability of vaccinating in current season
Current season	Previous season	n	Did previously vaccinate	Did not previously vaccinate
2011	2010	457	0.80	0.14
2012	2011	361	0.87	0.17
2015a	2012	361	0.78	0.27
2016	2015	2,124	0.86	0.14
2017	2016	1,808	0.91	0.16
2017	2010	980	0.84	0.33

^aNo surveys on influenza vaccination were administered for seasons ending in 2013 or 2014.

The longitudinal nature of the data set allowed us to examine the consistency of individuals’ behavior across multiple seasons. To do so, we analyzed data from the subsets of respondents who completed each pair of surveys that were separated by ≥2 years. At the extreme, the likelihood of vaccinating in the 2016–2017 season conditioned on having vaccinated in the 2009–2010 season (the longest interval in the data set; Table 2, bottom row) was 0.84, and the likelihood of vaccinating after having not done so was 0.33. These values were not symmetrical (χ ²₍₁₎ = 34.05, p < .0001); respondents were far more likely to repeat the decision to vaccinate over a six-season span than to repeat the decision not to. Supplementary Table 4 contains conditional likelihoods for all pairwise combinations of seasons. As the elapsed time between two seasons increased, the likelihood of repeating the decision to vaccinate remained constant, whereas the likelihood of vaccinating after having not done so increased. These results address the first research question; respondents had a tendency to repeat earlier decisions, and the tendency to repeat vaccination over longer timescales was more persistent than the tendency to repeat nonvaccination.

Population Heterogeneity

To explore heterogeneity in the sample, we examined the subset of 214 respondents who completed surveys in all six seasons. From these 214 individuals, we computed the distribution of the number of times individuals were vaccinated over the 6 years during which we administered surveys (Fig. 2, left panel, black line). Most respondents always vaccinated (six times out of six seasons) or never vaccinated (zero times out of six seasons), but about 46% sometimes vaccinated (from one to five times out of six seasons). The right panel of Fig. 2 (black line) shows the proportion of respondents who switched vaccination decisions from zero to five times across consecutive surveys. Most participants never switched decisions, but some (46%) switched at least once. We conducted the same analysis for respondents completing two to five surveys, getting very similar results (Supplementary Fig. S1). These results further speak to the first research question; some individuals’ vaccination decisions were unwavering across six seasons, whereas others’ decisions were far more variable.

Fig. 2.

Likelihood of respondents and of mathematical models vaccinating from one to six seasons (left panels) and likelihood of respondents and of mathematical models switching behaviors from zero to five times (right panels).

Mathematical Models of Vaccination Behavior

The preceding analyses show that people tended to repeat the prior year’s behavior and, correspondingly, that some people had a very high probability of vaccinating and others had a very low probability of vaccinating. Yet the analyses do not resolve which behavioral model best accounts for the data—one where individuals have a stable propensity to vaccinate versus one where they have a stable propensity to repeat the prior behavior. To answer this question, we applied two mathematical models to the data from the 214 respondents who completed surveys in all six seasons. The vaccinate model (Fig. 2, red line) treats each individual as coming from one of three subgroups, each with a different probability of vaccinating. For example, the various subgroups might have a strong propensity to vaccinate, a weak propensity to vaccinate, or a strong propensity to not vaccinate. In this case, three parameters were estimated corresponding to the vaccination rates by subgroup. The repeat model (Fig. 2, blue line) treats people as coming from one of two subgroups, each with a different probability of repeating the previous season’s behavior. For example, one subgroup might have a very strong propensity to repeat the previous season’s behavior, and one might have a weak propensity to repeat the previous season’s behavior. In this case, two parameters were estimated corresponding to the repeat rates by subgroup. Details on model construction, validation, and additional results are provided in the Supplementary Material.

Table 3 shows estimated vaccination rates for three subgroups in the vaccinate model. The average of the three subgroups’ vaccination rates, weighted by the percentage of individuals assigned to each, resembled the observed vaccination rate. The table also shows estimated repeat rates for two subgroups in the repeat model. The average of the two subgroups’ repeat rates, weighted by the percentage of individuals assigned to each, resembled the observed repeat rate.

Table 3.

Model parameters and subgroup prevalence

Model	Parameter estimatesa	Percentage of individuals
Vaccinate	Vaccinatelow = 0.04	41.1
	Vaccinatemed = 0.61	34.6
	Vaccinatehigh = 0.99	24.3
Repeat	Repeatlow = 0.61	45.4
	Repeathigh = 0.99	54.6

^aVaccinate refers to probability of vaccinating; repeat refers to probability of repeating.

To facilitate comparison with the survey data, Fig. 2 presents the expected number of individuals vaccinating from zero to six times along with the expected number of individuals switching behavior from zero to five times based on the vaccinate and repeat models. Both models accounted for the observed number of individuals vaccinating from zero to six times (Fig. 2, left). However, the vaccinate model underpredicted the number of respondents switching behavior only once (Fig. 2, right). This relates to an important difference between the two models. The vaccinate model predicts that after deviating from their predominant behavior, people will revert back to it producing a total of two switches. Conversely, the repeat model predicts that, after deviating from their predominant behavior, people will stick with the new behavior producing only one switch. In accord with this outcome, the goodness of fit statistics (negative log-likelihood; Supplementary Material) indicated that the repeat model provided a better account of respondents’ sequences of decisions (repeat = 300, vaccinate = 320) while using one fewer parameter. These results also hold for data from respondents with two to six seasons of data (Supplementary Fig. S1). These results address our second and third research questions; the data are most consistent with a model where people have a stable propensity to repeat the previous year’s decisions and where some people are uncompromising in their behavior, while others have only a moderate propensity to adhere to past decisions.

Determinants of Switching Behavior

For many individuals, behavior was nearly deterministic. But what factors may have caused people who occasionally switched to do so? We hypothesize that having contracted influenza may have caused individuals to switch. Individuals who reported contracting influenza after having not vaccinated may become more inclined to vaccinate in the future. Conversely, individuals who reported contracting influenza after having vaccinated may become less inclined to vaccinate in the future. This could occur if individuals wrongly attributed having contracted influenza to the vaccination itself or if their experience caused them to lose confidence in vaccine efficacy.

Using data from respondents who completed surveys across successive seasons, we computed the probability of vaccinating in the current season conditioned on the vaccination behavior and influenza outcome from the previous season (Fig. 3). Individuals overwhelmingly repeated the previous season’s behavior. However, contracting influenza increased the probability of vaccinating for individuals who had not vaccinated in the previous season by about 9 percentage points, whereas it decreased the probability of vaccinating for individuals who had vaccinated in the previous season by about 6 percentage points. In other words, contracting influenza increased the likelihood that individuals would switch behavior from one season to the next.

Fig. 3.

Probability of vaccinating in current season conditioned on previous season behavior and outcome.

We used mixed-effects logistic regression to analyze vaccination decisions in the current season, treating the previous season’s vaccination decision (vaccinate or not), influenza outcome (flu or not), and their interaction as predictors along with age and gender. This analysis was longitudinal and featured all observations from respondents with two or more seasons’ worth of data. Propensity to vaccinate was greater for individuals who vaccinated during the previous season (b = 3.60, z = 40.376, p < .001), and for individuals who reported contracting the flu during the previous season (b = 0.69, z = 4.661, p < .001). The main effects were qualified by a significant interaction (b = −1.03, z = 4.804, p < .001) such that individuals who vaccinated and contracted the flu were less likely to vaccinate than individuals who vaccinated and did not contract the flu.

To explore the nature of the interaction, we separated individuals based on whether or not they had vaccinated during the previous season. Among those who did previously vaccinate, contracting the flu did not significantly reduce the propensity to do so in the next season (b = −0.25, z = 1.528, p > .1). Among those who did not previously vaccinate, contracting the flu did increase propensity to vaccinate in the next season (b = 0.60, z = 4.074, p < .001). These results relate to all three research questions; after experiencing a negative outcome (i.e., contracting influenza), individuals have an increased likelihood of deviating from the prior year’s behavior.

Discussion

Despite the fact that past behavior is one of the strongest predictors of future behavior [7, 8], including for influenza vaccination [10, 11, 13, 14], the vast majority of studies on influenza vaccination behavior involve reports from a single flu season. These studies cannot address two questions about the multiple-season dynamics that play out at the individual level. First, do people have a stable propensity to vaccinate or, alternatively, a stable propensity to repeat previous decisions? Second, do people differ in the strength of their propensity to repeat previous decisions? The answers to these questions have important implications for how we characterize patients’ behavior and how we design interventions.

To address these questions, we analyzed individual- and group-level vaccination dynamics from a longitudinal survey database that spanned six influenza seasons over 8 years. We found that people behaved as though they had a fixed propensity to repeat their decision from the previous season. Although the propensity to repeat the previous decision was extremely strong for most people, it was weak for others. We elaborate on this point and the study’s four main findings next.

First, respondents repeated the previous year’s behavior with 85% probability on average. This is remarkably consistent with the finding by Chapman et al. that 85% of workers in a corporate setting who accepted the flu shot in the previous influenza season did so again versus 17% who did not previously accept the flu shot [11]. This is also compatible with the notion that intentions, which underlie deliberate health behavior, are stable over time [17–20]. Ouelette and Wood analyzed the relationships between current decisions, past decisions, and intentions for annual behaviors [7]. They found that the correlation between current decisions and (current) intentions was about twice the correlation between current decisions and past decisions. The correlation between current and past vaccination decisions in our study was substantially larger than for other types of deliberate decisions analyzed by Ouelette and Wood, indicating that the relationship between behavior and intention, if measured for influenza vaccination, would also be extremely high.

The seemingly equal propensity to repeat the decision to vaccinate or not did not hold across longer time spans. Although 85% of respondents who vaccinated in the first year of the survey did so again 8 years later, only 66% of respondents who did not vaccinate in the first year still refrained after 8 years (with a relatively monotonic decline across increasingly longer time spans). In other words, vaccination is more stable over long time periods and is the “stickier” of the two behavioral states. A potential explanation for this asymmetry is seen in Fig. 3. Contracting influenza caused some individuals to change their behavior in the next season. Yet, the change in vaccination propensity was greater for nonvaccinators (i.e., a 9% gain) than for vaccinators (i.e., a 6% loss). The slight asymmetry could reflect the fact that even when vaccination does not prevent influenza, it reduces disease severity. Additionally, if perceived risk and severity of the disease increases with respondents’ age, along with the probability that a provider recommends vaccination [24], this could reinforce the behavior of vaccinators while encouraging nonvaccinators to begin vaccinating.

Second, our comparison of two alternate mathematical models revealed that individuals were better characterized as having a stable propensity to repeat the prior year’s decision rather than as having a stable propensity to vaccinate. The subtle distinction arises from the fact that when individuals deviated from past behavior, they tended to repeat the new behavior. Paired with the earlier finding that vaccination is a “stickier” state than nonvaccination, this suggests that, once individuals transition from nonvaccination to vaccination, they will tend to remain in that state. This lends a hopeful note (and key challenge) for interventions that target individuals who have historically avoided vaccination but could potentially be convinced to switch.

Many social cognition theories postulate a relationship between intentions and behavior [17–20], and they explain deviations between intentions and behavior in terms of intentional instability or the presence of barriers that block the intended behavior. Self-reports of vaccination behavior supported a mathematical model in which people perpetuate new behaviors after switching to them. This is consistent with the notion that behavioral change was accompanied by intentional change. This underscores the importance of interventions that affect intention (i.e., provider recommendation). Our survey and mathematical modeling results do not exclude the importance of interventions that reduce barriers to vaccination, some of which are among the most effective reported in the literature [14]. The elimination of barriers may allow individuals to vaccinate, which may reinforce the decision to do so again.

Third, survey results along with mathematical models revealed substantial heterogeneity in the sample. This is seen most clearly in the repeat model where about half of individuals had a 99% chance of repeating the previous year’s behavior and half had only a 61% chance. This means that the annual 85% probability of repeating the previous year’s decision is not representative of any single individual. Rather, some individuals are nearly deterministic in their behavior and others frequently switch. In our sample, the stability of vaccination decisions across years did not appear to vary by gender, race, education, or age (although vaccination propensity did). An important question for future research is why individuals differ in the stability of their behavior and whether those differences relate to intentional stability, vaccination barriers, or both. Either way, individuals who change behavior frequently may be more receptive to interventions as they have at least demonstrated intent to vaccinate in the past.

Fourth and finally, the propensity to vaccinate in the current season depended on whether or not the individual reported having had the flu in the previous season but in a nonobvious way. Those who had not vaccinated and reported having the flu were more likely to vaccinate in the future. This likely reflects recognition of the foregone benefits of vaccinating. More surprisingly, those who had vaccinated and reported having the flu were somewhat less likely to vaccinate in the future. This may reflect a change in their perception of the value of vaccinating relative to its costs. The interaction between past behaviors and outcomes warrants further exploration and, given the exploratory nature of the finding, replication.

Our results and conclusions must be considered in light of four study limitations. First, vaccination behavior was measured using self-report rather than actual medical records and, hence, may be susceptible to self-report biases. However, self-reports of vaccination behavior have high agreement with medical records and high sensitivity and specificity [25, 26]. One study did find only moderate agreement between self-reports and records of flu vaccination along with a bias toward overreporting [27]. However, that study asked respondents whether they had ever received the flu vaccination. Due to the vaccine’s dichotomous nature and its yearly periodicity, self-reports about vaccination behavior in the previous 12 months, as in our study, may have especially high agreement.

A second limitation of the study is that it is a secondary analysis of existing survey data, which was collected through multiple studies on overlapping (but not identical) subsamples. Hence, the data available for each individual varied. Nevertheless, the results were remarkably stable across subsamples.

A third limitation is that we did not examine the effects of other external factors, such as access to care, that may have reasonably influenced vaccination decisions. Relatedly, social activities are a powerful determinant of intentions [14]. For example, most vaccination decisions take place in the context of a provider–patient dyad, ideas about vaccination and its value (i.e., contagions) are spread through social networks, and social norms based on the communal benefits of vaccination influence individuals’ decisions. The subset of social processes involving norms have inspired game-theoretic models of how rational agents should approach the vaccination dilemma—perversely, if vaccination uptake is high, an individual may have less incentive to vaccinate. Our mathematical model revealed meaningful individual differences yet did not specify the origin of those differences. The factors listed above could partially underlie the stability of vaccination over time, and future studies should explore these potential contributors. A final limitation is that, although our mathematical model provides a useful characterization of the sample heterogeneity, it does not specify the actual behavioral mechanisms that give rise to decisions.

A logical question is whether models that do contain behavioral mechanisms, such as agent-based models, can account for the heterogeneity seen in our sample, both in terms of individuals’ predominant tendencies and the strengths of those tendencies. Such models have been shown to give rise to habitual vaccination and nonvaccination behavior [28]. Richer behavioral models may be needed to capture the confluence of experiential and social factors (i.e., altruism) on vaccination behavior. We leave these final two issues—to identify the behavioral mechanisms of vaccination choice and to unify them with psychological, contextual, and sociodemographic and physical factors—for future research.

Compliance with Ethical Standards

Authors’ Statement of Conflict of Interest and Adherence to Ethical Standards The authors report no conflicts of interest.

Authors’ Contributions M.M.W, A.M.P., and R.V., S.A.N., D.P.K., and C.A.G. conceived of the research questions; M.M.W., A.M.P., and R.V. analyzed the data and created the model; M.M.W., A.M.P., and R.V. wrote the article; S.A.N., D.P.K., and C.A.G. reviewed and edited the model. M.M.W, A.M.P., and R.V., S.A.N., D.P.K., and C.A.G. conceived of the research questions; M.M.W., A.M.P., and R.V. analyzed the data and created the model; M.M.W., A.M.P., and R.V. wrote the article; S.A.N., D.P.K., and C.A.G. reviewed and edited the model.

Ethical Approval All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.

Informed Consent Informed consent was obtained from all individual participants included in the study.