What determines individuals who are at-risk for Huntington’s disease to pursue genetic testing? An optimal expectations approach
Mihai Vicol
Student number: 11441720
Faculty of Economics & Business Economics, University of Amsterdam
Bachelor Thesis
Supervisor: Davide Pace
Statement of Originality
This document is written by Mihai Vicol who declares to take full responsibility for the contents of this document.
I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it.
The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
Table of contents Abstract ... 4 Introduction ... 5 Research Question ... 6 Data Description ... 10 Econometrics ... 12 Literature Review ... 17 Results ... 20 Conclusion ... 25 Bibliography ... 27
Abstract
This paper is concerned with individuals who are at-risk for Huntington’s disease (HD), a severe neurological disorder without any cure, which slowly leads to the breakdown of one’s neural cells. Although testing for HD is widely available, low rates of genetic testing are observed among the at-risk for HD population. Oster et al. (2013) analysed the behaviour of a large sample of individuals who are at-risk of having HD, the PHAROS group, and found a positive relationship between the severity of the disease, as determined by professional neurologists, and test rates. The authors observed the behaviour of untested individuals and the ways it changes once symptoms begin to escalate. They found that untested patients tend to exert overly optimistic behaviour regarding their chances of not carrying the disease, despite contrary objective evidence. Furthermore, they showed that, contrary to the neoclassical framework, information avoidance is optimal for individuals who are at-risk for HD. However, there seems to be a threshold of experienced symptoms which, if crossed, causes information avoidance to become unfeasible simply because the objective evidence overwhelms one’s longing to feel optimistic about her health prospects. Accordingly, this paper discusses part of the results of Oster et al. (2013) in light of the optimal expectations framework they used, with specific focus on overconfidence, information avoidance and their effects on the individuals in the PHAROS group.
1. Introduction
Huntington’s disease (HD) is a hereditary neurological disorder that causes progressive damage, and ultimately the breakdown, of its bearer’s neurons. It currently affects one in every 10,000 to 30,000 individuals in the world, depending on the country of provenance. It is hereditary in the sense that if one of the parents has developed HD, there is a 50% probability that the offspring also inherits the disease. The genetic mutation responsible for causing the disease is that of the huntingtin gene, prompted by a repetition of more than 36 times of the CAG trinucleotide. The evolution of HD consists of five different stages. The amount of time a particular person spends in any of the stages varies and is dependent on individual characteristics. Early symptoms most often include difficulty in concentration or in making decisions, depression and a decrease in one’s ability to coordinate her body.
The advancement of the disease brings about motor and behavioural signs such as chorea - the sudden and involuntary movements of the muscles of the neck, face, trunk and extremities -, hypokinesia – a severe, most often complete loss of one’s ability to control her muscles -, grinding of teeth or impairment of speech. Most often the onset of HD occurs in the interval of 30 to 50 years of age, but there are also exceptions to this rule. For instance, in the case of juvenile HD the disease can occur at ages of 20 or less; at the opposite side of the spectrum, some patients begin experiencing HD - like symptoms at ages of 70 or more. Although special cases are possible, such patients who develop HD either very early or very late represent a small percentage of the whole HD population, averaging at around 10% (Almqvist et al., 2001). Naturally, as individuals age without clear symptoms of HD, the probability that they carry the disease considerably decreases. Life expectancy after the onset of the disease varies and depends to a large extent on the age of the individual, but is generally between 10 and 30 years. At the moment there is no medical treatment that has proved to stop the evolution of HD. However, a genetic test has become available in 1993. By taking the test, individuals who are at risk of developing the disease can find out with certainty about their gene status.
This paper continues as follows. In section 2 we address and discuss the research question, as well as offer a brief overview of the framework of optimal expectations. Section 3 describes in detail the data used by Oster et al. (2013) in the original experiment and presents summary statistics about the patients who took part in the clinical trial. In section 4 the econometric framework used by the same authors is explained, together with all the necessary assumptions. In section 5 we review previous empirical research in medicine, psychology and economics,
dealing with low test rates for life-threatening diseases, overconfidence and information avoidance. Section 6 presents some of the results of Oster et al. (2013) and discusses them in light of the previous literature. Section 7 concludes.
2. Research Question
In order to better understand what prompts such individuals to undergo testing, Oster et al. (2013) analysed the behaviour and decisions of 1,001 at-risk for HD individuals who agreed to take part, over a 9-year period, in the PHAROS clinical trial (“The Prospective Huntington At Risk Observational Study”). An important feature of the PHAROS group was that its members have not undergone testing upon recruitment, but could choose to do so throughout the period of the study. As the study progressed, the individuals who turned out to suffer from HD were increasingly more likely to develop new symptoms. All members of the PHAROS group were consulted and evaluated by neurologists at set intervals of time in order to observe the possible evolution of the disease and the extent to which its advancement prompts the participants to take up certain decisions they would have otherwise not embraced.
One of the ways in which the doctors from the trial quantified the development of HD was through assigning an evaluation score to each participant at every visit. This score reflected the stage of HD a patient was in at the time of examination. The evaluation score could have been between 0 and 4, with 0 indicating no signs of disease and 4 signalling certain signs of HD. As patients would develop more severe symptoms over time, the investigators could adjust the evaluation scores they allocated. Consequently, the main goal of Oster et al. (2013) was to check whether participants were more likely to test once their evaluation scores changed. They found that there is indeed a significant positive relationship between the two, especially so when the evaluation score changed from 3 to 4. More precisely, the most significant increase in test rates among patients was found among individuals who have undergone a change in the evaluation score from 3 to 4. However, the authors also found increases in test rates when the evaluation score increased from 0 to 1 or from 1 to 2.
At evaluation scores of 4 individuals usually experience severe symptoms such as involuntary body movements or a growing inability to coordinate their bodies. Therefore, it is likely that a non-negligible number of patients were aware that carrying the disease was a real possibility at that point. Combining this with the significant increase in test rates brought about by a rise
in evaluation scores from 3 to 4, we can infer that a considerable number of individuals with evaluation scores of 4 have undergone testing only to confirm what they already had a strong intuition about, or what is known as confirmatory testing. According to Oster et al. (2013), 30% of the symptomatic individuals in the PHAROS group underwent confirmatory testing. Likewise, the focus of this paper is on understanding the testing behaviour of individuals who are at-risk of developing HD. More precisely, we wish to study the relationship between the evaluation scores of individuals, as determined by the investigators, and the test rates among patients in different stages of HD. To that end, we present the findings of Oster et al. (2013) and discuss them in light of the existing literature.
In order to test whether increased evaluation scores lead to higher test rates, the authors only allowed individuals who were at-risk for HD to take part in the experiment. This is the case because people who are not at-risk for HD will not provide any useful data to answer the research questions. Although this study only involved patients who were at-risk for HD, the results of the experiment can be valuable for better understanding the behaviour of a wide range of individuals and to thus be able to devise incentive schemes that would lead to more desirable conduct. For instance, understanding the reasons behind low test rates and information avoidance in the case of patients who are susceptible for HD can prove useful for people suffering from other life-threatening diseases where test rates are similarly low but which, if discovered early, can lead to the recovery of the patient, such as the different types of cancer. Understanding what causes people to avoid testing in these situations and what can be done to increase test rates may lead to improved programs that better incentivize at-risk individuals to take up actions that are beneficial for their health.
Secondly, understanding how and in which circumstances overconfidence and information avoidance manifest themselves is relevant for a broad range of societal issues. For instance, the behaviour of at-risk for HD individuals who, despite objective evidence, still believe not to carry the disease resembles that of a great deal of investors in the moments preceding a financial crisis. Much like the HD patients, many investors consider threatening signs as rather circumstantial and more often than not they choose to ignore them. Furthermore, they are overly confident that the stock market or the housing prices will keep increasing in the future, despite clear signs that things are in fact likely to aggravate. If the beliefs of investors turned out to be wrong, they would suffer a utility loss because those beliefs determined them to take up suboptimal actions. Therefore, despite dealing with a seemingly isolated category of
individuals, the paper by Oster et al. (2013) showed that the type of behaviour exerted by at-risk for HD individuals serves as a starting point for understanding the phenomena of overconfidence and information avoidance in high-stakes situations where holding wrong beliefs about future states and thus acting in a suboptimal manner can lead to substantial drops in utility. Such situations are characterized by uncertainty and hence require some degree of risk-taking from the perspective of individuals. Choosing beliefs is especially risky in high stakes situations, for wrong beliefs have a significant negative impact on the utility of individuals through the suboptimal actions that are taken up. Understanding the ways in which people act in such circumstances is essential in order for regulatory authorities, for instance, to mitigate the potential losses and reduce the extent of risk taking such that societal welfare does not drop below a certain level.
Given that developing HD changes one’s outlook on life and future prospects significantly, it might seem that at-risk individuals would wish to know whether or not they carry the disease. However, evidence shows that only around 5% to 10% of the patients who are at-risk for HD choose to test. The first reason that comes to mind for the low observed rates could be the high price of testing, consultations and of ex-ante and ex-post counselling, all of which are mandatory for patients who wish to learn their gene status. Presently, the cost for an HD test alone ranges between €300 and €400, and can increase up to €700 or €800 when including necessary counselling and consultations. According to the standard neoclassical model, which posits that utility maximizing individuals with rational preferences act while taking into account all the available and relevant information, the only aspect that prevents patients from undergoing genetic testing is the economic cost: testing rates decrease as costs grow. In theory, if the testing costs were to drop to 0, all at-risk individuals would test. Hence, once the cost reaches 0, it is rational for everyone to undergo testing and irrational to do otherwise.
However, intuition has it that there might be something more than just the economic cost that makes individuals reluctant to test. For instance, it seems likely that individuals attach a certain utility on not knowing the state of the future if that state is likely to have a substantial negative effect on their well-being. In other words, people might be better off not testing and living with the uncertainty rather than testing and finding out they carry HD. The awareness of being HD - positive could negatively influence one’s outlook on life: it could cause the bearers to live with anxiety for the rest of their lives or prompt them to enter serious and long periods of depression, among others. Having these in mind, it seems that testing avoidance can be rational
from the individuals’ perspective. Within such a framework, as opposed to the neoclassical case, one’s utility would be directly affected by beliefs about the future state.
Oster et al. (2013) proposed that a model of optimal expectations would be better fit to explain the behaviour of individuals who are at-risk for HD. The optimal expectations model has several characteristics. According to Brunnermeier & Parker (2005), it posits that at any point in time when making decisions, individuals take into account both the utility they experience at the present moment and that which will be experienced into the future. Moreover, conditional on not being tested, one is able to choose her own beliefs regarding the likelihood of each future state being realised. Not to mention, these beliefs do not have to match the true probabilities of the states being realised. This set of beliefs grants its bearer anticipatory utility, a type of felicity that is experienced at the present time and which results from holding certain beliefs about future states. Gaining anticipatory utility is not mandated by the state actually being realised into the future, but in the individual at the present moment believing it will. Naturally, if an agent holds distorted beliefs about the future, that is, if she wrongfully believes that her preferred state is likely to be realised, she will incur a utility loss in the moment in which the realised state turns out to differ from the anticipated state.
However, the utility loss is not directly generated by the fact itself of holding wrong beliefs, but rather indirectly through the suboptimal action that is being taken up as a result of holding those wrong beliefs. For instance, a patient who was objectively deemed by the doctors to have HD but who subjectively believed up until some point not to have HD will not suffer a utility loss because of her prior beliefs per se when finding out she actually carries the disease. The utility loss comes from the actions she might have undertook during the period of uncertainty, when she wrongly believed not to have the disease. Such actions might have included getting a new job, planning to retire at a certain age, getting married or deciding whether to have children or not. Because the agent planned for such important decisions believing not to have HD, the acknowledgement of the actual gene status would reveal to her that some of these actions were in fact suboptimal and would thus lead her to experience a utility loss because of that.
Therefore, there is always a trade-off between the gain in anticipatory utility resulting from holding overly optimistic beliefs and the decrease in utility brought about by the desired state not being realised. As long as the gain in utility exceeds the loss, agents are better off being optimistic about the future. For some patients it could be that the option of being able to stay
optimistic about the future grants them so much utility such that the value of testing is negative. One constraint of the model is that once a set of beliefs has been chosen, agents are only allowed to act in accordance with those beliefs. For instance, an HD – positive individual who has not undergone testing but still strongly believes to have the disease is required by the model to make significant changes in her life that reflect the behaviour of a person who knows her faith is sealed.
All things considered, the optimal expectations model is able to offer a different and more realistic explanation for why many individuals choose not to get tested for HD. The reason is that testing takes away the ability of patients to hold subjective beliefs about the likelihood of future states. Subsequently, it takes away one’s potential gain in anticipatory utility that would result from believing to be healthy no matter what and replaces it with the certainty of the test result, where the probability of having the disease is not set by the individuals themselves anymore, but is rather outside of their grip. Regarded from this perspective, not wishing to know whether or not one carries the disease, or information avoidance, seems optimal from the perspective of the patients. Indeed, one of the contributions of Oster et al. (2013) was to show that information avoidance is consistent with a framework of optimal expectations.
3. Data Description
The 1,001 PHAROS participants were selected in a non-random manner from the population of individuals at-risk of developing Huntington’s disease. Every person who signed up for the trial would be accepted if she met certain conditions. Firstly, prospective participants had to have one parent or first-degree relative who developed HD. This is required because the study is concerned with individuals who are at risk of HD. The disease is hereditary and there is a 50% probability that the off-spring inherits it from the affected parent. This condition automatically excludes people with no family history of HD from taking part in the experiment. Secondly, one must have not undergone genetic testing for HD in order to be eligible for the clinical trial. Moreover, given that there is a 50% chance that the offspring inherits the genetic expansion from her parent, every patient in the PHAROS group theoretically has a 50% chance of developing HD at the start of the experiment. As the study progresses this percentage is likely to change for each individual. Secondly, patients must not have pursued genetic testing until the onset of the study. Thirdly, they wished to remain unaware of their gene status, for the
evaluation scores and motor scores as determined by the investigators would never be made available to the participants.
Fourthly, patients had to be aged between 26 and 55. According to Almqvist et al. (2006), individuals in this age interval were the most suitable for the purpose of the trial, for they were the most likely to develop HD during the planned evaluation period.
The main reason for which non-random selection was preferred over a randomized control trial among the HD population is that of ethical and legal considerations. One cannot, from neither point of view, be forced to take part in a clinical trial without consenting to do so. Hence, the PHAROS group is composed of at-risk individuals who agreed to participate in this project on a volunteer basis. The study took place in the US and in Canada from 1999 to 2008.
The PHAROS group is comprised of individuals that are relatively highly educated, averaging 14.9 years of education. Not to mention, there are considerably more female participants, at about 2/3 of the whole sample. The 2:1 female to male ratio has been observed in others experiment concerning HD, for instance Hogarth (1999) or Kirkwood et al. (2000). According to Almqvist et al. (2006), this imbalance stems from the fact that “women are more willing than men to contribute to observational clinical research that is unlikely to provide any direct benefits, a finding perhaps related to a greater interest by women in their reproductive fate”.
Patients are consulted by specialists for signs of developing HD once every 9 months. They are administered clinical tests of “motor and ocular performance, gait and involuntary movements such as chorea”. The investigators subsequently use the results of the tests to rank the individuals’ motor performance on a scale from 0 to 154. The mean motor score is 3.76, indicating no abnormalities. A patient is likely to have signs of HD if his motor score is at least 11. In addition to the motor score, after the evaluation investigators also express their medical opinion regarding the likelihood that patients have HD. They do so by assigning each individual a number on a scale from 0 to 4, where 0 signals normal behaviour and 4 represents “unequivocal signs of HD (> 99 percent confidence of having HD)”. The mean of this evaluation is 0.668, which again signals no abnormal behaviour in most of the sample. According to Almqvist et al. (2006) (who use the same data as Oster et al. (2013)), 92.3% of participants had a doctor evaluation risk of 0 or 1 at the beginning of the clinical trial, which suggests mostly normal and asymptomatic behaviour. It is desirable that most participants at the start of the study are asymptomatic or have only developed mild symptoms. This is the case
because we aim to measure the change in patients’ testing behaviour once symptoms escalate. It is important to note that the patients are not informed about the doctors’ evaluation and the results of the tests they have undergone
Participants are also asked at every visit about their perceived chance of having HD. The mean result is 42.7%. Considering that every individual in the PHAROS group has at least a 50% probability of developing HD, the realised perceived probability may be understated. This seems to be consistent with the intuition that patients prefer to be overly optimistic about the future. Moreover, only 5.7% of patients decided to test. Although every patient has agreed to a DNA test within the experiment, the results of this test are not made available to anyone. Therefore, the 5.7% represents the share of patients who have been tested outside the experiment. The low proportion likely indicates a preference for uncertainty among patients in the PHAROS group. Accordingly, not testing might provide patients with the opportunity to continue living a normal life.
Last but not least, patients were asked 6 times throughout the clinical trial whether they experienced important life experiences in the last year. These include marriage, divorce, getting a new job, getting pregnant or experiencing a major financial change, among few others. The idea is to observe whether a change in HD-like symptoms considerably influences the occurrence of such significant decisions.
It is important to note that all descriptive statistics reported by the authors represent the mean scores and percentages of the PHAROS group bundled together from the whole duration of the experiment. That is, they incorporate the aggregate behaviour of all patients and thus also include potential behavioural changes over the duration of the study. Accordingly, there are no separate summary statistics for a particular group of patients in a certain year.
4. Econometrics
The authors wanted to check whether patients are more likely to get tested for HD once their symptoms escalate. Accordingly, they used as the dependent variable the percentage of individuals who undertook testing since the last visit, while the investigators’ evaluation of symptoms served as the independent variable. Although the assessment of the investigators is not disclosed to patients and thus does not directly influence testing rates, higher motor and
evaluation scores are consistent with more serious symptoms that the patients themselves can potentially identify. Therefore, it is the patients’ own assessment of their symptoms which prompts them to undergo testing, and not the unobservable investigators’ scores. However, it is likely that the two are strongly correlated.
The regression equation the authors used to estimate the aforementioned relation is:
Test_rates =
β0 + β1 * last_uhdrq + β2 * male + β3 * age + β4 * education + εiThe term last_uhdrq indicates the investigators’ assessment of the patients’ likelihood of having HD at the last visit; it can take values from 0 up to and including 4. Test_rates, the dependent variable, represents the share of patients who undertook genetic testing since the last visit. εi is the error term. male, age and education are control variables used in order to isolate the effect of the independent variable on the dependent variable and make sure it is not driven by differences in sex, age or education. Leaving such variables out may lead to omitted variable bias and an either under- or overestimated coefficient for last_uhdrq. The coefficient would be under - or overestimated because on top of the effect that investigator scores have on test rates, it would also include the sum of the effects that the three control variables have on test rates. The direction of the bias is difficult to guess. Firstly, as individuals age without symptoms, they are less likely to undergo testing. This entails a negative relationship between age and testing behaviour if individuals do not end up having HD. However, patients who carry HD are more likely to develop symptoms as they get older. In this case, there is a positive relationship between age and testing behaviour.
Concerning education, Feinstein & Sabates (2006) found that there is a strong positive correlation between years spent in school, health awareness and preventive behaviour regarding one’s health. We could include one’s decision to test in the latter category. Given that education seems to positively influence testing behaviour and that most patients in the PHAROS group are highly educated, not including education as a control variable would lead to overestimated coefficients for the dummy variables. Lastly, although women are more likely to participate in clinical trials as shown in the Data Description section, there is no indication that they are also more likely to exert a substantially different testing behaviour when compared to men. After taking into account the theoretical effects of all three control variables on test rates, we might conclude that not including them can lead to overestimated coefficients for the dummy variables (considering that the negative and positive effects of age on test rates nullify
each other). However, in reality it is difficult to precisely identify the direction and extent of the bias.
The Stata command used by Oster et al. (2013) to run the regression is:
xi: reg tested i. last_uhdrq male age education, cluster(id)
The combination of xi at the start of the regression and the indicator variable i. last_uhdrq tells Stata to create one dummy variable for all but one values that last_uhdrq can potentially take. Therefore, there will be 4 new dummies created, from _Ilast_uhdrq_1 to _Ilast_uhdrq_4. The _Ilast_uhdrq_0 variable will be omitted as it represents the baseline category with which the others are compared. These are dummy variables and their values are either 0 or 1, depending on their evaluation scores. Each resulting coefficient indicates the difference between that particular group and the baseline group. For instance, the coefficient of _Ilast_uhdrq_3 represents the difference in testing rates between patients with an evaluation score of 3 and patients with an evaluation score of 0. If the authors only used a single categorical variable that took different values conditional on the state of the patients, say CatLast_uhdrq, (which took value 0 when there were no abnormalities, value 1 when the investigator was less than 50% certain the patients had HD, value 2 when the certainty of having HD as identified by the investigator is between 50% and 89%, value 3 when it is between 90% and 98% and value 4 when it is higher than 99%), the econometric model would have suffered. That is, using only one categorical variable implies that going from an evaluation score of 0 to a score of 1 must have the same effect on test rates as going from an evaluation score of 3 to a score of 4, for instance. This would falsely indicate a linearity in terms of the different categories of evaluation scores.
In fact, we expect to find a significantly larger (rather than identical) increase in test rates when the evaluation score changes from 3 to 4 as compared to when it changes from 0 to 1, simply because more patients from the former category are likely to be relatively more aware of their symptoms when compared to the latter group. This could be because patients who recently underwent a change in the evaluation score from 3 to 4 might have clearer signs of HD than patients from other groups and may thus be more likely to opt for confirmatory testing when compared to participants in other score categories. Oster et al. (2013) indeed found that a shift in the evaluation score from 3 to 4 brought about a much larger increase in test rates as contrasted to an increase from 0 to 1. Therefore, the use of indicator variables is necessary in
order to capture the heterogeneous effects that the different evaluation scores have on the test rates.
In addition to that, the use of a single categorical variable that takes different values for patients in different HD states can lead to heteroskedasticity. When CatLast_uhdrq were low, almost no patients would choose to test because they think it unlikely to carry the disease, thus the low variance of the error terms. However, when CatLast_uhdrq was high, there would be a considerably higher variation in testing since some patients would firmly believe there is a high probability that they carry HD and thus would opt for testing, while others would still remain optimistic about their gene status and not undergo testing. Therefore, we would have a comparatively wider spectrum of observations when CatLast_uhdrq is high than when it is low. As a result, the variance of the residuals would not be constant across observations. The econometric model used to estimate the relationship between the testing rates and the doctors’ evaluation scores is OLS. The first assumption of the OLS method is that the model should be linear in parameters, although it need not be linear in variables. Since our regressors are dummy variables (for instance _Ilast_uhdrq_3 = 1 if the doctor’s evaluation score = 3 and 0 otherwise), they meet the linearity assumption by definition. However, if we were to assume a linear relationship between variables that are, for example, in a quadratic relationship, this would result in standard errors and coefficients for the dummy variables that are inaccurate. The second OLS assumption states that the error term has to have a conditional mean of zero. That is, conditional on the x’s that we observe in our sample, the expected value of the error term has to be 0 on average. In other words, the model has to be correct on average. If this condition is not met, the expectation of the OLS estimator would be different from the true population coefficient. It says that if patients with a particular evaluation score consistently choose to undergo testing because of another reason that is not included in the regression but is correlated with the evaluation score, the results may be biased. As mentioned above, this could potentially pose a problem because the patients choose to test as a result of their own assessment of symptoms, which is subsequently correlated with the doctors’ evaluation scores. The third OLS assumption entails there should be no serial correlation of the error terms. That is, consecutive error terms have to be independent of each other. To account for serial correlation, the command cluster(id) is used at the end of the regression in Stata. Since id indicates a personal identification number that is unique for each participant in the trial, the clustering is by individual. Given that throughout the duration of the experiment the same
subject is observed multiple times, it is unreasonable to believe that the error terms are independent and identically distributed across time for one particular individual. Therefore, clustering allowed for the standard errors to be correlated within individuals, but independent and identically distributed across individuals.
When clustering, it is the standard error (and the t and p values) that might change, but the coefficients stay the same. If the condition for serial correlation is not met, the OLS estimator is no longer efficient. According to Williams (2015), inefficiency leads to the standard errors estimated by the OLS regression to be smaller than the true standard errors. Subsequently, the coefficients of the parameters turn out to be more accurate than is the case in reality. This can lead us to sometimes reject a true null hypothesis, resulting in a type I error.
An estimator, say β1, is more efficient than β2 if the variance of β1 is smaller than the variance of β2. In practice, the most efficient estimator is the one that diverges the least from the true population parameter. However, although an estimator, say β1, is more efficient than β2, it could be that β1 is also slightly more biased when compared to β2. That is, it might only contain the true population parameter in a relatively small number of cases. In other words, the expected value of the estimator would not equal the true value of the parameter. This trade-off between unbiasedness and efficiency needs to be taken into account when choosing an estimator. We use the Durbin-Watson statistic to test for serial correlation of the first order. For higher order serial correlation, we use the Breusch-Godfrey test.
Fourthly, OLS requires the independent variables not to be correlated with the error term. There are two conditions that need to be met in order for omitted variable bias to occur. Firstly, there must be a relationship between the independent variable and the omitted variable. Secondly, the latter has to exert an effect on the dependent variable. It is reasonable to expect no omitted variable bias within the PHAROS group. That is, it seems unlikely that a separate unobserved variable, particular to a subgroup of individuals with a certain evaluation score, explains part of the variation in testing rates and correlates with the evaluation score. Fifthly, OLS requires no multicollinearity. In our case, this condition entails that we should not include in our regression independent variables that are a linear function of other independent variables. Given the nature of the independent variables in our model – dummy variables indicating the score categories patients are in – there is no threat of imperfect multicollinearity. Accordingly, we could either have perfect multicollinearity or no multicollinearity at all. The threat of perfect multicollinearity is the reason for which we excluded _Ilast_uhdrq_0 from our regression. It is
also the reason for which we only use one dummy variable for the sex of the patients, instead of two. If we were to include those variables in our regression, Stata would either automatically drop one of the dummies from the regression or would not run the command and instead return an error. Having accounted for these measures, the assumption of no multicollinearity is met. Next, another assumption is that of homoscedasticity. In other words, the variance of the error terms should be constant across observations. According to Astivia & Zumbo (2019), if this condition is not met and heteroscedasticity is present, we would have standard errors that are too low and hence narrow confidence intervals that do not allow us to reject the null hypothesis at the initially agreed upon alpha level. Accordingly, we may run the risk of an increase in the evaluation score that has a significant impact on test rates and that is not being picked up by our regression. In other words, we would reject a true null hypothesis, a type I error. Furthermore, had heteroscedasticity been present, the coefficients of the dummy variables in our model would have been biased in the sense that the subset of data with the highest error variance would have received a relatively larger weight when compared to those with a lower error variance. Heteroskedasticity can be tested using the Breusch-Pagan test.
Finally, although the random sampling assumption is not met because of ethical and legal concerns, Oster et al. (2013) showed that their sample is likely to be representative of the whole population, for testing rates that they found are not only similar to testing rates in other studies (for instance Sholson & Young (2011)), but also to those in the population, between 5% and 10%. If the only potential sources of difference between participants and non-participants are ultimately reflected in test rates, then the sample could indeed be representative of the whole population.
5. Literature Review
Genetic testing is rather uncommon. Roberts et al. (2004) found hypothetical testing rates among participants with a family history of Alzheimer’s to be only 24%. Hypothetical test rates indicate the willingness of individuals to undertake testing on a theoretical level, without them actually needing to take up the action. They are usually considerably higher than realised test rates, for they require no active commitment (for instance Ropka et al. (2006)). According to Lerman et al. (1999), but 40% of patients with a family history of colon cancer chose to undergo genetic testing. The same authors also found that only 57% of the patients who underwent
testing for colon cancer subsequently decided not to be informed about the results. Sholson & Young (2011) documented testing rates of below 10% in patients who were susceptible for Huntington’s disease. Especially in the case of genetic testing, potential anxiety could determine at-risk individuals to avoid essential information about their health (Golman, Hagmann & Lowenstein, 2017). Therefore, it seems that when it comes to testing for life-threatening illnesses, a considerable share of individuals prefers to live with uncertainty as long as there is a probability that testing can bring about the negative outcome along with the associated undesirable feelings. By testing, one loses the ability to choose her own beliefs about the likelihood of future states. Thus, it is possible that testing could take away one’s ability to live life as before the verification, an option which many at-risk individuals seem to greatly value. In the words of Golman, Hagmann & Lowenstein (2017), “avoiding the diagnosis cannot help one avoid the disease […], but it can help one avoid the stress and anxiety”.
Oster et al. (2013) found that at-risk for HD patients who have not undergone testing behave almost identically in terms of important life decisions such as marriage, divorce or pregnancy when compared to patients who are sure not to carry the disease, despite expected differences in symptoms. However, the same authors have found that the likelihood of undertaking genetic testing increased with the gravity of the symptoms. Koszegi (2003) argued that people who develop suspicious symptoms are more reluctant to seek medical help because of anxiety and “fears of bad news”. Caplan (1995) also found that women with increasingly severe symptoms of breast cancer wait significantly longer before visiting a doctor as opposed to women with constant, less serious signs of illness. Hence, it seems testing avoidance is a necessary condition for a large share of individuals who are at risk for various life-threatening illnesses and who wish to maximize their utility. Indeed, the novel insight and main contribution of the paper by Oster et al. (2013) which we replicate is that if the patients seek to maximize their utility, it is optimal for them to avoid any information that does not contribute towards this goal.
Bénabou & Tirole (2016) termed the types of behaviour discussed in the previous paragraph as instances of “strategic ignorance”. That is, most individuals prefer to actively avoid information that can potentially lead to life-long feelings of anxiety and depression. It is more comfortable from a psychological point of view for individuals who are susceptible of carrying life-threatening diseases to live with uncertainty rather than to take up testing and risk enduring a hopeless existence.
However, more often than not humans expect that nothing bad can come their way; it is other people who are supposed to suffer from misfortunes, not themselves. There seems to be an innate perception of optimism towards the future, characteristic to a great deal of individuals, which keeps them hopeful and boosts their present felicity. Indeed, people are more likely to take into consideration the positive information they come in contact with, while at the same time disregarding material that is likely to bring about negative consequences and that could cause a potential utility loss in the future (for instance Garrett et al., 2016). Bénabou & Tirole (2016) argued that although large amounts of overconfidence are perilous to most individuals, some overconfidence is beneficial for it gives one confidence and hope. In other words, it seems as though it is embedded in and constructive for most individuals to exhibit some degree of overconfidence about the future. Weeks et al. (2012) discovered that almost 70% of patients in late stages of lung cancer and 80% of those in a similar stage of colorectal cancer failed to understand that chemotherapy is unlikely to cure their disease. Instead, they were overly optimistic about their chances of recovery once they engaged in chemotherapy.
In another study, 90% of individuals with a driving license in the U.S considered themselves to have superior driving skills than half of the same country’s drivers (Svenson, 1981). Furthermore, Preston & Harris (1965) found that even after suffering serious car accidents, a significant proportion of drivers were still overly optimistic about their driving skills, as if in a sturdy state of negation. Korn et al. (2014) compared patients suffering from depression to healthy patients and found that the latter group is more likely to exert a larger dose of optimism when compared to the former. That is, individuals who perceive themselves to be healthy, such as patients in incipient stages of HD, are comparatively less likely to take undesirable information into account when making decisions and thus tend to behave in an overly optimistic manner. In the case of patients who are at-risk for HD, overconfidence is high when individuals perceive themselves to be healthy gradually decreases as symptoms begin unravelling. According to Camerer (1997), overconfidence derives itself from an inner conflict between the desire to “feel good about oneself” and the need of being realistic. Generally, the former impulse takes precedence over the latter and is able to dominate it, hence the overly optimistic behaviour that is observed in most experiments.
Weinstein (1980) found that people perceived negative events as more probable to happen to others than to them. Moreover, according to him there is a strong propensity of believing that one has a below average chance of experiencing unwanted events, such as being diagnosed
with HD. Furthermore, he argued that the direr the consequences of a negative event, the higher the inclination to believe that the probability of experiencing that event is below average for a given individual. This would be implausible if people were not to derive anticipatory utility about future states. Since HD is a degenerative neurological disorder and has no cure, it is constructive for many to believe not to have it, despite objective evidence. This is so because a confirmed positive diagnosis would radically change the future prospects of people.
An important contribution of Oster et al. (2013) was to consider information avoidance, rather than cost, as an essential motive for which individuals who are at risk for HD do not test. According to Golman, Hagmann & Lowenstein (2017), it might be possible that patients designate the economic cost as the official reason for avoiding testing, despite their real motives being different. To counteract the latter claim, Oster et al. (2013) showed that 60% of the participants in their sample who chose not to undergo testing stated that “preference for living with uncertainty, as opposed to the economic cost, is an important reason for not testing”. Anderson et al. (2019) subsequently analysed the subjects in the PHAROS group and found that the patients’ main reasons for testing avoidance were the absence of an effective remedy and the “inability to undo knowledge”. Oster et al. (2013) were the first to show that the decisions of patients not to undergo testing is in fact an optimal response given the circumstances. More particularly, they proved that information avoidance from the part of patients who are at-risk for HD is rational and can be incorporated within a framework of optimal expectations.
6. Results
One of the goals of Oster et al. (2013) was to determine whether there exists a relationship between the testing rates of individuals and the evaluation score of investigators. Figure 1 presents their results. On the vertical axis we have the proportion of individuals who underwent testing since the last visit, while the horizontal axis displays the evaluation score of investigators at the patients’ last visit. The investigators give their assessments of the severity of the disease in numbers ranging from 0 to 4; each number further translates into a category (0 – Normal, 1 – Nonspecific abnormalities, etc). The solid line emphasizes the mean testing rates among patients for each separate category of symptoms. The dotted line shows the coefficient of last_uhdrq from the regression discussed in the Econometrics section. The two lines differ because the dotted, as opposed to the solid line, also accounts for the effect of the
male, age and education control variables on test rates. Therefore, while the solid line captures both the effects of the control variables and of the investigators’ evaluation scores on test rates, the dotted line only captures the latter effect. As a result, the difference between the two lines indicates the effect of the control variables on test rates. * indicates that the results are significant at the 5% level. ** indicates that the results are significant at the 1% level.
Figure 1. The relationship between evaluation scores of patients and test rates. Reprinted from Oster et al. (2013).
Both lines show a positive relationship between testing rates and the doctors’ evaluation scores: the higher the evaluation scores, the more people pursue testing. The figure reveals that testing rates among patients with an evaluation score of 2 are significantly different at the 5% level from those among patients with an evaluation score of 1. However, the largest difference in test rates, of more than 2%, is found when comparing individuals with an evaluation score of 3 to individuals with an evaluation score of 4. According to this finding, the testing behaviour of patients with a score of 4 is significantly different, at the 1% level, when compared to that of patients with a score of 3. Hence, out of all the patients in the PHAROS group, those with an evaluation score of 4 have the highest probability of testing, of about 4%. Oster et al. (2013) also found that testing is prompted by a change in the evaluation scores. The most significant
hike in test rates is found when the evaluation score rises from 3 to 4. This change of evaluation scores brings about an almost 4% increase in test rates, a finding significant at the 10% level. Firstly, it would be surprising that patients chose to take up testing despite being nearly certain to have the disease. Given that there is no known cure for HD, it would be rather curious that a great deal of individuals chose to test only to have the gruesome confirmation of what they were already highly suspicious about, together with the associated feelings of anxiety and depression. However, the optimal expectations model allows individuals to choose their own subjective probabilities about the likelihood of carrying the disease. That is, it allows the patients’ own assessments to be different from the actual probabilities of having HD as determined by the investigators’ evaluation scores. The findings of Oster et al. (2013) are indeed consistent with this rationale. As is observed from Figure 2, where the horizontal axis indicates the motor score of patients as determined by the investigators and the vertical axis shows the patients’ own subjective probability of having HD, most individuals tend to ignore the objective evidence when considering their own chances of having HD, thus being overly optimistic about their chances of not carrying the disease. We can see that among the patients with an objective probability of having HD very close to 100%, the average probability of carrying the disease as perceived by the patients is only 50%. A notable aspect is that regardless of the actual probability of being HD-positive, the subjective probabilities of most patients tend to gravitate around the 50% mark.
Lastly, it seems that the largest share of patients who subjectively believe to have no risk of carrying HD at all is found among individuals with objective probabilities of having HD higher than 99%. Despite overwhelming evidence, it seems that the change in symptoms consistent with a change in evaluation scores from 3 to 4 triggers a defensive mechanism that allows some patients to not only avoid, but to completely ignore signals that can be potential bearers of bad news. Bénabou & Tirole (2016) argued that most individuals who might find themselves in life threatening situations often deem it beneficial to engage in strategic ignorance, that is, in the act of taking into account disproportionately more positive information, despite both positive and negative information being equally encountered and accessible. One contribution of Oster et al. (2013) was to show that in circumstances of great imminent danger, such as the ones HD patients with an evaluation score of 4 find themselves into, individuals may enter a state where they completely block out (or regard as inconclusive) all negative information and only allow for non-threatening clues to enter their utility function.
Weeks et al. (2012) reached a similar conclusion when analysing the behaviour of patients in late stages of lung and colorectal cancer, although this was not necessarily the goal of their research. The aforementioned phenomenon could be an extreme instance of strategic ignorance. One track for future research would be that of establishing whether such extreme strategic ignorance is observed and persists in other experimental settings, as well as to pinpoint the specificities of the circumstances under which it emerges. It might seem counterintuitive that an evaluation score of 4 brings about both higher testing and higher denial among patients. Nevertheless, it is plausible that patients who reach such a late stage of HD could split into two different groups according to their mental fortitude.
Having to bear the substantial psychological burden of very likely carrying the disease, some of them may simply collapse and would have to face the grim reality, thus taking up testing; it might also be that at this point the anticipatory utility they derive is not high enough to compensate them for the cost of the desired state not being realised. However, others may derive so much anticipatory utility from believing not to have the disease even in this late stage of HD such that admitting anything short of being healthy is simply not an option. Hence, such patients enter the only phase consistent with their beliefs, one of complete negation; entering this extreme phase might also be coupled with individuals exerting a substantial degree of psychological endurance that the former group could not have exerted. An interesting track for future research would be that of ascertaining whether highly intense psychological situations could have diametrically opposed effects on a group of individuals.
Secondly, it is interesting to note that a significant percentage of patients only choose to test once they reach an evaluation score of 4. At that point, the objective probability of carrying HD is higher than 99%. According to this finding, it seems that information avoidance in the case of untested individuals who are at-risk for HD is optimal only to the extent to which it is plausible for patients to believe not to carry the disease. Mild symptoms can be more easily ignored than acute symptoms. For instance, it may be rather difficult for a patient to encounter severe chorea or hypokinesia and still remain overly optimistic about her chances of not having HD. On the other hand, mild signs of chorea or brief losses of concentration can be much more easily disregarded and considered circumstantial instead of acknowledged as signs of HD. Hence, Oster et al. (2013) contributed to the existing literature not only by proving that information avoidance can be optimal from the patients’ perspective within a framework of optimal expectations, but also by showing that there is an extent to which agents deem it beneficial to avoid information. In other words, there seems to be a threshold of experienced symptoms which, if crossed, causes information avoidance to become unfeasible simply because the objective evidence overwhelms one’s longing to feel optimistic about her health prospects.
The findings of Oster et al. (2013) further reinforced the idea that humans are innately overconfident, and can continue to do so even under the direst of circumstances. Similar results were found by Svenson (1981), Preston & Harris (1965) and Camerer (1997), among many others. However, as shown above, after a certain point is crossed, being overly optimistic simply becomes unfeasible. This explanation seems to be in line with that of Bénabou & Tirole (2016), according to which moderate amounts of overconfidence can be beneficial, while substantial ones can be rather detrimental to the well-being of individuals. Therefore, two interesting tracks for future research would be that of studying whether such overconfidence shifts actually occur in further experimental settings, as well as that of drawing clearer boundaries for the threshold at which overly optimistic behaviour becomes impractical. Finally, the study by Oster et al. (2013) confirmed the low testing rates for genetic diseases found in the previous literature. Most importantly, though, they provided an optimal expectations framework within which, far from being irrational, it is both rational and optimal for agents to avoid testing.
7. Conclusion
The aim of this paper was to present and discuss part of the results obtained by Oster et al. (2013). More specifically, the focus of the present writing was on explaining why is it that patients with high evaluation scores are more likely to take up testing for Huntington’s disease, as compared to those with low evaluation scores. To that end, we have shown the roles that overconfidence and information avoidance play when considering the behaviour of individuals who are at-risk for HD. While the neoclassical framework deemed information avoidance as a sign of irrationality from the part of individuals, the framework of optimal expectations that Oster et al. (2013) used was able to incorporate information avoidance and classified it as an optimal course of action for patients. However, the main contribution of this paper was to signal the possibility that information avoidance could be optimal for at-risk for HD patients only to a certain extent. Hence, it might be optimal only as long as individuals found it plausible to believe not to have the disease, that is, as long as the symptoms were so moderate such that they allowed the patients to believe them to be circumstantial, rather than clear signs of HD. Finally, we have to consider an important limitation of the paper by Oster et al. (2013). In their experiment where they analysed the same PHAROS data that Oster et al. (2013) used, Almqvist et al. (2006) computed that over the duration of the experiment only 96 individuals out of the whole PHAROS group would manifest clear signs of HD (and a change in evaluation scores from 3 to 4), indicated by a diagnosis rating of “definite HD”. This is less than 10% of the patients’ pool and might not be representative of the whole PHAROS group. This could be the case because, firstly, the sample of individuals who reach an evaluation score of 4 is small. Hence, basing the results solely on the 96 individuals can lead to reduced statistical power, large standard errors and an inflated effect of the evaluation scores on test rates.
Secondly, this could cause a selection problem because it is possible that this minority of individuals could be considerably different when compared to the 90% who are yet to develop clear signs of HD. In other words, it is likely that the behaviour, once they attained an evaluation score of 4, exerted by individuals who have not yet reached an advanced stage of HD, would be dissimilar to the behaviour of the 96 patients on which the results of Oster et al. (2013) are based upon. Most importantly, the 90% could be different in terms of testing behaviour. They might lack the overconfidence that the 10% had and thus have a more realistic view on the situation, which would translate in higher test rates. It is also possible that the anticipatory utility most of the patients derive from being overly optimistic about the future
once their disease advances might be more moderate when compared to that of the 96 individuals, thus making the cost of testing psychologically smaller, yet again implying increased testing at higher evaluation scores. More concisely, the 10% who showed clear signs of HD throughout the duration of the experiment can be a minority in terms of exerted testing behaviour. If this is the case, it might lead to reaching a different conclusion and therefore potentially nullifying the results of Oster et al. (2013).
Bibliography
Almqvist, E. W., Elterman, D. S., MacLeod, P. M., & Hayden, M. R (2001). High incidence rate and absent family histories in one quarter of patients newly diagnosed with Huntington disease in British Columbia. Clinical Genetics, 60(3), 198-205.
Almqvist, W. E. (2006). At Risk for Huntington Disease: The PHAROS (Prospective
Huntington At Risk Observational Study) Cohort Enrolled. Archives of Neurology 63(7), 991-996.
Anderson, K., Eberly, S., Marder., K. S., Oakes. D., Kayson, E., Young, A., & Shoulson, I. (2019). The choice not to undergo genetic testing for Huntington disease: Results from the PHAROS study. Clinical Genetics, 96(1), 28-34.
Astivia, O. L. O., & Zumbo, B. D. (2019). Heteroskedasticity in Multiple Regression Analysis. What it is, How to Detect it and How to Solve it with Applications in R and SPSS. Practical Assessment, Research & Evaluation, 24(1).
Bénabou, R., & Tirole, J. (2016). Mindful Economics: The Production, Consumption, and Value of Beliefs. Journal of Economic Perspectives, 30(3), 141-164.
Brunnermeier, M. K., & Parker, J.A. (2005). Optimal Expectations. American Economic Review, 95(4), 1092-1118.
Camerer, C.F. (1997). Progress in Behavioral Game Theory. Journal of Economic Perspectives, 11(4), 167-188.
Caplan, L.S. (1995). Patient delay in seeking help for potential breast cancer. Public Health Reviews, 23(3), 263–274.
Feinstein, L., & Sabates, R. (2006). The role of education in the uptake of preventative health care: The case of cervical screening in Britain. Social Science & Medicine, 62(12), 2998-3010.
Garrett, N., Lazzaro, S. C., Ariely, D., & Sharot, T. (2016). The brain adapts to dishonesty. Nature Neuroscience, 19(12), 1727-1732.
Golman, R., Hagmann, D., & Lowenstein, G. (2017). Information Avoidance. Journal of Economic Literature, 55(1), 96-135.
Hogarth P; Huntington Study Group. A study of the feasibility of clinical research in the Huntington’s at-risk population. Paper presented at: 18th International Meeting of the World Federation of Neurology Research Group on Huntington Disease; August 31, 1999; The Hague, The Netherlands. Abstract 105.
Kirkwood, S. C., Siemers, E., Hodes, M. E., Conneally, P. M., Christian, J. C., & Foroud, T. (2000). Subtle changes among presymptomatic Huntington’s disease gene carriers. Journal of Neurology, Neurosurgery & Psychiatry, 69(6),773-779.
Korn, C. W., Sharot, T., Walter, H., Heekeren, H.R., & Dolan, R. J (2014). Depression is related to an absence of optimistically biased belief updating about future. Psychological Medicine, 44(3), 579-592.
Koszegi, B. (2003). Health Anxiety and Patient Behaviour. Journal of Health Economics, 22(6), 1073-1084.
Lerman, C., Hughes, C., Trock, B. J., Myers, R. E., Main, D., Bonney, A., Abbaszadegan, M. R., Harty, A. E., Franklin, B. A., Lynch, J. F., & Lynch, H. T. (1999). Genetic Testing in Families with Hereditary Nonpolyposis Colon Cancer. Journal of the American Medical Association, 281(17), 1618-1622.
Malmendier, U., & Taylor, T. (2015). On the Verges of Overconfidence. Journal of Economic Perspectives, 29(4), 3-8.
Oster, E., Shoulson, I., & Dorsey, E. R. (2013). Optimal Expectations and Limited Medical Testing: Evidence from Huntington Disease. American Economic Review, 103(2), 804-830.
Preston, C. E., & Harris, S. (1965). Psychology of drivers in traffic accidents. Journal of Applied Psychology, 49(4), 284-288.
Roberts, J. Scott, Barber, M., Tamsen, M. B., Adrienne, L. C., Farrer, L. A., Larusse, S. A., Post, S. G., Quaid, K. A., Ravdin, L. D., Relkin, N. R., Sadovnick, A. D., Whitehouse, P. J., Woodward, J. L., & Green, R. C. (2004). Who Seeks Genetic Susceptibility Testing for Alzheimer's Disease? Findings from a Multisite Randomized Clinical Trial. Genetics in Medicine, 6(4), 197-203.
Ropka, M. E., Wenzel, J., Philips, E. K., Siadaty, M., & Philbrick, J. T. (2006). Uptake rates for breast cancer genetic testing: a systematic review. American Association for Cancer Research, 15(5), 840-855.
Svenson, O. (1981). Are We All Less Risky and More Skilful than Our Fellow Drivers? Acta Psychologica 47(2), 143-148.
Weeks, J., Catalano, P. J., Cronin, A., Finkelman, M. D., Mack, J. W., Keating, N. L., & Schrag, D. (2012). Patients’ expectations about effects of chemotherapy for advanced cancer. The New England Journal of Medicine, 367(17), 1616.
Weinstein, N.D. (1980). Unrealistic Optimism About Future Life Events. Journal of Personality and Social Psychology, 39(5), 806-820.
