Who gets duped? The impact of education on fraud detection in an investment task

The literature on financial education and financial literacy suggests that they have positive effects on financial decision making. Lusardi and Mitchell (2014) reviewed studies that showed higher financial literacy was linked to better financial behavior. However, some researchers (Willis 2011; Hastings et al. 2013) challenged this relationship. Kaiser and Menkhoff (2017) and Kaiser and Menkhoff (2020) conducted meta-analyses of financial literacy studies and found evidence that financial education improved financial literacy and behavior. Despite some conflicting results, the general consensus is that more education and literacy enhance financial decision making. Therefore, it is reasonable to expect that greater financial understanding would also reduce the vulnerability to financial fraud.

There are few academic papers on financial fraud, but some non-peer reviewed papers have been produced by organizations such as the Pension Research Council (e.g., Kieffer and Mottola 2016), which explored how individual characteristics influenced the likelihood of being defrauded. Knutson and Samanez-Larkin (2014) examined how personality traits affected skepticism and found that investors with lower impulse control were more prone to financial fraud, while cognitive ability and risk attitude did not matter. Anderson (2016) and Anderson (2013) found that consumer literacy and numeracy were associated with the ability to identify and avoid fraudulent advertising and schemes, respectively. Recently, some peer-reviewed papers have emerged on this topic. Andreou and Philip (2018), Engels et al. (2020), and Wei et al. (2021) all found that higher financial literacy helped individuals detect fraud in their financial transactions.

Fraud is a form of deception, a topic that has been investigated more extensively in other social sciences. Levine (2014) has outlined what he calls Truth Default Theory (TDT), a comprehensive theory of deception and deception-detection. One key assertion of TDT is that people presume honesty in communication. When two people communicate, both people assume that the other person is telling the truth. While Bernie Madoff and the data presented on financial fraud provide obvious exceptions to this assumption, even a brief bit of introspection reveals its overall plausibility. Human beings communicate with each other constantly, and it is beyond our cognitive facilities to constantly monitor every bit of communication for deception. TDT argues that it is evolutionarily adaptive for humans to assume honesty, and Levine (2019) provides significant evidence in support of this argument.

If humans assume honesty, how is deception ever uncovered? TDT (illustrated in Fig. 1) argues some events push people into a state of suspicion, in which statements are evaluated for their honesty. Potential trigger events include the speaker having a motive for deception (like personal gain) or a lack of correspondence between communication content and some knowledge of reality. Once in the suspicious state, the person actively evaluates the veracity of the statements, taking account of the internal logic of the statements, and the correspondence between the statements and other known facts. If the statements fail to have internal logical consistency or contradict other information, the person updates his/her beliefs that the communication is honest. If this updating results in the level of perceived veracity falling below some threshold, then the statements are evaluated as dishonest.

TDT also helps to provide context for research by Zhang et al. (2015), who examined how cues could be used to help increase investors’ skepticism. In Zhang et al.’s decision task, survey respondents from Amazon MTurk were asked to play the role of a financial advisor and recommend a mutual fund for a hypothetical client. The survey participant had to choose between five different fictionalized funds, one of which was based on Madoff’s fund. Zhang et al. found that 68% of their respondents recommended the Madoff fund; however, when respondents were asked about which fund was most suspicious, this cue reduced the number choosing the Madoff fund to 51%. In terms of TDT, the cue acted as a trigger that helped to move some participants into a state of suspicion.

TDT helps explain the results presented earlier regarding the role of financial literacy. First, having more financial literacy could make people more likely to become suspicious in a situation involving fraud, and to investigate the fraudulent opportunity further. Second, having more financial literacy could also make people more skilled at detecting fraud once they are suspicious. In this paper, we present a decision task that is very similar to the one used by Zhang et al. (2015), but with students of economics and finance as participants.

The works by Andreou and Philip (2018), Engels et al. (2020) and Wei et al. (2021) represent all the published work in the fields of economics and finance on individual, non-professional investors and fraud detection. All those papers rely upon self-report surveys to investigate fraud detection. In particular, survey participants were asked if they had detected fraud in the past, and then this behavior was correlated with other collected variables like education, financial literacy, gender, etc. Importantly, all this data was self-reported. From the surveys we are able to learn if an individual thought that someone had tried to defraud them; however, we do not know if what the individual detected was actual fraud, nor do we know if other fraud attempts were successful. The current paper addresses this issue by presenting survey respondents with a decision task (based on Zhang et al. 2015) in which the nature of the fraud is known, allowing us to draw stronger conclusions about fraud detection. Furthermore, we advance the field of fraud detection by using TDT as our model – we are unaware of any papers in economics or finance that use this framework to understand fraud detection.

The sample for this study consisted of undergraduate students from the College of Charleston.¹ Students were recruited from a variety of finance and economics backgrounds, ranging from no prior coursework to extensive coursework. Students were offered extra credit in their courses in return for completing the survey. All students gave consent before starting the survey. Anonymity was guaranteed; credit was offered only for participation, not performance.

At the College of Charleston, the undergraduate economics and finance classes from which our sample was selected are taught in sequence, starting with Principles of Microeconomics, which introduces market dynamics and resource allocation. The next class is Principles of Macroeconomics, which focuses on aggregate economic analysis. The third class, Business Finance focuses on corporate finance fundamentals, leading to Intermediate Business Finance where students apply these principles using financial software. The sequence culminates in Applied Portfolio Management, where students actively manage diverse asset portfolios, integrating knowledge from all previous courses into real-world financial analysis and decision-making. Because enrollment in these sequenced classes is mutually exclusive, there was no chance that a student might be asked to take the survey twice.

A total of 479 students attempted the decision task, with 430 finishing the survey. Of those students, 390 undergraduates filled out all fields required for our analysis. The full dataset and survey instrument are available at https://​doi.​org/​10.​17632/​6d4y9g5jt4.​2 (Blackwell 2022).

The basic structure of the survey was as follows:

For the main decision task, survey participants were asked to make an investment recommendation from among four different hypothetical funds (whose returns are shown in Fig. 2 with S&P 500 for comparison).² All the funds are based on real investments, although the fund names and dates are disguised (i.e. the data do not necessarily span between 2009 and 2014). One of the funds, “Fortitude Investments,” is based on Madoff’s fund. The main decision task was taken almost verbatim from the task presented by Zhang et al. (2015).

The participants were provided basic information about average returns and volatility for each fund. The subjects were then offered the opportunity to review more information about each fund. Choosing this option gave participants specific information about the fund’s investment strategy and auditors, with the Madoff fund’s auditor description listed as:

Fortitude uses SA & Associates, CPA for their auditing purposes. SA & Associates was established 15 years ago. The chief auditor was formerly a VP at Fortitude Investments.

This information provided the only additional opportunity to trigger the participants’ suspicion before submitting their recommendation. We collected data on the number of funds for which each participant asked for additional information as well as which funds they examined.

After submitting their recommendation, participants rated each fund based on how suspicious or unethical they found the fund to be, completed the Big Five Personality Inventory (John et al. 1991, 2008), and submitted information on their socio-economic and educational background. The post-decision data is discussed further below.

Descriptions of all variables used in our analysis are given in Table 1, with summary statistics for the analysis sample provided in Table 2. Our primary outcome measure is a binary variable indicating if the participant recommended the Madoff Fund or an alternative. We assume that participants who did not pick the Madoff fund did so because they detected potential fraud from that mutual fund. Participants provided demographic data by self-reporting their age, gender, and race. Because there may be a difference in the effect of additional information depending on whether the Madoff Fund is one of the funds being compared, we split this variable into those who included the Madoff Fund in their examination and those who excluded it.

For educational information, participants indicated which sequenced economics and finance classes they have completed. We used this data to create cumulative course indicators. This means the total effect of education for a participant who has completed Business Finance will also include the effect of Principles of Micro- and Macroeconomics, and coefficient estimates indicate the marginal effects of completing each course.

The Big Five Personality Inventory (BFI) measures participants on the five traits of extraversion, conscientiousness, agreeableness, neuroticism, and openness. The BFI consists of statements with which each individual could agree or disagree. An example BFI statement is: “I am someone who is talkative” (John et al. 1991). A person who agrees with this statement would rank highly for extraversion. The American Psychological Association Dictionary definitions for the BFI traits are included in Table 1.

To measure skepticism, we followed Zhang et al. (2015) by asking students to rank how suspicious they found each fund. To avoid triggering scrutiny due to asking the question, the question was left until after participants had submitted their recommendation. To ensure that these scores are comparable across observations, we created the variable SUS by dividing the score assigned to the Madoff Fund by the average of all scores provided by each participant.

Finally, our dataset includes the time taken to complete the survey in seconds. While almost all the 390 undergraduates who filled out all required fields took between 3 and 30 min to complete it, 20 participants completed the survey in less than 3 min. These participants chose all five fund options in nearly equal proportions, suggesting that all their answers may have been randomly selected to complete the survey as quickly as possible. To avoid introducing random noise into all variables used, we drop these 20 observations from our primary analysis sample.

In the remaining 370 observations of our analysis sample, the median duration was 434 s, or slightly over 7 min. The mean duration was 1,997 s (more than 33 min), reflecting a small number of extreme outliers. We include a dummy variable for participants who took more than 30 min to control for these outliers.³

According to TDT, deception and fraud are extraordinarily difficult to detect under most circumstances. Only those who are triggered to scrutinize have a real chance to detect fraud, and even then, only if enough inconsistencies can be identified. In the context of our investment task, this suggests a two-stage relationship between education, skepticism, and detecting the fraud of the Madoff Fund. In the first stage, the level of skepticism is determined. In the second stage, participants choose, conditional on their level of skepticism, to recommend either the Madoff Fund or an alternative.

If the participant has not been triggered to scrutinize for possible deception, they will take the information presented at face value in the second stage. In this “trusting” case, we would expect additional information about the fund options to reinforce the benefits of the high-return, low-risk Madoff Fund, increasing the likelihood of recommending it. Similarly, more economics and finance coursework should enable the participant to recognize the desirability of these fund traits, also increasing the likelihood of recommending the Madoff Fund.

If the participant has been triggered to scrutinize, however, we would expect the opposite effects. Since deception detection relies on recognizing inconsistencies between the deception and other known information, additional information about the fund options and additional background in economics and finance should provide more opportunities to detect the deception, which should reduce the likelihood of recommending the Madoff Fund.

Mathematically, TDT suggests a model for fund choice such as the following:where f(·) is increasing in both education and information, g(·) is decreasing in both, and controls include demographic and personality factors which may also affect the choice of fund.

Since the functional form of the binary choice of fund is dependent on the level of skepticism, the first stage involves identifying what determines suspicion. TDT does not make sharp predictions about what determines initial suspicion, but as it seems likely that individuals may differ in levels of innate skepticism, we control for personality and demographics. While we would not expect education to have a significant effect on innate skepticism directly, familiarity with economics and finance should make participants more confident in requesting additional financial information. We hypothesize that the effect of education on SUS should be mediated by information gathering.

TDT suggests a structural break in the probability of recommending the Madoff Fund, with the coefficients related to education and information changing between those who are trusting and those who become skeptical. However, there is no obvious a priori value of SUS that should be used to test such a break. Since TDT claims that most individuals have a high threshold, a high level of suspicion may be required to ensure sufficient ‘state’ skepticism. On the other hand, the context of the survey and the type of students who select into economics and finance courses may suggest a high unobservable innate skepticism, requiring only a modest level of suspicion to trigger scrutiny.

The simplest econometric approach to searching for a breakpoint is a threshold linear probability model. This extends the method of ordinary least squares (OLS) regression, which chooses coefficients that minimize the sum of squared residuals (SSR), to the choice of breakpoint. The data are sorted by relative suspicion, split into high and low suspicion groups at all possible values of the threshold variable, and linear probability regressions are run for each possible split. The model that provides the lowest SSR is then reported.⁴

The linear probability model results are provided in Table 4. The threshold for SUS is 0.57, placing less than 15% of the analysis sample in the trusting group. Extroverted and female participants are still less likely to choose Madoff, while conscientious participants are more likely to do so.

Gathering more information seems consistent with TDT. For the trusting group, asking for more fund information without comparing it to the Madoff Fund increases the likelihood of choosing Madoff. For the skeptical group, the coefficient is similar in magnitude but opposite in sign, indicating that gathering more information makes choosing Madoff less likely for those who are scrutinizing.⁵

The effect of education is less obvious, with no statistically significant education coefficients in our sample. This may be due to the small sample, as expanding the sample to include all undergraduates with the required data suggests that completing the Business Finance course has a positive and significant (5% level) effect, increasing the likelihood of choosing Madoff. For the skeptical group in the expanded sample, having completed the same course reduces the likelihood of choosing Madoff, although this effect is only significant at the 10% level.

To properly determine if there is a difference in the effect of education between the trusting and skeptical groups, we must test for differences in corresponding coefficients in our analysis sample. The χ²-statistic for difference in coefficients is 5.00, which is statistically significant at the 5% level. Although we cannot confidently say completing Business Finance makes choosing Madoff more likely for the trusting group and less likely for the skeptical, there is a difference between the groups in the direction suggested by TDT.

None of the other education variables show a significant difference in coefficients. This suggests that the skills that differentiate the choices of trusting and skeptical students are not acquired after a single introduction to economic and finance coursework. This seems reasonable since only basic financial concepts are covered in the two economics courses studied. Key models such as the CAPM are not covered in detail until Business Finance.

One possible shortcoming of this analysis is the use of linear probability models, which have a variety of known shortcomings. One popular alternative for binary choices is the logit model, which we also present in Table 4. We consider versions with no break (pooled) and with the same break in coefficients suggested by the threshold model.

The signs and general significance of gender, extraversion, and conscientiousness match the threshold model, both in the pooled model and the model with the break. In the pooled model, both completing Business Finance and gathering more information reduce the likelihood of choosing Madoff, although the effect of the course is only significant at the 10% level.

The break is highly statistically significant (χ² = 1,370), providing strong evidence that skepticism alters some or all the effects of education and information gathering. We again see statistically significant differences (and signs that match TDT) in the effect of the Business Finance course and gathering information between the trusting group and the skeptical group.

The magnitudes of the coefficients on Business Finance are very close, and we cannot reject the null that they sum to zero (χ² = 2.2). If financial education increases the likelihood of already-skeptical students avoiding a Ponzi scheme, it is difficult to identify such an effect in our data. However, gathering information, whether excluding or including information about the Madoff Fund, seems to not just cancel out but actively reduce the likelihood of choosing Madoff (χ² = 19.4 and 4.3, respectively).

Finally, the effect of the Intermediate Business Finance course in the break model deserves some comment. In both groups, the coefficient seems to have a dampening or corrective response to that of the previous course. In the case of the trusting group, we can reject the null that the coefficients sum to zero (χ² = 5.6) at the 5% level, suggesting that students with more practice using workhorse finance models are less inclined to jump at a fund that promises high returns for free. However, we are unable to reject the null that the effect among skeptical students sums to zero (χ² = 0.1).

Given the evidence that skepticism has an important effect on fund recommendation in our decision task, we now consider our first-stage hypotheses. Results of an OLS regression of SUS on our demographic, personality, education, and information variables are presented in Table 5.

We see that none of the demographic variables are statistically significant, and among the BFI only extraversion is significant (at the 10% level). Among the education variables, only the Principles of Microeconomics course is significant, also at the 10% level.

The negative sign on this coefficient may be concerning since we would not like to think an economics course makes students less adept at detecting fraud. However, this may reflect any number of driving forces. Since students in this course have only just been introduced to basic ideas like volatility, such students may simply associate high risk with suspiciousness.

In contrast to the other groups of explanatory variables, asking for additional information has a strong and clear effect on the perceived relative suspiciousness of the Madoff Fund. When the Madoff Fund is included among the funds for which more information is considered, each additional piece of information increases how suspicious the Madoff Fund seems. This act of comparing the Madoff Fund’s information to multiple alternatives could be viewed as financial due diligence and may represent the participants actively scrutinizing returns that seem too good to be true.

These results broadly align with the interpretation of the SUS variable as capturing ‘state’ rather than ‘trait’ skepticism. To examine our hypothesis that the effect of education on ‘state’ skepticism should be mediated through information gathering, we consider a regression of requests for additional information on education and our controls. Since our dependent variable is count data, we use Poisson and ordered logit regressions.

Of our controls, only gender is statistically significant in both models. Female participants are less likely to ask for more information than male participants, which is curious given that female participants are also less likely to choose Madoff.

The impact of education on information gathering is present but is more complex than expected. We do find that students with a finance background are more likely to request additional information, but the effect does not become statistically significant until students have completed Intermediate Business Finance. Even more unexpectedly, this effect seems to be reduced or canceled entirely for students who have also taken the Applied Portfolio Management course (χ² = 0.34 for the Poisson regression and 0.57 for the ordered logit).

Likely causes for this bounce back are not immediately obvious. Perhaps these students have such a high degree of confidence that they do not think the additional information will be necessary. This finding is similar to what the professional skepticism literature shows: less experienced auditors are more skeptical than more experienced ones. However, this does not mean that more experience or knowledge reduces skepticism. It could mean that more advanced students have learned to select the most relevant information for their investment decisions and ignore the less important ones. Or perhaps the bounce back is simply the result of normal variation among a small number of participants who fall into these two highest education categories (18 for Intermediate and 32 for Portfolio). Whatever the explanation, the results seem to align with TDT-based predictions, but only up to a point.