Mobile Apps and Financial Decision Making* (2024)

Search for other works by this author on:

Article history

1. Introduction

Does the availability of new technology equip consumers to make better financial decisions? Ostensibly, if people have easier access to information, they should be able to avoid mistakes (Stango and Zinman, 2009; Jørring, 2019). However, while the rate of technology adoption is straightforward to quantify (Anderson, 2015; Carlin, Olafsson, and Pagel, 2019), measuring its economic impact is challenging. It is difficult to find settings, especially natural experiments, where it is possible to calibrate how technology affects people’s behavior and their outcomes.

In this paper, we investigate how access to a mobile financial app affects consumer financial decision making. We use individual, transaction-level data from a financial aggregation platform in Iceland. The platform allows users to link all of their checking, savings, and credit card accountsx, and view all spending, income transactions, and account balances in one place. A considerable fraction of adults in Iceland use this service and the user population appears to be representative of the overall population (Olafsson and Pagel, 2018; Carvalho, Olafsson, and Silverman, 2019).

We exploit an exogenous shock that made accessing the online platform easier. Before November 2014, access to the personal financial management software was possible only via an Internet browser. However, on November 14 2014, a mobile application was released, which gave users easier and remote access to their bank account information. This caused a discontinuous jump in the propensity for users to log in to the platform (Figure1).

Our dataset contains individual-level, time-series information about the frequency, and method of access to bank information (computer versus mobile app), demographics, expenditures by category, income, use of consumer credit (credit cards and checking account overdrafts), and resultant financial outcomes (consumer debt and bank fees).

To make welfare statements, our primary object of interest is the frequency of non-sufficient fund (NSF) fees. Twenty-seven percent of our subjects (3,681 individuals) incurred an NSF fee at some point during our sample period. When a consumer attempts to make a purchase with a debit card and exceeds her overdraft limit, she incurs an NSF charge, but the purchase is denied. This represents an unambiguous mistake because the consumer suffers a cost with no benefit. This contrasts with overdraft interest or late fees (which we also analyze): while it is likely better for consumers to avoid such costs,1 the ability to borrow money conveys a benefit.2

We analyze two specifications. First, we estimate standard intention-to-treat (ITT) regressions, in which economic outcomes are regressed on an indicator for whether a month is post-November 14, 2014. We include individual fixed effects to control for all time-invariant individual characteristics and month-fixed effects to take care of seasonal variation. We perform additional regressions where we interact demographic characteristics with the treatment variable to characterize who in the population benefited the most from the new technology. We also provide evidence that no other confounding events appeared to take place simultaneously when the app was released.

Second, we perform a difference-in-differences (DiD) estimation. The control group is composed of people who were signed up automatically for the platform through their internet bank, but never accessed it before the app was released. We compare their behavior to the treatment group who used the platform before the app was available and were more responsive to the introduction of the new technology.

The release of the app was associated with a drop in NSF fees in both the ITT and DiD specifications. Using 12-, 18-, and 24-month windows around the app release in the ITT specification, we document a 14.1%, 26.8%, and 38.4% decrease in NSF fees per individual. As we discuss in the paper, treatment-on-the-treated (TT) calculations imply an average annual decline of $4.46– $4.79 in NSF fees at 12-months following the app release (Baker, Gruber, and Milligan, 2008; Havnes and Mogstad, 2011).3 As a basis for comparison, NSF fees are typically $8–$9 in Iceland. Our demographic analysis indicates that women, high income individuals, and Baby Boomers were more likely to benefit from the app release. Additionally, individuals who incurred higher fees before the app introduction benefited more from the new technology.

The aggregation platform is exclusively for informational purposes and did not provide nudge-based interventions or any financial actions (e.g., paying bills). It follows that the likely mechanism by which the app lowered NSF charges was by making access to information less costly and promoting more frequent information acquisition. It is possible that the app may have made financial information more salient, but our setting does not provide an opportunity to assess the impact of salience.

The results regarding overdraft interest and late fees are mixed. Our ITT estimates show that use of overdrafts and late fees increased in the months following the app introduction. However, the DiD specification shows a relative reduction in these costs for treated individuals compared with controls. As we discuss in the paper, however, our analysis regarding these types of costs are limited by the fact that we did not have information about individual overdraft interest rates in the data, and interest rates were known to be changing during our sample period. As such, we cannot make unambiguous conclusions about these types of costs.

Previous papers provide evidence that technology can improve welfare. Technology may teach inexperienced consumers to save, avoid fees, and better cope with unanticipated shocks (Breza, Kanz, and Klapper, 2020). It may provide mechanisms in developing economies that improve consumption-smoothing and risk-sharing during economic shocks (Jack and Suri, 2014; Suri, 2017). Technology may also be used to increase the attention that consumers pay to managing their personal finances (Stango and Zinman, 2014; Karlan, Morten, and Zinman, 2017; Medina, 2021; Bursztyn et al., 2019).

In contrast to much of the previous literature, we study a developed country where our subjects have higher income with comparatively more financial experience. This makes us more confident that we are studying how technology reduces information frictions, rather than how it affects general financial education. The simple fact that many subjects were previous users of a lower-tech desktop version of the interface, and then received access to a newer version that lowers information costs, allows us to say more about the precise mechanism through which this technology affects consumer behavior.

The remainder of the paper is organized as follows. In Section 2, we describe the data and provide summary statistics. In Section 3, we explain our identification approach, report our main results, and discuss their robustness. Finally, Section 4 provides concluding remarks.

2. Data and Summary Statistics

We use data from Iceland that are collected by Meniga, a financial aggregation software provider to European banks and financial institutions. The platform allows bank customers to aggregate information about all of their bank accounts and credit cards across multiple banks in one place by consolidating their data from various sources.

The platform automatically records all bank and credit card transactions, including descriptions, as well as balances, overdrafts, and credit limits. Figure2 displays screenshots of the app’s user interface. The left screenshot shows the app’s financial feed including bank account information, the middle one shows the transactions interface, and the right one shows background characteristics that the user provides in the app’s settings.

The app’s interface did not change over our sample period and the desktop interface was the same in appearance and functionality. Upon logging in, users first see the financial feed interface. After that, they can tap on “Show Only Transactions” to see the transactions interface. To edit their settings, they can tap on “More” and “Edit Profile.”

Later versions of the interface (offered after our sample period) flagged certain events after individuals logged into the app. These might include unusually high transactions, money deposits, or low balances. Figure3 shows an example screenshot of the feed a consumer might observe with flagged events and a screenshot of the default options for messages and merchant offers. When the app was introduced in Iceland in 2014, there were no merchant offers. It was not until 2017 that Meniga expanded their merchant offer features. Also, the app did not send push notifications, so that individuals would have to log in first to see messages or flagged data.4

Using this data from Iceland has two main advantages when studying the effect of new technology on financial decision making. First, Icelandic consumers use electronic means of payments almost exclusively,5 which implies that the data capture the financial lives of the users better than aggregation data in other settings. Second, this dataset is more representative of the underlying population than data obtained from other financial aggregator apps. This is because: (i) all adult individuals in Iceland need to have a bank account6; (ii) bank accounts in Iceland are personal and cannot be shared, even within households; (iii) the majority of Icelanders use internet banking7; (iv) all individuals with an online bank account can access Meniga with one click without signing up on a separate homepage; and (v) Meniga is the only financial aggregation platform in Iceland.

In January 2017, 262,846 Icelandic individuals were older than 16  years of age. Our sample contains 52,545 users (approximately 20% of the population). For our analysis, users with incomplete records are excluded subject to three criteria. First, they must have observable bank account balances and credit limits. Second, key demographic information about the user must be available (age, sex, and postal code). Finally, the consumption stream of each user must be credible, which we validate by requiring at least five food transactions per month in at least twenty-three of the last 24  months. After imposing these requirements, we are left with 13,411 active users with complete records. We confirm that the demographic and economic characteristics of our study population are not different from those who were excluded and are representative of the adult Icelandic population.8

For purposes of analysis, we divide the user population into three generations. Baby Boomers were born between 1946 and 1964, members of Generation X were born in 1965–80, and Millennials were born between 1981 and 2000.

We collect a monthly panel of individual logins, bank fees, and credit use from November 2011 to January 2017. The data include how many times each individual logs in via the mobile app versus through a desktop computer. We observe the following types of bank fees and charges.

1. NSFs: The bank charges this fee when a debit card transaction is attempted and there are insufficient funds or the overdraft limit is exceeded in the consumer’s current account. As discussed above and in Carvalho, Olafsson, and Silverman (2019), incurring such a fee is consistent with a mistake made by a consumer. The consumer incurs the cost of attempting a purchase, but is denied a benefit in obtaining a product.

2. Late fee charges: Fees assessed for paying bills after their due date.

3. Overdraft interest: Overdrafts are the primary form of high-interest, unsecured consumer debt in Iceland. An overdraft occurs when withdrawals from a current account exceed the available balance. This means that the balance is negative and the bank provides credit to the account holder at an interest rate that is partly based on an individual’s affordability and credit history. At the time of the app release in 2014, the average overdraft interest rate was approximately 13% (see Figure4), which is a mark-up of 7% over the central bank policy rate (CBPR).

TableI reports the distribution of first-time mobile app users over time. As can be seen, 76.6% of people who ever used the mobile app during our sample period started using it during the first 12 months after it was released (i.e., by October 2015). This corresponds to 33.7% of the individuals in our sample. Given the large scale adoption of the technology, it is unlikely that a selection bias is present: that is, it is unlikely that such a large group of people changed their financial decision making for other reasons coincidentally with the date of the app release.

TableII displays summary statistics for our sample for the period before the app release, where we distinguish between adopters of the app and non-adopters. As can be seen, adopters and non-adopters have comparable characteristics, but adopters tend to be a bit younger and incur slightly lower bank fees. But, they are similar in terms of total income, liquidity, and cash holdings.

However, as we outline below, our empirical set-up compares people to their previous selves before the app release and we include individual-fixed effects. As such, these slight differences do not preclude an analysis of the effect of the app release on consumer decision-making. Another potential concern is that any effect that we might identify would be due to one specific fraction of the population, such as younger individuals. We investigate this below, but there does appear to be a quick and broad adoption of the app by all generations (Figure5).

3. Empirical Analysis

3.1 ITT Specification

Our empirical strategy exploits the discontinuity in the ease of accessing the financial aggregation platform that arose when the mobile application was released on November 14, 2014. As we investigate in Section 3.3, to our knowledge, no other confounding event took place around the same time. The timing of the app release is plausibly exogenous to individual characteristics, but caused some individuals to log in more often and is thus a valuable source of identifying variation. We exploit this to estimate the effect of more convenient, lower-cost access to financial information.

We estimate standard ITT regressions, in which bank fees and other outcomes are regressed on an indicator for whether the month is post-November 14, 2014, individual-fixed effects, and month (1–12)-fixed effects to control for seasonality. Specifically, we run various regressions based on the specificationwhere i is an individual identifier, t represents time (month-by-year), $Yi,t$ is individual i’s outcome of interest at time t, and PostNov14 is a dummy variable that takes the value 1 in periods after the release of the mobile app and zero before. The parameter ρ_i is an individual-fixed effect, ξ_m is a month-fixed effect, and the vector $Xi,t$ includes controls including demographic characteristics. Standard errors are clustered at the individual level.9

Note that some outcomes, including overdraft interest and late fees, are tied to the CBPR (and hence vary with the business cycle) which fluctuated considerably during the sample period. This implies that overdraft interest and late fees in the same month but in different years are not comparable. Dealing with this empirically is challenging. First, we do not have individual-level data on the interest rates actually paid by each consumer. Because of the disparity between central bank rates and overdraft interest rates, including a monthly central bank rate in our regressions would not lead to more precise or meaningful estimates. Second, we cannot add year–month-fixed effects to the regression because these would be collinear with the post-November 14, 2014 dummy variable. As such, our estimates regarding the effect of the app on overdraft interest and late fees should be considered with some reservation.

TableIII shows that the amount and number of NSF charges incurred by individuals in our sample dropped significantly. Using statistics from TableII as a baseline, the average amount of NSF fees for adopters was 29 ISK before the release. Given the coefficients in TableIII, this corresponds to a 14.1%, 26.8%, and 38.4% decrease in NSF fees per individual, using a 12-, 18-, and 24-month window around the app release. In our sample, 3,681 individuals incurred an NSF charge at some point during the sample period, which corresponds to 27.4% of our sample. Given the prevalence of these fees within the population and the fact that they represent a deadweight loss to consumers, this appears to be meaningful.

Columns (3–6) of TableIII show the change in the use of overdrafts and late fees after the app was introduced. All of the coefficients are positive and significant. Indeed, using 12-, 18-, and 24-month windows around the app release, the increases in overdraft interest were 7.2%, 13.1%, and 18.9%. At first glance this may appear to not support our hypothesis that the new technology improved consumer welfare. But, that conclusion would not be reliable. As this was a time of economic expansion, using debt may have been rational. There may be a benefit of debt that outweighed the added costs that people incurred. Whether the coefficients were negative or positive, it would be impossible to make a welfare statement about the effect of the app.

In contrast, for the reasons already outlined, we are able to make welfare conclusions about NSF fees. To investigate this further, we analyze which demographic groups benefited the most from the app introduction. We split our sample by gender, generation (Baby Boomers, Generation X, and Millennials), and into three income terciles (1 = lowest income and 3 = highest income). Additionally, we consider pre-November 14 2014 behavior: we split the sample into pre-app bank fee terciles (1 = lowest and 3 = highest) and pre-app platform login terciles (1 = lowest and 3 = highest). We re-run the regression in (1) and interact the PostNov14 indicator with these variables.

TableIV shows that women, Baby Boomers, and higher income individuals benefited more from the app.10TableV shows that individuals who, prior to the app release, incurred higher bank fees benefited more. The coefficients in TableV do suggest that people who logged in more prior to the app release benefited more, but the estimates are not statistically significant.

The ITT regression compares average outcomes in the post-period to average outcomes in the pre-period. To investigate this further, we replace the PostNov14 dummy variable in Equation (1) with indicators for the number of months from the mobile app release. We plot our year–month coefficients for the various financial outcomes in Figure6. By inspection, it is straightforward to see that the frequency and amount of NSF fees dropped following the app introduction. However, the change in overdraft usage and late fees is equivocal.

3.2 DiD Specification

A potential weakness of our analysis so far is that there is a single treatment (the availability of the new technology) that affected all people in our sample at the same point in time. As we discuss in Section 3.3, there are no other major events that occurred around November 14, 2014 that confound our estimates. Notwithstanding, we estimate the effect of the app release with a DiD approach to check the robustness of our findings.

The control group is defined as people who were signed up automatically for the platform through their Internet bank, but never accessed the platform before the app was released.11 In practice, signing up through an Internet bank and linking all accounts to the platform requires one click, but does not require logging into the platform or using it. We identify these individuals from their login behavior before the app introduction. One hypothesis is that the app release was less meaningful for these individuals: they had not used the aggregation service beforehand and their propensity to log in is arguably less likely to be influenced by the app release. As such, the treatment intensity of the app release would be greater for individuals who had used the aggregation platform in the prior period.

This is indeed what we see in the data. As can be seen in Figure1, upon the release of the mobile app, the average number of logins through the mobile app by those who had logged in via the desktop platform prior is three times as high as the average number of logins by those that had never logged in before. Also, while admittedly not statistically significant, the coefficients from TableV do suggest that people who had logged in more prior to the app release did benefit more from the app release.

We, therefore, use these two groups to evaluate the effect of the greater intensity of the treatment (easier access to information) that the mobile app provided. We estimate the following regression model:where treatment status is defined as $Xi,t$ is a vector of control variables (month-of-year, gender, age, and area of residence) and everything else is defined as before. The effect of the treatment variable T_i on the outcome variable $Yi,t$ is captured by β. Standard errors are clustered at the individual level.

The results in TableVI show that individuals that had logged into the aggregation platform prior to the release of the app tended to incur less NSF fees. Additionally, compared with people who never used the technology, release of the app led the treatment group to pay relatively lower late fees and overdraft interest. While this latter result supports our hypothesis that easier access to information lowers the cost of finance charges, we cannot make an unambiguous welfare statement for the reasons already mentioned.

To better appreciate the magnitude in the drop in NSF fees, we use the definition of treatment and control groups in the DiD specification and calculate a TT effect. In contrast to randomized-controlled trials, defining treatment and control groups can be challenging in empirical field studies and often require some ingenuity (Baker, Gruber, and Milligan, 2008; Havnes and Mogstad, 2011). In our setting here, we, therefore, define the difference in the probability of treatment between the treatment and control groups in two ways. The first is the difference in the propensity to log in to the platform with the app: 0.0412. The second is the difference in the propensity to log in to the platform in any manner: 0.0443. We use each of these to compute our TT measures.

As in Havnes and Mogstad (2011), the parameter of interest, β, captures the average causal effect of being treated in our DiD specification. Consistent with both Baker, Gruber, and Milligan (2008) and Havnes and Mogstad (2011), we also interpret β as an ITT effect, since our DiD model estimates the reduced-form relative impact of the new technology on the treatment group.

As such, using our DiD estimate from TableVI, the monthly TTs are computed as $0.40 and $0.37, respectively. On an annual basis, this is equivalent to a $4.79 and $4.46 drop in NSF fees. For completeness, we also perform the same procedure using our ITT estimate from TableIII. The monthly TTs are $0.99 and $0.92, respectively, which are equivalent to a $11.90 and $11.07 drop in NSF fees on an annual basis. Given that the typical NSF fee in Iceland is $8–$9, both sets of estimates appear to be meaningful.

3.3 Discussion

One concern that always arises in event studies like this is whether individuals change their behavior for a reason that is outside of the model or due to confounding events that took place simultaneously. In our setting, the concern might be that the choice to start using the app is driven by a separate decision to change attention to personal finances (as a means to change financial behavior). If this were the case, though, we would expect consumers to increase their frequency of logins before and after November 14 2014. Surely, there is nothing special about this date in particular. But, if rising attention to personal finances occurred via a separate choice, we would not expect to see the sharp jump that exists in the data.

Another concern might be whether our analysis could be confounded by other events around November 2014, such as regulation changes or macroeconomic events. The only regulatory change that we are aware of is a court ruling on December 14, 2014, that addressed deceptive merchant fees.12 However, this did not involve consumer fees directly and the percentage changes in merchant fees were very small. So, we do not think that this is material to our analysis.

We also reviewed several macroeconomic indicators and conditions around the event and found only one that might be important. Around the time of the app release, there was a small decrease in interest rates in Iceland (Figure4). Potentially, this could affect overdraft interest rates and, therefore, the interest expense that consumers in our sample paid. However, the most important effect that we observe in the data is a reduction in NSF fees, which are not influenced by interest rates.

We also ran a Factiva economic news search for Iceland during November 2014 to screen for confounding events. Many articles did report that the Icelandic central bank took another step in reducing interest rates in November 2014. However, we did not find any other confounding events. The interest rate had been trending downward over 2014 and 2015 (although overdraft interest payments trended upward). Without doubt, other macroeconomic conditions, such as personal and government consumption, output, and economic sentiment, changed over that period as well, but not discontinuously in November 2014.

Finally, another concern might be that the financial crisis of 2008 affected Iceland differently than other countries where consumer debt is prevalent and this might limit the external validity of our findings. However, the effect of the crisis was not that different in Iceland and the country had more or less recovered a few years after its onset (Olafsson, 2016). The country actually experienced high GDP growth and low unemployment during our entire sample period.

Additionally, Iceland is similar to many other economies, including the USA, with regard to the use of consumer debt. Individuals who started using the app within 12 months in Iceland held approximately $4,872 in overdrafts conditional on having overdraft debt. As previously mentioned, in Iceland individuals typically pay off their credit card in full and use overdrafts to roll over high-interest unsecured debt. Nevertheless, they still enjoy substantial liquidity because they have additional borrowing capacity before hitting their limits: a liquidity of $11,694 on average. In comparison, in the USA, the average credit card debt for individuals who roll it over is approximately $4,000 and individuals also enjoy substantial liquidity (Survey of Consumer Finances). Therefore, it appears that our results can be generalized to the USA and other European countries with relatively large consumer debt holdings, such as the UK, Spain, and Turkey.

4. Concluding Remarks

In this study, we analyze the effect of the release of a mobile app by a financial management platform. The release ostensibly eases consumers’ plight to gather information and make good choices. We quantify these effects and find a reduction in NSF fee payments in response to the reduced cost of accessing information more often. This drop was more pronounced for women, Baby Boomers, and individuals with higher income.

As we discuss, because the incurrence of NSF fees in Iceland does not bring any benefit for consumers, the decision to try to make a payment that results in an NSF charge is always dominated by the decision to avoid making the payment. Therefore, this drop represents a welfare benefit to consumers. In light of results from Stango and Zinman (2009) and Jørring (2019), this implies that the new technology helped consumers avoid making mistakes.

Mobile apps continue to develop faster than academics or practitioners can evaluate them. The results in this paper hopefully encourage future study in this area as financial technology evolves and affects peoples’ lives.

Data Availability

The data underlying this article are proprietary and belong to Meniga, they were provided through a research collaboration agreement that prevents us from sharing the data with other researchers.

Conflict of interest: All authors declare that they have no conflicts of interest in relation to this manuscript.

We thank numerous seminar participants, and our discussants and conference participants at the AFFECT Conference University of Miami, University of Kentucky Finance Conference, Santiago Finance Conference, Cerge-Ei Prague, 6th ITAM Finance Conference, and the AEA. This project has received funding from Danish Council for Independent Research, under grant agreement no 6165-00020. This project has benefited from funding from the Carlsberg Foundation. We are indebted to Meniga and their data analysts for providing and helping with the data. We also thank Fedra De Angelis Effrem and Andrea Marogg for outstanding research assistance.

1

Stango and Zinman (2009) and Jørring (2019) find that incurring late fees and paying overdraft interest are often associated with financial mistakes. The authors observe each consumer’s balance at the time the fees are incurred and confirm whether they were avoidable. Unfortunately, in our dataset, information about fees and interest payments is only available at a monthly frequency, which precludes a similar exercise.

2

In Iceland, checking account overdrafts are the way most consumers roll over unsecured high-interest consumer debt. Individuals pay interest on overdrafts, but there is no discrete overdraft fee.

3

For our TT computations, we use the definition of treatment and control groups in the DiD specification and scale our DiD estimate in two ways: (i) with the difference in the propensity to use the app and (ii) with the difference in the propensity to log in to the platform in any manner. We also perform the same procedure with our ITT estimate; the TT computation yields an annual drop in NSF fees of $11.90.

4

It is important to note that during our sample period, the platform looked and functioned differently than what is portrayed in advertisem*nts on Meniga’s homepage and in their international demo. During our sample period, users in Iceland have to log in initially to see all messages and warnings. A company statement from Meniga about user-based events is available from authors upon request.

5

ATM withdrawals make up approximately 1% of spending transactions by amounts or transactions volume.

6

There are many reasons for this. For instance, checks are not used in Iceland and if individuals want to receive salary payments or state benefits, they need a bank account.

9

We repeat our analysis by (i) clustering at the month-by-year level; (ii) double-clustering at the individual level and at the month-by-year level; and (iii) clustering at the area of residence level. Standard errors do not change much across the different clustering schemes and the statistical significance is robust. Results are available upon request.

10

For all of these regressions, multiplying the percentage of the population in each category by the appropriate coefficient yields the unconditional estimate in TableIII.

11

An ideal control group would be a subset of individuals who had no access to the technology. However, our entire sample did have access to the platform, so this is not a feasible control group in our setting.

References

© The Author(s) 2022. Published by Oxford University Press on behalf of the European Finance Association. All rights reserved. For permissions, please email: journals.permissions@oup.com

Issue Section:

Article

Email alerts

Citing articles via

More from Oxford Academic