Time Preferences, Health Behaviors, and Energy Consumption

Time Preferences, Health Behaviors, and Energy Consumption David Bradford, University of Georgia, [email protected] Charles Courtemanche, Georgia St...
1 downloads 0 Views 308KB Size
Time Preferences, Health Behaviors, and Energy Consumption

David Bradford, University of Georgia, [email protected] Charles Courtemanche, Georgia State University and NBER, [email protected] Garth Heutel, University of North Carolina at Greensboro and NBER, [email protected] Patrick McAlvanah, Federal Trade Commission, [email protected] Christopher Ruhm, University of Virginia and NBER, [email protected]

Growing evidence from economics and from psychology suggests that people make intertemporal trade-offs in a way that is not found in standard economic models. The standard assumption made in economic models is that consumers discount the future in a time-consistent manner; the relative weight that I place on consumption one year from now relative to now is equal to the weight that I place on consumption ten years from now relative to nine years from now. However, many laboratory and econometric studies suggest that consumers are timeinconsistent, and in particular exhibit present bias: the weight I place on consumption now relative to a year from now is greater than the weight I place on consumption nine years from now relative to ten years from now. If consumers are time inconsistent and exhibit present bias, important policy implications arise across a broad range of issues. Environmental and energy policy is affected, since consumers may be less willing to invest in energy-efficiency technologies (hybrid cars, home energy improvements) if they do not account for future savings as much as the standard model assumes that they do. Health policies are affected by the fact that consumers may underinvest in health, e.g. eating too much, exercising too little, or failing to buy insurance, since they are present biased. Policies designed to encourage retirement savings need also to accommodate the fact that consumers are time-inconsistent. In each of these examples, economists have suggested

that an explanation for consumer behavior (e.g. underinvestment in energy efficiency, high obesity rates, lack of retirement savings) may be present bias. The purpose of this paper is to provide direct evidence of the link between timeinconsistent preferences and consumer choices over health behaviors and energy consumption behaviors. We develop a survey that elicits each individual's time preferences and behavior in these markets that have been predicted to be impacted by present bias. We measure time preferences by asking questions about intertemporal tradeoffs, paying out one of the choices for randomly-selected respondents to mitigate hypothetical bias. We compute both the β€œbeta” (present bias) and β€œdelta” (time consistent) components of a quasi-hyperbolic specification, allowing for an empirical test of the association between time inconsistency and outcomes including body mass index, obesity status, dietary and exercise habits, smoking, drinking, health insurance status, preventive health care utilization, car model (to see its fuel efficiency), use of energy-efficient technologies in the home, and amount of retirement savings. A prior literature exists that merely tests whether or not consumers' time preferences are time-consistent. Laboratory evidence goes back to Thaler (1981), and laboratory evidence combined with neurological evidence in McClure et. al. (2004) support the existence of present bias in preferences. Other studies look at consumer behavior in the market for evidence of present bias. Individuals' choices about exercising (Dellavigna and Malmendier 2006), doing homework (Ariely and Wertenbroch 2002), participating in welfare programs (Fang and Silverman 2009), and eating (Ruhm 2012) all show evidence of time inconsistency and present bias. Buyers of cars seem to underweight future gasoline costs (Allcott and Wozny 2012). Obesity rates and body-mass index (BMI) seem to be correlated with present bias (Courtemanche, Heutel and McAlvanah 2011).

A paper similar in scope to this proposal is Chabris et. al. (2008), who measure discount factors of subjects and test for correlations between the discount factor and market outcomes. They find that discount factor predicts behavior like exercise, BMI and smoking. However, they do not attempt to separate out a present bias in discounting from a time consistent discount factor. The distinction between these two types of discounting can have important policy implications. In particular, present bias may justify government intervention in a way that is not justified if decisions are based on time-consistent preferences (O'Donoghue and Rabin 2006). Chabris et. al. (2008), along with another paper that similarly connects laboratory-calculated discount factors with market outcomes, Burks et. al. (2012), do not examine individuals' choices about energy efficiency investments, retirement savings, or health insurance. Preliminary results from a small pilot of the survey provide some evidence of a link between time inconsistency and these behaviors: smokers are significantly more likely to exhibit present bias, and those with present bias are significantly less likely to have well-insulated homes.

Survey Design An online survey was conducted using Qualtrics software. A panel of 200 respondents was purchased from Qualtrics Panels. The survey was conducted online December 5-7, 2012. We asked four sets of questions of the respondents. The first set consists of demographic questions (e.g. age, income, race, education). The second set of questions is used to calculate individual discount factors and test for present bias. We asked each respondent whether he would prefer a smaller payment more quickly or a larger payment after a longer wait (these questions are called multiple price list

(MPL) questions). There were three payout time pairs to choose among: now vs. one month from now, now vs. six months from now, and six months from now vs. seven months from now. For each of the three payout time pairs, the larger payment (after the longer wait) was $30, and the smaller payment ranged from $29 to $8. Each respondent was asked 22 such questions. 1 Using these questions, we can calculate a monthly discount factor for each of the three payout time pairs; call these 𝛿0,1 , 𝛿0,6 , and 𝛿6,7. (That is, 𝛿0,1 is the discount factor calculated using the respondent's answer to the MPL questions about payoffs now vs. one month from now.) 2

Several individuals' responses to the MPL questions are missing or are such that we cannot calculate one or more of the discount factors; these individuals are dropped. 3 We take the average of all of these three discount factors and call it π›Ώπ‘Žπ‘£π‘” . This assumes time-consistent

discounting. However, we allow for time-inconsistent discounting by noting that a respondent can have a different value for 𝛿0,1 and 𝛿6,7. Under time-consistent discounting, these two values

are equal to each other. If 𝛿0,1 < 𝛿6,7, then consumers are present-biased. Imposing a quasi𝛿

hyperbolic structure to the subjects' discounting, the present bias discount factor π›½π‘žβ„Ž = 𝛿0,1 and 6,7

the long-run discount factor π›Ώπ‘žβ„Ž = 𝛿6,7 . To combat hypothetical bias, we paid a random subset

of respondents based on their responses to these questions. 4

The third set of questions asks respondents about health-related economic decisions. These include questions about smoking (have you ever smoked, do you currently smoke, how many cigarettes per day do you currently smoke), wearing sunscreen, sexual behavior, and health 1

The time periods of the questions were the same as those used in Meier and Sprenger (2010). We adjusted the dollar values of the payments downward to reflect our budget (and rounded each to the nearest dollar integer). 2 The discount factor is calculated based on the question where the respondent switched from preferring the later payout to preferring the earlier payout. This method is described in more detail in Meier and Sprenger (2010, 199200). 3 For example, some respondents switched from the earlier to the later payments more than one time within a single price list. 4 20% of respondents were randomly selected to receive a payment. For each respondent chosen, one of the 22 MPL questions was randomly chosen to be the payout question. Payments were Amazon.com gift cards.

insurance. These questions were predominantly drawn from the US Center for Disease Control and Prevention's Behavioral Risk Factor Surveillance System (BRFSS) 2011 questionnaire. 5 The fourth and final set of questions asks respondents about energy use decisions. These include questions about transportation (what is your primary form of transportation, what is the fuel economy of your car) and home energy use (does your home have energy-efficient light bulbs, have you had an energy audit of your home conducted). These questions were predominantly drawn from questions asked in the US Energy Information Administration's 2009 Residential Energy Consumption Survey. 6

Preliminary Results The primary empirical question is to identify statistically significant correlations between time preferences and these decisions or behaviors. Table 1 provides summary statistics of several of the variables described above (means, standard deviations, and counts). We present the statistics for the entire sample (of those who responded to the question), and separately for those who are present-biased (π›½π‘žβ„Ž < 1) and those who are not present-biased (π›½π‘žβ„Ž β‰₯ 1). We

present a t-statistic for a test of a difference in means in the fourth column. The sum of the count of present-biased responses and non-present-biased responses is not the count of all responses because some respondents' MPL responses were inconsistent and did not yield a value for π›½π‘žβ„Ž .

The first panel of Table 1 presents the demographic variables. Age is in years, income is

in thousands of dollars, and the rest of the variables are binary indicators. White is one if the race/ethnicity question was answered "White (non-Hispanic)." (The other options were AfricanAmerican (non-Hispanic), Hispanic, Asian, and Other.) College is one if the highest level of 5

That survey is available here: http://www.cdc.gov/brfss/. That survey is available here: http://www.eia.gov/consumption/residential/. Some additional questions were taken from the survey designed for Attari et. al. (2010). 6

education completed is college graduate or postgraduate degree. There are no significant differences in means across present-biased and non-present-biased individuals. The second panel of Table 1 presents the discount factor variables described above. The average monthly discount factor using all three sets of MPL questions π›Ώπ‘Žπ‘£π‘” is 0.8779, which is quite low for a monthly discount factor but consistent with previous literature finding low

discount factors when using MPLs (Meier and Spregner 2010) (Frederick, Loewenstein and O'Donoghue 2002). There is no significant different in π›Ώπ‘Žπ‘£π‘” between present-biased and non-

present-biased individuals. 7 The long-run discount factor π›Ώπ‘žβ„Ž is significantly higher for present-

biased individuals. This is not what we would expect, and it is likely a result of the way in which present bias was calculated (π›Ώπ‘žβ„Ž is in the denominator of the calculation for π›½π‘žβ„Ž ). Of course, π›½π‘žβ„Ž is significantly lower for present-biased individuals.

The next panel of Table 1 examines differences in health outcomes and behaviors between the two groups. Respondents were asked if they would say that their health "in general" is excellent, very good, good, fair, or poor. The variable "health status" is that response on a scale from 0 to 4 where 4 is excellent and 0 is poor. The variable "good health" is a binary indicator equal to one if the response is excellent, very good, or good. Neither of these variables is significantly different between present-biased and non-present-biased individuals. "Ever smoked" is one if the respondent has smoked at least 100 cigarettes in his or her entire life; "regular smoker" is one if the respondent now smokes every day or some days (vs. not at all). Present-biased individuals are twice as likely to be regular smokers as are non-present-biased individuals, significant at the 95% level. Non-present-biased individuals are more likely to have

7

Although not reported in Table 1, we find that 50.6% of our respondents are present-biased, somewhat higher than the 36% found in Meier and Sprenger (2010). We find future bias (π›½π‘žβ„Ž > 1) in 19.8% of the sample, compared with 9% in Meier and Sprenger (2010).

health insurance and to answer that they use sunscreen always or nearly always when they are outside, but neither difference is significant. This panel supports our intuitive priors about how present bias affects health behaviors and outcomes, but statistical significance is limited because of small sample size. The last panel of Table 1 examines energy consumption behaviors. Each of these variables is a binary indicator. "Alternate transportation" equals one if the respondent's primary method of transportation to work or to school is anything other than "car (driving by yourself)" (i.e. car (driving with others/carpooling), public transportation, bicycle, walk, or other). Presentbiased individuals are slightly more likely to use alternate transportation, but the difference is not quite statistically significant. "High fuel economy" is one if the respondent's vehicle's fuel economy is 31 miles per gallon or greater. "Large vehicle" is one if the respondent's vehicle is anything other than "car" (i.e. station wagon, pickup truck, sport utility vehicle, minivan, or other). Present-biased individuals are less likely to have a high fuel economy vehicle and more likely to have a large vehicle, but the differences are not quite significant (although the difference for fuel economy is almost significant at the 90% level). "Home well insulated" equals one if the respondent says that his or her home is well insulated, rather than adequately insulated or poorly insulated. "Less energy than average" is one if the respondent claims that his or her residence consumes less energy than the average residence in the neighborhood. Presentbiased individuals are about 14 percentage points less likely to live in a well-insulated home, and this difference is significant at the 95% level. There is no significant difference between presentbiased and non-present-biased individuals in whether they use less energy than average. As with the results about health behaviors, the summary statistics on energy consumption largely confirm

our priors (present-biased individuals are less likely to make energy-saving investments), although statistical significance is limited. We next turn to regression analysis. Table 2 presents OLS regression results for the health outcomes and behaviors. The dependent variable in each regression is one of the health outcomes from Table 1. The right-hand-side variable of interest is a binary indicator for present bias. Odd-numbered columns include just this indicator and a constant; even-numbered columns also add in demographic controls. As was seen in the table of summary statistics, the only outcome for which present bias enters significantly is being a regular smoker. Adding in demographic controls does not substantially affect the coefficient or its significance. The coefficient on present bias in all other columns follows intuition but is insignificant. Table 3 presents similar regressions, this time looking at the energy consumption variables that appear in Table 1. Again, most of these coefficients on present bias are insignificant although they generally follow intuition. Only the coefficients on present bias in the regressions for having a well-insulated home are significant. (The coefficients on present bias in the high fuel economy regressions are significant at the 85% level.) Table 4 explores alternate measures of time preferences and how they affect the significant results found in Tables 2 and 3. We focus just on the indicators for being a regular smoker (columns 1–4) and having a well-insulated home (columns 5–8). The first two columns regress the indicator variable on π›Ώπ‘Žπ‘£π‘” , the measure of the discount factor that imposes time-

consistency. It does not enter significantly in any column, although it is of the expected sign in columns 1, 2, and 5 (we expect more patient individuals to be less likely to smoke and more likely to have well-insulated homes). The next pair of columns controls for both π›Ώπ‘žβ„Ž and π›½π‘žβ„Ž ,

the discount factors of a quasi-hyperbolic specification. In these regressions, only the

coefficients on π›½π‘žβ„Ž enter significantly, and their significance drops when demographic controls

are included (the p-value on the coefficient on π›½π‘žβ„Ž in column 4 is 0.192; in column 8 it is 0.196).

The results of these alternate specifications imply that when time preferences are measured the

standard way (i.e. ignoring present bias), there is no relationship found between them and these behaviors. Only when time-inconsistency is allowed do we see the expected relationships, and there is only a significant relationship between the present bias discount factor π›½π‘žβ„Ž and the

behavior, not between the long-run discount factor π›Ώπ‘žβ„Ž and the behavior. Of course, these results could also arise simply because the sample sizes are so small.

All of the other columns of Tables 2 and 3 can be replicated using these alternative measures of time preferences. Most of these regressions yield insignificant coefficients. One interesting exception is the regression for alternate transportation when controlling for π›Ώπ‘Žπ‘£π‘” . In

those regressions (whether controlling for demographics or not), the coefficient on π›Ώπ‘Žπ‘£π‘” is

significantly negative, indicating that more patient individuals are less likely to use alternate transportation (anything other than a car).

References

Allcott, Hunt, and Nathan Wozny. "Gasoline Prices, Fuel Economy, and the Energy Paradox." NBER Working Paper, November 2012. Ariely, Dan, and Klaus Wertenbroch. "Procrastination, Deadlines, and Performance: SelfControl by Precommitment." Psychological Science 13, no. 3 (May 2002): 219-224.

Attari, Shahzeen, Michael DeKay, Cliff Davidson, and WΓ€ndi Bruine de Bruin. "Public Perceptions of Energy Consumption and Savings." Proceedings of the National Academy of Sciences 107, no. 7 (September 2010): 16054–16059. Burks, Stephen, Jeffrey Carpenter, Lorenz Gotte, and Aldo Rusticini. "Which Measures of Time Preference Best Predict Outcomes? Evidence from a Large-Scale Field Experiment." Journal of Economic Behavior and Organization 84, no. 1 (September 2012): 308-320. Chabris, Christopher, David Laibson, Carrie Morris, Jonathon Schuldt, and Dmitri Taubinsky. "Individual laboratory-measured discount rates predict field behavior." Journal of Risk and Uncertainty 37 (2008): 237-269. Courtemanche, Charles, Garth Heutel, and Patrick McAlvanah. "Impatience, Incentives and Obesity." NBER Working Paper, 2011. Dellavigna, Stefano, and Ulrike Malmendier. "Paying Not to Go to the Gym." American Economic Review 96, no. 3 (June 2006): 694-719. Fang, Hanming, and Dan Silverman. "Time-Inconsistency and Welfare Program Participation: Evidence from the NLSY." International Economic Review 50, no. 4 (November 2009): 1043-1077. Frederick, Shane, George Loewenstein, and Ted O'Donoghue. "Time Discounting and Time Preference: A Critical Review." Journal of Economic Literature 40, no. 2 (June 2002): 351-401. McClure, Samuel, David Laibson, George Loewenstein, and Jonathan Cohen. "Separate Neural Systems Value Immediate and Delayed Monetary Rewards." Science 306 (October 2004): 503-507.

Meier, Stephan, and Charles Spregner. "Present-Biased Preferences and Credit Card Borrowing." American Economic Journal: Applied Economics 2, no. 1 (January 2010): 193-210. O'Donoghue, Ted, and Matthew Rabin. "Optimal Sin Taxes." Journal of Public Economics 90, no. 10-11 (November 2006): 1825-1849. Ruhm, Chris. "Understanding Overeating and Obesity." Journal of Health Economics 31, no. 6 (December 2012): 781-796. Thaler, Richard. "Some Empirical Evidence on Dynamic Inconsistency." Economics Letters 8, no. 3 (1981): 201-207.

Table 1 – Summary Statistics All

Present Biased

Not Present

t-test of

Biased

difference of means (p-value)

Demographics Age

Female

White

College

Income ($1000)

51.46

51.02

52.88

0.9295

(13.36) [200]

(12.38) [87]

(13.83) [85]

(0.3539)

0.5330

0.5581

0.5238

–0.4467

(0.5002) [197]

(0.4995) [86]

(0.5024) [84]

(0.6557)

0.7990

0.8160

0.8118

–0.0725

(0.4018) [199]

(0.3897) [87]

(0.3932) [85]

(0.9423)

0.3850

0.3678

0.3647

–0.0421

(0.4879) [200]

(0.4850) [87]

(0.4842) [85]

(0.9665)

53.45

55.90

53.43

–0.3451

(43.21) [185]

(38.25) [83]

(51.58) [77]

(0.7305)

Discounting π›Ώπ‘Žπ‘£π‘”

0.8779

0.8801

0.8756

–0.2780

(0.1037) [165]

(0.0787) [84]

(0.1249) [81]

(0.7814)

π›Ώπ‘žβ„Ž

0.8723

0.9402

0.8071

–5.980***

(0.1634) [183]

(0.0905) [87]

(0.1865) [85]

(0.0000)

π›½π‘žβ„Ž

0.9702

0.8267

1.117

9.195***

(0.2527) [172]

(0.1349) [87]

(0.2611) [85]

(0.0000)

Health Outcomes and Behaviors Health Status

Good Health

Ever Smoked

Regular Smoker

2.370

2.299

2.318

0.1284

(0.9474) [200]

(0.9358) [87]

(0.9662) [85]

(0.8980)

0.8250

0.7816

0.8235

0.6871

(0.3809) [200]

(0.4155) [87]

(0.3835) [85]

(0.4929)

0.5303

0.5698

0.5357

–0.4441

(0.5003) [198]

(0.4980) [86]

(0.5017) [84]

(0.6575)

0.2525

0.3448

0.1786

–2.499**

(0.4356) [198]

(0.4781) [87]

(0.3853) [84]

(0.0134)

Have Health

0.8050

0.7816

0.8235

0.6871

Insurance

(0.3972) [200]

(0.4155) [87]

(0.3835) [85]

(0.4929)

Regularly Use

0.2929

0.2644

0.2892

0.3599

Sunscreen

(0.4563) [198]

(0.4436) [87]

(0.4561) [83]

(0.7199)

Energy Consumption Behaviors Alternate

0.2261

0.2759

0.2143

–0.9321

Transportation

(0.4194) [199]

(0.4995) [87]

(0.4128) [84]

(0.3526)

High Fuel

0.2112

0.1571

0.2647

1.553

Economy

(0.4094) [161]

(0.3666) [70]

(0.4445) [68]

(0.1228)

Large Vehicle

0.4506

0.4789

0.4118

–0.7917

(0.4991) [162]

(0.5031) [71]

(0.4951) [85]

(0.4299)

Home Well

0.3618

0.2644

0.4118

2.058**

Insulated

(0.4817) [199]

(0.4436) [87]

(0.4951) [85]

(0.0412)

Less Energy than

0.4700

0.4713

0.4824

0.1447

Average

(0.5004) [200]

(0.5021) [87]

(0.5027) [85]

(0.8851)

Notes: In the first three columns, the mean is in no parentheses, the standard deviation is in parentheses, and the count of nonmissing values is in brackets. ** indicates significance at 5% level; *** indicates significance at 1% level. Variables are described in the text.

Table 2 – Health Outcomes and Behaviors Regressions (1) (2) Health Status Health Status Present Bias

-0.0188 (0.146)

172 0.000

158 0.097

172 0.003

158 0.037

170 0.001

156 0.138

(7) Regular Smoker

(8) Regular Smoker

(9) Have Health Insurance

(10) Have Health Insurance

(11) Regularly Use Sunscreen

(12) Regularly Use Sunscreen

0.166** (0.0663)

0.174** (0.0670) -0.00194 (0.00258) -0.0205 (0.0683)

-0.0419 (0.0610)

-0.0591 (0.0638) 0.000694 (0.00196) 0.100 (0.0640)

-0.0248 (0.0691)

-0.0268 (0.0686) -0.00701** (0.00281) -0.198*** (0.0710)

Income ($1000)

Female

0.536*** (0.0547)

0.0539 (0.0768) 0.00608* (0.00309) 0.129* (0.0768) 0.229** (0.102) -0.191** (0.0841) -0.00108 (0.00110) 0.0712 (0.169)

0.824*** (0.0416)

College

Age

-0.0419 (0.0610)

2.318*** (0.105)

White

Present Bias

(5) (6) Ever Smoked Ever Smoked

-0.0531 (0.0640) -0.00392* (0.00230) -0.0234 (0.0664) -0.0427 (0.0817) 0.0520 (0.0657) 0.000946 (0.000615) 0.999*** (0.112)

Female

Observations R-squared

(4) Good Health

-0.0850 (0.150) -0.0162*** (0.00617) -0.102 (0.154) 0.0161 (0.226) 0.345** (0.170) 0.00300 (0.00198) 2.935*** (0.341)

Age

Constant

(3) Good Health

0.0341 (0.0767)

White College Income ($1000) Constant

0.179*** (0.0420)

0.118 (0.0934) -0.148** (0.0735) -0.000168 (0.00100) 0.246* (0.131)

0.824*** (0.0416)

0.0978 (0.0921) 0.106 (0.0645) 0.00124 (0.00129) 0.560*** (0.124)

Observations 171 157 172 158 R-squared 0.036 0.087 0.003 0.076 Notes: Heteroskedasticity-robust standard errors are in parentheses. *** p