Discount rate heterogeneity among older households: a puzzle?
 2.6k Downloads
 1 Citations
Abstract
We put forward a method for estimating discount rates using wealth and income data. We build consumption from these data using the budget constraint. Consumption transitions yield discount rates by household groups. Applying this technique to a sample of older households, we find a similar distribution to those previously estimated using field data, though with a much lower mean than those found using experiments. Surprisingly, among this older population, patience is negatively correlated with education and numeracy. This goes against the positive correlation found for younger populations in experiments and some field studies. We discuss potential explanations for this result.
Keywords
Time preference Discount rate ConsumptionJEL Classifications
D12 D31 D91 E211 Introduction
In many situations, individuals make decisions that involve a comparison of present and future circumstances. They must decide how much to invest in education, how much to save for retirement, how much to invest in health, etc. In each case, these decisions are based on some assessment of the potential welfare at different periods under different scenarios. Following Samuelson (1937), economists have largely adopted a discountedutility model which assumes that preferences over time can be condensed into one major parameter, the geometric discount rate (see Frederick et al. 2002 for a critical review and Hall 2010 for a review of recent research developed using this approach).
To estimate discount rates, both field data and experiments are found in the literature. Experimental studies are by far the most numerous. Among the 42 studies surveyed by Frederick et al. (2002), 34 use experimental methods. A typical approach is for individuals to be offered a menu of (real or hypothetical) choices between a quantity of money now and a different quantity of money at some point in the future. Respondents’ choices are used to estimate a discount rate.
Our paper fits into a much smaller literature that estimates discount rates using field data on aspects of behaviour and a lifecycle model of consumption and saving. A typical way to estimate preference parameters in such models, though not the one that we will take, has been to solve numerically the intertemporal optimisation problem that the agents in a particular population are assumed to face. Estimates of parameters such as the discount rate are chosen such that the model’s predictions are close, in some metric and according to some data, to those seen in reality. Such studies vary in the extent to which heterogeneity in the discount factor is admitted into the model. Some papers assume homogenous discounting behaviour, like French (2005) and Edwards (2013) where discounting is exponential and Laibson et al. (2007) where discounting is quasihyperbolic. More flexibility was allowed by Attanasio et al. (1999) who estimate a version of the lifecycle model where the discount rate varies stochastically with the composition of the household while even more is allowed by Samwick (1998) and Gustman and Steinmeier (2005) who estimate a different discount rate for every household.
These papers fully specify a lifecycle model and solve it. The method we employ does not do this but rather uses the firstorder condition to that solution—the Euler equation. We first generate longitudinal observations on consumption using a procedure introduced by Ziliak (1998) and Browning and LethPetersen (2003). This involves calculating consumption using comprehensive and high quality data on assets and income and the intertemporal budget constraint. Our resulting distribution of consumption is shown to be remarkably similar to that derived from the UK’s household budget survey. Using the Euler equation and consumption transitions at the household level, we estimate average discount rates for groups of households. Such an approach has typically been precluded in the past by the absence of good quality panel data on consumption—a problem discussed in detail by Browning et al. (2003).
Our approach has some parallels with papers that have previously relied on the Euler equation to estimate parameters, in particular the elasticity of intertemporal substitution. Estimation in this manner was carried out by Campbell and Mankiw (1989) and Attanasio and Weber (1995) among others. For a lively criticism of this approach, see Carroll (2001) and for a defense see Attanasio and Low (2004). Our approach differs from these papers in three principal ways. First, we are able to use consumption transitions at the household level rather than relying on aggregate or cohortlevel data. Second, our use of household rather than cohort level consumption data allows us to use the exact Euler equation in our estimation, rather than relying on Taylor series approximations. Third, we do not assume that the discount rate is the same for each individual in our sample, nor do we assume that it is unchanging across the lifecycle.
We apply the procedure outlined above to a representative sample of older English households using the English Longitudinal Survey of Ageing (ELSA). We show, unsurprisingly, that there is substantial heterogeneity in discounting in that population. The typical levels of discount rates that we estimate are of similar magnitude to those estimated in other papers based on the lifecycle model of consumption and saving. These rates imply substantially less discounting than is implied by the results of experimental studies.
Our most surprising result is that discount rates tend to rise with education and levels of numerical ability (i.e. those with less education and those who are less numerically able tend to be the most patient). This result is contrary to that found in the literature that measures the extent to which individuals discount future income streams (see for instance Warner and Pleeter 2001; Harrison et al. 2002; Dohmen et al. 2010). These papers differ in their empirical approach—the first uses data on the choices of departing military personnel over whether they will take their severance payment in a lumpsum or in the form of an annuity payment, while the second and third papers use laboratory experiments. The literature using field data offers less conclusive evidence. Gourinchas and Parker (2002) solve a lifecycle model and estimate discount rates by matching simulated consumption data to those observed in the data at different ages and, similar to us, finds that people with more education are less patient. On the contrary, Cagetti (2003), implementing a broadly similar procedure as that used by Gourinchas and Parker, but using data on assets instead of consumption, finds evidence suggesting the more educated are more patient. This is also found by Lawrance (1991), using an approach based on a loglinearised Euler equation and data on transitions in food expenditure.
We discuss our somewhat puzzling result below and raise the possibility that prior evidence, driven largely by choices individuals made at younger ages is not applicable to the discounting between periods at older ages.
The rest of this paper is structured as follows. Our empirical approach is outlined in Section 2. In Section 3, we describe the data and explain how we calculate consumption from assets and income. Results are presented in Section 4 and discussed in Section 5. Section 6 concludes.
2 Theory and empirical approach
 (i)$$ p_{\tau} c_{i\tau} + {\sum}_{j} p^{j}_{(\tau+1)} X^{j}_{i(\tau+1)} = e_{i\tau} + d_{i\tau}+ {\sum}_{j} r^{j}_{\tau} p^{j}_{\tau} X^{j}_{i\tau} + {\sum}_{j} p^{j}_{\tau} X^{j}_{i\tau} \;\; \; \forall \;\tau $$(1)
 (ii)$$ X^{j}_{i(\tau+1)} \ge b^{j}_{i(\tau+1)} \; \;\; \forall \; \tau, j $$(2)
All features of the model other than the instantaneous utility function are allowed to vary freely over time—in particular, we do not assume that the discount rate is timeinvariant.
This equation forms the basis for our empirical approach. We group households according to particular characteristics (such as their education level, numerical ability, age and marital status) and estimate the expectation in the equation above by using the sample average of the quantity in square brackets among households of a particular group. In using the sample average to estimate the expectation term, we need to assume that there are no differential shocks across households that lead to systematic differences in the term \(\frac {p_{t}}{p_{t+1}} \left (\frac {c_{t}}{c_{t+1}}\right )^{\gamma }\).
It is worth pointing out how the change in consumption (the central observable quantity that enters the Euler equation) identifies the discount rate. The faster is consumption growth, all else being equal, the lower is the discount rate (that is, the more patient is the individual concerned). Households who are patient tend to forsake current consumption for future consumption—and therefore exhibit consumption growth. The converse is also true. Those who are impatient tend to prefer current consumption to future consumption. They therefore have lower (or even negative) consumption growth.
We cannot identify the coefficient of relative risk aversion (γ in Eq. 6) and we assume that is does not vary across individuals. The assumption of a coefficient of relative risk aversion that does not vary across households is a strong one (Outreville 2015 surveys the empirical literature which has shown differences in risk aversion exist by education group). However, our data does not contain sufficient individual variation in interest rates to warrant identification. We set γ equal to 1.25. This is consistent with the range of elasticities of intertemporal substitution estimated by Attanasio and Weber (1993) on UK data and very close to that obtained by Gustman and Steinmeier (2005). We have generated results assuming alternative values of γ. While the mean discount rates are sensitive to these values, the ranking of households’ discount rates is the same for all positive choices of γ that are constant across households.
Finally, it is worth making explicit two restrictions implicit in the use of Eq. 6. These relate to liquidity constraints and changes in the utility function due to changing household composition or changing labour supply.
First, recall that the Euler Eq. 6 only holds at equality when individuals are not liquidity constrained. If we use Eq. 6 to estimate the discount rate for a group containing liquidity constrained individuals, the estimate will be biased downwards. Concerns about the presence of liquidity constraints in our case are mitigated by the fact that we work on a population over the age of 50, at a point in the lifecycle where most have accumulated some liquid wealth—almost 95 % of our sample have positive gross liquid asset holdings. As a check that our results are not being driven by liquidity constraints, we have confirmed that there are no substantial changes in focusing only on those with liquid assets above a certain minimum level (see Section 5).
Second, when individuals leave or join a household between periods, the assumption of a constant instantaneous utility function at dates t and t+1 does not make sense; we therefore do not include households whose composition changes between waves of data. Further, as evidence points to changes in consumption patterns around retirement (Banks et al. 1998; Wakabayashi 2008), we exclude from our sample households where some member left the labour market during the period covered by our data.
The previous two paragraphs have outlined two exclusions from our estimating sample. Some further exclusions are necessary due to the fact that we are not able to calculate consumption satisfactorily for every household in our sample. The extent of these exclusions is outlined in the A. To deal with the fact that those omitted are unlikely to be a random subsample of our overall sample, we generate weights representing the probability of each household being included in our sample and, in our results we attach a weight to each household of the inverse of this probability. These probabilities are estimated as functions of marital status, education, age, income quintile and wealth quintile. Our results will, therefore, be representative of the entire population aged 50 and over if the selection into our sample can be adequately modelled as a function of these characteristics.
3 Data
Our data come from the English Longitudinal Study of Ageing (ELSA). ELSA is a panel survey that is representative of the English population aged 50 and over. It started in 2002, and individuals have been reinterviewed every 2 years since then—our main results use data from the first three waves. The purpose and form of the survey is similar to the Health and Retirement Study (HRS) in the US and the Survey of Health, Ageing and Retirement in Europe (SHARE) in 20 European countries. The first wave was conducted between April 2002 and March 2003 and sampled 12,099 individuals (of whom 11,391 were core sample members; the remainder was individuals aged under 50 who were the partners of core sample members). There are 7894 benefit units (i.e. a single person or couple along with any dependent children) where each member of the couple is a sample member. Our sample is drawn from these benefit units.
While ELSA contains questions on some components of expenditure (food, domestic fuel and clothing) it does not, unfortunately, contain data on total expenditure, which, approximating consumption, is needed to estimate the discount rate. In fact, there is no nationally representative longitudinal survey that collects total expenditure in the UK and such data is rare internationally.^{2} This lack of comprehensive longitudinal data on expenditure has proved something of an obstacle to bringing Euler equations to data. The literature has either relied on aggregate data, or, following Browning et al. (1985), has used repeated crosssections to form a quasipanel of birth cohortlevel average expenditure. An alternative approach was suggested by Skinner (1987) and refined by Blundell et al. (2008). It involves estimating the relationship between food (and possibly other items of) expenditure and total expenditure using a household budget survey. As generalpurpose panel surveys often contain data on food expenditure, this estimated relationship can be used to impute total expenditure.
We proceed in a different manner: following Ziliak (1998) and Browning and LethPetersen (2003), we use the rich data on income and assets that is contained in ELSA to back out expenditure from the intertemporal budget constraint. In all the following, we will equate total expenditure (excluding mortgage repayments) with consumption. The rest of this section summarises this procedure—further details are given in the A—and shows a close correspondence between features of the resulting distribution of consumption with those that are obtained using the UK’s Household Budget Survey.
3.1 Calculating consumption using longitudinal data on assets and income

Flow variables representing consumption c _{ t }, noncapital income e _{ t }, transfers d _{ t } and the asset yield, including any capital gain or loss, \({r^{j}_{t}}\), are measured over the entire two year period;

Stock variables \({X^{j}_{t}}\) represent holdings of assets at the beginning of the period (i.e. at the time of the first interview); \(X^{j}_{t+1}\) represents asset holdings at the beginning of period t+1 or equivalently at the end of period t (i.e. at the time of the next interview);

Asset prices \({p^{j}_{t}}\) and \(p^{j}_{t+1}\) represent asset prices at the time of the first and second interview;

The overall price level p _{ t } represents the average price level in the period between the two interviews.
Equation 7 is the equation that we use to calculate consumption between two waves of the survey. Having calculated consumption in this manner, we make one further adjustment and subtract mortgage repayments (both capital components and interest) from the resulting quantity. While these represent cash expenditure on housing, they are not generally indicative of consumption of the flow of housing services.
Some, but not all of the quantities on the righthandside of Eq. 8 can be directly read from the ELSA data. In each wave the value of each asset held (p ^{ j } X ^{ j }) is recorded, as are noncapital income (e), capital income (q ^{ j } p ^{ j } X ^{ j }) and lumpsum transfers (d) in the period prior to the interview (where ‘period’ in the case of most forms of income and transfers represents 12 months). ELSA respondents are asked for details on 16 different financial asset and (nonmortgage) debt instruments.^{3} These include various type of savings products, bond holdings and equity holdings. See Section A in the Appendix for more detail and Table 10 in that section for summary statistics on holdings in each asset.
Our data does not record capital gains on assets held between the two waves (\(\frac {p^{j}_{t+1}  {p^{j}_{t}}}{{p^{j}_{t}}}\)), nor does it contain data on income and transfers for a period of approximately one year (recall that ELSA sample members are surveyed approximately every two years)  two objects that appear in Eq. 8. The majority of assets held by the population in our sample are in safe forms—so there is no capital gain to be considered for these assets. For equity holdings, we assume a capital gain (or loss) in line with the change in the FTSE index between the two interview dates. Estimating income in the missing year is facilitated by exploiting the longitudinal aspect of the survey data—we interpolate linearly between income in year y and income in year y+2 to obtain income in year y+1. Finally, we assume that there are no lumpsum transfers in the missing year and we exclude from our sample those households where it is likely that some member received a lumpsum transfer (due to retirement, redundancy or the death of a spouse or parent). We give further details about all of these assumptions and their implications in the A.
3.2 Comparing consumption in ELSA and in the EFS
In this section, we compare the distribution of consumption estimated in the manner described above with that estimated using the Expenditure and Food Survey (EFS).^{4} The EFS is the UK’s household budget survey, and is used to calculate the commodityweights of the UK’s inflation indices (the Consumer Prices Index and the Retail Prices Index). The data is collected annually, throughout the year and the survey is designed to be nationallyrepresentative. Respondents are asked to record all purchases over a 2week period in a diary and also to complete a questionnaire that seeks information on infrequentlypurchased items. The combination of the diary and the questionnaire allows a comprehensive measure of consumption to be calculated.
Figure 2 shows that there is a close correspondence between the distributions in both shape and location. The correspondence is closest at the bottom of the distribution (i.e. up to annual consumption of £10,000). At this point, the distributions diverge somewhat—with the distribution of consumption in ELSA lying to the right of that in the EFS. This divergence, which represents a tendency for consumption to be greater in ELSA than the EFS in the upper half of the distribution, is consistent with the fact that consumption in the EFS (grossed up to national levels) is known to underrecord the level of consumption calculated as part of the National Accounts with the degree of underreporting thought to be greater for those who have higher levels of consumption (see Brewer and O’Dea 2012).
3.3 Summary statistics
Summary statistics
(A)  (B)  

All  Our  All  Our  
Households  sample  Households  sample  
Age groups (props.)  Annual income (£1000s)  
50–59  34.45  34.13  Mean  17.01  17.14 
60–69  26.78  25.29  p10  5.08  5.14 
70–79  23.84  23.95  p25  7.58  7.70 
80 +  14.93  16.63  p50  12.69  13.30 
Total  100.00  100.00  p75  21.10  21.58 
p90  32.36  31.09  
Education (props.)  Net liquid wealth (£1000s)  
Low  54.70  54.65  Mean  73.28  45.00 
Middle  28.57  28.24  p10  0.00  0.00 
High  16.66  17.10  p25  1.61  2.00 
Total  100.00  100.00  p50  13.50  11.52 
p75  61.55  51.60  
p90  169.00  121.36  
Numerical Ability (props.)  Net housing wealth (£1000s)  
1(Lowest)  11.89  9.73  Mean  114.11  117.03 
2  41.22  40.97  p10  0.00  0.00 
3  29.95  32.53  p25  0.00  20.00 
4(Highest)  15.47  16.07  p50  90.00  95.00 
Missing  1.47  0.70  p75  161.00  175.00 
Total  100.00  100.00  p90  250.00  260.00 
Marital status (props.)  Sample size  
Single  7.19  8.42  7,894  1,504  
Marr/cohab  52.08  54.32  
Widowed  24.12  26.23  
Sep/Div  10.48  11.03  
Other  6.15  0.00 
4 Results
This quantity would be equal to the discount rate if c _{ t+1} and p _{ t+1} were perfectly forecasted by households at date t. Our presentation of the distribution of this quantity which we refer to below as the ‘expost’ discount rate is a useful preliminary step.
Our use of four waves of ELSA data gives us up to three observations on consumption for each household and therefore up to two observations on the expost discount rate. Figure 3 shows two distributions of the expost discount rates (trimming the bottom 10 % and top 10 % of the sample). The median discount rate is approximately −3 % in the earlier period and is 0 % in the later period. These median expost discount rates are low relative to estimates of the discount rate found in the literature that estimate such rates using field data and very low relative to those found in the experimental literature.
Figure 3 shows substantial heterogeneity around these medians. This depicts the distribution of the discount rate, perturbed by two phenomena. First, realisations of stochastic variables will differ from their expectations. Second, our consumption data is likely to include some measurement error. Figure 3 also shows the distribution of the geometric mean of our two successive observations on the discount rate. The variance of this distribution is substantially smaller than the variance of either crosssectional distribution. This could be due to some combination of averaging over time of the discount rate for each family and to a diminished effect of measurement error once we take a time average.
The most striking difference between our results and those of Samwick and GS is the substantial number of households that we find with expost discount rates of less than 5 %. We find approximately 60 % here in this portion of the distribution compared to approximately 40 % in the distribution of discount rates in the two papers based on the US allage population. We find a larger share of households with negative expost discount rates (approximately 50 % of the sample). This compares to approximately 10 % of those in Samwick’s sample (see his Fig. 3; note that GS do not report the proportion with negative discount rates).
In our two crosssectional distributions, we find a similar share of households in the right tail of the distribution (those with a discount rate greater than 15 %) as do Samwick and GS, and find less mass in the region of 5 to 15 %. On our average measure, the mass in the left tail of our distribution increases, largely at the expense, relative to either crosssectional distribution, of that in the right tail.
The models of both Samwick and GS assume a discount rate for a particular household that does not change over time. If discount rates do vary over time, their estimates represent some average of the lifetime sequence of discount rates. Therefore, the large mass that we find in the left tail of the distribution could be reconciled with the estimates of Samwick and GS if households have higher discount rates at younger ages than at older ages (at which point they enter our population of interest).
Median expost discount rate by household characteristics
Age  \(\hat {\rho }\)  σ  Marital status  \(\hat {\rho }\)  σ  Education  \(\hat {\rho }\)  σ  Numerical ability  \(\hat {\rho }\)  σ 

50–59  −2.2  2.4  Single  0.1  5.3  Low  −3.2  1.0  1 (Low)  −2.9  2.0 
60–69  −4.6  1.9  Married  −2.1  2.1  Mid.  −1.8  2.1  2  −3.2  1.1 
70–79  −2.5  1.4  Widowed  −3.1  1.7  High  6.5  5.6  3  −0.8  2.7 
80 +  0.5  2.5  Sep./Div.  −4.8  2.1  4 (High)  −1.3  4.0  
All  −2.3  1.0  All  −2.3  1.0  All  −2.3  1.0  All  −2.3  1.0 
We use a grouping estimator to estimate the average exante discount rate for groups defined by age, marital status, levels of education and numerical ability. The estimator is based on Eq. 6. It estimates the expectation in that equation by its sample analogue for a particular group and weights the results to account for possibly nonrandom selection into our sample. For each group, we trim the sample, removing those in the first and tenth decile of consumption growth. Unlike Alan and Browning (2010), we do not explicitly account for the role of measurement error. However, our approach does not involve assuming (as does the vast majority of work in this area) that preference parameters remain the same over the whole of the lifecycle.
Mean exante discount rate by household characteristics
Age  \(\hat {\rho }\)  σ  Marital status  \(\hat {\rho }\)  σ  Education  \(\bar {\rho }\)  σ  Numerical ability  \(\bar {\rho }\)  σ 

50–59  −0.0  1.4  Single  1.2  2.2  Low  −2.6  0.7  1 (Low)  −2.9  1.4 
60–69  −1.8  1.2  Married  2.0  1.2  Mid.  0.8  1.4  2  −2.4  0.8 
70–79  −0.7  1.1  Widowed  −3.2  1.0  High  6.7  3.0  3  1.6  1.5 
80 +  −0.5  1.7  Sep/Div  −3.8  1.2  4 (High)  2.3  2.4  
All  −0.9  0.6  All  −0.9  0.6  All  −0.9  0.6  All  −1.0  0.6 
The magnitude of the differences between the education groups is large.^{8} For example, the difference between the point estimates of the mean discount rate for the ‘low’ education group and those in the ‘high’ education group is over 9 percentage points. This means that, over this period and in this population, if those with the most education are to exhibit the same saving behaviour at the margin as those with the least, the former group will require a safe return of 9 percentage points greater than the latter group.
Mean exante discount rate for groups defined by pairwise combinations of both education and numerical ability
Low education  Med./High education  

Low numeracy  −3.1  1.0 
(0.8)  (1.8)  
High numeracy  −1.5  4.6 
(1.5)  (1.9) 
Sensitivity of mean exante discount rate for groups defined according to education
Education  (1)  (2)  (3)  

\(\bar {\rho }\)  σ  \(\bar {\rho }\)  σ  \(\bar {\rho }\)  σ  
Low  −1.0  0.9  −1.5  0.9  −0.6  1.3 
Mid.  3.9  1.8  2.7  1.8  3.0  2.6 
High  10.5  3.6  3.9  4.2  −0.3  6.8 
All  2.1  0.8  −0.5  0.8  −0.3  1.2 
We might want to confirm that our relatively small sample, combined with our trimming of the largest 10 % increases in consumption and largest 10 % decreases in consumption does not materially affect our results. To investigate this, we solve and simulate behaviour from a simple lifecycle model. We specify a lifecycle model where agents make a consumption and saving choice every year from the age of 20 to 100. In each period until the age of 65 they receive income which follows an autoregressive process with autocorrelation of 0.95 and variance of 0.02. We specify that there are equal numbers of ten types of agents, each with a different discount rate. Discount rates range from −2.5 to 7.5 % with a mean of 2.5 %.
After solving (using a backwards recursion) for the set of consumption functions, we carry out the following procedure 499 times. First, we draw a sample of 200 agents (20 from each discount rate type). Second, we simulate their consumption behaviour using the calculated consumption functions. Third, we take one consumption transition between two sequential years for each individual. The age for the transition is chosen as a random age between 50 and 80; this age differs for each individual. We then trim the largest and smallest 10 % of consumption transitions and estimate a discount rate (just as we do in our estimation). The mean (across 499 simulations) estimated discount rate is 2.37 % with a 95 % confidence interval of [2.16 %, 2.59 %] which contains the truth. This simulation reassures us that estimation of the discount rate in sample sizes of the type that we have is possible with reasonable precision.
5 A puzzling result?
Our results are puzzling in light of the studies, particularly those using experimental designs, that find that low educated individuals tend to lack patience compared to higher educated. A primary difference between this paper and most of the rest of the literature is the fact that results in the latter come from samples of individuals who tend to be much younger than ours. Work using lifecycle models (e.g. Samwick 1998; Gustman and Steinmeier 2005) typically focuses on working age individuals while the experimental literature often uses samples of students. In contrast, our sample comprises older households in England, aged 50 and above. Frederick et al. (2002) noted at that time that ‘no studies [had] been conducted to permit any conclusions about the temporal stability of time preferences’ and, while Bishai (2004) does investigate how time preference changes over the age range 14 to 37, we are not aware of any studies that look at the discounting behaviour of the oldest households. We now briefly consider some other explanation for our results.
First, consider survival probabilities. Households in the model discount the future for two reasons—first their ‘pure’ rate of time preference, and second their expectation of being alive at each period in the future. Differential survival probabilities could explain our result on education in the following manner. Suppose that all education groups had the same mean ‘pure’ discount rate, but that one group had a longer lifeexpectancy. This group would appear to be more patient. A negative correlation between education and life expectancy could explain, at least some of, our results. In our data, however, it is those with lower education levels that tend to have shorter life expectancies. ELSA respondents are asked the following question: ‘What are the chances that you will live to be X or more?’, where X depends on their current age. We run a simple linear regression of the responses to this question on dummies for our education groups and age dummies (the latter to account for the possibility that those in different education groups are distributed differently across ages). We find that those in the middle education consider themselves to be 3.1 percentage points more likely to live to the age referenced in the question than those in the low education group, and those in the high education group are 5.4 percentage points more likely. Differential life expectancies, therefore, would seem to work in the opposite direction from our puzzle.
Second, if it were the case that those with more education had access to higher pretax safe rates of return, then our assumption that everyone has the same pretax rate of return would generate a downward bias in the estimates of discount rates for the more educated relative to the less. That those with more education might face (through greater financial literacy) higher rates of return is plausible. This would render our result a conservative one—and the actual gap between the discount rate of those with less and more education could be greater than that which we find.
Mean exante discount rate for groups defined according to education and level of positive gross liquid assets
Education  Level of positive liquid gross assets  

Base  ≥ 2,500  ≥ 5,000  ≥ 10,000  
\(\bar {\rho }\)  σ  \(\bar {\rho }\)  σ  \(\bar {\rho }\)  σ  \(\bar {\rho }\)  σ  
Low  −2.6  0.7  −5.2  1.0  −5.3  1.3  −5.2  1.8 
Mid.  0.8  1.4  −0.2  1.8  −0.4  2.1  −1.9  2.4 
High  6.7  3.0  8.0  3.4  8.7  3.4  6.9  3.6 
All  −0.9  0.6  −2.4  0.9  −1.0  1.1  −1.2  1.4 
Mean exante discount rate between two different set of years
Education  2002/04–2004/06  2004/06–2006/08  

\(\bar {\rho }\)  σ  \(\bar {\rho }\)  σ  
Low  −2.6  0.7  0.7  0.7 
Mid.  0.8  1.4  −0.2  1.3 
High  6.7  3.0  12.4  2.5 
All  −0.9  0.6  2.0  0.6 
Discount rates and housing price growth
(1)  (2)  (3)  

Mid Education  1.33  −0.04  −0.06 
High Education  9.62***  9.87**  9.87** 
Housing wealth effect  −0.02  
Constant  −3.16***  −3.41*  −3.42* 
Observations  1,503  986  986 
As a final comment, we acknowledge the role that measurement error could play in our results. Note that classical measurement error, for example with a constant variance error that is multiplicative with consumption, will affect the level of the estimated discount rates but not affect the relative position of the groups. One needs a nonstandard type of measurement error (for instance multiplicative with higher variance among those with more education relative to those with less) to undo our results.
6 Conclusion
This paper puts forward a method for estimating individual discount rates using field data. We build consumption panel data from the intertemporal budget constraint and panel data on income and wealth. This householdlevel panel data is used, with the Euler equation, to estimate discount rates for groups defined by socioeconomic characteristics.
We show, unsurprisingly, that there is substantial heterogeneity in discounting in our sample which is drawn from a population of older households. But surprisingly, we find that, among this older population, households with less education and lower numerical ability exhibit greater patience than, respectively, those with more education and greater numerical ability. This result, which is robust to differential housing wealth shocks, differential mortality and differential incidence of liquidity constraints is somewhat puzzling as it is the opposite to that found in investigations of time preference for younger households.
Footnotes
 1.
Alternatively, we could specify a liquidity constraint that ensures that total debts are no greater than a certain quantity.
 2.
A notable exception to this is the Spanish ‘Encuesta continua de presupuesto familiares’, a diarybased longitudinal survey of expenditure. The Panel Study of Income and Dynamics in the US, since 1999, has collected expenditure data that covers approximately 70 % of average total expenditure.
 3.
There are also questions on housing wealth, physical wealth and pension wealth.
 4.
This survey has, since 2008, been known as the Living Costs and Food Survey. However, the data that we show is from years prior to this, so we make use of the older name.
 5.
Those who did not complete UK school leaving exams (typically taken at the age of 18) have ‘low’ education, those who completed these but have no thirdlevel education are in the ‘middle’ education group, and those with any postschoolqualification are in the ‘high’ education group. Numerical ability is tested using six questions in ELSA—our categorisation follows that in Banks et al. (2010). Marital status is defined covering the first four waves of ELSA  those whose status changed are in the ‘other’ group. The age, education and numerical ability of the couple are taken as that of the older/more educated/more numericallyable of the couples. More details on the education and numerical ability characteristics are given in Appendix A.
 6.
These results are not directly comparable with those in Alan and Browning (2010) where the distributions are presented graphically. Additionally, that paper discusses how their results relate to those in Samwick (1998). Alan and Browning (2010) restrict the discount rate to be greater than 0. We find, as do GS and Samwick, evidence of households with negative discount rates.
 7.
While we refer to (the absence of) a relationship with age, we are not able to separately identify associations of discount rates with each of age and cohort. Exploiting a longer panel than we have access to would allow us to separately identify age and cohort effects if we assumed that there were no time effects in discount rates.
 8.
The differences are also statistically significant. The difference between the mean discount rates of the low and middle education groups are significant at the 5 % level and those between the high and low education group are significant at the 1 % level.
 9.
There are also questions on housing wealth, physical wealth and pension wealth.
 10.
This imputation procedure, carried out by the ELSA team is called a conditional hotdeck. Given that we know the, say, cash holdings of a particular household are between £a and £b—that individual is assigned a random draw from the empirical asset distribution of those who report their assets exactly as between £a and £b and have the same characteristics along some dimensions: here it is age and household composition.
 11.
A benefit unit is a single adult or couple along with any dependent children that they have. Relatively few members of our sample have dependent children, so our results can be thought of as representing single adults and couples.
 12.
These are largely bank savings bonds (i.e. effectively fixed term savings accounts) rather than government or corporate bonds.
 13.
One potential reason for this is the different methods by which the data on food spending is gathered. The ELSA data is taken from respondents’ responses to a question that asks how much they spend on food in a typical month. The EFS data is taken from respondents’ spending diary entries.
Notes
Acknowledgments
Cofunding is acknowledged from the European Research Council (reference ERC2010AdG269440—WSCWTBDS) and the Economic and Social Research Council (Centre for Microeconomic Analysis of Public Policy reference RES544285001). Thanks to James Banks, Richard Blundell, James Browne, Rowena Crawford, Mariacristina De Nardi, two anonymous referees and to seminar participants at the Institute for Fiscal Studies, the Paris School of Economics and the Economic and Social Research Institute of Ireland for valuable comments. Any errors are our own.
Compliance with ethical standards
Funding
This study was funded by the European Research Council (reference ERC2010AdG269440 – WSCWTBDS) and the Economic and Social Research Council (Centre for Microeconomic Analysis of Public Policy reference RES544285001).
Conflict of interest
The authors declare that they have no conflict of interest.
References
 Alan S, Browning M (2010) Estimating intertemporal allocation parameters using synthetic residual estimation. Rev Econ Stud 77(4):1231–1261CrossRefGoogle Scholar
 Attanasio O, Banks J, Meghir C, Weber G (1999) Humps and bumps in lifetime consumption. J Bus Econ Stat 17(1):22–35Google Scholar
 Attanasio OP, Low H (2004) Estimating euler equations. Rev Econ Dyn 7 (2):405–435CrossRefGoogle Scholar
 Attanasio OP, Weber G (1993) Consumption growth, the interest rate and aggregation. Rev Econ Stud 60(3):631–649CrossRefGoogle Scholar
 Attanasio OP, Weber G (1995) Is consumption growth consistent with intertemporal optimization? Evidence from the consumer expenditure survey. J Polit Econ 103(6):1121–1157CrossRefGoogle Scholar
 Banks J, Blundell R, Tanner S (1998) Is there a retirementsavings puzzle?. Am Econ Rev 88(4):769–788Google Scholar
 Banks J, O’Dea C, Oldfield Z (2010) Cognitive function, numeracy and retirement saving trajectories. Econ J 120(548):F381–F410CrossRefGoogle Scholar
 Bishai DM (2004) Does time preference change with age?. J Popul Econ 17 (4):583–602CrossRefGoogle Scholar
 Blundell R, Pistaferri L, Preston I (2008) Consumption inequality and partial insurance. Am Econ Rev 98(5):1887–1921CrossRefGoogle Scholar
 Brewer M, O’Dea C (2012) Measuring living standards with income and consumption: evidence from the UK. Ifs working paper 12/12Google Scholar
 Browning M, Crossley TF, Weber G (2003) Asking consumption questions in general purpose surveys. Econ J 113(491):F540–F567CrossRefGoogle Scholar
 Browning M, Deaton A, Irish M (1985) A profitable approach to labor supply and commodity demands over the lifecycle. Econometrica 53(3):503–544CrossRefGoogle Scholar
 Browning M, LethPetersen S (2003) Imputing consumption from income and wealth information. Econ J 113(488):F282–F301CrossRefGoogle Scholar
 Cagetti M (2003) Wealth accumulation over the life cycle and precautionary savings. J Bus Econ Stat 21(3):339–53CrossRefGoogle Scholar
 Campbell JY (2000) Asset pricing at the millennium. J Financ 55(4):1515–1567CrossRefGoogle Scholar
 Campbell JY, Mankiw NG (1989) Consumption, income and interest rates: Reinterpreting the time series evidence. In: NBER macroeconomics annual 1989, volume 4, NBER chapters. National Bureau of Economic Research, Inc., pp 185–246Google Scholar
 Carroll CD (2001) Death to the loglinearized consumption euler equation! (and very poor health to the secondorder approximation). The B.E Journal of Macroeconomics 0(1):6Google Scholar
 Dohmen T, Falk A, Huffman D, Sunde U (2010) Are risk aversion and impatience related to cognitive ability?. Am Econ Rev 100(3):1238–1260CrossRefGoogle Scholar
 Edwards RD (2013) The cost of uncertain life span. J Popul Econ 26(4):1485–1522CrossRefGoogle Scholar
 Frederick S, Loewenstein G, O’Donoghue T (2002) Time discounting and time preference: a critical review. J Econ Lit 40(2):351–401CrossRefGoogle Scholar
 French E (2005) The effects of health, wealth, and wages on labour supply and retirement behaviour. Rev Econ Stud 72(2):395–427CrossRefGoogle Scholar
 Gourinchas P.O., Parker JA (2002) Consumption over the life cycle. Econometrica 70(1):47–89CrossRefGoogle Scholar
 Gustman AL, Steinmeier TL (2005) The social security early entitlement age in a structural model of retirement and wealth. J Public Econ 89(23):441–463CrossRefGoogle Scholar
 Hall RE (2010) Fowardlooking decision making. Princeton University PressGoogle Scholar
 Harrison GW, Lau MI, Williams MB (2002) Estimating individual discount rates in Denmark: a field experiment. Am Econ Rev 92(5):1606–1617CrossRefGoogle Scholar
 Laibson DI, Repetto A, Tobacman J (2007) Estimating discount functions with consumption choices over the lifecycle. NBER Working Papers 13314, National Bureau of Economic Research, Inc.Google Scholar
 Lawrance EC (1991) Poverty and the rate of time preference: evidence from panel data. J Polit Econ 99(1):54–77CrossRefGoogle Scholar
 ONS (2009) Wealth in great britain main results from the wealth and assets survey 2006/08. Technical report Office for National StatisticsGoogle Scholar
 Outreville JF (2015) The relationship between relative risk aversion and the level of education: a survey and implications for the demand for life insurance. J Econ Surv 29(1):97–111CrossRefGoogle Scholar
 Samuelson P (1937) A note on measurement of utility. Rev Econ Stud 4:155–161CrossRefGoogle Scholar
 Samwick AA (1998) Discount rate heterogeneity and social security reform. J Dev Econ 57(1):117–146CrossRefGoogle Scholar
 Skinner J (1987) A superior measure of consumption from the panel study of income dynamics. Econ Lett 23(2):213–216CrossRefGoogle Scholar
 Wakabayashi M (2008) The retirement consumption puzzle in Japan. J Popul Econ 21(4):983–1005CrossRefGoogle Scholar
 Warner JT, Pleeter S (2001) The personal discount rate: evidence from military downsizing programs. Am Econ Rev 91(1):33–53CrossRefGoogle Scholar
 Ziliak JP (1998) Does the choice of consumption measure matter? An application to the permanentincome hypothesis. J Monet Econ 41(1):201–216CrossRefGoogle Scholar