Asset allocation and age effects in retirement savings choices
The authors gratefully acknowledge the support of HESTA, GESB, UniSuper, and STA in the conduct of this research. We would also like to thank Jacqui Whale for her excellent work on four complex databases and extracting the data for the analysis. We also thank the very helpful comments of two anonymous referees, Rose Kushmeider and other participants of the 2007 Washington Area Finance Conference, and Don Ross and other participants at the 2007 Annual Meeting of the Academy of Financial Services.
Abstract
We examine the asset allocation decisions of members of three large Australian retirement savings funds. Superannuation Guarantee legislation in 1992 made Australian employees compulsory investors by requiring employers to contribute a fixed proportion of earnings to a superannuation fund on behalf of employees. A majority of these employees can choose an investment strategy for these contributions. We examine how actual investment strategy and asset allocation choices of members change with age in view of the conventional wisdom that individuals allocate less to risky assets as they age and investments theory which provides conflicting advice on the issue.
1. Introduction
Ageing populations and the forecast increased demands on government age pensions (Guest and McDonald, 2000) have made the accumulation of retirement savings by individuals a significant policy issue in the developed world. In Australia, a central component of the policy response has been to boost mandated employer-funded retirement savings contributions on behalf of employees. Widespread employer contributions emerged in 1986 for employees covered through the industrial bargaining system and became compulsory for all employees1 in 1992 with the introduction of the Superannuation Guarantee. The assets controlled by superannuation funds in Australia are growing rapidly, from $228.3 billion in June 1995 to $1.05 trillion in December 2008 (Australian Prudential Regulation Authority, 2009), even accounting for the 15 per cent drop in assets in 2008 from the global financial crisis. The major change in aggregate asset allocations by superannuation funds over the past two decades has been an increase in equity exposure at the expense of short- and long-term interest bearing securities and property, from 32 per cent of total assets invested domestically in 1988 to 52 per cent in December 2008.2
Members of these funds are assuming greater responsibility in choosing the investment strategy applied to their savings with a shift from defined benefit to defined contribution plans and an increase in the investment choices available (Gerrans et al., 2006). A majority of Australian workers face two choices about what happens with their superannuation guarantee contribution, distinguished by who offers the choice. The first is offered by the employer and is the choice of the superannuation fund to which the employee wishes her superannuation contributions be directed. With the passage of the Superannuation Legislation Amendment (Choice of Superannuation Funds) Act 2004, this choice became mandatory for 5.2 million of a potential 9.5 million employees who previously did not have this choice (Clare, 2005).3 The second is offered by the superannuation fund and is the choice of investment strategy for contributions or accumulated balances. This offer of investment strategy choice is not mandatory but superannuation funds offer these choices as an integral selling feature of their products. This paper focuses on the second choice, and specifically how the investment strategy and asset allocation selected for future contributions may change with member age. Common advice from funds and advisers is to reduce exposure to growth assets towards retirement. The following is typical:
If you are approaching retirement, and you are accessing your super over time, you should consider investing in asset classes such as shares and property securities, to increase your potential for growth. If you are close to retirement and need to withdraw all your super, then a low risk, low return might be appropriate (Vanguard Investments, 2005, p. 12–13).
Economic and investment theory has identified the conditions necessary for asset allocations to be independent of age (Samuelson, 1969; Merton, 1969; Mossin, 1968) given assumptions concerning market completeness, return distributions, individual utility functions and absence of labour income. Examinations of individual or household actual asset allocations rarely have access to complete datasets of all aspects of wealth, a feature previously noted:
A striking feature of much of this literature is that it pays scant attention to the most important non-human assets available to individuals or households approaching retirement – housing and social security. We believe that little of relevance can be said about asset allocation unless these are included in the analysis (Iwaisako et al., 2004, p.4).
However, the implicit assumption that individuals consider each of these forms of savings as part of a total portfolio is contestable (Thaler, 1980). Even if we agree with the general message of Iwaisako et al., given the limited Australian evidence on asset allocation within superannuation by age we believe we can make a contribution to the literature with access to a very large database of member investment choices.
Three not-for-profit superannuation funds have allowed access to their membership data to enable examination of investment strategy decisions. The three funds have combined assets of $23.7 billion and 1.3 million members. Fund 1 was one of the first industry funds to offer choice to its members in 1995. Choice was introduced to Fund 2 members in July 1997. Fund 3 members were first offered choice in April 2001.
The next section of this paper reviews the literature related to the influence of age on the formulation of an investment strategy and asset allocation. Section 3 examines the current level of investment choice offerings in Australian superannuation funds. Section 4 reviews the data for the three funds used in the analysis. Section 5 provides analysis of investment choice and equity asset allocation by age and the final section concludes with future analysis suggestions for the data.
2. Literature and hypotheses
2.1. Investment asset allocation and age: theory
Samuelson (1989) summarises the development of views regarding individual portfolio asset allocation with age noting that the conventional or folk wisdom suggests that investors should be more risk tolerant when young and progressively decrease exposure to relatively risky equities in favour of lower risk cash and fixed interest securities. Samuelson’s (1969) early work demonstrated that a rational wealth maximiser with constant relative risk aversion will keep a constant proportion in equities over time. Samuelson (1991) subsequently demonstrated that the conventional wisdom could be supported if the assumption of a random walk for securities is replaced with mean reversion and negative serial correlation, and this held even when the assumption of risk aversion was retained.
Samuelson (1994) argues that a desire for a minimum level of retirement wealth will imply an optimal investment strategy of declining equity allocation with age. McNaughton et al. (1999) argue that this view is mistaken, based on the exclusion of the differential return to risky and safe assets, and suggest that, in fact, an increasing equity allocation with age is more appropriate.
Bodie et al. (1992) and Bodie (2003) suggest two reasons why the proportion of risky assets should theoretically decline with age. The first flows from the fact that when we are young, typically a large proportion of our total wealth is in human capital. Further, this human capital is usually less risky than equity. With a large proportion in lower risk human capital, to get more risk into the individual’s total wealth portfolio the proportion invested in equity is increased. With age the proportion that human capital comprises of total wealth (human capital plus financial capital) declines. Therefore, the individual decreases their proportion of equity held. The second reason relates to the flexibility of their labour supply. That is, the ability to work longer hours or take a second job. If we accept that this flexibility is greater when younger and subsequently declines, the proportion of risky assets will be greatest when younger and subsequently decline. Bodie (2003) also notes though that the opposite can be true. If you are a young entrepreneur with a large proportion of risky capital, it may be optimal to have a lower proportion of risk equity and to increase the proportion with age.
Other reasons suggested for asset allocation varying with age include changing risk aversion with age (Ballente and Green, 2004) and information costs (Haliassos and Bertaut, 1995). A Delphi study of financial educators and financial planners by Greninger et al. (2000) reports a relatively weak consensus of opinion on asset allocation with 60 per cent agreeing that the proportion of conservative assets should be increased closer to retirement and 20 per cent indicating that it was never prudent to do this.
The theoretical literature, while mixed, provides a more consistent view that equity allocation will decline with age. It is muted, however, on the age relationship with other individual asset classes, including property, cash and fixed interest. This gives rise to the first two related hypotheses:
H1: Individual risky asset class allocations (equity and property) decline consistently with age.
H2: Individual conservative asset class allocations (fixed interest and cash) increase consistently with age.
2.2. Investment and portfolio asset allocation and age: empirical evidence
Guiso et al. (2000) provide an international perspective (US, UK, Netherlands, Germany and Italy) on age and the decision to invest in equity or not (participation decision) and, conditioned on owning equity, how much to allocate to equity and other asset classes (allocation decision). In terms of the participation decision, Guiso et al. (2000) find a humped profile of equity versus age. On the allocation decision, Guiso et al. (2000) argue that with the exception of the Netherlands, equity allocation over the age profile is generally flat after controlling for other variables. Iwaisako et al. (2004) use Japanese data and report a positive age impact on equity participation, flattening at the highest age group, for different household types. They also find that while equity portfolio allocation increases with income, proportionally more property is held. Agnew et al. (2003) find a decreasing trend in equity allocation of 93 basis points each year using a linear specification and a peak allocation at age 32.5 years using a non-linear specification. Their sample of 7000 401(k) accounts came from one plan between 1995 and 1998. Agnew et al. (2003, p. 203) note that as their sample covers ‘only a period of roughly 4 years, it is more natural to estimate and interpret models with age and time effects only’ and thus exclude cohort effects.
Ameriks and Zeldes (2004) examine equity allocation by age using a sample of 16000 TIAA-CREF 403(b)4 accounts between 1987 and 1996, and five waves of the Survey of Consumer Finances (SCFs) between 1983 and 1998. It is notable that the analysis of the SCFs data is based on stock of assets and therefore change in allocations may be attributable to relative performance of asset classes. The TIAA-CREF data permit analysis of future contributions allocations which are not influenced by asset class performance in themselves. Ameriks and Zeldes (2004) estimate conditional and unconditional equity allocations to explore identification problems with time, cohort and age effects. They suggest that estimation of an age relationship must include time effects and they reject their results including age and cohort effects only as implausible.5 Overall, they conclude that households do not decrease equity shares with age. This result is stronger when they examine equity allocations conditioned on the account having an equity allocation. However, they also find both hump-shaped profiles and declining equity allocations by age, sensitive to year, using the TIAA-CREF contributions data.6 The hump effect they suggest is due to the probability of owning equity rather than equity allocation differences. Guiso et al. (2000) make a similar point in a summary of international evidence suggesting that the variation observed in allocation to risky assets is because of variation in proportion of households holding risky assets as conditional allocations appear much more stable. McCarthy (2004) also finds strong cohort effects with a household in 2000 holding more risky assets than a comparable household in 1980 though the cause is open to interpretation. Agnew et al. (2003) suggest that it is more appropriate to model the decision of equity ownership and allocation jointly as the ‘same variables determine whether to hold equities and how much equities to hold’ (Agnew et al., 2003) which is the approach adopted in this paper.
The Australian Stock Exchange (2004) suggest direct ownership of equity (equity participation) increased from 52 per cent to 55 per cent between 2000 and 2004 though they suggest that the allocation to asset classes has not changed significantly between 2002 and 2004. The Reserve Bank of Australia (2003), utilising data from the 2002 HILDA survey, report a humped profile of the proportion of households owning equity by household age. For those aged 16–34, the proportion is 29 per cent rising monotonically to 53 per cent for those aged 55–64 and declining to 34 per cent for households aged 75 or more. The median equity ownership by age also has a humped profile rising from 7 per cent for households with a 16–34 year old reference person to 48 per cent for 65–74 year old before declining to 30 per cent for households with a reference person 75 or more. This oldest age group, which exhibits a decline in both equity ownership and equity allocation, is largely absent in the current sample which is focussed on workers making superannuation contributions.
In summary, the empirical literature identifies support for non-monotonic patterns in equity allocation, with several suggesting a peak in allocations followed by a decline with age. This suggests an alternative age related hypothesis:
H3: Equity allocation exhibits a humped shape age profile.
2.3. Non-age related factors and asset allocation: controls
Empirical evidence and theory suggest other characteristics that may impact on individual asset allocations which should be controlled for when investigating the role of age. Income and wealth have shown to be positively related to equity allocation (Guiso et al., 2000; Agnew et al., 2003; Ameriks and Zeldes, 2004; Calvet et al., 2008). Males have generally been shown to allocate more to risky assets (Bajtelsmit et al., 1999; Bernasek and Shwiff, 2001; Gerrans and Clark-Murphy, 2004). Ameriks and Zeldes (2004) highlight the need to control for time effects when examining allocations by age. The impact of the rate of return achieved by an individual also requires control. Thaler and Johnston (1990) suggest a ‘house-money effect’ where prior gains (losses) increase (decrease) the level of risk taking or allocation to risky assets.
3. Superannuation investment strategy choice in Australia
The level of investment strategy choice offered in Australia varies by category of fund. Overall 62 per cent of funds offer investment strategy choices although if this is measured by the assets of funds offering choice the figure is 93 per cent. The proportion of funds offering choice is lowest for Corporate funds (53 per cent of funds or 93 per cent of assets) and highest for Industry funds (86 per cent of funds or 93 per cent of assets). The average number of choices in funds offering choice is 45 with retail funds offering 112 choices on average and corporate funds seven choices (Australian Prudential Regulation Authority, 2009).
Although increased opportunities for choice have been made available, the majority of members do not exercise choice and remain in the default strategy of the fund. The total assets in the default option of funds across the superannuation industry ranges from 40 per cent for retail funds to 70 per cent for Industry funds. This is not a perfect measure of member activity or degree of engagement in superannuation as the default option may be a member’s ‘choice’ and they therefore do not need to change investment strategy, and second, it does not identify the proportion of members or member accounts. Fry et al. (2007) suggest loss aversion as an explanation for the lack of fund choice in Australia, which may similarly be used to explain the lack of investment choice by members.
4. Data description
Each of the funds analysed in this paper offer members the choice of readymade options, with a specified investment strategy, or a do-it-yourself (DIY) strategy where members choose their own investment strategy. The three funds’ investment options are summarised in Table 1.
| Fund 1 | Fund 2 | Fund 3 | ||
|---|---|---|---|---|
| Asset classes | Equity | International sharesAustralian shares | International sharesAustralian sharesAustralian sustainable sharesInternational sustainable shares | International equityAustralian equity |
| Property | Property | Listed property | Property | |
| Fixed interest securities | Fixed interest | Australian fixed interestInternational fixed interest | Fixed interest securitiesInflation linked bonds | |
| Cash | Cash | Cash | Cash | |
| Other | InfrastructurePrivate equityAbsolute return strategies | |||
| Readymade options | Capital guaranteed | Cash plan | ||
| Cash plus | Low risk plan | Conservative plan | ||
| Core pool | Balanced plan | Balanced plan | ||
| Shares plus | Shares plus | Growth plan | ||
| Eco pool | ||||
| Overseas shares pool | ||||
| Australian shares pool |
Funds 1 and 2 allow members to choose both individual asset classes and readymade options in a DIY mix, whereas Fund 3 only allows selection of asset classes in a DIY choice. The asset classes and readymade options offered by each fund can be regarded more as generic labels rather than offering a definitive guide to the asset class exposure each contains. The default investment option in each fund, for those members not making a choice, is comparable in asset allocation to the aggregate asset allocations discussed in the Introduction. The main asset class allocation is equity which varies in allocation between 54 per cent and 57 per cent. Property varies between 8 per cent and 12 per cent, cash between 2 per cent and 4 per cent. Fixed interest allocation is notably different in one fund with an allocation of 25 per cent, whereas the other two, it is 10 per cent. The remainder asset class is ‘other’ which varies between 5 per cent and 20 per cent. The ‘Other’ category includes infrastructure, private equity and absolute return strategies.
A database of all member choices, as choice was introduced by the funds, is available. Default behaviour is particularly evident in each fund and the following section summarises choice activity for each of the funds. Fund 1 commenced choice in July 1995 and until December 2006, 70 613 members made 81 274 changes. Fund 2 introduced choice in July 1997 and up to December 2006, 24 354 members made 32 059 changes. Fund 3 members have faced the same set of investment choices since choice inception and since the introduction of choice in July 2001 until December 2006, 23 528 members made a total of 39 047 investment strategy changes.
However, the complete menu of options for members has only been available since 2001/2002. A breakdown of member choices over this period is presented in Table 2. In Fund 1, between July 2001 and December 2006, 66 490 members made changes with 20 548 of these electing the DIY option. From January 2002 to December 2006, 24 354 Fund 2 members made 32 059 changes. In Fund 3, 23 528 members made 39 047 changes. These statistics relate to actual choices and do not include the initial default allocation. The analysis presented in the next section focuses on the DIY asset class choices which have complete member demographic, balances and contributions data.
| Changes | ||||||||
|---|---|---|---|---|---|---|---|---|
| Records | Members | Records | Members | Records | Members | |||
| Fund 1 | Fund 2 | Fund 3 | ||||||
| Panel A: Total member choices | ||||||||
| July 2001–December 2006 | January 2002–December 2006 | July 2001–December 2006 | ||||||
| DIY | 22 966 | 20 548 | DIY | 8063 | 6519 | DIY | 18 646 | 6363 |
| Pooled/mixture | 52 487 | 47 809 | Pooled/mixture | 23 996 | 18 976 | Pooled/mixture | 20 401 | 18 567 |
| Total | 75 453 | 66 490 | Total | 32 059 | 24 354 | Total | 39 047 | 23 528 |
| Panel B: Sample with complete data July 2001–December 2006 | January 2002–December 2006 | July 2001–December 2006 | ||||||
| DIY | 13 712 | 12 135 | DIY | 4645 | 3562 | DIY | 10 940 | 4103 |
| Pooled/ mixture | 34 003 | 30 565 | Pooled/mixture | 13 963 | 10 666 | Pooled/mixture | 12 507 | 11 429 |
| Total | 47 715 | 41 365 | Total | 18 608 | 13 464 | Total | 23 447 | 14 673 |
- This table summarises the number of accounts and unique members within each fund who made an investment choice over the analysis period. DIY refers to members who constructed their own asset allocations from available asset classes. Pooled/mixture refers to choices which involved fund readymade options, that is where the strategic asset allocation is nominated such as ‘growth’ or ‘balanced’. Panel A summarises all changes that were made and Panel B summarises the sample of data with complete data records which are utilised in the paper for analysis. DIY, do-it-yourself.
5. Results
5.1. Analysis of asset allocations by key demographics
Preliminary univariate analysis of the data examined the distribution of mean asset allocations by: fund and year; age, balance and contributions quintiles; and gender. To assist analysis the asset classes, presented in Table 3, are grouped into aggressive (equity and property) and conservative (cash and fixed interest) asset classes. Two tests comparing means are conducted. The first, reported in the Table 3, are t-tests of equality of mean asset class allocation between each variable group and a base group. The second, not reported separately here, is a one-way anovaF-test of equality of mean asset class allocations across all groupings of a variable, conducted separately for each fund. Several results emerge.
| n | Aggressive | Conservative | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Equity | Property | Fixed interest | Cash | |||||||||
| Mean | SD | Med. | Mean | SD | Med. | Mean | SD | Med. | Mean | SD | ||
| Age | ||||||||||||
| <30 | 5950 | 45.50** | 37.00 | 50 | 19.14** | 23.81 | 10 | 12.92** | 22.35 | 0 | 17.07* | 31.37 |
| 30–37 | 5989 | 52.64 | 37.83 | 55 | 20.52 | 26.02 | 10 | 11.53 | 21.49 | 0 | 11.12 | 25.67 |
| 38–45 | 6431 | 50.63 | 36.92 | 50 | 21.65 | 26.17 | 15 | 12.22 | 21.68 | 0 | 10.79 | 24.74 |
| 46–53 | 6649 | 48.81 | 37.45 | 50 | 20.84** | 26.32 | 10 | 11.89** | 20.85 | 0 | 14.00* | 28.73 |
| ≥53 | 4278 | 45.03 | 39.17 | 50 | 18.86 | 26.36 | 10 | 13.55 | 24.07 | 0 | 18.43 | 33.89 |
| Year | ||||||||||||
| 2001 | 1674 | 71.95 | 27.92 | 80 | 13.82 | 17.61 | 10 | 8.85 | 14.66 | 0 | 4.01 | 14.37 |
| 2002 | 2262 | 42.51 | 33.67 | 40 | 23.15 | 23.29 | 20 | 16.93 | 23.81 | 10 | 11.71 | 23.03 |
| 2003 | 5228 | 32.23 | 32.89 | 25 | 27.19 | 26.33 | 20 | 21.93 | 26.51 | 15 | 13.42 | 24.11 |
| 2004 | 4534 | 46.76 | 35.29 | 50 | 20.08 | 24.17 | 12 | 15.23 | 23.34 | 0 | 10.83 | 23.58 |
| 2005 | 6743 | 53.82** | 36.88 | 55 | 19.27 | 26.03 | 10 | 9.48 | 20.13 | 0 | 12.98 | 29.28 |
| 2006 | 8856 | 52.93 | 40.47 | 60 | 17.67 | 26.98 | 0 | 6.87 | 17.52 | 0 | 19.14 | 35.32 |
| Gender | ||||||||||||
| Male | 13 646 | 54.06 | 38.42 | 60 | 19.25 | 26.70 | 10 | 11.15 | 21.82 | 0 | 12.59 | 28.49 |
| Female | 15 651 | 44.16 | 36.41 | 50 | 21.24 | 24.87 | 15 | 13.38 | 22.05 | 0 | 15.18 | 29.12 |
| Contributions | ||||||||||||
| <1290 | 5860 | 43.93 | 37.89 | 45 | 20.38 | 25.54 | 10 | 13.31 | 23.36 | 0 | 16.18 | 30.56 |
| 1290–2699 | 5860 | 43.65 | 37.13 | 45 | 20.78 | 25.71 | 12 | 13.40 | 22.65 | 0 | 16.34 | 30.32 |
| 2690–3977 | 5860 | 46.96 | 37.56 | 50 | 21.06 | 26.36 | 12 | 13.01 | 22.57 | 0 | 14.49 | 29.02 |
| 3978–5824 | 5859 | 54.46** | 36.37 | 60 | 20.89 | 25.23 | 15 | 11.66 | 20.71 | 0 | 9.91 | 24.05 |
| ≥5824 | 5858 | 54.86 | 37.83 | 60 | 18.46 | 25.86 | 10 | 10.32 | 20.23 | 0 | 12.94 | 29.35 |
| Balance | ||||||||||||
| <5923 | 5860 | 39.85** | 37.23 | 40 | 21.31 | 26.77 | 10 | 13.78 | 24.64 | 0 | 17.93 | 32.20 |
| 5923–18 070 | 5860 | 43.98 | 35.92 | 45 | 21.66 | 24.73 | 15 | 14.32 | 22.38 | 0 | 13.70 | 26.22 |
| 18 071–43 255 | 5859 | 50.65 | 36.57 | 50 | 20.66 | 24.89 | 15 | 12.04 | 20.18 | 0 | 12.07 | 26.13 |
| 43 256–95 119 | 5859 | 56.58 | 35.86 | 60 | 20.13 | 24.54 | 12 | 12.07 | 21.60 | 0 | 8.34 | 21.85 |
| ≥95 120 | 5859 | 52.78 | 40.23 | 60 | 17.81 | 27.54 | 0 | 9.50 | 20.42 | 0 | 17.83 | 34.83 |
| Fund | ||||||||||||
| 1 | 13 712 | 39.09 | 34.39 | 40 | 22.59 | 25.22 | 20 | 14.90 | 23.46 | 0 | 13.59 | 26.73 |
| 2 | 4645 | 57.75** | 38.97 | 65 | 16.47 | 22.93 | 8 | 15.26 | 25.96 | 0 | 10.52 | 22.82 |
| 3 | 10 940 | 57.09 | 38.15 | 65 | 19.10 | 27.24 | 10 | 7.90 | 16.88 | 0 | 15.92 | 33.22 |
| Total | 29 297 | 48.77 | 37.68 | 50 | 20.31 | 25.76 | 10 | 12.34 | 21.97 | 0 | 13.97 | 28.86 |
- This table provides summary asset allocation information for DIY choices used in the regression sample in Table 5. The summary statistics relate only to records where complete data is available. Tests of differences in means are conducted in each category relative to the underlined and italicised group. All differences are significant at the 99% confidence level, except those marked with * which are significant at the 95% confidence level and those marked with ** which are not significant. Median values are not reported for Cash as it is 0 for every group. All figures shown are percentages. DIY, do-it-yourself; SD, standard deviation.
In previous studies, income and wealth have generally been shown to have a positive relationship with the proportion of aggressive assets held. The present database does not have direct information on either variable though a proxy for income is the employer contributions paid to the fund on behalf of employees and a proxy for wealth is the member fund balance. The F-tests generally reject the equality of means across each variable for each asset class within each fund. This conclusion is supported when a comparison of means is made between each grouping and the nominated comparison group of a variable. With few exceptions, mean asset allocations are significantly7 different. The only exception to this is allocation to property and fixed interest across balance and contributions quintiles for Fund 2. The t-Tests of mean asset allocations are significantly different by gender across all asset classes with the largest difference in equity allocations. Males (females) have a higher (lower) mean allocation to equity and a lower (higher) allocation to cash and fixed interest securities. The pattern for property is reversed, however, with females having a significantly higher allocation.
5.2. Analysis of asset allocation by age
An examination of the allocation to conservative asset classes indicates a generally positive age8 trend with some evidence of a U-shaped age profile for cash. F-tests reject equality of means across all age groupings for each asset class. However, t-tests do not reject a difference in mean allocations for equity, property, and fixed interest between the youngest and oldest age groups. Similarly, no significant difference is supported for property and fixed interest for the two oldest age quintiles. This provides preliminary evidence rejecting hypothesis two. Allocations to fixed interest and cash do not decline consistently with age.
A further breakdown of mean age asset allocations, by gender and fund provides some additional striking results for cash allocations. Fund 3 female members across each age quintile have much lower cash allocations, whereas the youngest female Fund 1 members have extremely high cash allocations in each year. This suggests possible fund specific effects, such as how investment options are described and presented to members or, may influence allocations. Alternatively, further member characteristics such as job tenure or employment industry may also be important. Further analysis is warranted of this group of members but this is beyond the scope, and data availability, of this paper.
Examination of the aggressive asset classes indicates mixed patterns by age. Although allocations to property do not vary by more than 2.8 per cent age points across age groups, the F-test rejects equality of means. The F-test also rejects equality of allocations by age for equity and suggests preliminary evidence of a humped age profile. This preliminary evidence rejects hypothesis one. Allocations to equity and fixed interest do not increase consistently with age. Further, the data lend support to hypothesis three of a humped age profile for equity.
It is interesting to examine the relative allocations by the two highest age quintiles. Arguably, it should be the second oldest age quintile, which across the funds is those aged 46–53 years, where members become more engaged in their retirement planning and interested in their superannuation asset allocation decisions. Examining conservative (aggressive) allocations in the two oldest age quintiles across the three funds, in each year indicates that in 14 of the 17 instances, the oldest quintile has a higher (lower) allocation to conservative (aggressive) assets than the next younger quintile.
5.3. Regression estimation
To further investigate the relationship between asset allocations and age, a regression was estimated to relate the asset class allocations within member DIY choices to a range of explanatory variables. A preliminary summary of the range of allocations made by members is presented in Table 4.
| Fund 1 | Fund 2 | Fund 3 | Total | Fund 1 | Fund 2 | Fund 3 | Total | |
|---|---|---|---|---|---|---|---|---|
| Property | Equity | |||||||
| 0 | 34.62 | 48.25 | 46.53 | 41.23 | 30.86 | 18.26 | 22.01 | 25.56 |
| 1–19 | 14.58 | 13.78 | 15.73 | 14.88 | 3.27 | 8.27 | 1.91 | 3.55 |
| 20–39 | 26.84 | 23.83 | 19.19 | 23.50 | 14.24 | 4.97 | 6.27 | 9.79 |
| 40–59 | 16.58 | 9.52 | 9.15 | 12.68 | 21.60 | 14.57 | 13.93 | 17.62 |
| 60–79 | 2.33 | 1.01 | 2.67 | 2.25 | 12.86 | 12.59 | 17.19 | 14.43 |
| 80–99 | 0.67 | 0.39 | 0.97 | 0.74 | 5.67 | 8.57 | 10.60 | 7.97 |
| 100 | 4.38 | 3.23 | 5.77 | 4.72 | 11.51 | 32.77 | 28.08 | 21.07 |
| Sub-total | 13 712 | 4 645 | 10 940 | 29 297 | 13 712 | 4 645 | 10 940 | 29 297 |
| Cash | Fixed interest | |||||||
| 0 | 61.17 | 69.39 | 64.69 | 63.78 | 53.86 | 58.73 | 69.42 | 60.44 |
| 1–19 | 14.71 | 12.64 | 16.12 | 14.91 | 13.27 | 13.82 | 13.52 | 13.45 |
| 20–39 | 11.73 | 4.67 | 3.41 | 7.50 | 18.84 | 12.81 | 11.17 | 15.02 |
| 40–59 | 4.94 | 9.11 | 2.18 | 4.57 | 8.93 | 6.61 | 3.31 | 6.46 |
| 60–79 | 0.63 | 0.30 | 1.07 | 0.74 | 1.15 | 2.39 | 1.08 | 1.32 |
| 80–99 | 0.31 | 0.06 | 0.96 | 0.51 | 0.31 | 0.69 | 0.44 | 0.42 |
| 100 | 6.53 | 3.83 | 11.57 | 7.98 | 3.63 | 4.95 | 1.06 | 2.88 |
| Sub-total | 13 712 | 4 645 | 10 940 | 29 297 | 13 712 | 4 645 | 10 940 | 29 297 |
- This table provides a breakdown of effective asset allocations within selected allocation bands for members making DIY choices. The weightings are effective asset allocations reflecting the strategic allocation indicated by the fund in their annual reports to each investment option and the member’s investment option allocation.
The table clearly identifies that censoring of the data is present as members are only allowed positive asset class allocations to a maximum of 100%. Of the total 29 297 choices made, a large proportion for each asset class has allocations of zero. The highest is 64 per cent for cash and the lowest 26 per cent for equity. Similarly, a large proportion (21 per cent) has a 100 per cent allocation to equity with 8 per cent having the same allocation to cash. As has been identified, previous work has generally approached this in two ways. The first is by separating the member’s decision to allocate to equity, using a probit or logit binary selection model, from the decision of how much to allocate to equity using the standard Tobit regression (e.g. Ameriks and Zeldes, 2004, although they employ an OLS regression). The second approach models the decisions jointly (e.g. Agnew et al., 2003). We follow the joint modelling decision with the intention of exploring the separate decision modelling in further research. The sensitivity of equity to age has been estimated using a generalisation of the censored least absolute deviations (CLAD) estimator developed by Powell (1984) to accommodate the data censoring. A Tobit regression was also estimated, to investigate sensitivity of results to estimator choice, but is not reported because although the results in this instance are qualitatively the same the CLAD-estimator is robust to non-normality and heteroscedasticity, whereas the Tobit estimator is consistent only when all distributions are normal and homoscedastic (Huberman and Jiang, 2006). The CLAD estimator’s standard errors are calculated using bootstrap resampling using 500 random samples, applied after convergence.9
, which is the unobserved unrestricted preferred allocation, whereas the data available is 
the observed or restricted allocation P of member i at time t. The relationship between the observed and latent variable is specified in (1):
()5.4. Regression results
The parameter estimates from regressions of asset allocations are presented in Table 5. The allocation to equity is significantly higher, and the allocations to cash significantly lower, for males. Property allocation is significantly less for males but it is the cash allocation which is pronounced. Income, as proxied by contributions (Logconts), has a significant positive relationship with property allocations but is not significant for the other three asset classes. Wealth, as proxied by accumulated balance (Logbalance), has a positive (negative) relationship with equity and fixed interest (property and cash) allocations.
| Equity | Property | Fixed interesta | Casha | |||||
|---|---|---|---|---|---|---|---|---|
| Age (base >53) | ||||||||
| <30 | 11.9638** | −0.9553 | −10.9200** | 0.1964 | ||||
| (1.3661) | (0.7992) | (1.7755) | (1.627) | |||||
| 30–37 | 15.7332** | 0.9135 | −11.5277** | −4.1221* | ||||
| (1.2162) | (0.6819) | (1.7177) | (1.7649) | |||||
| 38–45 | 11.7818** | 1.6732* | −9.1714** | −4.4296* | ||||
| (1.2542) | (0.7143) | (1.7928) | (1.8111) | |||||
| 46–53 | 6.6654** | 1.8285** | −5.5823** | −4.1620* | ||||
| (1.2105) | (0.6572) | (1.7217) | (1.683) | |||||
| Age | 1.5740** | 0.9452** | −0.3076 | −1.6349** | ||||
| (0.2345) | (0.0941) | (0.2583) | (0.3007) | |||||
| Age2 | −0.0242** | −0.0111** | 0.0083* | 0.0206** | ||||
| (0.0029) | (0.0012) | (0.0034) | (0.0038) | |||||
| Male | 3.2624** | 3.3922** | −1.8503** | −1.6795** | −1.6133 | −1.2898 | −31.0594** | −4.4452** |
| (0.6969) | (0.684) | (0.4342) | (0.4430) | (0.9889) | (0.9712) | (7.3569) | (1.0458) | |
| Logconts | 0.1935 | −0.7887 | 2.281** | 3.7711** | 0.9465 | 0.8022 | −0.6418 | 0.0451 |
| (0.9031) | (0.9657) | (0.5137) | (0.5162) | (1.0109) | (0.8612) | (0.9409) | (1.0169) | |
| Logbalance | 7.7256** | 8.1001** | −2.0892** | −5.4716** | 3.8447** | 3.8253** | −4.4799** | −5.3664** |
| (0.7886) | (0.8192) | (0.5047) | (0.4311) | (1.0834) | (0.9897) | (1.0793) | (1.0641) | |
| Year (base 2001) | ||||||||
| 2002 | −25.9774** | −25.6743** | 8.2662** | 9.1217** | −11.7783 | 8.4095** | 3.9464 | −35.7603 |
| (1.4766) | (1.586) | (0.7393) | (0.8401) | (9.4591) | (3.228) | (2.7199) | (23.6634) | |
| 2003 | −38.6348** | −38.6151** | 11.6118** | 14.2943** | 0.0115 | 20.8314** | 7.7257** | −31.1254 |
| (1.3002) | (1.3279) | (0.6752) | (0.8163) | (9.4975) | (3.0613) | (2.7654) | (23.6197) | |
| 2004 | −17.1145** | −17.3427** | 6.5185** | 6.8949** | −1.7109 | −33.8553 | ||
| (1.4989) | (1.4953) | (0.8655) | (0.8673) | (2.9712) | (23.0176) | |||
| 2005 | −9.3849** | −9.6225** | 0.4873 | 4.8190** | −18.7167 | 8.8103 | −20.5971** | |
| (1.3437) | (1.5175) | (0.6145) | (0.918) | (15.2905) | (7.9700) | (7.5532) | ||
| 2006 | −8.2380** | −7.9929** | −8.0127** | −0.4306 | 1.5506 | −49.9571* | ||
| (1.3305) | (1.3876) | (1.4602) | (0.8539) | (3.3906) | (24.2112) | |||
| Old6mthret (%) | −0.5295** | −0.4720** | −0.0856 | −0.2028** | −1.0717** | −1.1466** | −0.0692 | −0.3095* |
| (0.1336) | (0.1351) | (0.0524) | (0.0753) | (0.1293) | (0.1148) | (0.1189) | (0.102) | |
| Fund (base Fund 3) | ||||||||
| Fund 1 | −9.4188** | −9.1268** | 0.9029 | 0.7771 | −3.5907 | −3.2240* | 6.9512** | 7.8951** |
| (0.7963) | (0.9054) | (0.5322) | (0.5097) | (1.4276) | (1.3232) | (1.3503) | (1.1407) | |
| Fund 2 | 8.9257** | 9.2805** | −8.843** | −4.0517** | −3.5607** | −4.4848* | 25.0008** | −3.8615 |
| (1.1243) | (1.1845) | (0.6985) | (0.8375) | (2.2189) | (2.1700) | (6.7104) | (3.6146) | |
| Constant | 27.5649** | 17.646** | 11.807 | 7.1017** | 8.9361* | −21.4669** | 18.9003** | 86.7351** |
| (3.6165) | (5.2916) | (2.1787) | (2.612) | (9.9758) | (5.8426) | (6.0157) | (−1.6349) | |
| Pseudo R2 | 0.0757 | 0.0763 | 0.0427 | 0.0281 | 0.0322 | 0.0334 | 0.0174 | 0.0130 |
| Censored (=0) | 7487 | 12 077 | 17 707 | 18 686 | Censored (=0) | 7487 | 12 077 | 17 707 |
| Uncensored | 15 638 | 15 838 | 10 746 | 8272 | Uncensored | 15 638 | 15 838 | 10 746 |
| Censored (=100) | 6172 | 1382 | 844 | 2339 | Censored (=100) | 6172 | 1382 | 844 |
| Total | 29 297 | 29 297 | 29 297 | 29 297 | Total | 29 297 | 29 297 | 29 297 |
- *95 percent confidence interval; **significant at the 99 percent confidence level. aSelected year dummies dropped because of co-linearity. This table presents the results of censored quantile regressions of asset class allocations in DIY superannuation investment choices. The values reported are marginal effects of the latent variable, that is the unconstrained allocation to each asset class. These are calculated at the mean values for continuous variables (log of contributions (Logconts), log of accumulated balance (Logbalance), performance of the old investment choice over the previous 6 months (Old6mth)). For categorical variables, the marginal effects are relative to the oldest age quintile, Females, 2001 and Fund 3 respectively. An additional regression is estimated for each asset class to investigate non-linearity, including age as a continuous variable and age-squared. The pseudo R2 value is measured using 1 − (sum of weighted deviations about estimated quantile)/(sum of weighted deviations about raw quantile) (Jolliffe et al., 2000). Standard errors shown in brackets are bootstrap estimates from 500 re-sampled replications. The Cash regressions are relative to the sixtieth percentile rather than median because of non-convergence.
In terms of aggregate allocations to aggressive and conservative asset classes, these results are somewhat mixed. For allocation to equity, the positive relationship aligns with previous evidence on gender and wealth discussed in Section 2. There is little, however, to compare the results relating to cash, fixed interest and property allocations, particularly in retirement savings decisions, as much of the asset allocation literature focuses on allocation to an aggregate of ‘risky’ assets or equity. The results suggest that relying on aggregated ‘aggressive’ and ‘conservative’ classifications can mask impacts at the individual asset class level. For example, while equity allocation may increase with wealth it may be at the expense of property exposure. Similarly, the reduction in cash allocations is offset by increasing allocation to fixed interest.
Asset allocation is significantly different by fund membership though it differs by specific asset class. Significant time effects are also present though this depends on asset class. The allocation to equity is significantly less in each year relative to 2001 which may reflect the poor performance of Australian equity, and international equity more notably, classes in the second half of 2001 through to the end of 2002 relative to other asset classes. In contrast, allocation to property is higher in each year with the exception of 2006. Cash and fixed interest allocations are significantly different though no consistent relationship emerges with the relationship depends on year and specification.
The impact of age has been investigated using quintiles with the reference group being members older than 53, the oldest member group, and in a second specification using age and age-squared to further explore a non-linear relationship. The combined effect of the coefficient for age and age-squared suggests a humped equity and age profile with allocations peaking at 34. This humped relationship is confirmed when examining the quintile coefficients. Each younger age quintile has a higher allocation than the oldest quintile with the allocation highest for the 30–37 years group relative to members aged more than 53. This evidence is comparable with Agnew et al. (2003) who report similar results with a peak allocation marginally earlier at 32.5 years. These results implicitly assume no cohort effects. This is a surprising result given that at that age the member’s investment horizon to retirement is over 30 years with a further 20 years life expectancy.
Allocation to property also suggests a humped profile, if somewhat flatter. The two youngest age quintiles allocation is not significantly different to the oldest quintile whereas the third has a larger allocation, peaking with the second eldest quintile, 46–53 years. The age and age-squared specification suggest allocation to property peaks at age 43. Viewed with the results for equity, this again reinforces an early decline for exposure to aggressive asset classes. The regression and univariate results provide strong evidence to reject hypothesis one and fails to reject hypothesis two for equity. The evidence for property is less compelling and rejects hypothesis one. Property allocations do show some evidence of an age-related peak but the evidence more supports a flat age relationship.
Allocation to fixed interest largely increases with age with either specification. Using the quintile specification, there is a marginal drop in allocation for the second youngest age group. This therefore supports hypothesis two. Allocations to cash, however, reject hypothesis two as the results suggest a ‘U-shaped’ profile with the lowest allocation in the 38–45 quintile, or age 44 as suggested by the alternate specification. The surprising result for cash is that there is no significant difference in allocations between the youngest and oldest quintile. This result does not appear driven by a particular fund or gender as mean allocations broken down by fund and gender support the result.
6. Conclusions and further analysis
It is acknowledged that in examining asset allocation by age, it is ideal to have information about the individual’s ‘non-human assets’ (Iwaisako et al., 2004) and investment in human capital as well as information on partner wealth, neither of which is possible in this paper. However, superannuation is a specific long-term investment vehicle of increasing importance to Australian employees and existing analysis of the behaviour of a large number of members from different funds is limited.
The results suggest that the fund itself may have an impact on asset class allocation levels. Large differences are evident between funds suggesting that possible information or framing effects may be important. Regression results suggest that member characteristics of gender, age and wealth proxy can in combination explain significant differences in asset allocation as well as common effects including the year the decision was made and the fund the member belonged to. There is support for a humped shape age-profile for aggressive asset class allocations and a ‘U-shaped’ age-profile for defensive asset class allocations. This, however, masks more pronounced individual asset class relationship.
Property allocations appear least sensitive to age, peaking at age 43. Allocation to equity is significantly related to age with the allocation increasing until mid-thirties and then declining. A shift away from equity when the superannuation investment horizon remains close to 30 years is surprising though consistent with US research. Fixed interest allocations generally increase with age. Allocations to cash suggest a surprising age relationship with its ‘U-shaped’ profile. The lowest age quintile has comparable allocations to the oldest with a bottoming of allocations in the middle age quintile of 38–46, bottoming at age 44.
Taken together, it could be suggested that asset class allocation adjustment by age is achieved at the extreme of the risk classes, equity and cash. The relative risks of these asset classes are perhaps most clearly understood. However, the high-cash allocation of the lowest age group remains puzzling and warrants further investigation. More work is also needed to explore the impact of member characteristics including which number decision it is that a member is making, an issue unexplored in the literature. That is, do the allocations differ for the first, second, third, etc. decisions rather than calendar time and is this consistent by age? Gerrans et al. (2006) identify an apparent different nature to the first decision a member makes and any further changes in respect of the possible influence of historical performance of asset classes. Further analysis will also focus on better controlling member characteristics and pay more attention to the distribution of allocations to asset classes, in particular, the prevalence of extreme allocations, i.e. 0 and 100 per cent allocations.
