Monday, April 20, 2015

Spending Typically Declines as We Age

The most common assumption of retirement spending strategies is that real (inflation-adjusted) spending from savings will be flat throughout retirement, yet most studies of actual retiree household expenditures show that constant real spending is atypical. For most retirees, expenditures decline pretty consistently as we age.

Two of my favorite studies on this topic are David Blanchett's Estimating the True Cost of Retirement (2013, PDF) and Sudipto Banerjee's Expenditure Patterns of Older Americans, 2001-2009 (2012, PDF). The results of the studies are quite similar – not surprising since they used the same databases – but each provides unique information.

Blanchett christened his findings the "retirement spending smile", though be forewarned that if you Google "Blanchett smile", you will find a multitude of photos of Cate Blanchett's lovely face with poor David nowhere to be found. (It wasn't a terrible disappointment.)

Following is a chart of the "smile" from Blanchett (2013). (A quick note: you can double-click any chart in my posts to see a larger version. Also, while burnt orange text indicates a link to another website, yellow text indicates a mouse-over. Hover your mouse over the link for more information.)

There are three things I should note about the chart. First, the term "Experience" labeling the y-axis is an "auto-incorrect" for "Expenditures." Second, the smaller smile was added because of limited sample sizes for some tests. Pay more attention to the 30-year smile. My third point is a larger issue.

I suspect that some readers interpret the spending smile as showing that spending is high in early retirement, becomes lower until age 75 and then returns to nearly the level of early retirement near age 90, but this is not a graph of total annual spending. It is a graph of the annual real change in consumption for a typical retiree. In other words, it shows a decrease (and very rarely an increase) in spending at say, age 61 compared to age 60. It shows not the change of spending, but the rate of that change.

The rate of the decrease changes throughout retirement, but because these rates are nearly always negative (below the zero percent line on the y-axis in Blanchett's chart above), spending constantly decreases, but at different speeds. Banerjee shows the data in terms of total spending instead of the rate of annual change in spending and this point is more clear in his chart:

Reconstructing annual total expenditures from Blanchett's annual rate of change data for a retiree with a $100K annual spending target, we see a chart below that is similar to Banerjee's.

Mathematically speaking, the Banerjee curve is an annual spending function and the Blanchett smile curve is the derivative of the spending function. Banerjee shows the spending curve for a typical retiree while Blanchett shows the acceleration of that curve. Both show that expenditures generally decline with age, as have earlier studies. Blanchett additionally shows that expenditures drop more rapidly each year of early retirement and drop more slowly each year of late retirement, but both show that the amount of spending almost always declines.

Medical expenses late in life can increase expenditures significantly, but both studies appear to show that even when medical expenses do increase expenditures at older ages, they are lower than early retirement spending in real dollars.

The Banerjee chart and the Blanchett annual expenditure chart are not identical. Banerjee shows a steeper decline. Part of the reason for this may be, as Blanchett suggests, that he scrubbed the data to eliminate data points that seemed unreasonable, while Banerjee appears to have used the entire dataset.

Another reason is that Blanchett shows that rates of spending decline vary for undersavers and oversavers, while Banerjee provides a single rate of decline for all households.  Regardless, both studies find that typical retiree expenditures decline as we age. They do not remain constant in real dollars as spending strategies generally assume.

Why is this important? It should be obvious that when we try to estimate an amount of our savings that we can safely spend in the current year we must make some assumption about our future spending patterns. Spending strategies assume that our expenditures in real dollars will remain flat throughout retirement. If our actual spending will increase over time, we can safely spend less in the current year than these strategies predict, and the reverse is true if our expenditures will actually decline after we retire.

A 30-year retirement with level real spending of $100,000 a year would cost about $2.4M if we discount future expenses at 2%. Assuming Blanchett's findings for a retiree with a spending target of $100,000 a year, the same retirement would cost about $2.1M. Using the Banerjee 2012 finding that expenditures tend to decline about 2% annually, that retirement would cost only about $1.8M.

The following chart shows the expected annual spending and cost of an initial $100K annual retirement using all three assumptions:

Future spending is difficult to predict with any accuracy, but a spending strategy that assumes flat real spending throughout retirement, as nearly all do, will underspend early in retirement if the retiree's expenditures decline over time as Blanchett, Banerjee and several other researchers believe they commonly do. In these examples, Blanchett predicts a 12.5% less expensive retirement and Banerjee forecasts 25% less. From another perspective, that means a worker would need to save 12.5% or 25% less to fund retirement.

To quote Blanchett, "While many retirement income models use a fixed time period (e.g., 30 years) to estimate the duration of retirement, modeling the cost over the expected lifetime of the household, along with incorporating the actual spending curve, results in a required account balance at retirement that can be 20% less than the amount required using traditional models."

How does this impact our retirement plan? Clearly our future spending trend assumption has a significant impact on both how much we need to save and how much we can "safely" spend in the current year. Unfortunately, like assuming many other critical retirement unknowns such as future market returns and the length of our own retirement, choosing a future spending assumption is both critical and challenging.

In my next post, Retirement Spending Assumptions and Net Worth, I'll explore these two papers to see what they tell us about how we should choose.


  1. This article does a nice job highlighting how unrealistic models are when they assume constant real spending throughout retirement. Mathematically, these models appeal to academics because of their analytic simplicity, but it’s time to retire them in favor of more realistic spending assumptions.

    The way I think about the shape of the spending curve in retirement is as a sum of several curves, with each curve possibly having different embedded inflation assumptions. This is very easy to do with spreadsheet modeling, where recurring expenses can each be inflated at a different rate over the retirement planning window. For instance, I use one rate (CPI + a modest buffer) for general non-medical non-discretionary floor expenses, another (below-CPI) figure for discretionary expenses (i.e., things I expect to spend less on in real terms as I age), another for medical expenses (above CPI), and another (CPI) for annual adjustments to the tax tables, the standard deduction, personal exemption amounts, etc. due to inflation.

    How the expense curves add up to determine the overall expense curve will depend upon what proportion of total expenses each segment represents at the start. So if discretionary expenses initially represent a fairly large share for a retiree and they are assumed to grow more slowly than non-discretionary expenses, the groundwork will be in place to produce Blanchett’s smile pattern. Over time, expenses with above-CPI rates of increase (e.g., medical) will become a larger proportion of total expenses and help to turn the right side of the smile upwards. Because of the changing proportions of categories that inflate at different rates, the overall rate of expense inflation is never constant, and I think it is a mistake to plan with a single overall constant rate.

    The current trend toward requiring more out-of-pocket medical expenses than in the historic windows reflected by Blanchett’s and Banerjee’s (and Bernicke's) research has me thinking that the right side of the smile is likely to turn upwards faster than in the past, at least for retirees with moderate means.

    1. Since the trend toward higher OOP expenses is increasing for all ages, wouldn't that shift the entire spending curve upward? Is there a reason to believe that the acceleration of spending would increase, and that it would accelerate faster in late retirement? It seems to me that what you describe would raise the spending curve (a good point) but not affect the Blanchett smile (acceleration of spending). Am I missing something?

      Thanks for contributing.

    2. The rate of utilization of medical services typically increases as one ages, often dramatically late in the retirement years. So we have a medical inflation rate (typically well above the CPI) that applies to an increasingly large proportion of one's overall expenses.

    3. Actually, out-of-pocket medical expenses have been declining since about 2000, and throughout the study data period from 2001-2009, as a percent of total expenditures. That could change, of course, in either direction.

      While utilization of medical services does increase with age, from about 7% at age 60 to about 20% of total expenditures at age 90, other expenses decline. The Blanchett smile is a function (derivative) of total expenditures, so any shift to the smile will result from a net of these changes.

      I created a spreadsheet here to model the changes assuming a 10%, 20% and 30% increase in medical expenses above those in the Blanchett paper. If you look at the chart at the bottom of the worksheet you will see they have very little impact on the shape of the smile, and the change is about the same on the left and right.

      Please take a look and let me know if you see any problems with my analysis. It appears to me that significant increases in OOP health care expenses impact spending significantly, but apparently not acceleration of spending.

      I could be mistaken, but it is an issue that interests me, so I look forward to your comments.

      While the acceleration curve ("smile") is interesting from an academic perspective, I fear that it confuses the issues for the average do-it-yourself-er or even planner and I suggest they focus on the spending curve, instead.

  2. By the way, I like your approach to expense modeling.

  3. How would one model expenses for technology and services that don't exist today so far into the future? They’re completely unknown and unforeseen today. We forget how things change - but can easily remember by looking back 10, 20 or 30 years. Disruptive change always happens - an example being telecommunications where my home phone line in the wall has been replaced by a cell phone; internet is not dial up; and television is now paid, not free. All those expenses are quite different than before. Medicine is quite different today, as are its' expenses, compared to the past as well – and that changing trend will continue into the future; how, no one knows.

    An easier approach is simply to calculate how much money is needed to support one's Standard of Individual Living (SOIL) (individual to accentuate fact that everyone is different) and then live within one's means provided by those resources into the future (via withdrawal rate calculations each year as you've written in prior blogs). Some expenses go away while others emerge. One can still assume, if they like, a flat income curve, or a declining curve to model income over time. This removes a potential behavioral bias where lots of details provide the illusion of accuracy where none really exists. In the end, the solution depends on the income the resource of money may provide; how that money is spent will change.

    Another small point, similar to rote use of the replacement ratio (Aon Consulting and Georgia State University), is that an average of many individuals masks the truth behind each of them. In the expenditure case - the trend of the many may mask specifics on some individuals who buck the trend. What if you’re one of the general trend buckers?

    The difference between constant, versus declining, overall total requirements may be prudently viewed as a reserve sum of money for uncertainty and error. On the other hand, if one was not able to save what traditional models suggest, one could certainly spend more than suggested early one - just needing to realize in this case they're more likely to be forced to reduce spending later; rather than the "luxury" of choice had they saved more.

    As always Dirk, a very good post which does get to the main point you're making here, spending is not a constant. I would also add that WHAT one spends money on is not a constant either. The important objective is to have the total resources sufficient to afford the SOIL one desires throughout retirement - how those resources are spent will change. Always enjoy your posts and offline conversations!

    1. Larry, instead of a Standard of Individual Living (SOIL), shouldn't that be an Individual Standard of Living (ISOL)? :-)

      All good points.

      Indeed, my point is that retirement spending is not constant (and, as you say, not even very predictable), but worse, that most spending rules assume it will be.

      Thanks, as always, for contributing to the conversation. Always a pleasure.

  4. Dirk, SOIL goes back to my book "Wealth Odyssey," sold on Amazon or elsewhere. It's an acronym that fits other acronyms in that work. But, you get the idea as to what it conveys.

    And yes - that's the problem with "rules" of any kind in general; they ignore exceptions which in and of themselves almost seem to be the rule!

    So for those seeking a rule - recalculate each year given the facts known at that time. Going by calculations even a couple of years old seems risky in and of itself.

    1. It's OK, Larry. You are hardly the first person to to twist a name to result in a catchy acronym! Poking fun, but you get the idea across.

      I heartily agree with your last paragraph and the point of my post is that when you choose a rule to use in that annual recalculation, be aware of its assumptions and whether or not they apply to you. I'll be talking about that more as this series of posts progresses.

      To my readers: I have not yet read Larry's book, but because I am a firm believer in his approach to retirement planning, I will recommend it on that basis until I do. You can find it here.

  5. Do the expenditures of the old sink endogenously, stemming from their preferences, or are they just a response to getting poorer with age?

    If the former, that's useful info. If the latter, haven't we entered a circular argument?

    1. Great question and the Blanchett paper tries to delve into that. I will talk more about it in my next post. The answer is that it depends on the individual household's economic situation and we can only make assumptions about general cases.

      Households with high spending and low net worth will likely see declines in spending over time when they realize they're "getting poorer with age." Households with low spending and high net worth will likely see increases in expenditures over time when they realize they are able to spend more than they have been.

      Households whose spending is more appropriate to their net worth will likely see lower spending changes over time.

      Some spending declines are, as you say, endogenous. Increased spending on healthcare is an example of non-discretionary increases, but a decline in spending due to a less active lifestyle is also a factor. International air travel, as an example, typically declines during our 70's and domestic air travel typically declines in our 80's. The amount of money we spend as we age changes, as do the things we spend it on.

      So, in summary, I'd say spending changes as we age both because of life style changes and as a result of over- or under-spending. But more in my next post.

  6. All of the charts seem to be based on household spending.

    Reading this analysis, I'm having a hard time getting my head around the fact that the real-world household size for a retired couple shrinks from two to one somewhere along this curve.

    For that household, this logically leads to a one-time, step-function reduction of between 0 and 50% in household spending.

    Wouldn't a large data set of such individual household curves result in aggregate curves that look much like the smiles published by the researchers?

    1. First, I'd like to address your final phrase, "the smiles published by the researchers." Although this question has been researched many times and in many countries, only one researcher, David Blanchett, published a "smile" (and, frankly, I'm beginning to wish he hadn't). The smile is a curve of the acceleration of spending, not of spending itself. (See Retirement Spending Assumptions and Net Worth.)

      Annual spending does not decline in early retirement, level off, and then increase in late retirement in the shape of a smile. It typically trends downward in the shape of a playground sliding board, as in the graphs in the post above, throughout retirement.

      You are correct that at some point a 2-person household will lose one spouse and typically spending declines about 25%, or as Laurence Kotlikoff says, "Two can live as cheaply as 1.6."

      The households in the CAMS data, as best I can tell, include one-person households, 2-person households and no doubt a small percentage of larger households. Spending will probably decline for multi-person households after a spouse dies. I believe households had to respond to all five waves over the decade to be included, in which case 1-person households that didn't survive that decade would have been excluded. According to census data, about 47% of people over age 60 live as a married couple.

      I don't quite understand your question, so I suspect this may not answer it. Are you asking if a household eventually losing a spouse explains the gradual decline in retirement spending found in every study?

      If you can ask your question in a different way, I'll take another shot at answering it.

      Thanks for joining the discussion!

    2. Thank you for your response. I now see I was not very clear.

      Yes, I'm asking if a household eventually losing a spouse explains (or at least influences) the original researchers' conclusion of a gradual decline in average household spending. I haven't read the references in detail, but it would seem there is a non-trivial error potential in applying average retiree household spending (based on a variety of household sizes) to an individual case with a known household size. For a couple, my intuition tells me that using average of household spending levels that not normalized to examine only households starting out as couples may not be an satisfactorily accurate indicator of spending trends for a particular retired couple.

      Since posting my question, I created a simple spreadsheet based on the 1.6 / 1 Kotlikoff ratio you mentioned. The columns are 5 year increments between age 60 and 95. I then created 7 rows, each representing a household of two that declines to a household of one during the 35 years. In the age 60 column, each household has two persons spending 1.6X. At age 95, all of the households consist of one person spending X. In between, each column represents one fewer household of two persons, e.g. at age 80 there are 4 couples and 3 singles in my data set. For each row, that household’s spending is a step function consisting of periods of spending 1.6X followed by periods on spending X.

      Summing up the total spending for each column and dividing by the 7 households in the data set gives an average household spending level per household at a particular age. This method yields an age 90 average household spending at a level of 68% of the age 60 average household spending, which is remarkably similar to the Blanchett and Banergee curves.

      Is this a coincidence, an invalid comparison or a potential issue with the researcher’s methods?

      Of course, the answer is critical only to the extent that a couple chooses to apply a spending reduction method when estimating future expenses. I am looking forward to your upcoming post on that subject.

    3. If I understand the spreadsheet you are describing, you have omitted mortality. As I explained, 53% of the households are single and wouldn't be affected by what you describe. For the other 47%, you need to consider the probability at each age that one spouse died the previous year (or 5 years in your spreadsheet) but the other did not. If you want to email your spreadsheet to, I'll take a look.

      The more important question that you raise, I think, is whether and how the researchers accounted for this. I don't have enough information to answer that but I will try to find out.

    4. From David Blanchett, "For my analysis, I only include those households that have the same number of members for each of the five periods. In theory someone could get remarried and I would include them, but I excluded scenarios where a spouse died over the 10 year period."

      So, loss of a spouse does not contribute to the declining expenses in Blanchett's study. If you reduce the change of expenditures in your spreadsheet by multiplying the amount of the change at each age by the probability at each age that one but not both spouses has died, I don't think your curve will look much like the spending curves from these research papers.

      Great question, though. Thanks.