August 22, 2024 – Welcome back after a long, relaxed summer break this year. Today’s post – suggested by a reader – deals with an intriguing question: how does early retirement impact my life expectancy? You see, since retiring six years ago, I’ve always assumed I’m doing my body a favor. I live a more relaxed life, exercise regularly, and sleep more every night. So, I’ve always believed my 2018 early retirement decision should improve my health and increase my life expectancy, right? Well, shockingly, the academic research on this topic is mixed. True, some studies seem to indicate early retirement reduces mortality. However, several studies point in the opposite direction. The commonly referenced rationale for this paradoxical result is that early retirees ostensibly often have reduced social contacts and physical and intellectual stimuli, increasing the risk for physical and mental decline.

Bummer! Was it a mistake to retire early? I did notice more gray hair in the last six years. Have you noticed a decline in my mental capacities? Am I becoming a curmudgeon? Be honest, everybody! Well, not so fast. There are many reasons not to be too concerned about (early) retirement’s alleged adverse health effects. Let’s take a look…

Before we get started, though, let me point you to three recent podcast appearances of mine.

But now, back to today’s content…

I’m not an expert in this field, but a reader who suggested this topic gave me a reading list to review. Here’s a sample of papers I found. I should state that I read through them without replicating any of the studies. I categorize them by the estimated impact on life expectancy and provide a brief summary of the results:

Zulkarnain, Alice and Rutledge, Matthew S., How Does Delayed Retirement Affect Mortality and Health? (October 5, 2018). Center for Retirement Research at Boston College, CRR WP 2018-11 , Available at SSRN: https://ssrn.com/abstract=3261325. This paper uses a dataset from the Netherlands and finds that delaying retirement reduces mortality for men. The paper finds no link between retirement timing and the incidence of diabetes or depression. Put differently, the change in mortality must come from conditions other than diabetes and depression. There is no discernible effect on women’s mortality.

Andeas Kuhn, Stefan Staubli, Jean-Philippe Wuellrich, and Josef Zweimüller, Fatal Attraction? Extended Unemployment Benefits, Labor Force Exits, and Mortality. NBER Working Paper No. 25124. Available at https://www.nber.org/system/files/working_papers/w25124/w25124.pdf. The paper finds that for men, there is a 0.2-year reduction in the age at death for each year a person retires earlier. The effect is more prominent in blue-collar workers and workers with low work experience. For white-collar workers and workers with high work experience, the effect is less prominent and statistically insignificant. There is no discernible effect on women.

Stephanie Behncke. Does retirement trigger ill health? Health Econ. 2012 Mar 21(3):282-300. doi: 0.1002/hec.1712. Epub 2011 Feb 14. PMID: 21322085. Available at https://pubmed.ncbi.nlm.nih.gov/21322085/. The paper uses data from the English Longitudinal Study of Ageing (ELSA) and finds that early retirement negatively impacts various health indicators.

Cristina Bellés-Obrero, Sergi Jiménez-Martín, Han Ye. The Effect of Removing Early Retirement on Mortality. September 2022. IZA DP No. 15577. Available here: https://docs.iza.org/dp15577.pdf. Retiring later increases the mortality risk. The effects are worse for low-skilled workers as well as in “physically and psychosocially demanding jobs.”

Bound, John and Waidmann, Timothy, Estimating the Health Effects of Retirement (October 1, 2007). Michigan Retirement Research Center Research Paper No. UM WP 2007-168, Available at SSRN: https://ssrn.com/abstract=1082047. For men, retirement has a slightly positive health impact.

Why is there such disagreement in empirical studies? Let me list a few reasons:

A significant concern is that the retirement decision is endogenous. Specifically, it is often linked to an individual’s health. Thus, it’s frequently not clear if a person died early because they retired early or because they knew of their underlying health condition, and that caused them to retire early to enjoy their last few years of reduced life expectancy. Just running a blind OLS regression of mortality on early retirement may falsely attribute mortality and bad health to early retirement when the true causality goes the other way around.

In a different area, in pharmaceutical research, you want to avoid studies where people endogenously pick what drug they take because a new and effective drug might appear ineffective or even harmless if patients with more severe cases of a condition are more willing (desperate?) to try it. Thus, you want to create a study where the subjects are exogenously and randomly assigned a drug vs. placebo. Statistically speaking, that’s the gold standard to avoid this endogeneity issue and other problems. The analog in early retirement vs. mortality research would have been to take a sample of people, force some of them to retire at one age, the rest at a later age, and then monitor their health and mortality for the next few years. That would be impractical, unenforcible, and not to mention unethical. So, mortality researchers often have to rely on the crummy datasets they can get their hands on.

Of course, researchers are aware of this endogeneity issue and have found ways to correct it. Every paper I’ve seen in this field claims to account for endogeneity using a so-called instrumental variable (IV) approach. Without getting too geeky, IV means finding another explanatory variable, Z, with a causal relationship with the endogenous independent variable, X, (work vs. retirement decision). But the IV variable should impact the dependent variable, Y, (mortality or health) only indirectly through its correlation with X. For example, Zulkarnain and Rutledge (2018) use a variable DWB (“Doorwerkbonus”), which is the eligibility for a temporary Dutch work incentive bonus program that likely correlates with the retirement decision but not directly with a subject’s health. In the first regression stage, the authors estimate the work status of each subject using the DWB variable as one of the independent variables. In a second-stage regression, the authors then use not the actual but the estimated(!!!) work vs. retirement status as an independent variable to account for health and mortality effects five years later.

In any case, the endogeneity problem is present in every study. For example, Zulkarnain and Rutledge (2018) report that in all naive OLS regressions (i.e., ignoring the IV methodology), retirement consistently had a detrimental impact on health and mortality for both males and females. However, the effect is weaker when they use the 2-stage IV method. However, the effect remains statistically significant for males.

Still, there is no guarantee that you will eliminate all of the endogeneity with your IV methodology. Depending on how well your IV approach works, you may remove the endogeneity and find that early retirement improves health, as seen in some papers. Or you may only partially remove it, which may explain some negative health results of early retirement. When I studied and practiced economics, I ranked these empirical studies regressing everything in the kitchen sink and waiving the magic IV wand as somewhere between intellectual lightweights and economic junk science. So, I’m not too concerned about those empirical results.

Also, endogeneity is not the only headache, which brings me to this next point…

Mortality is a complex issue, and trying to account for cross-sectional differences with one single explanatory variable, like your retirement date, is rife with problems. Explaining one outcome with one factor and ignoring others can lead to paradoxical results. It is so paradoxical that it got its own name: Simpson’s Paradox, named after British statistician Edward Simpson.(To be clear, I’m not claiming that the studies ran univariate regressions, but missing variables in multivariate regressions create the exact same Simpson’s Paradox.)

The paradox may present itself in many ways, but one classic is this numerical example. Imagine we have two cohorts, A and B. Within each cohort, we observe variables X and Y. We could think of X as the retirement age and Y as the age at death. The top two panels show that X and Y are negatively correlated in both subsamples. The R^2 (goodness of regression fit) is around 0.8 in each subsample, and the slope parameter, about -0.20 in each sample, is significantly negative in each case, with a t-stat in the double digits (not reported in the Excel charts here, but I did calculate it separately).

But notice what happens if we merge the two cohorts into one sample as I did in the bottom panel: Now we get a positive slope of +0.5! Still, a decent R^2. The t-stat of the slope is still above 10. What happened here? The two subsamples have vastly different mean values for their X and Y variables. Thus, the slope of the joint OLS regression is mainly impacted by the location of the point clouds in the scatterplot rather than the true underlying relationship in X and Y within the subsamples.

This data paradox is indeed a headache for mortality researchers. Imagine Cohort A is comprised of construction workers, while Cohort B is comprised of college professors. Construction work is very physically demanding, often necessitating a lower retirement age. Moreover, the job also takes a toll on your body and may reduce your life expectancy. On the other hand, college professors tend to work much longer and have a higher life expectancy. Thus, effects from other variables like the type of job may improperly impact our early retirement slope estimate if not correctly accounted for.

Of course, data scientists will tell you that you need not worry about Simpson because we can cure the Paradox if we simply “control” for all those other pesky factors. And what econometricians mean by “controlling” is that they throw an extensive range of additional explanatory variables into the regression equation: gender, education, health status, industry, marital status, income, wealth – you name it, whatever is included in your database. And I should stress that all of the empirical studies referenced above control for several other possible mortality factors. You’d be laughed out of the room if you tried to sell a univariate regression linking mortality to only the retirement age.

So, researchers have done their duty to alleviate all concerns about that pesky Simson’s paradox after throwing all the obvious variables into the statistical kitchen sink, right? Wrong! While all the mortality researchers hope they haven’t missed anything crucial, they can’t control for every conceivable additional mortality factor. Some unknown and/or unobservable effects might still lurk in your sample and could tilt your results. Not all databases have all the necessary series. You may still get nonsensical slope estimates for the early retirement variable.

But let’s assume the retirement researcher has thought of everything and has access to all the data series. They don’t, but even if they did, there is still a way to mess it up, which brings me to the next issue…

Another potential concern about controlling for everything in the kitchen sink is that including the other variables will not fix the problem if done improperly or incompletely. To illustrate this, let’s look at another numerical example, again with completely made-up data.

Imagine two cohorts again, A and B, each displaying a significant statistical relationship between variables X and Y. However, the slopes are different this time: -0.20 in Cohort A and +0.20 in Cohort B. If we aggregate the sample, we are left with a big nothing-burger.

We can also study the detailed OLS regression results; see the table below. In Model 3, where we merge the two samples without any dummy variables, the R^2 is now only 0.0001, the slope is close to zero, and its t-stat is no longer significant. No surprise here! However, even including a dummy variable for Cohort A in Model 4 does not cure our problem. Both the dummy variable and the X slope are statistically insignificant. The R^2 is only 0.07. Only in model 5, where we have an intercept, a dummy for Cohort A, and two separate slopes for the two cohorts, would we recover the information from the two cohorts again. Notice that the intercept plus Cohort A dummy sum up to the actual intercept in Model 1: 64.63+21.07=85.70.

Thus, to fix this issue, we’d need to include two slope parameters, one for Cohort A and one for Cohort B, in addition to the dummy. That way, we’d effectively model two separate intercepts and two separate slopes, as observed in the data. Forcing the slope to be the same when it’s clearly not, will give you misleading results if the two types are included in the regression equation only through separate dummy variables. And that’s exactly what the retirement mortality researchers are doing. At least in most of the papers I found.

Of course, the researchers indeed run separate regressions, usually for males vs. females or white-collar vs. blue-collar jobs. But the “different slope” setup can and should appear with many other variables. For example, it’s conceivable that every industry and job duty should have different marginal impacts of early retirement on mortality. The same may hold for different wealth and income levels, pre-retirement health status, etc. Forcing the early retirement impact to have the same slope for all while only including some dummy variables can give you nonsensical results.

But the problem is likely even worse, which leads me to the section…

Dummy variables and separate slopes are only feasible if the Cohort A vs. Cohort B status is observable. What if the two cohorts are due to unobservable psychological differences? Say, Cohort A are the productive and healthy folks who flourish in retirement; they improve their health status and increase their life expectancy. In Cohort B, on the other hand, we have the “sloths” who lose their social circles, avoid physical and intellectual stimuli, and die earlier due to depression and diabetes.

And what if the Sloth vs. Productive early retirement status is utterly uncorrelated to all the other observables? It can happen to men and women, educated and uneducated retirees, high-income and low-income retirees, high-wealth and low-wealth retirees, etc. It would render the empirical research all but useless. Ironically, this would explain the wide range of empirical results, i.e., depending on whether you oversample or undersample the two unobservable types, you get positive or negative outcomes due to early retirement.

Just as a thought experiment, let’s assume that the statistical studies point to an increase in mortality due to early retirement. Say, someone comes up with the definitive proof, methodologically correct and without any data flaws, that early retirement indeed reduces everyone’s life expectancy. My response would be, “So what!?” I mostly retired to increase my quality of life. The “Quality over Quantity” maxime works here as in many other areas! In other words, imagine that back in 2018, I had a life expectancy of 40 more years if I kept working, and by retiring early, I would have reduced that by three months (as estimated in one of the studies). Would I have still retired? Heck, yes! Assuming I would have worked 2,500 hours a year, that’s 50,000 hours behind my office desk; that’s the equivalent of almost six years of 24/7 at the office. A three-month reduction in my life expectancy is well worth that tradeoff.

I worry very little about this empirical early retirement mortality research. First of all, the empirical results are all over the map. Linking a single variable, early retirement, to some of the most complex observables like health and mortality creates various statistical challenges. As a former economist and data scientist, albeit in a different subfield of economics, i.e., macroeconomics and finance, I always thought that most of these purely empirical data-mining exercises in large cross-sectional databases were economic junk science. It’s a bit like sausage-making; the final result can be tasty (intriguing and newsworthy results), but you’d lose your appetite if you witnessed the actual sausage-making.

Second, even in papers that “prove” a negative impact of early retirement, the results are often not 100% consistent across the board. For example, according to one study, the effect is not even statistically significant for white-collar workers. Well, I was a white-collar worker, so I should be safe. Case closed!

Third, even if someone could conclusively prove to me that, on average, early retirement indeed leads to adverse health outcomes, like depression, diabetes, high blood pressure, etc. I don’t care. I’m not an average retiree. My decision and recommendation to others would remain the same: retire early and don’t be that lethargic sloth who wastes away in early retirement. Do something useful with your early retirement. No statistician with their IV mumbo jumbo can convince me that early retirement is bad for your health if you stay fit (mentally and physically), volunteer in your community, make new friends, travel extensively, etc. And, even in the worst case, where you indeed have a slightly reduced life expectancy, you at least increase the quality of life during early retirement relative what you would have done in a corporate job.

In that spirit, everybody, have a fulfilling early retirement and hopefully a long life as well!

Title Picture credit: pixabay.com