The emergence of comparable micro-data from multiple countries makes it possible to take advantage of cross-country variation in pension policies as valid instruments for identifying the effect of retirement on cognition.
The data come from three different surveys: the U.S. Health and Retirement Study (HRS), the English Longitudinal Study of Ageing (ELSA), and the Survey of Health, Ageing and Retirement in Europe (SHARE), which interviewed households from 11 countries in its first wave. All are nationally representative surveys of the population age 50 or older (age 51 or older for the HRS) and collect a wide array of information.9
The HRS was first collected in 1992 and the other two surveys were designed to be closely comparable, with the explicit objective of facilitating cross-country comparisons. The first wave of ELSA was collected in 2002 and the first wave of SHARE in 2004. We use data from 2004, the first year for which we have data available from all three surveys. Sample sizes run about 20,000 individuals in the HRS, a little over 9,000 in ELSA, and between 1,000 and 3,000 in each of the SHARE countries.
Our analysis is cross-sectional, comparing the cognitive performance of the retired with that of the not-retired in 2004 and using national pension policies as instruments to correct the endogeneity bias resulting from retirement’s being a self-chosen state. More specifically, we seek to estimate the relationship between cognition of individual i in country j as a function of a constant term and the individual’s retirement status plus the error term. We treat retirement status as an endogenous variable and estimate the equation using variables characterizing national policies as instruments, as described later in more detail.
provides intuition for how the estimated relationship identifies the parameter of interest. For the moment, assume that only the “disengaged lifestyle” mechanism is operative. That is, assume that cognition is only affected by the state of being retired, measured by the dummy variable, Rij , which is equal to one if person i in country j is retired and zero otherwise. illustrates the hypothesized relationship. The solid line depicts the path of cognitive aging that would be followed by a person who never retires—or, more realistically, by a person who retires at a late age—and the two dotted lines indicate the paths that would be followed by a person who retires early at age A1 or at a later age A2 . The steeper slope of the dotted lines reflects the negative effect on cognition caused by leaving the more stimulating work environment.
Identification of Retirement Effect on Cognition
The data presented earlier in can be interpreted in terms of the diagram in . Suppose that A1 corresponds to age 50-54 and A2 to age 60-64 and assume that A1 is the average age of retirement in country 1 and A2 is the age of retirement in country 2. At age A1 , workers in both countries have the same level of cognitive ability, C0 . However, the cumulative effects of retirement implies that by the time they reach age A2 , the cognitive abilities of workers in country 1 will have fallen below those in country 2.
With a different interpretation, the diagram in can also illustrate the “on-the-job retirement” hypothesis. Imagine that workers cut back on their rate of human capital investment 10 years before they expect to retire, causing an increase in the rate of cognitive decline equal to the difference in the slopes of the solid and dotted lines in . We could then interpret A1 as age 40-44 and A2 as 50-54. When workers are evaluated at age 60-64 (denoted by A3 in the figure), the cognitive score of persons aged 60-64 relative to 50-54 year-olds in country 1 will be lower than in country 2.
Our estimations rely on a comparison between retirees and working individuals. This illustration also shows that our estimation strategy picks up both potential effects of retirement on cognition, the “unengaged lifestyle” effect and the “on-the-job retirement” effect, but that it cannot distinguish between them.
We use the same measure of cognitive performance as Adam et al. (2007)
– a score ranging from 0-20 reflecting the total number of words remembered in exercises of immediate and delayed recall of the same 10 words. Retirement status is derived from respondent reports of whether they are currently working for pay. A person is considered “retired” if that person is not working for pay and “not retired” otherwise. We include all men and women, irrespective of prior work history, in view of the fact that the lack of a stimulating work environment would also apply to people with limited or no attachment to the labor force. This implies that we will estimate the average “treatment effect” of retirement (or not working) in the entire population. To get a sense of the magnitude of the effect that is present in the data shows the average cognitive score for each country by the fraction not working for 60 to 64 year-old males and females pooled. We consider a narrow age-band as a way of conditioning on age in this visual representation. As indicated by the fitted line there is a systematic relationship between the average cognitive score and the fraction not working across countries, suggesting that on average being retired is associated with a lower memory score of about 4.9 points on a 20 point scale.
Cognition by percent not working for pay, 60-64 year-old men and women, weighted
To the extent that different levels of cognition across countries might influence the timing of retirement we cannot attribute a causal interpretation to the patterns shown in . We use variation in public pension policies to address this endogeneity issue.
Our instrumental variables estimation, as always, can be viewed as a two-stage process. In the first stage, we use retirement status as our dependent variable and national pension policies as the explanatory variables: whether the individual has reached the age of eligibility for early retirement benefits in the public pension system and whether the individual has reached the age of eligibility for full retirement benefits. We also investigate specifications that use as additional instruments the number of years to or since reaching the age of eligibility for early retirement benefits (eligibility age minus current age); and the number of years to or since reaching the age of eligibility for full retirement benefits. These distance variables pick up the strong financial incentives to retire inherent in most public pension schemes as shown in Gruber and Wise (1999)
. All instruments vary across country and sex (see table in Appendix
We focus our analysis on the age range where we have maximum variation in the instrumental variables. As a result our analytical sample includes persons aged 60 to 64 from the United States, England, and all other European countries included in the first wave of SHARE, that is, Sweden, Denmark, the Netherlands, Germany, Switzerland, Austria, France, Belgium, Spain, Italy, and Greece.
In the second stage, we use the cognition score as the dependent variable, and as the explanatory variable, we use the retirement of each individual as predicted from the first-stage regression. In other words, our explanatory variable in this second stage includes only variation in retirement behavior that can be explained by national policies, which should offer an unbiased estimate of the effect of retirement on cognition.
Before we present the regression results we provide a graphical representation of our identification strategy. shows to what extent national pension policies drive cross-country variation in the fraction not working for pay and in average cognitive scores. The top left graph in plots for each country the percent of 60 to 64 year-olds who are not working against the percent of 60 to 64 year-olds who are eligible for early pension benefits. The slope of the fitted line conveys that the higher the fraction eligible for benefits, the higher the fraction not working for pay. The R-squared of this simple regression is 0.49. The effect is larger for eligibility for early than for full pension benefits, as is shown in the top right graph. Similarly, the bottom horizontal panel plots for each country the average cognitive score of 60 to 64 year-olds against the percent who are eligible for early (left graph) or for full (right graph) pension benefits. In this case the fitted regression line implies that the higher the fraction eligible for pension benefits the lower the average cognitive score in the population. Again, both the size of the effect and the R-squared of the regression are larger for eligibility for early than for full pension benefits. To arrive at the effect of interest – the effect of retirement on cognition – we can read off the instrumental variables estimate directly from these graphs by dividing the slope coefficient for cognition by the slope coefficient for the fraction working for pay. It is (−0.036*100/0.523=) −6.88 when using eligibility for early retirement benefits as the instrument, and it is (−0.011*100/0.228 =) −4.82 when using eligibility for full retirement benefits as the instrument.
Graphical illustration of the IV-estimation
The results from micro-level estimation where we use both eligibility for early and for full retirement benefits as instruments are closely comparable. shows the results.10
As expected, the indicators of eligibility for early and for full pension benefits are very strong predictors for retirement in the first stage regression.
Instrumental Variable Estimation of the Effect of Retirement on Cognition
We find a large and significant effect of retirement (or more precisely of “not working for pay”) suggesting that retirement is associated with a reduction in the memory score of about 4.7 points on a scale from 0 to 20 compared to those who continue working. The average score in the sample is just under 10 and the standard deviation is 3.3. So the estimated effect amounts to just under 1.5 times the standard deviation of the cognitive score in our analytical sample. The results are robust to alternative specifications. For example, inclusion of the distance to and from the eligibility age for early and for full pension benefits gives a point estimate of −4.3 for the effect of retirement on the cognitive score. Also controlling for age in the most rigorous manner, that is, by estimating the same relationship separately for each single year of age from 60 to 64 and averaging the five point estimates we find an effect of −5.7. Extending the age range of our analytical sample to encompass all ages where there is variation across countries in the eligibility for pension benefits we re-estimated the relationship for all 55 to 65 year-olds. We added controls for age (age and (age/10)2 ). We obtain a point estimate of −4.4 (or −4.5 when adding to the instruments the distance to and from the age of eligibility for early or full pension benefits). The estimates are strongly significant at the 1-percent confidence level across all specifications.
The estimated effect of retirement on cognition proxied by performance on word recall exercises is very large indeed. It is important to keep in mind that our framework does not measure solely the immediate effect of the onset of retirement. Instead it captures the effect of a much more extended process of retirement that may start – at least for some – already on the job according to our “on-the-job retirement hypothesis” and continues well into retirement. Retired individuals in our sample differ in their timing of retirement and will be at different distances from the path of cognitive decline without retirement as depicted in . Some may have retired just recently while others may have been retired for many years. We do not investigate the issue of duration here.
Coe and Zamarro (2008)
use the same SHARE data to investigate the effect of retirement on various health measures, including performance on the word recall test. While they employ the same strategy to deal with the endogeneity of retirement their analytical framework is very different. They apply a regression discontinuity approach attempting to isolate the effect of the onset of retirement, holding everything else the same, and do not find any effect of retirement on cognition. However, since retirement may take some time to affect cognitive function, their model is answering a different question than ours.
Can the very large effects we identify be plausible? Our findings are based on a powerful natural experiment that involves a major change in lifestyle (work to retirement) and that is unmatched in size (populations from multiple countries) and duration (several decades) when compared to common laboratory experiments conducted on small – usually selected – samples over short periods of time. Therefore we believe that this finding deserves serious attention. One would expect the effect to vary across individuals. Future research will have to investigate the role of education, environment and other factors.