Parallel Progress in Perceived Age and Life Expectancy

Abstract

Background

Human life expectancy continues to rise in most populations. This rise not only leads to longer lives but also is accompanied by improved health at a given age, that is, recent cohorts show a reduction of biological age for a given chronological age. Despite or even because of the diversity of biomarkers of aging, an accurate quantification of a general shift in biological age across time has been challenging.

Methods

Here, we compared age perception of facial images taken in 2001 over a decade and related these changes in age perception to changes in life expectancy.

Results

We show that age perception changes substantially across time and parallels the progress in life expectancy. In 2012, people aged more than 70 years needed to look 2.3 years younger to be rated the same age as in 2002.

Conclusions

Our results suggest that age perception reflects the past life events better than predicts future length of life, that is, it is written in your face how much you have aged so far. We draw this conclusion as age perception among elderly individuals paralleled changes in life expectancy at birth but not changes in remaining life expectancies. We suggest that changes in age perception should be explored for younger age classes to inform on aging processes, including whether aging is delayed or slowed with increasing life expectancy.

Human life expectancy has risen in long-lived populations by about 2 years per decade (1,2). This continuous and exceptional rise in life span started around one and a half centuries ago and not only led to longer lives but has also been accompanied by increased health span (3–6). That is, with increasing life expectancy, we observe improved health and a reduction in biological age measured by improved scores on biomarkers of health and aging at a given chronological age. Middle-aged people nowadays show improved health compared to former times (7,8), and improvements are not limited to younger people or the most recent cohorts. Even the oldest old (>90 years) score better on biomarkers of aging when born in more recent cohorts compared to earlier cohorts (8,9). Such findings of shifting biological age indicate that people in recent times (i) age slower, (ii) delay aging, or (iii) begin life at a lower level of biological age, but these effects are not mutually exclusive (6,8).

Despite the qualitative evidence for prolonged healthy lives and improved physiological and cognitive function of recent birth cohorts, quantifying the relationship between rising life expectancy and improved biological age has been challenging (5,6,10). Biomarkers of aging provide the main tool to investigate biological age. These biomarkers are extremely diverse and target different levels of biological organization; famous examples include telomere length, the epigenetic clock, allostatic load, but also nonbiochemical measures, such as grip strength and other physiological and functional measures (11–13). The challenge of how to best evaluate biological age is illustrated by the many different (healthy) aging indices that have been developed, and the controversy surrounding how to weigh the contributing biomarkers (14–18). Correlating different biomarkers of aging against each other illustrates that they capture different aspects of the aging process because they do not necessarily correlate well with each other (13). This weak correlation among biomarkers supports the understanding that aging is a complex process. Numerous of these biomarkers are accurate descriptors of biological age for certain ages but not for others. For instance, grip strength declines steadily from age 50 onward but is uninformative for younger ages (19). One challenge in evaluating biomarkers is, how to convert the unit a biomarker is measured in—microunits per milliliter for insulin, kilograms for grip strength, or some scoring of cognitive tests—to changes in hazard risk, related mortality, and consequently life expectancy. In consequence, relating—in a general way—the biological age to the biological process of aging as aimed at with hundreds of different biomarkers remains difficult (13). To define the underlying aging process in relationship to the biological and chronological age, one also needs to account for partial correlation between average chronological aging processes, the interplay among the different biomarkers, and the biomarker environment interaction (20). Though we aim at a general integrated measure of biological age, we have to face that most biomarkers seem to target specific aspects of the aging process (13).

One of the many biomarkers of aging is perceived age, that is, how old one looks, which is associated with survival, and physical and cognitive functioning (21,22). For people aged 70 years and older, even after correcting for sex and chronological age as well as other partly correlated biomarkers—for example, cognitive scores, strength scores, grip strength, and the Mini-Mental State Examination—annual mortality risk increased by 3% with each year deviation in perceived age from chronological age (21). Younger people aged 30–70 years had relatively bad health if their perceived age was judged substantially older than their chronological age and perceived age correlated with other biomarkers of aging in a follow-up study comparing the same individuals at age 26 and 38 (23,24). Perceived age might, therefore, be an integrative and relatively general biomarker of aging with the advantage compared to many other biomarkers: both perceived age and chronological age are measured in the same units, age in years, and can, therefore, be directly compared to each other and to life expectancy.

Here, we used the change in perceived age of the same facial images assessed over a decade to evaluate how this measure of biological age changed over that decade and how this change relates to shifts in cohort life expectancy. For our evaluation, we used facial images taken in 2001 (February–April) and collected two rounds of perceived age on these images; the first round was collected in November 2002 and the second round was collected in December 2012. Obviously, the images have not changed over the 10 year period between the round of assessments. The obvious and naive expectation would be that the perceived age should not change, which would indicate that the biological age—measured as perceived age—did not change over a decade. However, as we argue here, if the process of aging changes with time and increased life expectancy, we expect a change in biological age over time and such change should be reflected in perceived age. Therefore, changes in age perception are expected to track the change in biological age across time and we expect that the same images evaluated at the end of 2002 would be judged to be younger and closer to their chronological age compared to the evaluation 10 years later. Our aim was to test the hypothesis that the difference in perceived age between the two assessments quantifies the change in biological age and parallels the process of changed cohort life expectancy over this decade.

Methods

We captured facial images of 238 Danish women and 144 Danish men aged more than 70 years (Supplementary Figure 1) in a standardized way between February 2001 and April 2001 (details in (21)). In November 2002, each of 20 raters assessed the perceived age of each individual in the 382 images without any knowledge of the participants chronological ages. In December 2012, this assessment was repeated on the same images by nine raters (one rater was excluded as an outlier as this rater consistently rated much higher compared to the others). Four raters in the 2012 assessment also rated in the 2002 assessment (see Supplementary Figure 5). Note, our results were robust to the rater selection, that is, when we restricted our analysis to the four raters that rated in both assessments, results did not change qualitatively, the model comparison results were very similar to those presented in Table 2, and the quantitative changes in Akaike’s information criterion (AIC) differences were very small (restricted analysis not presented). In our analysis and investigation, we focus on average perceived age for each image over the 20 raters in 2002 or the 9 raters in 2012 but also evaluate variance in perceived age within images and among raters by including random effects of image ID, rater ID, and rater age. These random effects were included to account for bias in the raters that might explain the shift in age perception. We do not expect much systematic bias, but some variance among individual raters. We derived this expectation from previous studies that showed that female geriatric nurses (expected to be experts in age rating of older women), younger male student teachers (expected to be the worst raters of older women), and peers (older women of the same age as the study group) did not differ in their accuracy and ability to rate ages (21). In our data, young and old raters in average rated slightly younger compared to mid-aged raters that tended to slightly overestimated ages (analysis not presented). Combined, these findings suggest that neither age nor sex of the raters significantly influences age perception.

For the core part of the study, to address how perceived age has changed over time, we used linear models with difference between chronological and perceived age as response variable and chronological age, sex, and assessment year as fixed explanatory factors. We included image ID, rater ID, and rater age as random effects. We set up and ran models in the program R (25) and selected among models based on AIC, with a difference of more than 2 in AIC being considered to provide better support to the model with the lower AIC (26). We evaluated model assumptions using diagnostic plots, which suggested that a Gaussian error structure was appropriate for all models. To explore characteristics of the data and to detect potential biases in the data, we formulated additional models for which we used (i) chronological age as response variable and sex as explanatory variable, (ii) perceived age as response variable and sex and year as explanatory variables (Supplementary Table 1), as well as (iii) the difference in the deviance between chronological age and perceived age in 2002 and the deviance between chronological age and perceived age in 2012 as response variable, and sex and chronological age as explanatory fixed factors (Supplementary Table 2).

Our approach lowers the variance in change in biological age over this 10-year period (2002–2012) compared to the actual background population in 2002 and 2012. In real life, some individuals might have for instance just gone through some severe illness when their image was taken, and such events would increase the variance in change in biological age. Excluding this variation sharpens the focus on the perception that changed over the decade of investigation.

Results

Mean chronological age of the 238 Danish women and 144 Danish men aged more than 70 years of whose facial images were taken in 2001 was 75.9 years. Both sexes showed similar chronological age distributions (men mean: 75.5; women mean: 76.2; Null model: no difference between sexes AIC = 60,010, model with sex differences AIC = 60,010; Table 1; Supplementary Table 1 and Supplementary Figure 1). Perceived mean age in 2002 and 2012 was lower for men than for women (Table 1; Supplementary Figure 1 and Supplementary Table 1). Perceived age between the assessment years was well correlated (both sexes: 0.88; men: 0.82; women: 0.90). The difference between the sexes in perceived age between 2002 and 2012 was slightly more pronounced for women than for men (Sex*Year interactive model best supported, Supplementary Table 1). The distribution of chronological age of the sample images did not differ among the sexes, but perceived age differed between men and women and between the assessment years.

Table 1.

Mean and Standard Deviation of Chronological and Perceived Age, Deviance Between Chronological and Perceived Age, and Change in Life Expectancies in Denmark

	Mean ± SD (minimum, maximum)
Parameters	Both sexes	Men	Women
Chronological age	75.9 ± 4.6 (70, 90)	75.5 ± 4.8 (70, 90)	76.2 ± 4.5 (70, 89)
Perceived age 2002	76.5 ± 5.6 (54, 100)	76.3 ± 5.4 (54, 98)	76.6 ± 5.7 (55, 100)
Deviance between perceived age 2002 and chronological age	0.5 ± 6.4 (–26, 26)	0.7 ± 6.7 (–24, 26)	0.4 ± 6.2 (–26, 19)
Perceived age 2012	78.8 ± 5.5 (58, 100)	78.1 ± 5.0 (59, 94)	79.2 ± 5.7 (58, 100)
Deviance between perceived age 2012 and chronological age	2.9 ± 5.5 (–16, 21)	2.6 ± 5.5 (–16, 20)	3.0 ± 5.6 (–16, 21)
Period life expectancy 2003*	77.7	75.3	80.1
Period life expectancy 2013*	80.3	78.3	82.3
Change in period life expectancy 2003–2013	2.6	3.0	2.25
1913 Cohort life expectancy	64.9	61.8	68.1
1923 Cohort life expectancy	67.2	63.9	70.6
Change in cohort life expectancy (1913–1923 cohort)	2.3	2.1	2.5

Note: All measurements are in years.

*Closest period estimates (January 2003 and January 2013) to the perceived age estimates (November 2002 and December 2012) (28).

Mean perceived age in November 2002 was rated to be about half a year higher than the chronological age of the images taken in spring of 2001, whereas mean perceived age in December 2012 was rated to be almost 3 years older than the chronological age and 2.3 years older compared to the 2002 assessment (Table 1). Note that age perception might already have shifted between the time, the images were taken in spring 2001, and the first rating in November in 2002; such shift (fast adjustment in age perception) could explain the high mean perceived age ratings in 2002. The deviance between chronological and perceived age differed slightly more for women than for men among the years (Tables 1 and 2; Figure 1). The deviance between chronological and perceived age between the assessment years was well correlated (both sexes: 0.91; men: 0.92; women: 0.92). To this end, the main finding is that perception of age changed substantially over a decade in both men and women (Tables 1 and 2).

Table 2.

Statistical Test on Deviance Between chronological and Perceived Age as a Function of Sex, Rating Year (Year), and Chronological Age

Model #	Factors	Slope ± SE	Quadratic ± SE term	df	AIC
0	Null model (intercept only)	—	—	5	60,153.7
1	Sex	—	—	6	60,155.2
2	Year	—	—	6	60,152.1
3	Sex + year (additive)	—	—	7	60,153.7
4	Sex*year (interactive)	—	—	8	60,130.9
5	Chronological age	–0.66 ± 0.04	—	6	59,950.3
6	Sex + chronological age	–0.66 ± 0.04	—	7	59,951.6
7	Sex*chronological age	–0.59 ± 0.05 (women) –0.77 ± 0.08 (men)	—	8	59,951.7
8	Year + chronological age	–0.66 ± 0.04	—	7	59,948.8
9	Year*chronological age	–0.73 ± 0.04 (2002) –0.51 ± 0.01 (2012)	—	8	59,752.9
10	Sex + year + chronological age	–0.66 ± 0.04	—	8	59,950.1
11	Sex + year*chronological age	–0.73 ± 0.04 (2002) –0.52 ± 0.01 (2012)	—	9	59,754.2
12	Year + sex*chronological age	–0.59 ± 0.05 (women) –0.77 ± 0.08 (men)	—	9	59,950.2
13	Sex*year*chronological age	–0.65 ± 0.05 (women 2002) –0.85 ± 0.08 (men 2002) –0.46 ± 0.02 (women 2012) –0.60 ± 0.04 (men 2012)	—	12	59,745.7
14	Chronological age+(chronological age)2	–0.49 ± 1.18	–0.001 ± 0.008	7	59,960.2
15	Year*chronological age + (chronological age)2	–0.55 ± 1.18 (2002) –0.34 ± 0.02 (2012)	–0.001 ± 0.008	9	59,762.8
16	Sex*year*chronological age + (chronological age)2	–0.63 ± 0.46 (women 2002) –0.83 ± 0.08 (men 2002) –0.44 ± 0.19 (women 2012) –0.59 ± 0.03 (men 2012)	–0.0001 ± 0.008	13	59,755.7

Note: All models include the same random effects, rater ID, image ID, and rater age. Slopes and quadratic terms are only reported for the continuous factor chronological age. The main mean differences are reported in the main text and illustrated in Figures 1 and 2 and Supplementary Figure 1. Lower Akaike’s information criterion (AIC) values indicate better statistical support for the model. Bold AIC values highlight the best supported model with one, two, or three (or more) factors, respectively.

Figure 1.

Smoothed distribution of deviance between chronological and perceived age in 2002 and 2012 for (A) both sexes combined, (B) men, and (C) women. Positive deviances show higher perceived age compared to the chronological age whereas negative values show the opposite. The vertical dark gray dashed line marks zero deviance, ie, the exact mean chronological age, the dashed light grey line marks the mean deviance between the chronological age and the perceived age in 2002, and the darker grey hatched line marks the mean deviance between the chronological age and the perceived age in 2012.

The deviation between the chronological age and the perceived age depended on the actual chronological age in men and women (Figure 2; Table 2). Younger aged persons were more often rated to have higher perceived age than their chronological age (positive deviances), whereas older persons were more often rated to look younger than their chronological age (negative deviances). This chronological-age-specific shift was found for both men and women. The dependency on chronological age was more pronounced in 2002 than in 2012 (Year*Chronological age estimates, Table 2), and chronological age had a slightly stronger influence on how deviance changed between 2002 and 2012 in men compared to women. The relationship between the response variable (ie, the deviance between chronological and perceived age) and chronological age was linear. Models with curvilinear terms (quadratic terms) were not better supported than such with only linear terms (Table 2). Such chronological age-dependent age perception has been frequently described before for similar data and explained as regression toward the mean (27).

Figure 2.

Chronological age plotted against average perceived age for perceived age assessed in 2002 and 2012 for both sexes combined (A), men (B), and women (C). The dashed gray lines mark the theoretical one to one chronological to perceived age relationship. The thin gray lines connect the perceived age estimates on each image rated in 2002 and 2012. For better visualization of the data a jitter and small offset between the two ratings is added to the chronological age.

Because the interpretation and visualization of three-way interactions (sex*year*chronological age, Table 2) is challenging, we also illustrate the difference in the deviance between chronological and perceived age in 2002 and 2012 along with the chronological age axis (Supplementary Figure 2). Such difference in the deviance corresponds to the length of the thin grey lines in Figure 2; these lines increase in length with increasing chronological age. These differences in deviance not only depended on the chronological age but also differed among the sexes (women mean: 2.6 ± 0.2; men mean: 1.9 ± 0.2), but it is less clear whether there are important interactive effects between sex and chronological age, or curvilinear (quadratic) chronological age effects (Supplementary Table 2).

Our main interest here was on the shift in age perception across time, but not only the time (2002–2012) has changed, also the raters differed (4 raters were involved in both sessions of rating) and the individual perception among raters varies. To account for such variability and quantify its effect, we included the age, the rater ID, and the image ID as random effects in our statistical models. The latter was also done to account for nonindependence of rating the same image by multiple raters in 2002 and 2012. The largest fraction of variance in the deviance between perceived and chronological age is found related to the image ID, that is, the difference in age perception among different raters rating the same image (variability in rated age within images; 39% random effect variance of image ID). This uncertainty in rating the age is not determined by the raters as the variance explained by the different raters is separated out and accounts for 14% (random effect variance raters; Supplementary Figure 5). These 14% illustrate that some raters tend to overestimate age and some tend to underestimate age, but compared to the 39% within image variability, it shows that the uncertainty of rating the age mainly arises within the raters. The age of the raters explained another 11%, and the remaining (residual) variance explained 36%. The main (fixed) effects of sex, chronological age, or the assessment year were robust and did not depend on the combination of the three random effects included. Overall, the greatest uncertainty (39%) comes from within individuals rating the age of a person (image), followed by the tendency of some raters to over- or underrate ages (14%).

Discussion

We show, as hypothesized, that the perception of age changes across time, which implies that age perception adjusts to the shift in biological age or at least the biological age to chronological age conversion changes over time. For instance, a person aged 70 years in 2012 needed to look substantially younger (2.3 years) to be perceived the same age as a 70-year-old in 2002. Therefore, biological age—derived from perceived age—changed by 2.3 years over the 10 year period. This change is close to the change in life expectancy at birth in Denmark over that same period, which has risen by 2.6 years (Table 1) (28). However, such similarities might not be meaningful as life expectancy at birth is a period estimate across all age classes and not only limited to older cohorts. More interesting is that cohort-based life expectancy changed by similar amounts (Table 1) (28). Unfortunately, cohort life expectancies for later cohorts that build the bulk of our study sample are not yet available as many members of more recent cohorts are still alive (28). Nonetheless, it is not expected that shifts in cohort life expectancy for more than 70 years old over that decade would be significantly different.

Even though the shift in perceived age over the investigated decade agrees well with shifts in life expectancy, a more relevant measure for comparison might be the change in remaining cohort life expectancy, for example, how many additional years can a 70-year-old of the 1942 cohort (being 70 in 2013) expect to live compared to a 70-year-old from the 1932 cohort. Remaining cohort life expectancy has risen between 2003 and 2013 for 70-year-old men and women by 0.84 years, for 80-year-old women by 0.72 years and for men by 0.84 years; and for 90-year-old women by 0.42 years and for men by 0.3 years (29). Hence, perceived age has changed much more than remaining life expectancy and is more closely associated with shifts in cohort life expectancy at birth.

Our results suggest that age perception reflects past life events better than it predicts future length of life. This implies that the past aging history of an individual, such as physiological aging processes of accumulation of oxidative damage, is better reflected in the facial features than the future potential of life, hazard risks, and the resulting remaining life expectancy. Under this point of view, we might want to consider how survivorship, the probability of living to a certain age, has changed over the 10 year time period. Survivorship to age 70 in 2003, that is, proportion of the 1932 cohort living to at least age 70, was 0.68 in women, 0.57 in men, and 0.63 when combining both sexes (28). Similar survivorships of individuals born 10 years later, that is, comparable survivorships of the 1942 cohort, were associated with 76.1-year-old women, 75.6-year-old men, and 75.8 years old when both sexes are combined (polynomial projections after age 71 based on 1942 cohort death rates extracted from the Human Mortality database; (28)). We observe larger shifts (5.6–6.1 years) among similar survivorship levels over the 10 year period compared to the perceived age shifts. Hence, age perception does not directly relate to shifts in accumulated survival but rather to changes in life expectancy at birth. Substantially, more people reached age 70 in the 1942 cohort compared to the 1932 cohort. Our sample only included individuals aged more than 70 years. Hence, the people whose facial images were used in our study had already passed substantial parts of their life and progressed a good part along their individual aging trajectories. These past trajectories seem to leave their marks in the faces and dominate the age perception. Such findings support our argument that perceived age accurately captures the biological age of an individual as described by how far the body has declined in function over its past life.

Taking our results together, we find ourselves in a somewhat conflicting situation in that our average age rating is quite accurate and adjusts over time, but at the same time reveals a chronological age-related bias. This age bias, overestimating the age of younger people and underestimating older people, could be explained by regression toward the mean, or by selective processes as related to frailty arguments, where the most frail and oldest looking individuals die first and only the most robust and young looking individuals make it to old age (30). Regression toward the mean is an effect where raters consciously or subconsciously realize the age range of the sample images and subsequently tend to rate closer to the sample mean, that is, they overestimate younger individuals and underestimate older individuals. If regression to the mean occurs, we expect that the variance in age perception to decrease with increasing numbers of images rated; however, such convergence to the mean would lead to increased variance of the deviance between perceived and chronological age with increasing numbers of images rated. At the onset of rating images, the raters have not yet developed a sense of what the age range of people they rate on is, such sense will develop with increasing number of images rated on and raters should regress more and more toward the mean. However, we did not detect such pattern of convergence toward the mean with increasing number of images rated on, the variance in deviance does not increase (Supplementary Figure 4). As for other explanations, frailty models suggest that in an aging cohort, the population is composed more and more of robust individuals because the most frail individuals are more likely to die earlier in life. Such arguments have been used to explain why humans approach a mortality plateau at very old ages (31). If this type of selection bias occurs between the ages of 70 and 90 years, it could explain the chronological age-specific biases we observe in our study. Age perception seems to adjust over time, but it is less clear why raters would not be able to adjust to a selective bias as explained by frailty models. Such adjustment to the selective bias should happen for all cohorts and not just for specific cohorts.

Even though we find that the average perceived age approaches the average chronological age across all images rated, the power of perceived age as a biomarker of aging might rather be seen in its potential to explain deviances from chronological age. A biomarker of aging that only reveals chronological age is relatively uninformative, except for special cases when one does not know the chronological age of an individual as for example, for a blood spot in a crime scene or for wild animals (11,32). Perceived age holds great potential as a biomarker for biological age because biomarkers of aging should inform about biological age differences beyond chronological age, and because perceived age is moderately correlated to chronological age, it shows the potential to provide exactly such additional information (33). This potential of perceived age as biomarker is confirmed by previous findings that show that perceived age corresponds to health measures, cognitive function, and various hazard risks and might be a better predictor of mortality hazard compared to chronological age (21–24). Our study focuses on perceived age and does not include health measures or mortality hazard risk (beyond relating to the overall change in life expectancy). This lack of information might be limiting our conclusions. However, the 382 individuals this study is based on are a subsample of the 1,826 individuals of a previous study that showed how perceived age predicts mortality risk to a larger degree than chronological age (21). That study also illustrates how perceived age correlates well with other biomarkers of aging (21). The 382 individuals were those for which high-quality images were available and that build intact pairs of twins (27,34).

Our study shows that rating the age of an individual is associated with substantial uncertainty (Supplementary Figure 5); therefore, perceived age should always be rated by multiple raters to overcome individual tendencies of raters and uncertainties of rating a given person. Previous studies showed that ratings on perceived age were little biased by characteristics of raters with respect to age and sex, these studies also showed that ratings on perceived age based on 10 or 20 raters were accurate and comparable (21). Our main finding lays in the substantial shift that age perception underwent between 2002 and 2012, which was observed in almost all individuals (Supplementary Figure 3). This degree of generality is remarkable and suggests that the perception of biological age does not rely on specific characteristics only exhibited by a number of individuals, but rather that perceived age is a general and integrated biomarker of aging with the contemporary background population as the reference.

Comparing our findings to other investigations on shifts in biological age across time reveals important similarities and some differences. Levine and Crimmins (35) for instance investigated shifts in biological age over a 20 year time period in the United States. They used chronological age and eight biomarkers on a large sample to calculate biological age based on a frequently applied algorithm (36). They found overall shifts in biological age (~4 years, ~1.8 years per decade), which relates to the shift in US life expectancy at birth over that time period of 3.77 years. Contrasting our results, they showed that men improved more than women, whereas in our study men improved slightly less than women. Levine and Crimmins partly explained these sex-specific differences by sex-specific changes in behaviors (smoking, obesity, and medication) in the United States over their investigated time period. In agreement with our data, they also found greater improvement (greater decline in biological age) with increasing age, though their data span a much larger age range compared to ours. The general trends in shifting biological age that can be shown using different biomarkers highlight the potential powerful insights we can gain on altered aging processes, though the details associated with increased life expectancy and health spans remain largely to be explored, as need differences among populations to be considered (20).

One of the riddles behind increased life expectancy is whether it occurs because aging is delayed, that is, the deterioration in function starts at a later age in more recent cohorts the deterioration process and therefore aging is slowed down, or individuals begin adult life at lower biological ages. A slow downed aging process, or reduced rate of aging, would be expressed in a slower deterioration process in more recent cohorts, which relates to the concept that the rate of aging changes over time. If rate of aging changes, it should also leave its mark on biological age (2). Our cross-sectional approach is not well suited to powerfully distinguish between a slow downed aging process or delayed aging process. If the biological aging process would be delayed and not slowed, we would expect to see a shift in the intercept in the deviance between chronological and perceived age between 2002 and 2012 (Figure 2). We see a substantial shift in the mean perceived age that could suggest a shift as predicted by delayed aging, but we also see a change in the slope of the deviance between 2002 and 2012 with respect to chronological age (Figure 2 and Supplementary Figure 2), which could indicate a change in the rate of aging. Future investigations of shifts in perceived age, assessed over longer time periods (multiple cohorts) and not limited to older individuals, could provide much better insights to resolve the controversy about the altered process of aging than the data we analyzed (2,35).

Conclusions

We show that perceived age evaluated by multiple raters is likely a good and general biomarker of aging that can be used to track changes in biological rather than chronological age. Perceived age reflects past aging processes better than it predicts future outcomes of aging. Extending perceived age investigations to younger ages (<70), as suggested in the context of other biomarkers of ageing (24), would provide better means to shed light on a variety of controversial topics in demography and ageing research, including whether increased life expectancy is achieved by delayed aging, by lower starting levels of biological age, or slowed aging. The advantage of a direct comparison of age perception shifts and changing life expectancy makes perceived age a very interesting and useful biomarker that is in addition quite cost-efficient. However, this direct comparison also challenges us for interpretation of results. For instance, we would not be able to differentiate whether shifts in perceived age relate to changes in remaining cohort life expectancy or to changes in life expectancy at birth. We might want to aim for the use of qualitative biomarkers that can be directly related to other well recorded shifts in demographic parameters for advancing our understanding of the aging process and its consequences across time—perceived age might be one of them.

Funding

This work was supported by the Max-Planck Society; within the Danish Aging Research Center by the VELUX Foundation; and the Longitudinal Study of Aging Danish Twins within which the initial photos were collected, this collection was based on grants from the US National Institutes of Health (NIA P01 AG008761).

Ethical Approval

The study was approved by the regional scientific ethical committee in Denmark (Case No VF20040241).

Data Availability

Data cannot be made available for legal reasons. The informed consent does not comprise making the individual level data available and sharing those with other researchers.

Conflict of Interest Statement

The authors have declared that no competing interests exist.