Title: Terminal Decline and Ergodicity in Life Satisfaction
1Terminal Decline and Ergodicity in Life
Satisfaction
- Ryne Estabrook
- September 13, 2007
- Design and Data Analysis
Gerstorf, D., Ram, N., Estabrook. R., Schupp, J.,
Wagner, G.G. \ Lindenberger, U. (2008). Life
Satisfaction Shows Terminal Decline in Old Age
Longitudinal Evidence from the German
Socio-Economic Panel Study (SOEP). Developmental
Psychology, 44, 4, 1148-1159.
2Terminal Decline
Terminal Decline and Ergodicity In Life
Satisfaction
- Terminal decline has long been studied as a
late-life cognitive phenomenon. - Kleemeier, 1962 see Bäckman MacDonald, 2006
for review. - Terminal decline can be defined in a number of
ways - All definitions describe negative (or more
negative) change. - Some differentiate between terminal drop
(accelerated or curvilinear declines) and
terminal decline (level differences). - Often a methodological distinction.
- These two phenomena often termed together as
terminal change.
3Terminal Decline
Terminal Decline and Ergodicity In Life
Satisfaction
- Evidence of terminal decline over a range of
cognitive functions. - Areas of cognitive functioning relatively
unaffected by age show the most pronounced and
consistent terminal decline effects. - White and Cunningham (1988) hypothesis.
- Vocabulary, crystallized abilities, episodic
working memories. - Differences between survivors and decedents found
anywhere from two to twelve years prior to death. - Rationales include pre-clinical dementia and
cardiovascular functioning. - Primary and secondary aging (Busse, 1969).
- No difference between biological and experimental
mortality.
4Multivariate Terminal Decline
Terminal Decline and Ergodicity In Life
Satisfaction
- Religious Orders Study (Wilson, et al, 2003)
- Found terminal decline in episodic memory,
working memory, speed, and general cognitive
abilities. - Terminal decline found in virtually all
individuals, but at highly variable rates. -
- Individual aging (Lövdén, et al, 2005)
- Used cluster analysis to characterize several
response profiles. - Shows terminal decline in block design, episodic
memory semantic memory as transitions through
programs. - Both characterize individually varying forms of
terminal decline.
5Non-Cognitive Decline
Terminal Decline and Ergodicity In Life
Satisfaction
- Terminal change in neuroticism
- Neuroticism shows long-term changes in adulthood.
- These changes predict mortality (Mroczek Spiro,
2005). - Terminal decline also noted in self-reported
health depression. - (Kaplan Camacho, 1983 Idler Angel, 1990
Wolinsky Johnson, 1992 Berg, 1996). - Terminal decline in well-being
- Self-reported well being showed stronger declines
over distance to death than chronological age. - Population-level change point of 4 years prior to
death identified. - (Gerstorf, Ram, Röcke, Lindenberger, Smith,
2007).
6Life Satisfaction
Terminal Decline and Ergodicity In Life
Satisfaction
- Life satisfaction is an interesting construct for
terminal decline. - Life satisfaction measures are typically very
stable in late adulthood - (Diener, Lucas, Scollon, 2006 Filipp, 1996
Horley Lavery, 1995 Kunzmann, Little, Smith,
2000). - Well being may be both a consequence and source
of successful aging. - (Bales Baltes, 1990).
- Socio-emotional selectivity theory would not
terminal decline in life satisfaction. - (Carstensen, L.L., Isaacowitz, D.M., Charles,
S.T.,1999)
7Life Satisfaction
Terminal Decline and Ergodicity In Life
Satisfaction
- If terminal decline in life satisfaction is
exhibited - Supports previous studies on terminal decline in
well-being. - Paves way for a gestalt view of terminal decline.
- Allows for future research in cognitive-emotive
links in tertiary aging. - If terminal decline in life satisfaction is not
exhibited - Supports socio-emotional selectivity theory and
other theories prescribing constancy in adult
personality.
8Life Satisfaction Data
Terminal Decline and Ergodicity In Life
Satisfaction
- German Socio-Economic Panel Study (SOEP).
- Annual nationally-representative panel survey of
German population. - Data collected annually from 1984-2005.
- Participants recruited via random walk method.
- Recruitment in 1984, 1990, 1995, 1998, 2000,
2002. - Participants approached in representative public
places. - 60-70 response rate.
- 4-14 attrition rate (small gifts results for
participation). - Face-to-face interviews
- Mail questionnaires for 10 of multi-occasion
respondents.
9Life Satisfaction Data
Terminal Decline and Ergodicity In Life
Satisfaction
- Mortality Sample
- Must be 70 at the first measurement occasion,
- Must provide mortality information (N1637).
- Age
- Chronological Age
- Mean 78.20
- SD 5.88
- Range 70-100
- Distance-to-Death
- Mean -6.32
- SD 4.22
- Range -22 to -1
10Life Satisfaction Data
Terminal Decline and Ergodicity In Life
Satisfaction
- Life Satisfaction
- Wie zufrieden sind Sie gegenwärtig, alles in
allem, mit ihrem Leben? - How satisfied are you with your life, all
things considered? - 0-10 Likert Scale (0unsatisfied, 10very
satistfied). - Standardized to a T-metric on full sample.
- Full Sample Mean 7.02, SD 1.55.
- Mortality Sample Mean 6.75, SD1.57.
11Research Questions
Terminal Decline and Ergodicity In Life
Satisfaction
- Is the data better explained by chronological
aging or distance-to-death? - Is there evidence of terminal change?
- Are there individual differences in levels and
rates of change in life satisfaction? - Are individual differences in changes in life
satisfaction related to one another?
12Research Question 1
Terminal Decline and Ergodicity In Life
Satisfaction
- Is the data better explained by chronological
aging or distance-to-death? - Is there evidence of terminal change?
- Compare fit of mixed effects models from life
satisfaction across chronological and death ages. - Fit multiple functional forms
- Linear Mixed Effects
- Quadratic Mixed Effects
- Linear-Linear Spline Mixed Effects
- Fixed Change Point
- SAS PROC MIXED/NLMIXED
13Multi-Phase Mixed Effects
Terminal Decline and Ergodicity In Life
Satisfaction
- Alternatively, (linear-linear) spline model.
- (Cudeck Harring, 2007 Cudeck Klebe, 2002
Willett Singer, 2003 Wang McArdle, in
press). - Fixed Effects
- LifeSat ß0 ß1(time-ß3) e, timelt ß3
- LifeSat ß0 ß2(time-ß3) e, timegt ß3
- Benefits
- Better characterizes multiple change processes.
- More easily interpretable than higher order
parametric models. - Drawbacks
- Random change points are difficult to fit
temporarily omitted.
14Terminal Decline and Ergodicity In Life
Satisfaction
Chronological Age
Life-Satisfaction by Chronological Age, Full
Sample
15Terminal Decline and Ergodicity In Life
Satisfaction
Chronological Age
Life-Satisfaction by Chronological Age, N150
16Chronological Age
Terminal Decline and Ergodicity In Life
Satisfaction
- All models provided much improved fit over a
no-change model. - All models feature (increasingly) negative slope
over time. - Linear-linear spline provides the best model fit.
17Terminal Decline and Ergodicity In Life
Satisfaction
Distance-to-Death
Life-Satisfaction by Distance-to-Death, Full
Sample
18Terminal Decline and Ergodicity In Life
Satisfaction
Distance-to-Death
Life-Satisfaction by Distance-to-Death, N150
19Distance-to-Death
Terminal Decline and Ergodicity In Life
Satisfaction
- Linear-linear spline again provides the best
model fit. - All distance-to-death analyses show superior fit
to parallel chronological age analyses. - Evidence of declines related to distance-to-death
rather than chronological age.
20Research Question 2
Terminal Decline and Ergodicity In Life
Satisfaction
- Are there individual differences in rates of
change in life satisfaction? - Are individual differences in changes in life
satisfaction related to one another? - Compare fit of fixed effect and mixed models
- Linear-Linear Spline Mixed Effects
- Fixed Change Point
- SAS PROC NLMIXED
21Multi-Phase Mixed Effects
Terminal Decline and Ergodicity In Life
Satisfaction
- Random effects significantly improve model fit.
- Log likelihood feasible because models are
nested.
22Multi-Phase Mixed Effects
Terminal Decline and Ergodicity In Life
Satisfaction
Random Effects Covariance Matrix for Multi-Phase
Growth Model on Distance-to-Death in Life
Satisfaction
plt.001
23Multi-Phase Mixed Effects
Terminal Decline and Ergodicity In Life
Satisfaction
Y48.14-0.64(time), time lt -4.19 Y48.14-1.94(ti
me), time gt -4.19
24Results
Terminal Decline and Ergodicity In Life
Satisfaction
- Distance-to-death models provided better fit than
all comparable chronological age models. - Life satisfaction shows sharper decreases closer
to death. - Mixed-effects models indicate individual
variability in declines and relationships between
level and rates of change. - No relationship between pre-terminal and terminal
slopes.
25Random Change Points
Terminal Decline and Ergodicity In Life
Satisfaction
- Presented models include random effects for
intercept and both linear slopes, but a fixed
change point. - Random change points are notoriously difficult to
estimate with standard methods. - Usually fit with SAS PROC NLMIXED.
- Include First-Order Taylor Series and adaptive
Gausian quadrature methods. - (Cudeck Klebe, 2002 Cudeck Du Toit, 2003
Hall et al., 2000). - Sensitive to initial values
- Inability to satisfy convergence criteria.
- Biased point and standard error estimates
26Bayesian Estimation
Terminal Decline and Ergodicity In Life
Satisfaction
- Bayesian Estimation with Gibbs Sampling
- Identical fixed effects model, random effects for
all parameters. - (Wang and McArdle, in press)
- Benefits
- Estimable.
- Less sensitive to misspecification of starting
values. - Drawbacks
- Incomparable to other types of models.
- Potentially sensitive to poor priors.
- Time.
27Random Change Points
Terminal Decline and Ergodicity In Life
Satisfaction
- Changes to Model
- Inclusion of random change points
- Appropriate additions to parameter covariance
matrix - Non-informative priors
- Changes to Dataset
- Dataset restricted to 400 individuals with 12 or
more observations. - Individuals with few occasions tended to be
mischaracterized by the model, particularly with
change points beyond their time series. - Population parameters do not significantly differ
with this restriction.
28Terminal Decline and Ergodicity In Life
Satisfaction
Random Change Points
- Inclusion of random change points improves model
fit. - While the DIC is analogous to AIC, the two cannot
be directly compared, preventing direct fit
comparisons between the Bayesian and Gaussian
models.
29Random Change Point
Terminal Decline and Ergodicity In Life
Satisfaction
- Gaussian method is on same (restricted) sample as
Bayesian models. - Use of Bayesian model results as initial values
for SAS PROC Mixed did not yield convergence.
30Terminal Decline and Ergodicity In Life
Satisfaction
Random Change Points
31Random Change Point
Terminal Decline and Ergodicity In Life
Satisfaction
- Inclusion of random change points improves model
fit. - Random change point model has the lowest residual
variance of any model so far. - Change points has significant variance and
covariances with other parameters. - Parameters changed with the inclusion of random
change points - Terminal slope is .85 T-units steeper.
- Change 1 unit on Likert scale every 7 years.
- Alternatively, .5 units for the average decline
phase. - Average change point is 1 year closer to death.
32Terminal Slope Question
Terminal Decline and Ergodicity In Life
Satisfaction
- Why did the terminal slope decrease with the
inclusion of random change points? - Misspecification problem
- Would variation in change points lead to
misspecification of the terminal slope? - Ergodic problem
- What if some individuals are not undergoing
linear-linear change?
33Misspecification Problem
Terminal Decline and Ergodicity In Life
Satisfaction
- How does fixing the change point affect the
terminal slopes? - Misspecification of change points at the
individual level should bias slope estimates. - For individuals with change points closer to
death than the population (fixed) change point - Terminal slope will be biased in the direction of
the pre-terminal slope. - Pre-terminal slope estimates will be unaffected.
- For individuals with change points further from
death than the population (fixed) change point - Pre-terminal slope estimates will be biased in
the direction of the terminal slope. - Terminal slopes will be unaffected.
- Such bias will affected fixed parameters.
34Terminal Decline and Ergodicity In Life
Satisfaction
Misspecification Problem
Intercept N(109.84, 9.25) Pre-Slope N(-1.019,
0.07) Post-Slope N(-5.035, 1.41) Change Point
N(-12.44, 8.70)
35Terminal Decline and Ergodicity In Life
Satisfaction
Misspecification Problem
- Fixed change point model yields biased estimate
of pre-terminal slope, terminal slope and change
point. - Intercept unaffected.
true value outside of confidence interval
36Misspecification Problem
Terminal Decline and Ergodicity In Life
Satisfaction
- Both pre-terminal and terminal slopes biased.
- As theorized, each biased in the direction of the
other. - This created a covariance between the slopes
- rmodel .395
- rdata -.112.
- Location of change point biased as well.
- Change point closer to center of time-series
(t-15) - True -12.43, Sampled -12.72.
- True -15.40, Sampled -15.37
- True -18.30,Sampled -17.62.
- Intercept unaffected.
37Ergodic Problem
Terminal Decline and Ergodicity In Life
Satisfaction
- Assumptions of ergodicity are critical to any
model, particularly mixed-effects models. - This assumption stipulates that processes
occurring at the population level are
representative of individuals in that population. - In our dataset, several types of ergodic problems
may occur - Classic case some individuals arent
characterized by spline model. - Competing risks problem some individuals would
undergo terminal change, but dont get the chance
to. - Misspecification problem (fixed change point, for
example).
38Ergodic Problem
Terminal Decline and Ergodicity In Life
Satisfaction
- Evidence of ergodic problems in Life Satisfaction
data - µchange point -3.54
- s change point 4.90
- Some individuals have change points outside of
their time series. - Possible causes
- Some individuals are better characterized by
linear change. - Artifact created by distribution specified in
prior. - Carry out a final simulation, where the
distribution of change points extends beyond the
observed data. - Most closely mirrors the existing data.
- Compare fixed and random change-point models.
39Terminal Decline and Ergodicity In Life
Satisfaction
Ergodic Problem
- Fixed change point model yields biased estimates
of terminal change and change point fixed
effects. - Intercept and pre-terminal slope unaffected.
true value outside of confidence interval
40Ergodic Problem
Terminal Decline and Ergodicity In Life
Satisfaction
- As predicted, terminal slopes and change points
are biased. - Similarities between this simulation and observed
data suggest problems found in analyses are, at
least in part, artifactual. - Not definitive proof that random model is
completely unbiased. - We should exercise caution in analyzing
relationships between model parameters, either
interrelations or relationships to covariates. - This is not a solution to more traditional
ergodic problems. - Some individuals might be better characterized by
continuous change. - Some spline models may capitalize on residual
structures.
41Substantive Interpretations
Terminal Decline and Ergodicity In Life
Satisfaction
- Life satisfaction shows evidence of terminal
decline. - All population models fit significantly better
over distance to death than over chronological
age. - Spline models, particularly mixed effect spline
models, provided superior fit, suggesting
pre-terminal and terminal decline phases. - Terminal decline in life satisfaction shows
individual differences. - Intercepts, rates of change and locations of
change point all show individual differences and
covariances. - Models that dont allow for individual
differences, particularly in change point, show
misfit and multiple patterns of bias.
42Substantive Issues
Terminal Decline and Ergodicity In Life
Satisfaction
- Violations of ergodic assumption create problems
for terminal decline research. - Terminal decline faces additional issues from
competing risks. - Studies of terminal decline may face an difficult
problem. - Change point estimation will tend to be biased
towards the center of data collection. - There is a limit to the number of observations
between change point and death. - This effect is more pronounced if individuals
with few occasions cannot be included. - Increasing observations typically means fewer
participants, which is problematic for an
individually-varying phenomenon. - Terminal change may take on a more complicated
functional form that power and degrees of freedom
issues may obscure.
43Substantive Directions
Terminal Decline and Ergodicity In Life
Satisfaction
- Multivariate approaches to terminal decline
- Cognitive, emotion and personality.
- Cognitive-emotive links in terminal decline.
- If terminal decline occurs across psychological
domains, it stands to reason they may occur
together. - Alternate ways of conceptualizing aging time.
- Both chronological age and distance-to-death are
expressions of time in relation to an
individually-varying event (birth, death). - Refinement of model
- Still characterize change from one process to
another. - Create a model that functions conceptually in
continuous time.
44Methodological Directions
Terminal Decline and Ergodicity In Life
Satisfaction
- Refinement of the spline model to avoid the
problems detailed here. - Avoid biased parameter estimates.
- Particularly, get valid estimates of change point
distribution. - Create methods to distinguish between different
types of change. - Sensitivity to individuals yet to undergo
terminal change.
45Acknowledgements
Terminal Decline and Ergodicity In Life
Satisfaction
- Denis Gerstof, Penn State/Max Planck Institute
- Nilam Ram, Penn State/Max Planck Institute
- Jürgen Schupp, German Socio-Economic Panel Study
- Gert G. Wagner, German Socio-Economic Panel Study
- Ulman Lindenberger, Max Planck Institute
- Peggy Wang, University of Virginia
- Johnny Zhang, University of Virginia
- John Nesselroade, University of Virginia
- NIA T32 AG20500-01