5.3
5.3
5.3
The structural equationsThe structural equations The structural equationsThe structural equations
Structural parameters of the health and engagement equations (4) and (6) are
provided in Table 4. The latent poor-health index (ββ) is higher for non-white women and is non-linearly associated with age of the respondent.22 SES-related
differences in health are clearly evident by looking at the magnitude and signifi-
cance of the coefficients associated with educational attainments, social class and
home-ownership. Highly educated individuals have significantly better average
health than otherwise similar less educated people. Similarly, high social class
and housing wealth are also associated with better health.
T TT
TABLE ABLE ABLE ABLE 4:4:4:4:Structural parameters of the health (ββ) and engagement (πβ) equa-
tions
Covariate β
β πβ
Coefficient S.E Coefficient S.E
Female 0.050** 0.024 -0.260*** 0.045
Spline age 65-74 -0.006* 0.001 -0.016 0.011
Spline from age 74+ a 0.017*** 0.003 -0.042*** 0.008
Married/cohabiting 0.032 0.028 0.149** 0.066
Completed education before 14 years old 0.186*** 0.028 -0.055 0.066 Completed education after 19 years old -0.245*** 0.051 -0.172 0.117
Non-white 0.543*** 0.083 -0.865*** 0.154
Home-owner -0.302*** 0.031 0.137* 0.072
Social class: manual worker 0.177*** 0.027 -0.187*** 0.061
Notes: * p < 0.05, ** p < 0.01, *** p < 0.001. (a) collapsed at 90. Estimates in the table refers
to model D specification (see later in the text). The model also includes dummy variables on region of residence, if living in urban area (>10K), and from which HSE cross-section the eligible person was drawn. Standard errors were clustered at household level.
22 Sample membersβ age enters in all structural equations in the form of a spline with a knot at the
median age (74). It accounts for possible non-linearities in its relation with health (equation (4)), engagement (equation (6)) and retention to the study (equation (2)).
46
Engagement with the scope of the survey (πβ) is lower for older non-white women and higher for partnered individuals. One might expect a SES-engage-
ment gradient, in line with the idea that high SES individuals are more likely to
perceive the social benefits of complying with the scope of the survey (Uhrig,
2008), but our results are not consistent with that. Home-owners show a higher
level of engagement (significant at 10% level) and low social class is negatively
associated with engagement (p-value <0.001) but, surprisingly, we do not find
any significant association of πβ with level of education. This result is consistent with findings in Jenkins
et al.
(2006) where education was not significantly asso-47 T
T T
TABLE ABLE ABLE ABLE 5:5:5:5:Estimates of retention in the survey in wave 1 of ELSA
Covariate Model A Model B Model C Model D
Coefficient S.E Coefficient S.E Coefficient S.E Coefficient S.E
Female -0.107*** 0.023 -0.111*** 0.023 0.001 0.032 -0.002 0.032
Spline age 65-74 0.001 0.007 0.001 0.007 0.011 0.009 0.010 0.008
Spline from age 74+ a -0.027*** 0.004 -0.027*** 0.004 -0.012** 0.006 -0.012** 0.006
Married/cohabiting -0.079** 0.036 -0.081** 0.036 -0.175*** 0.048 -0.177*** 0.048
Completed education before 14 years old -0.096*** 0.036 -0.095*** 0.036 -0.099** 0.048 -0.097** 0.048
Completed education after 19 years old 0.082 0.064 0.087 0.065 0.208** 0.082 0.211** 0.082
Non-white -0.230** 0.102 -0.226** 0.102 0.155 0.137 0.157 0.137
Home-owner -0.001 0.041 -0.003 0.041 -0.061 0.052 -0.063 0.052
Social class: manual worker -0.104*** 0.033 -0.104*** 0.033 -0.039 0.044 -0.040 0.044
ββ -0.042*** 0.016 -0.047*** 0.016 -0.016 0.021 -0.023 0.021 (ββ)2 -0.030** 0.013 -0.026* 0.015 πβ 0.521*** 0.030 0.520*** 0.03 πππ£(ββ, πβ) -0.072** 0.030 -0.073** 0.03 Free parameters 45 46 55 56 Log-likelihood -49783.925 -40620.272 -49434.928 -49433.378
Correction for non-normality factor 1.1286 1.548 1.1333 1.1321
AIC 99769.85 81452.544 99075.856 99074.756
BIC 100480.725 82198.611 99800.808 99806.746
Notes: * p < 0.05, ** p < 0.01, *** p < 0.001. (a) collapsed at 90. All models also include dummy variables on region of residence, if living in urban area
48
Structural parameters of the retention model (equation (2)) are provided in
Table 5 in four different variants.
Model A
is a reduced version in which latenthealth enters linearly and we do not control for latent engagement.
Model B
introduces a quadratic term for health aiming at testing its possible non-linear
relationship with retention in wave 1.
Models C
andD
introduce latent engage-ment in the set of covariates when ββ is entered linearly and in a quadratic form. For model A, we found results in line with previous research. Older married
women are less likely to remain in the study as well as the non-white population.
Lower education and social class are negatively associated with participation in
wave 1 of ELSA but we did not find any significant relationship with home own-
ership.
In model A, there is a positive relationship between health and survey partici-
pation: poor-health significantly reduces the likelihood of retention in wave 1 but
model B, which fits the data slightly better than model A, reveals a significant
U-shape relationship, meaning that, while unhealthy individuals are less likely to
remain in the study at follow-up, very healthy individuals do not show signifi-
cantly lower dropout probabilities with respect to the remaining sample popula-
tion. This is consistent with the view that healthy people might have concerns
about the time cost of participating and in taking part in medical and cognitive
tests which, apart from being time-consuming, may also be felt to be humiliating.
49
for retention when latent health enters in the model linearly and in a quadratic
form, respectively. Both specifications show that the latent engagement index
plays the most important role in explaining retention decision for wave 1, by
increasing the explained variance significantly, as shown from the goodness of fit
statistics available at the bottom of Table 5.
Controlling for engagement, the effect of other covariates is significantly weak-
ened. Gender, race, home-ownership and social class no longer play a significant
effect in explaining retention in the study. The coefficient associated with the
second spline of age23 is almost halved, whereas the effect of being partnered is
now about 2.2 times higher those obtained in models A and B. Similarly, the
effect of level of education is significantly increased, with an emerging significant
retention bias in favour of those more educated.
Controlling for engagement, latent health is no longer significantly related to
wave 1 participation when its effect is assumed to be linear (model C). The coef-
ficient of the quadratic term of ββ in model D is now significant only at the 10% level. Both specifications provide evidence that, controlling for the individualβs
engagement at baseline, health plays a minor role in explaining participation in
wave 1.
There is a small but significant correlation between ββ and πβ of about -0.07,
23 See previous footnote for a definition.
50
p<0.05, consistent with the idea that people in poor health face higher costs in
retrieving all the information required in formulating a response or are more
sensitive to privacy concerns (Beatty & Herrmann, 2002; Dunn
et al.
, 2004; Jen-kins
et al.
, 2006; Khoet al.
, 2009) and are less engaged. This has to be taken intoaccount when drawing conclusions from estimates in Table 5. A series of robust-
ness checks are confined in the Appendix. For example, constraining the correla-
tion between the two latent constructs to zero increases the importance of ββ in explaining retention in wave 1 but leaves other structural coefficients virtually
unaffected (see Appendix Table A1).