Diapositive 1 - Étude longitudinale du développement des enfants ...

bagimpertinentΠολεοδομικά Έργα

16 Νοε 2013 (πριν από 3 χρόνια και 9 μήνες)

62 εμφανίσεις

Lost but not forgotten : attrition in the Étude longitudinale du développement des
enfants du Québec (ÉLDEQ), 1998
-
2004

Julien BÉRARD
-
CHAGNON and Simona BIGNAMI
-
VAN ASSCHE


Département de démographie

Project financed by the Fonds québécois de recherche sur la société et la culture (FQRSC)

Background


Longitudinal surveys are increasingly used in the social sciences to describe behaviour dynamics, to identify the influence o
f
past on current behaviours, and to make
stronger causal inferences than it is possible with cross
-
sectional surveys.


In demography, longitudinal data are particularly relevant for the study of family transitions and life course analysis.


One of the main weaknesses of longitudinal surveys is that they are prone to attrition, that is, the loss of study subjects o
ve
r time that is due to respondents leaving the
study prematurely and permanently.

Why study survey attrition ?



Reduces the sample size (thus reducing the
power of statistical estimations).


Makes the study of small sub
-
samples harder.


Affects sample representativeness.


Might lead to selection bias.

Data


ÉLDEQ: Ongoing longitudinal survey aimed at studying
the factors influencing child development in Québec for a
cohort of 2120 children born in 1997
-
98.


The ÉLDEQ survey team has made lots of efforts to
minimize attrition and track down attritors.


Selected sample for the analysis: waves 1
-
8.

Attrition in the ÉLDEQ

Cumulative attrition rate (%), waves 1-8,
ƒLDEQ
0
5
10
15
20
25
30
1
2
3
4
5
6
7
8

Attrition is measured by the participation
to the main survey instrument (QIRI).


Attrition was very low in the first phase of
the survey (waves 1
-
5, before the sampled
children entered primary school), but high in
the second phase (waves 6
-
8).


Many factors (uncertainty about the
future of the survey, increasing length of
interviews, etc.) explain this finding.


Attritors’ characteristics


Attritors have different characteristics than the other respondents in the sample.


Overall, attritors distinguish themselves from non
-
attritors by their characteristics associated
with instability, poverty, immigration and social exclusion.


Type of family arrangement at the time of the survey for three
attrition groups (
p
<0,05)
71,4%
82,8%
82,8%
12,7%
7,4%
10,7%
15,9%
9,8%
6,5%
A
B
C
Intact family
Step-parents family
Single-parent family
Mother's first spoken language for three attrition
groups (
p
<0,01)
50,8%
81,0%
84,9%
7,9%
7,4%
7,4%
41,3%
11,6%
7,7%
A
B
C
French
English
Other
A : Wave 1 attritors

B : Wave 5 attritors

C : Non
-
attritors

Mother's highest diploma for three attrition groups (p<0,01)
32,3%
16,5%
15,1%
25,8%
27,3%
25,4%
21,0%
35,5%
29,6%
21,0%
20,7%
29,9%
A
B
C
No Highschool
Highschool
PS (except univ.)
University
Study objectives

1.
Compare the
characteristics of attritors

and non
-
attritors using chi
-
square tests and one
-
way ANOVA.

2.
Identify the
factors influencing the probability of attrition

using multivariate probit regressions.

3.
Measure
attrition bias

using BGLW tests for selected
variables.

Conclusions


Probability of attrition by mother's highest diploma,
wave 1
0,238
0,306
0,323
0,398
University
PS (except univ.)
Highschool
No highschool
Probability of attrition


The probability of attrition is modeled using a probit model with a set of
background and other individual characteristics as independent variables.


Most variables do not predict significatively the probability of attrition.

Attrition bias


Attrition bias is evaluated by means of BGLW tests (Becketti, Gould, Lillard and Welch,
1988) by regressing a selected variable of interest on a set of control variables plus a
dichotomous variable representing attrition in the following waves. The presence and
magnitude of attrition bias is inferred from the significance of the estimated coefficient for
attrition in this equation.


Attrition does not exert a signficant bias on most variables of interest (e.g. delay in child’
growth) with the exception of
mothers’ immigrant status

and
abortion
.


***

**

*

Probability of attrition by immigration status of the
mother, wave 1
0,214
0,343
Non-immigrant
Immigrant
***

Probability of attrition by whether the mother is
overprotective of her child (on a continuous scale
between 0 and 10), wave 1
0
0,1
0,2
0,3
0,4
0,5
0
1
2
3
4
5
6
7
8
9
10
Legend: *
p
<0,10; **
p
<0,05; ***
p
<0,01.

Notes: The household’s characteristics included in the models are: household income, household income squared, number of
siblings, whether home is owned). The individual characteristics of the mother included in the models are: age, highest
diploma, occupation. All probabilities are calculated using the mean score for continuous variables and the mode for discrete

variables.

Probability of attrition by whether the mother had an
abortion, wave 1
0,257
0,209
Mother had an abortion
Mother didn't have an abortion
**

Probability that the mother had an abortion by
attrition status (attrition bias), wave 1
0,322
0,385
Attritors
Non-attritors
**

Legend: *
p
<0,10; **
p
<0,05; ***
p
<0,01.

Note: The background characteristics considered for the BGLW tests are the same used for the probit models.

Mother's overprotection mean score by attrition
status (attrition bias), wave 1
4,58
4,38
Attritors
Non-attritors

Respondents’ attitude towards surveys (i.e. level of education) and geographic
mobility are the two most important factors associated with attrition.


Although attrition exerts important biases for univariate analyses, it does not
generally bias multivariate analyses.


The main effect of attrition for analyses of the ÉLDEQ data is to decrease the
sample size and thus reduce the power of statistical inferences.


Continuing efforts are made by the ÉLDEQ survey team to track down
respondents and thus limit attrition in future waves.


Future research should focus on the consequences of attrition for longitudinal
analyses of the ÉLDEQ data (survival analysis and multi
-
level analysis).


Researchers using longitudinal survey data should always check for attrition bias
in their analyses.

Probability that the mother is immigrant by attrition
status , wave 1
0,304
0,187
Attritors
Non-attritors
***