Longitudinal studies: Cornerstone for causal modeling of dynamic relationships
Dec 15, 2015
Longitudinal studies: Cornerstone for causal modeling of dynamic relationships
Illustrative examples from the Cebu Longitudinal Health and Nutrition Survey
• Prospective, community-based sample of 1983-4 birth cohort, follows mothers and index infant from urban&rural areas of Metro Cebu, The Philippines
• Bi-monthly surveys birth-2yr, follow-up surveys in 1991, 1994, 1998, 2002, 2005
• Extensive individual, household and community data
Types of longitudinal studies• Same individuals over time
• Common age at enrolment (e.g. birth cohort)• Life course studies, individual trajectories• Challenging to separate age vs time effects
• Eg, diet changes over time because kids get older or because there is a secular trend in dietary behaviors
• Different ages at enrolment • Panels/cross sectional time series: Different individual
over time, in common units (e.g. community, school, household) • Allow study of trends over time, but not individual trajectories
• Mixed: repeatedly study individuals, but with replacementEach poses different challenges for data collection
and analysis
Focus on cohort studies …repeated measures of the same individuals, over time allow for:
• Identification of sequence of events, providing basis for causal inference
• Comparison of inter vs intra-individual variation in susceptibility, behavior, health
• Response to shock or intervention differs between individuals
• Individual growth rates vary with age
Longitudinal Study Challenges
• Cost (time, $)• Attrition• Bias associated with repeated contacts
with individuals• observer effects• sampling bias amplified by repetition of surveys• panel conditioning: changes in response to
participation
Challenges of collecting longitudinal data Research priorities and funding opportunities change over time: funding infrequently covers more than 5 years at a time.
Example: Cebu Longitudinal Health and Nutrition Survey
Survey year
Focus Funder
1983-86 Infant feeding, growth, morbidity, mortality
NICHD, Ford Foundation
1991 Growth, school enrollment, IQ World BankNestle Foundation
1994 Family planning and women’s lives
USAID: Women’s Studies Project
1998 Adolescent Health Mellon Foundation
2002 Effects of health on young adult human capital
NIH-Fogarty ISHED
2005 Add biomarkers of CVD risk factors
NIH-Fogarty ISHEDObesity roadmap funds
Methodological challenges of collecting longitudinal data
• Technology for data collection and storage changes over time• Face to face vs. “ACASI”
• Measurement Issues• Change in personnel collecting data
• interobserver reliability is harder to maintain and measure over time • Change in how questions are asked
• e.g. Analysis reveals flawed question on round 1: do we change the question on round 2?
• Change in how questions are answered• different social climate or respondent knowledge gained over time (perhaps by
study participation) may affect veracity
• Who responds? Child vs mother? At what age does a child become the respondent?
• Change in meaning of indicators over time• E.g. wealth: TV vs computer vs. car over time
Dilemmas and choices….• Expanding the survey may
increase respondent burden and compromise participation rates• But… Failure to expand the
survey represents missed opportunities
• Follow-up of all migrants is desirable• But… Follow-up is costly and
not always feasible• Changing how a question
is asked eliminates comparability over time• But… keeping a flawed
question is bad science
Data collection challenges
• How often should participants be surveyed?
• Frequent measurement allows sequence of events to be identified• Pregnancy>>>quit school>>>marriage• Quit school>>>marry>>>pregnancy
• Respondent burden, “contamination” of sample
Analysis challenges
• Specialized techniques are needed to accommodate the strengths and weaknesses of longitudinal data
• Accounting for complexity• Accounting for changing inputs
across the lifecycle
Analysis challenges
• Accounting for differences in susceptibility• Example: parental investment may change
based on acquired characteristics of the child
• Example: developmental origins of adult disease: key premise is that prenatal factors alter response to subsequent exposures
• Intergenerational studies
Challenges: Selection bias related to attrition
• Loss to follow-up: Death, Migration, Refusal• May result in sample which is markedly
different from baseline sample in measured and unmeasured attributes
• Biased estimates may be obtained if the relationships of interest are fundamentally different in those remaining vs. lost, particularly when differences relate to unmeasured characteristics
Tools for handling selection bias• Heckman-type models estimate
likelihood of being in the sample simultaneously with outcome of interest
• Difficult to account for multiple reasons for attrition (with different potential for bias, e.g death vs migration)
Challenges: growth trajectories and functional forms• Ideally…we would like models to
accommodate• Non-linear “growth trajectories”• Differences in shape of trajectories at
different ages, and in the relationship of exposures to outcomes at different ages
02
46
8
0 .2 .4 .6 .8 1agey
weight diawk
02
46
8
0 .2 .4 .6 .8 1agey
weight diawk
02
46
81
0
0 .2 .4 .6 .8 1agey
weight diawk
02
46
8
0 .2 .4 .6 .8 1agey
weight diawk
Latent growth curves: A category of
Structural Equation Models
• Random intercepts and random slopes allow each case to have a different trajectory over time
• Random coefficients incorporated into SEMs by considering them as latent variables
• Capitalize on SEM strengths, including:• ML methods for missing data• Estimation of different non linear forms of trajectories,
including piecewise to identify different curve segments• Measures of model fit and • Inclusion of latent covariates and repeated covariates• Latent variables derived from multiple measured
variables• Account for bi-directional relationships
Data demands for econometric models• Detailed, time-varying, high quality
exogenous variables • Often this means community level
variables, so data collection cannot be limited to individual or household level information
What’s on the frontier for new longitudinal methods?
• ..”new data, methodologies, and tools from both inside and outside the social sciences are demonstrating real promise in advancing these sciences from descriptive to predictive ones”*
• “Longitudinal surveys” is one of 6 listed frontiers
• Improved statistical methods is another (but this section is about using the internet to conduct surveys!!)
*Butz WP, Torrey BB Some Frontiers in Social Science. Science June 2006
What is on the frontier??
• Addition of biomarkers• Overcoming squeamishness of social
scientists• Lack of laboratory facilities• What methodological improvements are
needed?• Innovative data collection and tracking
• Use of GPS and PDAs