TY - JOUR
T1 - Two-Method planned missing designs for longitudinal research
AU - Garnier-Villarreal, Mauricio
AU - Rhemtulla, Mijke
AU - Little, Todd D.
N1 - Funding Information:
This study was supported by grant NSF 1053160 (Wei Wu & Todd D. Little, co-PIs) and by the Center for Research Methods and Data Analysis at the University of Kansas (when Todd D. Little was director; 2009–2013).
PY - 2014/9
Y1 - 2014/9
N2 - We examine longitudinal extensions of the two-Method measurement design, which uses planned missingness to optimize cost-efficiency and validity of hard-to-Measure constructs. These designs use a combination of two measures: a "gold standard" that is highly valid but expensive to administer, and an inexpensive (e.g., survey-based) measure that contains systematic measurement bias (e.g., response bias). Using simulated data on four measurement occasions, we compared the cost-efficiency and validity of longitudinal designs where the gold standard is measured at one or more measurement occasions. We manipulated the nature of the response bias over time (constant, increasing, fluctuating), the factorial structure of the response bias over time, and the constraints placed on the latent variable model. Our results showed that parameter bias is lowest when the gold standard is measured on at least two occasions. When a multifactorial structure was used to model response bias over time, it is necessary to have the "gold standard" measures included at every time point, in which case most of the parameters showed low bias. Almost all parameters in all conditions displayed high relative efficiency, suggesting that the 2-Method design is an effective way to reduce costs and improve both power and accuracy in longitudinal research.
AB - We examine longitudinal extensions of the two-Method measurement design, which uses planned missingness to optimize cost-efficiency and validity of hard-to-Measure constructs. These designs use a combination of two measures: a "gold standard" that is highly valid but expensive to administer, and an inexpensive (e.g., survey-based) measure that contains systematic measurement bias (e.g., response bias). Using simulated data on four measurement occasions, we compared the cost-efficiency and validity of longitudinal designs where the gold standard is measured at one or more measurement occasions. We manipulated the nature of the response bias over time (constant, increasing, fluctuating), the factorial structure of the response bias over time, and the constraints placed on the latent variable model. Our results showed that parameter bias is lowest when the gold standard is measured on at least two occasions. When a multifactorial structure was used to model response bias over time, it is necessary to have the "gold standard" measures included at every time point, in which case most of the parameters showed low bias. Almost all parameters in all conditions displayed high relative efficiency, suggesting that the 2-Method design is an effective way to reduce costs and improve both power and accuracy in longitudinal research.
KW - Intentionally missing data
KW - Missing data
KW - Planned missingness
KW - Simsem
KW - Structural equation modeling
KW - Two-Method measurement
UR - http://www.scopus.com/inward/record.url?scp=84905637908&partnerID=8YFLogxK
U2 - 10.1177/0165025414542711
DO - 10.1177/0165025414542711
M3 - Article
AN - SCOPUS:84905637908
VL - 38
SP - 411
EP - 422
JO - International Journal of Behavioral Development
JF - International Journal of Behavioral Development
SN - 0165-0254
IS - 5
ER -