Browsing by Subject "Longitudinal data"

Now showing 1 - 4 of 4
  • Thumbnail Image
    Item
    Bayesian Modeling of Associations in Bivariate Mixed-Effects Models for Segmented Growth Curves
    (2018-11) Peralta Torres, Yadira
    Developmental processes rarely occur in isolation; often the growth curves of two or more variables are interdependent. In addition, frequently, growth curves do not portray a constant pattern of change. Different stages or segments of development are present in the data. Bivariate piecewise linear mixed-effects models (BPLMEM) are a useful statistical framework to simultaneously describe two processes that portray segmented linear trajectories and to investigate their associations over time. Interrelations between the growth curves are measured by assuming a joint distribution of the random-effects parameters of each outcome variable. Furthermore, associations in the outcome variables collected from the same subject should be taken into account when they are modeled jointly. This association is modeled by correlating the error variance parameters of each outcome variable. There are several drawbacks in the literature of bivariate piecewise mixed-effects models. An important limitation in the BPLMEMs literature is that researchers have assumed uncorrelated residual errors across the two longitudinal processes, which is something unlikely to hold in practice. Also, current modeling choices for the random-effects in bivariate piecewise mixed-effects model have shortcomings. For instance, researchers have unintentionally imposed dependencies among the elements of the covariance matrix associated with the random-effects; or they have modeled only few of its covariance parameters determined by the research interest. In addition, simulation studies using BPLMEMs are scarce. Little is known about the performance of bivariate piecewise mixed-effects models under different correlational scenarios of the random-effects parameters and the error variances. Furthermore, a criticism to the piecewise linear model is a hypothesized abrupt change between one linear segment to another because this performance may not be true for all empirical scenarios. However, although a smooth transition or adaptation period between linear segments might be more realistic, the piecewise linear model is extensively used in practice. Thus, it is natural to wonder under which scenarios this is an acceptable choice. The purpose of the present study was to develop a BPLMEM using a Bayesian inference approach allowing the estimation of the association between error variances and providing a more robust modeling choice for the random-effects. Furthermore, the performance of the BPLMEM was investigated via a Monte Carlo simulation study focusing on the strength of the associations of the error variance parameters and the growth curves (represented by the random-effects’ correlational structure). An additional purpose was to empirically characterize scenarios for which the piecewise linear model gives a reasonable approximation to an underlying smoothed segmented trajectory given by a quadratic bend joining the linear phases of growth. Lastly, the contribution of bivariate mixed-effects modeling approaches is illustrated by using a BPLMEM to investigate the joint development of mathematics and reading achievement and the association between their longitudinal trajectories. This constitutes a novel approach to examine associations between educational domains over time. Simulation results showed that the strength of the association between the growth curves and the sample size had a significant effect on the performance of the BPLMEM. Specifically, lower relative bias of parameter estimates and higher model convergence was related to a stronger correlational structure between the random-effects of the growth curves. Likewise, slightly higher coverage rates and better convergence were associated with a smaller sample size. In addition, it was possible to identify cases for which the piecewise linear model had an acceptable performance when the true underlying trajectory had an adaptation period or bend between linear segments. Scenarios with small or centered bends were accurately described by a piecewise linear model. Results from the illustrative example suggested that mathematics and reading achievement are positively associated all along their segmented trajectories and that the strength of such association decreases over time. In addition, evidence of the same patterns of association of reading over mathematics and mathematics over reading were found.
  • Thumbnail Image
    Item
  • Thumbnail Image
    Item
    Mixed effects models for size-attained data.
    (2012-01) Lendway, Lisa M.
    It is rare to have longitudinal data on the somatic growth of fish, that is, how their body length changes over time. In most temperate fish, scales or other hard parts, like otoliths or other bones, form annual rings or increments. Growth of the hard part can be measured, thereby giving a longitudinal record of hard part growth from cross-sectional data. Methods such as back-calculation and linear mixed-effects models have used the growth of hard parts to infer somatic growth. At times, it is not feasible to obtain the measurements of the hard part. Body length at time of capture is much easier to measure and reflects somatic growth, which is usually of more interest. In this thesis, I present a model that is based on a longitudinal approach but models length at time of capture, rather than the yearly body growth. It also allows for estimation of environmental impact on growth.
  • Thumbnail Image
    Item
    Panel conditioning in longitudinal social science surveys
    (2013-07) Halpern-Manners, Andrew
    Researchers who utilize data from longitudinal surveys nearly always assume that respondents' attributes are not changed as a result of being measured. Yet research in cognitive psychology, political science, and elsewhere suggests that the experience of being interviewed can spark important changes in the way respondents behave, in the attitudes that they possess, and in their willingness or ability to answer questions accurately when they are re-interviewed in subsequent waves. In this dissertation, I evaluate the severity of this problem in longitudinal social science surveys. Using a combination of observational and experimental data, I show that "panel conditioning" has the potential to affect a wide range of attitudinal and behavioral measures, including many items that are commonly used in sociological and demographic research. The causal mechanisms that give rise to these effects are discussed and a large-scale follow-up project is proposed.