Longitudinal data entails data generated from repeated measurements carried out on the same unit such as human beings, plots, or plants. Therefore, a lot of efforts go into longitudinal research studies in terms of making observations to ensure that the same observation/response involving one unit is measured repeatedly.
For instance, a researcher may wish to repeatedly measure the viral load on the same subject.
Thus, unlike cross-sectional research studies, longitudinal studies are designed not only to address scientific questions as to how the mean of individual responses varies across various treatments, but also to measure the change involving mean responses over time and the link between individual responses and time (Singer & Willett, 2003, p. 1-10). This review aims at describing the statistical models and methods for analyzing longitudinal data.
There are two major statistical models namely level-1 and level-2 models, which are suitable for longitudinal data analysis. However, some researchers prefer to treat the two models as a pair that is the core of the multilevel model for longitudinal data analysis.
In a level-1 model, one expects the research question to address inter-individual changes over a given period while in a level-2 model; predictors of inter-individual differences involving change are addressed (Singer & Willett, 2003).
Therefore, a scientific study suitable for longitudinal data analysis is characterized by multiple waves of data, a sensible metric for time, and a continuous outcome/response, which changes systematically over time. As a result, it is wise for researchers to conduct descriptive exploratory data analyses before using any statistical model.
Accordingly, the first step in longitudinal data analysis involves creating a longitudinal data set that fits the analysis. Here, there are two major formats for data arrangement namely a person-level and person-period data sets. In a person-level data set, the observations of each subject are organized into one record containing multiple variables for each measurement.
Conversely, in a person-period data set, the observations made on each subject are recorded in multiple records for each measurement. Arrangement of longitudinal data can be achieved through the use of different statistical software packages such as SAS, SPSS, and STATA (Singer & Willett, 2003, p. 20).
After data arrangement, the second step should entail exploratory analyses aimed at describing how each individual changes occur over time in the preferred data set. This step generates patterns of change for each individual through examining an empirical growth plot for that particular person.
Conversely, one can use various smooth trajectories to summarize the patterns observed for individual empirical growth records. In most cases, the ordinary least squares (OLS) regression is the most preferred method for smoothing the empirical growth trajectories.
Subsequently, the next step in data analysis entails assessing the inter-individual differences in the data set through examining and analyzing the entire set of smooth trajectories. Here, the entire set of individual trajectories is plotted on one graph to give the researcher the opportunity to observe the average growth trajectory from the logistic curves.
Additionally, through model fitting, a researcher frames various questions regarding changes in each data set. Therefore, the formal answers to these questions can be generated from multilevel modeling, simple analyses of estimated intercepts and slopes, examination of descriptive statistics such as means and standard deviations and analysis of correlation coefficients (Singer & Willett, 2003).
Finally, there is the need for researchers to explore the link between individual changes and the time-invariant predictors to uncover various systematic patterns involving inter-individual differences as a result of personal attributes. This can easily be done by graphically examining sets of smoothed trajectories for each individual and the link between OLS-estimated plots and their substantive predictors.
Overall, the availability of statistical software makes it easy to fit the multilevel model for longitudinal data analysis directly in order to correct for any biases made during the descriptive exploratory longitudinal data analyses.
Reference
Singer, J.D., & Willett, J.B. (2003). Applied longitudinal data analysis: Modeling change and event occurrence. New York: Oxford University Press. Retrieved from www.gse.harvard.edu/~faculty/singer/Papers/ch1&2.pdf