Longitudinal Data Analysis

Longitudinal data can be viewed as a special case of the multilevel data where time is nested within individual participants. All longitudinal data share at least three features: (1) the same entities are repeatedly observed over time; (2) the same measurements (including parallel tests) are used; and (3) the timing for each measurement is known (Baltes & Nesselroade, 1979). To study phenomena in their time-related patterns of constancy and change is a primary reason for collecting longitudinal data. Figure 1 show the plot of 50 participants from the ACTIVE study on the variable EPT for 6 times. Clearly, each line represents a participant. From it, we can see how an individual changes over time.

Growth curve model

Growth curve models (GCM; e.g., McArdle \& Nesselroade, 2003; Meredith & Tisak, 1990) exemplify a widely used technique with a direct match to the objectives of longitudinal research described by Baltes and Nesselroade (1979) to analyze explicitly intra-individual change and inter-individual differences in change. In the past decades, growth curve models have evolved from fitting a single curve for only one individual to fitting multilevel or mixed-effects models and from linear to nonlinear models (e.g., McArdle, 2001; McArdle & Nesselroade, 2003; Meredith & Tisak, 1990; Tucker, 1958; Wishart, 1938).

A typical linear growth curve model can be written as

\begin{eqnarray*} y_{it} & = & \beta_{0i}+\beta_{1i}\times time_{it}+e_{it}\\ \beta_{0i} & = & \gamma_{0}+v_{0i}\\ \beta_{1i} & = & \gamma_{1}+v_{1i}\end{eqnarray*}

where $y_{it}$ is data for participant $i$ at time $t$. For each individual $i$, a linear regression model can be fitted with its own intercept $\beta_{0i}$ and slope $\beta_{1i}$. On average, there is an intercept $\gamma_{0}$ and slope $\gamma_{1}$ for all individuals. The variation of $\beta_{0i}$ and $\beta_{1i}$ represents individual differences.

Individual difference can be further explained by other factors, for example, education level and age. Then the model is

\begin{eqnarray*} y_{it} & = & \beta_{0i}+\beta_{1i}\times time_{it}+e_{it}\\ \beta_{0i} & = & \gamma_{0}+\gamma_{1}\times edu_{i}+v_{0i}\\ \beta_{1i} & = & \gamma_{2}+\gamma_{3}\times edu_{i}+v_{1i} \end{eqnarray*}

GCM as a mulitlevel/mixed-effect model

A GCM can first be fitted as a multilevel model or mixed-effects model using the R package lme4.

To use the package, we would need to rewrite the growth curve model as a mixed-effect model. For the model without second level predictor, we have

\[ y_{it}=\gamma_{0}+v_{i0}+\gamma_{1}*time_{it}+v_{1i}*time_{it}+e_{it}. \]

For the one with the second level predictor, such as education, we have

\[y_{it}=\gamma_{0}+\gamma_{1}*edu+v_{0i}+\gamma_{2}*time_{it} +\gamma_{3}*edu_{i}*time_{it}+v_{1i}*time_{it}+e_{it}.\]

For demonstration, we investigate the growth of word set test (ws in the ACTIVE data set). In the current data set, we have the data in wide format, in which the 6 measures of ws are 6 variables. To use the R package, long format data are needed. For the long-format data, we need to stack the data from all waves into a long variable. The R code below reformats the data and plot them.

Unconditional model (model without second level predictors)

Fitting the model is actually straightforward using the lmer() function. The input and output are given below. Based on the output, the fixed effects for time (.214, t-value=11.59) is significant, therefore, there is a linear growth trend. The average intercept is 11.93 and is also significant. 

It is useful to test whether random-effects parameters such as the variances of intercept and slope are significance or not to evaluate individual differences. This can be done by comparing the current model with a model without random intercept or slope.

For example, to test the individual differences in slope for time. The random effects for time is .07. Based on ANOVA analysis, it is significant with p-value about 0. Therefore, there is significant individual difference in the growth rate (slope). This indicates that everyone has a different change rate. Note that in m1.alt, the random effect for time was not used.

To test the individual differences in intercept. The random effects for intercept is 21.11. Based on ANOVA analysis below, it is significant. Therefore, there is individual difference or individuals have different intercepts.

Conditional model (model with second level predictors)

Using the same set of data, we now investigate whether education is a predictor of random intercept and slope. Given there is individual differences in intercept and slope, we want to explain why. So, we use Edu as a explanatory variable. From the output, we can see that the parameter \(\gamma_{1} = .78\) is significant. Higher education relates to bigger intercept. In addition, the parameter $\gamma_{3} = -.022$ is significant. Higher education relates to lower growth rate of ws.

GCM as a SEM

In addition to estimating a GCM as a multilevel or mixed-effects model, we can also estimate it as a SEM. To illustrate this, we consider a linear growth curve model. If a group of participants all have linear change trend, for each individual, we can fit a regression model such as \[ y_{it}=\beta_{i0}+\beta_{i1}t+e_{it} \] where $\beta_{i0}$ and $\beta_{i1}$ are intercept and slope, respectively. Note that here we let $time_{it} = t$. If each individual has different time of data collection, it can still be done in the SEM framework but would be more complex. By writing the time out, we would have \[ \left(\begin{array}{c} y_{i1}\\ y_{i2}\\ \vdots\\ y_{iT} \end{array}\right)=\left(\begin{array}{cc} 1 & 1\\ 1 & 2\\ 1 & \vdots\\ 1 & T \end{array}\right)\left(\begin{array}{c} \beta_{i0}\\ \beta_{i1} \end{array}\right)+\left(\begin{array}{c} e_{i1}\\ e_{i2}\\ \vdots\\ e_{iT} \end{array}\right) \]

Note that the above equation resembles a factor model with two factors - $b_{0}$ and $b_{1}$ and a factor loading matrix with known factor loading matrix. The individual intercept and slope can be viewed as factor scores to be estimated. Furthermore, we are interested in model with mean structure because the means of $\beta_{0}$ and $\beta_{1}$ have their meaning as average intercept and slope (rate of change). The variances of the factors can be estimated - they indicate the variations of intercept and slope. Using path diagram, the model is shown in the figure below.

With the model, we can estimate it using the sem() function in the lavaan package. Because of the frequent use of growth curve model, the package also provides a function growth() to ease such analysis. Unlike the lme4 package, in using SEM, the wide format of data is directly used. The R input and output for the unconditional model is given below.

Note that the gcm() function works similarly as sem() function. Using this method, each parameter in the model can be directly tested using a z-test. In addition, we can use the fit statistics for SEM to test the fit of the growth curve model. Particularly for the current analysis, the linear growth curve model does not seem to fit the data well.

Fitting a conditional model is similar but one would need to use the predictor for the factors.