I am trying to conduct mediation analysis with longitudinal data (Health and Retirement Study), for individuals withing time (12 waves). The outcome is depression measured using the CESD-8 items (as discrete variable), the mediator is smoking (categorical variable, it could be i. non smoker, long time smoker who stopped smoking, long time smoker who continues to smoke, or based on frequency, also in categories) and the main predictort is time spent in retirement (categorical: 0=no retired, 1=0-3 years; 2=>3-6 years, 3=>6 years). The data is unbalanced, those who reported being fully retired at one wave were coded as fully retired in all following waves, retirement age was calculated based on the first year they stated they fully retired and birth year, then that was used to calculate time on retirement based on the HRS wave years. I am also controlling for covariates, such as health, age, gender, etc.
For a similar study, using alcohol use (days/week that drinks), I conducted genelarized linear models and used the nlcom command to calculate the indirect and total effects, with its CI, SE and p value. I do not seem to find information about using PARAMED with longitudinal data, and I read Vandeerweele article on three-way decomposition analysis. Thus, I would like to calculate the direct effect, total effect and indirect effects (pure indirect effect and mediated interactive effect). I need some guidance on how to do it using PARAMED or the nlcom command after computing the three mixed effect models.
My understanding is that with nlcom I can conduct a regression of retirement +covariates on depression and get the direct effect, then I can calculate a regression of retirement and the mediator (smoking) +covariates on depression and calculate the indirect and total effect using nlcom. For the three-way decomposition analysis I should conduct three regression models in order to be able to compute the 4 effects. Eq.1. retirement +covariates on depression; Eq2. retirement, smoking, retirement*smoking +covariates on depression, and Eq3. retirement+covariates on smoking.
Based on Eq1. I should have the direct effect of retirement, right?
Based on Eq2. I should be able to compute the pure and the mediated indirect effects?
if so, then why I need Eq3? How should I compute the total effect, by adding the direct, pure indirect, and mediated indirect (for interaction) effects together?
Or is there a way to compute this using PARAMED (the level 2 indicator is hhidpn, which is a unique ID for the individual). GSEM does not allow for interaction terms, so I cannot use that, as the three-way decompisition can only be used with models that allow interactions.
Your help and thoughts are greatly appreciated.
Some references, that I have read.
https://www.ncbi.nlm.nih.gov/pmc/art...C3563853/#APP1
https://www.stata.com/symposiums/bio...1_Bellavia.pdf
https://www.scinapse.io/papers/40003836
For a similar study, using alcohol use (days/week that drinks), I conducted genelarized linear models and used the nlcom command to calculate the indirect and total effects, with its CI, SE and p value. I do not seem to find information about using PARAMED with longitudinal data, and I read Vandeerweele article on three-way decomposition analysis. Thus, I would like to calculate the direct effect, total effect and indirect effects (pure indirect effect and mediated interactive effect). I need some guidance on how to do it using PARAMED or the nlcom command after computing the three mixed effect models.
My understanding is that with nlcom I can conduct a regression of retirement +covariates on depression and get the direct effect, then I can calculate a regression of retirement and the mediator (smoking) +covariates on depression and calculate the indirect and total effect using nlcom. For the three-way decomposition analysis I should conduct three regression models in order to be able to compute the 4 effects. Eq.1. retirement +covariates on depression; Eq2. retirement, smoking, retirement*smoking +covariates on depression, and Eq3. retirement+covariates on smoking.
Based on Eq1. I should have the direct effect of retirement, right?
Based on Eq2. I should be able to compute the pure and the mediated indirect effects?
if so, then why I need Eq3? How should I compute the total effect, by adding the direct, pure indirect, and mediated indirect (for interaction) effects together?
Or is there a way to compute this using PARAMED (the level 2 indicator is hhidpn, which is a unique ID for the individual). GSEM does not allow for interaction terms, so I cannot use that, as the three-way decompisition can only be used with models that allow interactions.
Your help and thoughts are greatly appreciated.
Some references, that I have read.
https://www.ncbi.nlm.nih.gov/pmc/art...C3563853/#APP1
https://www.stata.com/symposiums/bio...1_Bellavia.pdf
https://www.scinapse.io/papers/40003836
