Analysing Dyadic Paneldata

[Viktoria Bauer]

Hello everyone,

I have a dyadic paneldataset of couples including the satisfaction with the relationship for both partners (dependent variable) and values of how attractive both partners regard themselves and their partners (independent variable). The dataset is currently in a wide-format.

id pid wave sat3 psat3 per7i1 pper7i1 per7i2 pper7i2
1028000 1028101  8 8 9 4  3 6  5
1028000 1028101  9 8 7 .  . .  .
1028000 1028101 10 8 8 5  4 7  7
1028000 1028101 11 9 9 .  . .  .
1028000 1028101 12 9 9 4 -2 6 -2

I am currently wondering how I analyse this data properly. I first thought I would be able to analyse it in a long-format but I am not able to xtset the data since the time-defining variable occurs twice for each dyad.

Given that my data is in a wide-format, I have two dependent variable of interest, but I was unfortunately unable to find a Stata command which I can use for analyzing two dependent variables simultaneously for panel data.

However, just summarizing both scores into a new variable and including the independent variables of both partners just seems not suitable since the dependencies within the dyad are not taken into account.

I am thankful for any advise and support.

Kind regards,
Viktoria

[Clyde Schechter]

There is no generic approach to this. It depends on the specific research questions you want to answer. As you haven't said anything about this, the best I can offer is a few generic observations.

1. There are a few Stata commands that do accommodate multiple dependent variables. -mvreg- comes to mind, and although it is not a panel-data command, you can include indicators for your persons or dyads to emulate fixed-effects estimation of the linear regressions.

2. Depending on the research question, it may make sense to reshape the data long, and in your regression interact a partner vs respondent indicator variable with the explanatory variables of interest so as to get separate estimates for the effects among partners and responders.

3. Including a time variable in -xtset- is optional. You can just -xtset- the person id or a couple id as the case may be. By omitting the time variable you lose the ability to run models with auto-regressive structure or to use time series operators such as lag and lead. But most research questions can be answered without resort to those things anyway. So if you don't need those, just move on without a time variable in your -xtset- command. On the other hand, if you do reshape the data long, you can still use -xtset- with a time variable if you use a person-level identifier rather than the couple-identifier as the panel variable. However, see my next comment about the difficulties with viewing this data as panel data when working with person-level outcomes.

4. If you need analyses that treat person-level variables as outcomes, then you are not really working with panel data. This is a panel of couples, but it is not a panel of people. For people-level outcomes what you have is a three-level model: waves nested in (or crossed with) people nested in couples. So panel-data models may not be appropriate here, and may produce "answers" that are based on a model that misrepresents the data in important ways. Multi-level modeling commands (mixed, the me- commands) etc. would be more appropriate for this.

[Viktoria Bauer]

Thank you very much for this input!

My research question is in how far physical attractiveness predicts relationship satisfaction for couples who met in an online venue in contrast to couples who met in more traditional offline meeting venues. I am suggesting, that individuals within couples who met online reciprocally regard themselves and their partners as more equally in their physical attractiveness whilst individuals within couples who met offline are faced with higher mismatches, meaning that one individual is more likely to regard him- or herself as more attractive than the partner whilst the partner is more likely to regard him- or herself as less attractive in return.

[Clyde Schechter]

Your response in #3 actually sets out two research questions. And for present purposes, they are really very different in the way one might approach them.

Your first stated research question, uses relationship satisfaction as the outcome variable, and seeks to estimate an interaction between physical attractiveness and online vs offline meeting venues. While I suppose it is possible to combine the individual satisfaction scores of the two parties to the relationship into a single "couple satisfaction index," that seems difficult to do well. So probably this question would be approached by analyzing individual level data in a multi-level model.

The second question, seeks to determine whether couples who meet online have more similar levels of physical attractiveness ratings than those who meet offline. This would be approached by calculating, for each couple, the absolute value of the difference between their ratings of each other's physical attractiveness, and then doing a couple-level analysis of that outcome, with online vs offline meetup as the key explanatory variable. A panel-data analysis for this would be feasible as the person-level is obliterated by using an outcome variable that is defined at the couple level.

[Viktoria Bauer]

Again, thank you very much for your support, this really helps me!