Positive and Negative DV Values in Panel Data analysis, potential solutions

[William Nolan]

is it possible to regress a period average of your DV onto annual values of IVs. My DV is Chinese OFDI to African host countries and my IVs are GDP Inflation geodistance fdiopeness traderelationship landlock and institutional quality.

the reason for wanting to use a period average is to deal with volatile changes in fdi and smooth these out for each country. if this is possible/a correct thing to do what regression would this be.and what would the implications be

[Maxence Morlet]

I'm really not sure that I've understood what you want to do, so here come the questions:

- I presume you have panel data? Do you follow multiple countries over time? Please write down an equation with subscripts so we know what sources of variation you have. We need to know that to help you.

- Sources of variation are crucial, they allow to include additional vectors of fixed-effects to identify your coefficient of interest more precisely.

- You want to have variation in your data, it is generally a good thing. However, it is only a good thing when the variation is a genuine signal, and not noise.

- If you do follow multiple countries over time, the fixed effects estimator will basically do what OLS does but on a transformed model, where all variables are instrumented with their respective deviation within each individual from the individual mean (averaged over time). This means that time-invariant will have to be dropped as their coefficients cannot be identified.

[William Nolan]

Panel Data with the Dependent Variable Chinese OFDI to 42 African Countries from 2013-2020, Explanatory Variables are one-year lagged host country: GDP Inflation Geodistance (Time-Invariant) FDI Openess TradeRelationship Landlock (Dummy Variable) Institutional Quality and Institutional Distance

The issue is there is large within country variation for the DV, but the within country Ivs are relatively constant. For example one year in Angola Chinese FDI is 630mn but 2 years earlier it was -270mn. Therefore I was trying to gauge the merit in taking an average of the FDI to the host country over the 2013-2020 period and then regressing this onto the explanatory variables and what sort of regression this would be.

[William Nolan]

to be more specific my control variables are GDP Inflation Geodistance (Time-Invariant) FDI Openess TradeRelationship Landlock (Dummy Variable) and my two independent variables are institutional quality and institutional distance

[Maxence Morlet]

Correlated random effects is the way to go here, check out Wooldridge (2010) on how to implement it.

The thing is the coefficients you will get on your time-invariant regressors will be biased as they will be random effects coefficients.