How to add time-trendy variables to cross sectional analysis instead of using panel data analysis

[Maha Natsheh]

Hello everyone,

I made panel data for analyzing the determinants of economic performance for agricultural cooperatives measured by (Net surplus achieved) from 2021-2023,
after plenty of experiences, this is the most significant analysis:

My supervisor said that there are very large difference between (within R-squared) and (Between R-squared), in addition that consumption complex and grants in t-1 have adverse effect compared to logic. so he thought that these problems are because we have just 3 years in our data, and suggested to analyze the data by cross- sectional but by adding 2 time- trendy variables in the regression.

Please note that Hausman test choses fixed effect, on the other hand, we have important dummy variables which omitted directly in fixed effect choice due to collinearity, and they also insignificant in re experiments.
Also, we have important variables that available only for 2023.


My questions:
1) How to execute cross sectional analysis and how to convert variables in the other 2 years to time-trendy variables using Stata?
2) should I use long data form or what?
3) As I have some important variables just for 2023, could I make 2023 as the base year?

Looking forward your advices,

Thanks in advance!
Maha N.

[George Ford]

You got bigger problems: corr(u_i, Xb) = -1.0000 and the within R2 == 1.0.

[Maha Natsheh]

How could I solve this problem?

[George Ford]

The issue is probably related to the fact this is not a panel. You've got 25 observations and 16 groups. Do you have a lot of missing data? The fixed effects are likely the source of the problem.

Figure out why the model is predicting perfectly (probably the FE). If not, it may be one variable, but could be a combination of variables. Start with the simplest model and add variables and watch the R2 and corr.

If you want to add time, which you probably should, then include i.year as a regressor, but not sure that would work if the 25/16 is truly the data.

[George Ford]

I would recommend taking the time to understanding what you are doing. I get the sense you are running regressions without a base level of understanding what you are up to.

[Clyde Schechter]

With 25 observations in 16 groups, you are going to have nothing but trouble trying to fit so many variables in the model. Within each group you have a maximum of 3 observations, and even that, it seems, happens only occasionally, as the mean number of observations per group is only 1.6. These figures of 3and 1.6 limit what you can do with the model, beyond just the overall sample size of 25, because the fixed-effects model is only estimating within-group effects.

I'm guessing from the variable name that SqlogAss is the square of logAss. It makes no sense at all to fit a quadratic model through three or fewer data points. That's just fitting the noise in the data. Moreover, fitting 7 variables plus a constant to at most 3 data points per group also makes no sense at all. Adding yet more variables (time trends) to the model will only make matters worse.

This project needs serious rethinking. Either you need to get a much larger data set that can support this kind of modeling, or you need to come up with a far more parsimonious model that the existing data set can support.

I don't think I'm saying much that George Ford hasn't said above, just stating it more aggressively and adding some specific details.