Method to choose

[Padmavathi Bandaru]

I have dataset containing 3 variables(2 dependent 1 indepnedent) for 5 consecutive years for the same sample. I want 2 regreassion equations for each dependent variable. The data is non-linear. I did log-transformation but still the scatter plots still look non-linear
1. I don't want to take the log-transformed values because 2 of my variables have values less than 0.5, so taking logs make them all negative.
2. Can you suggest a methods to perform regressions on this data?

Thanks

[Carlo Lazzaro]

Padmavathi:
welcome to this forum.
The main issue here is that you have two simple (that is, with one predictor each) OLS.
What do you expect to retrieve from them?
I would stop at correlation and descriptive statistics.

[Padmavathi Bandaru]

Thnak you so much for your reply Carlo Lazzaro
Here is some additional info on my sutuation:

The variables I have are tariff, anti-dumping and emissions. I want to see how the emissions of a product effect the tariff and anti dumping over different years. I have the same product sample for different years. But can I perform Ols on panel and non-linear data? If not what regressions can I do?

When I graph the tariff and emissions or AD and emissions, the plots are too scattered and there is no trend - non-linear
the log transformation of one or all variables also doesn't help and don't want to because then all my emissions values and some tariff tariff are becoming negative because their original values are less than 0.5

[Carlo Lazzaro]

Padmavathi:
1) if you have panel data with a continuous dependent variable (tariff), go -xtreg-;
2) linearity is on coefficients, not variables;
3) a main potential concern is reverse causation-led endogeneity, though: while a change of 1 unit in emission can explain, other things being equal, variation in the conditional mean of teriff, the opposite may hold too. If that were the case, you should find out appropriate instruments to deal with this tricky issue.

[Padmavathi Bandaru]

Okay thank you so much Carlo Lazzaro

[Padmavathi Bandaru]

For the panel regression using -xtreg- for these variables, what would you suggest fixed effects or random effects?

[Padmavathi Bandaru]

Also, if I run reg only for a year individually the coefficient is negative and significant for all years . But when used xtreg (which include all years) the coefficient becomes positive and significant. Do you know what's happening here? (Also for previous question what is most suitable for this data fe or re?)

[Carlo Lazzaro]

Padmawathi:
1) you should start from -xtreg,fe-;
2) without sharing what you tyoed and what Stata gave you back (as per FAQ), it is impossible (for me, at least) to reply more positively.

[Padmavathi Bandaru]

Here is what I typed on stata and the type of coeffients I recieve each time
1. code : "xtreg tariff emissions, fe" gives small positive insignificant coefficient.
a)code: "reg tariff emissions if year == 2013" gives small negative significant coefficient.,
b)code: "reg tariff emissions if year == 2014" gives small negative insignificant coefficient and same for the rest of the years 15, 16, 17

2. code: "xtreg AD emissions, fe" gives small positive significant coefficient. reg AD emissions if year == 2017
a)code: "reg AD emissions if year == 2013" gives large negative significant coefficient and same for 2014
b)code: "reg AD emissions if year == 2015" gives large negative insignificant coefficient for the year 2016 and a small one for 2017

I want to know why the coefficients signs are changing when changed from individual year regressions to xtreg, and what to make of it?

[Carlo Lazzaro]

Padmawathi:
1) the way you share what you typed and what Stata gave you back should be improved (as per FAQ). Thanks.
2) that said:
a) it is not clear to me what you're expecting to get with one predictor only and, on the top of that, by limiting -regress- to one year only when you're dealing with a panel dataset;
b) your -xtreg,fe- code should include -i.year- among predictors;
c) your -regress- code is more similar to -xtreg,re- than -xtreg,fe-. To mimick -xtreg,fe- you should add -i.panelid- among the predictors (and -i.year-). Add -vce(cluster panelid)- to both -regress- and -xtreg,fe- codes if you have at least 30 panels. This way, you should obtain the same sample estimate of the coefficient (standard errors will differ, though);
d) use CODE delimiters to share what you typed and what Stata gave you back (see the FAQ). Thanks.

[Padmavathi Bandaru]

Thank you for your response Carlo
These are the new codes I used for regression now. (I have 39 panels so I added the vce(cluster panelid) )

xtreg tariff emissions i.naics_4 i.year, vce(cluster naics_4)

xtreg AD emissions i.naics_4 i.year, vce(cluster naics_4)

1. I have a paper that i read that did this analysis for only one year using OLS and got a negative and significant coefficients. Since I'm getting positive and significant coefficient for tariff and positive and insignificant for AD using this new code as you suggested. I'm trying to make sense of what's happening.

[George Ford]

#11 is not correct. i.naics is the fe. You could change xtreg to reg in this specification.

or

reghdfe tariff emissions, absorb(naics_4 year) cluster(naics_4)

or by year
reghdfe tariff i.year#c.emissions, absorb(naics_4 year) cluster(naics_4)


another option is mvreg.

mvreg tariff AD = emissions

To get year-by-year,

mvreg tariff AD = i.year#c.emissions

All of this is purely descriptive.

[George Ford]

I'd do it by year (at least as a check), since the coefficient's sign may change in such a simple model and you're trying to evaluate the prior study.

[Padmavathi Bandaru]

Thank you George. That really helped me a lot!!

[Padmavathi Bandaru]

Hi, Do you know how do I proceed with sensitivity analysis or robustness checks for this methodology? Thank you