0
$\begingroup$

I have data from 214 countries that range from 1990 to 2014. My dependent variables (I'm doing more than one regression) are just primary/secondary net enrollment rates for both sexes/males/females, so six variables. My independent variables are 'energy consumption (kWh per capita),' 'ict goods imports (% of total imports),' 'internet users (per 100 people),' '% with access to improved sanitation facilities,' 'health expenditure per capita PPP,' 'GNI PPP,' '#ofteachers in whatever I'm regressing,' and 'age dependency ratio, young.'

I'm using a panel data fixed effects model. My question is this — do I keep all of these independent variables as is? I have one professor who has suggested that I change them into growth rates, but the data has empty spaces in between different years in the same country.

Improve this question
$\endgroup$
2
  • $\begingroup$ A third option yet (see also the answer by @Danny) is consider working on transformed scales. Any variable increasing approximately exponentially might be better analysed on logarithmic scale. Any percent variable increasing from near 0 to near 100 might be better analysed on logit scale. $\endgroup$
    – Nick Cox
    Commented Apr 15, 2016 at 9:42
  • $\begingroup$ I think in your last paragraph you mean dependent variables. In my view, dependent and independent, although still very popular, are terrible terms, not least because so many people mix them up. Many other terms are available for dependent variables, including responses and outcomes. $\endgroup$
    – Nick Cox
    Commented Apr 15, 2016 at 9:45

2 Answers 2

Reset to default
1
$\begingroup$

It really depends on what questions you want to answer. You described your data and model, but not your questions.

If for example, you want to know 'what are the effects of increasing energy consumption by X kWh?' then you should leave the data alone. If, on the other hand, your question is 'what are the effects of increasing energy consumption 10%?' then you would want to convert to growth rates. Finally, if you want to compare the relative effects of energy consumption to, say, internet use then you would want to normalize your data. You do this by subtracting the mean from each value and then dividing by the standard deviation. This step will give unitless coefficients that are comparable between the different variables. Otherwise, you can't compare a regression coefficient that has units of kWh/person to a coefficient in connections per 100 people.

Improve this answer
$\endgroup$
1
$\begingroup$

If you don't know what questions you want to answer yet, it might be worth doing some bivariate plots of the data first. This would uncover some outliers, show you where the gaps in your data points are, whether the scales look ok and whether you might need to do some transformations, etc.

Improve this answer
$\endgroup$
1
  • $\begingroup$ I think you are right that in a dataset like this there would be outliers, but the general case is might be. $\endgroup$
    – Nick Cox
    Commented Apr 15, 2016 at 9:42

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.