Timeline for Why use logged variables?
Current License: CC BY-SA 3.0
5 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Aug 3, 2012 at 8:57 | vote | accept | Pr0no | ||
| Aug 2, 2012 at 12:13 | comment | added | Peter Flom | You don't want to take logs of percentages in this case: Taking the percentage does what the log would do. That is, it makes things ratios. Certainly you can take the log of some variables and not others. Taking the log doesn't require skewness charts, but usually variables that ought to be logged are right-skew. But the main thing is substance. If it makes no substantive sense to take the log, then don't do it. Instead, use statistical methods that work with skewed variables. SUBSTANCE comes first. | |
| Aug 2, 2012 at 11:53 | comment | added | Pr0no | Also, if I understand correctly, taking the logged variable isn't a free choice - it has to be argumented by skweness-charts (statistically)? And substantively is just more a logical reasoning for taking logged that actually provide proff for doing so? In other words, are there rules of thumb here, defining threshholds above which you should take logged version or is it a matter of interpretation? | |
| Aug 2, 2012 at 11:48 | comment | added | Pr0no | thank you for your quick reply. Two more questions come from your answer. First, if I have 3 properties for an object (stock tradingvolume, returns, and volatility) and take the logged version for one of them? What you say for company A and B tweets, might also count for their returns: if company A stock rises from 1 tot 1,50 then returns are (50%) 0.50. Company B needs an increase from 400 to 600 (200) for a similar %-return. And stemming from that: if returns are negative, LN(-0.50) obviously doesnt work. Is it then allowed to take -LN(0.50)? | |
| Aug 2, 2012 at 10:59 | history | answered | Peter Flom | CC BY-SA 3.0 |