Timeline for Regression coefficient has negative symbol but positive from the raw plot
Current License: CC BY-SA 4.0
16 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Nov 27, 2018 at 9:37 | comment | added | Glen_b | Many previous posts discuss this effect; searches on simpson's paradox or omitted variable bias. I haven't closed as duplicate because I think there's a more specific aspect to the question here (but other users may disagree with that assessment) | |
| Nov 27, 2018 at 1:45 | vote | accept | 89_Simple | ||
| Nov 27, 2018 at 1:19 | comment | added | user158565 | In 10 grouped scatterplots of x1 vs log(yields), it is hard to find the positive relation between them. Some of them implies negative relation. It means that when X2 is fixed, X1 and log(yield) has no relation or weak negative relation. It is what " x1 -0.07341 " means. | |
| Nov 27, 2018 at 0:45 | comment | added | Nick Cox |
x1 and x2 are both sharply bounded by maxima suggesting that they are originally fractions or percents of some maximum that is attainable (whereas a sharp minimum is not evident). I would explore using some transformation using folded root, just to try to see what is happening. See stats.stackexchange.com/questions/184247/… for what may be a similar case.
|
|
| Nov 27, 2018 at 0:24 | comment | added | 89_Simple | @whuber Thank you. I have added some additional comments to my question in line with what Ben explained below. Does it mean that the model is wrong and should not be used for any prediction purposes? | |
| Nov 27, 2018 at 0:23 | history | edited | 89_Simple | CC BY-SA 4.0 |
added 603 characters in body
|
| Nov 26, 2018 at 23:43 | comment | added | whuber♦ | Separating the observations into groups is too crude to provide insight into this relatively subtle effect. Instead, regress log(yield) and x1 separately against x2 and examine a scatterplot of the residuals on which a linear fit has been superimposed. You will then be able to see the negative trend and the reason for it. There is a slight but definite change in behavior for $x_1 \lt 0.62$ vs $x_1 \gt 0.62.$ This is entirely buried in the stack of points in the last scatterplot you show, making it impossible to see. | |
| Nov 26, 2018 at 23:23 | answer | added | Ben | timeline score: 3 | |
| Nov 26, 2018 at 23:05 | comment | added | 89_Simple | @user158565 I have done what you suggested but I cannot figure out how does that answer my question | |
| Nov 26, 2018 at 23:03 | history | edited | 89_Simple | CC BY-SA 4.0 |
added 456 characters in body
|
| Nov 26, 2018 at 22:39 | comment | added | 89_Simple | If there is negative corelation between x1 and x2, then I thought the regression coefficient would switch signs. Bit in this particular case, x1 and x2 are positively correlated, then why would their regression coefficients have opposite sign? | |
| Nov 26, 2018 at 22:37 | comment | added | 89_Simple | Thanks @RobertLong. I have added the link to the data. | |
| Nov 26, 2018 at 22:36 | history | edited | 89_Simple | CC BY-SA 4.0 |
added 8 characters in body
|
| Nov 26, 2018 at 19:24 | comment | added | Robert Long | It seems that the explanatory variables are highly collinear which probably explains it, but without more information or access to the data it is hard to say..... It is always good to provide the data.... | |
| Nov 26, 2018 at 18:32 | comment | added | user158565 | Separate the observations in to 10-15 groups according to the value of x2, for each group, generate scatterplot of x1 vs log(yield). Then you may get the idea what " x1 -0.07341" means. | |
| Nov 26, 2018 at 17:39 | history | asked | 89_Simple | CC BY-SA 4.0 |