t*****y
发帖数: 30 | 1 I am using linear multiregression to fit my data
ie y=ax1+bx2+cx3+dx4+ex5+fx6+g
The final R squre from the fitting is 81.94%. It is also the correlation
coefficient of experimental y and fitted y, which can be plotted.
I am wondering if there is nonlinearity buried in the data since 81.94% is
not very high.
But from the plot, I couldn't see very obvious curvature. But even if
the real relationship is nonlinear, it should be a curved plane in 6
dimensional space. I suppose I am not able to visul | r***a
发帖数: 58 | 2 First of all 81.94% is not low for a R square. Indeed, it is a decent one.
You should also look at the p-values of the coefficients, see if they
all are significant. If not, you can consider get rid of those non significant
ones one by one and fit the model. For sure, the less coefficent you have
in your model, the smaller R square you will get. Keep in mind that R square
is not the only criteria for model fitting.
You can also try qnorm plot to see if the residuas are normally distributed.
You
【在 t*****y 的大作中提到】
: I am using linear multiregression to fit my data
: ie y=ax1+bx2+cx3+dx4+ex5+fx6+g
: The final R squre from the fitting is 81.94%. It is also the correlation
: coefficient of experimental y and fitted y, which can be plotted.
: I am wondering if there is nonlinearity buried in the data since 81.94% is
: not very high.
: But from the plot, I couldn't see very obvious curvature. But even if
: the real relationship is nonlinear, it should be a curved plane in 6
: dimensional space. I suppose I am not able to visul
| t*****y
发帖数: 30 | 3 Thank you!
A set of coefficient looks like this:
Parameter Value Error t-Value Prob>|t| | r***a
发帖数: 58 | 4 p-value = Prob>|t|
if a p-value is greater than 0.05 (most commonly used criteria), the
coefficent is not significant. I will say C is not significant for sure,
get it out of the equation, run the regression again, so on so force
【在 t*****y 的大作中提到】
: Thank you!
: A set of coefficient looks like this:
: Parameter Value Error t-Value Prob>|t|
| t*****y
发帖数: 30 | 5 Thank you Ronna!
I am a chem major. These parameters are chemical structures related. I am
trying to find out how the structures influence the property Y. But from
statistics point of view, most of the parameters are not significant
(P-value>0.05) which makes me worry. Does it mean the model is nonsense?
But R-square is quite high >80% that means these parameters can be
regressed toward Y. Here, intercept is very significant. I will
interperate it as an system environmental related of course, it
【在 r***a 的大作中提到】
: p-value = Prob>|t|
: if a p-value is greater than 0.05 (most commonly used criteria), the
: coefficent is not significant. I will say C is not significant for sure,
: get it out of the equation, run the regression again, so on so force
| r***a
发帖数: 58 | 6 The more independent variables you include, the higher R square will be,
but it doesn't mean the more independent variables the better. Get rid
of the nonsignificant ones won't have too much effect on R square.
Statistics is about model fitting, not about stories telling. If the
coefficent is not significant, it is not, no matter how much you want it
to be, unless you change the dataset.
Sure, you can say which one is more important (significant), however, there
is no reason to include a coeffic
【在 t*****y 的大作中提到】
: Thank you Ronna!
: I am a chem major. These parameters are chemical structures related. I am
: trying to find out how the structures influence the property Y. But from
: statistics point of view, most of the parameters are not significant
: (P-value>0.05) which makes me worry. Does it mean the model is nonsense?
: But R-square is quite high >80% that means these parameters can be
: regressed toward Y. Here, intercept is very significant. I will
: interperate it as an system environmental related of course, it
|
|
