I have two different predictors "A" and "B" that can be used (not at the same time) to affect a response variable, "C".

Previous data has shown a very strong linear relationship between A and C, and B and C. When SLR models were fit to these relationships, both had R-squared values above 95%. Scatterplots confirm the highly linear relationships.

A claim has been made that when predictor "B" is applied, there is a 5% increase in the response "C", compared to when "A" is applied.

The two linear models fit to (A~C) and (B~C) have different slopes and intercepts. So for lower levels of A (or B), one predictor is associated with higher "C" measurements, but this is slightly reversed at higher levels.

Does anyone have an idea on how to test the 5% claim mentioned above? I was thinking it might be appropriate to create a model like this one:

$C = (U + V*X)*(1 + Y*B)$, where $U, V$ and $Y $ are constants.

X is the predictor level (either A or B). "B" in this model is an indicator variable such that B = 1 if B is being used, or B = 0 if A is used. If the 5% claim were true, we would expect Y to be .05.

The problem is that this model is not linear in the parameters, and as far as I know cannot be calculated using the common least-squares matrix operations.

Can anyone think of how to validate this model, or offer a better alternative to test the desired claim?

Thanks a lot for your help.

  • $\begingroup$ Are your predictors of the type Applied/Not applied, or can they be applied with different intensity level. $\endgroup$ – AlainD Aug 22 '18 at 8:10
  • $\begingroup$ I am sorry but I don't understand your question. $\endgroup$ – Greg Aug 22 '18 at 14:08

Adding 5% is tantamount to multiply by 1.05 [$x+5\%x = x + 0.05\cdot x = 1.05 \cdot x$].

So to test if there is a % increase on $C$, you should test the model $C \approx b^U \cdot c$, where $U = 1$ when $B$ is applied and $U = 0$ when $A$ is applied. You can linearize by taking the log on both sides $\log(C) \approx log(b) \cdot U + log(c)$.

  • The linear regression of the logs will give you $\log(b)$ and its p-sign, so that you will know of $B$ is significant or not.
  • The measure the percent increase is given by $p = (\exp(log(b))-1) \times 100$.
  • To test if it is equal to 5%, compute its Student by dividing the deviation by its Std-error (given by the regression) $t= \frac{ \log(b) - \log(1.05)} {std_{log(b)}}$.

If it is too sharp, ask questions preferably with the regression statistics.

  • $\begingroup$ Thanks for your idea. My main concern is the lowercase "c" in your model. If capital "C" is the response, then are you effectively modeling it with only a constant term, which differs whether A or B is used? As stated, there is a linear relationship between both C and A and C and B, so I feel that that should be represented somewhere in the model. $\endgroup$ – Greg Aug 22 '18 at 14:07
  • $\begingroup$ None that I know. If one factor (A) is addiive and the other (B) is multiiplicative, I do not know how to linearize both in the same model. You are stuck to study one after the other and compare the Student statistics. $\endgroup$ – AlainD Aug 23 '18 at 10:46

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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