1
$\begingroup$

I'm intending to do a response optimization of one response, $y$, having three predictor variables, $x_1$, $x_2$, and $x_3$. These variables are coded in the following manner:

A    B   C   y
-1.00000    -1.00000    -1.00000    66
 1.00000    -1.00000    -1.00000    80
-1.00000     1.00000    -1.00000    78
 1.00000    1.00000     -1.00000    100
-1.00000    -1.00000     1.00000    70
 1.00000    -1.00000     1.00000    100
-1.00000    1.00000      1.00000    60
 1.00000    1.00000     1.00000     75
-1.68179    0.00000     0.00000     100
1.68179     0.00000      0.00000    80
0.00000    -1.68179     0.00000     68
0.00000     1.68179     0.00000     63
0.00000     0.00000     -1.68179    65
0.00000     0.00000      1.68179    82
 0.00000    0.00000      0.00000    113
0.00000     0.00000     0.00000     100
0.00000     0.00000     0.00000     118
0.00000     0.00000     0.00000     88
0.00000     0.00000    0.00000       100
0.00000     0.00000     0.00000     85

What I have tried is (in Minitab)

Stat -> DOE -> Define Custom Response Surface

And choosing 3 responses, 6 center points, 1 replicate and alpha 1,682 and subsequently choosing

Stat -> DOE -> Optimize Response variable

and choosing Maximize and lowest 100 and target 118 I got that the maximum for $y$ is

101.711

and D=0.41

but A=0.6, B = -0.11 and C=0.15

which, if I've understood correctly, is invalid for this design.

When I've googled all I see is people who merely make surface plots and conclude in what path the response is optimized - however using three variables I do not know how to perform this.

$\endgroup$
  • $\begingroup$ Welcome to CrossValidated! You might want to consider either (1) asking this question on a Minitab-related site, where everyone will understand the terms you're using, or (2) writing it in a more general way so that non-Minitab users can help. (Also why do you think the optimum is incorrect?) $\endgroup$ – Scortchi - Reinstate Monica Nov 1 '13 at 17:12
  • $\begingroup$ Thank you! Oh, I was under the impression that CrossValidated was a subsite to StackExchange handling statistics and MiniTab. But I will post it in forum more specialized in these topics. I believe the optimial solution is wrong as it is not a point in the points I've constructed my fractional design (0,0,0), (1,0,0) and so on. I might have misunderstood though that they need to be in those points. $\endgroup$ – Cenderze Nov 3 '13 at 11:59
  • $\begingroup$ No, there's no reason the maximum or minimum of a response surface should happen to fall exactly on a design point. And your question will be fine for this site if you just add a little more context to make it understandable. E.g. what's 'D' ? What's a 'target' of 118 mean if you're just trying to maximize the response? $\endgroup$ – Scortchi - Reinstate Monica Nov 3 '13 at 12:40
  • $\begingroup$ Ah I see! Will have that in mind the next time. For now, the exercise is rather teoretic and the different effects are not defined. The target of 118 is merely that the highest value, y, defined was 118. $\endgroup$ – Cenderze Nov 3 '13 at 18:57
  • $\begingroup$ Found this. It seems that when you use 'maximize' in Minitab's response optimizer, 'target' is the response value above which all values are equally bad, & 'lower' is the value below which all values are equally good. So if you want to find the true maximum just put 'target' as anything implausibly high & 'lower' as anything implausibly low, I think. Strikes me as a little bizarre, but looks like your results should be fine (was 'C=0,15' a typo for 'C=0,13'?). $\endgroup$ – Scortchi - Reinstate Monica Nov 3 '13 at 19:44
2
$\begingroup$

This is a central composite design so I assume you're fitting a full second-order model for the mean response $\mathrm{E}(Y)$ on continuous predictors $x_1$, $x_2$, & $x_3$

$$\mathrm{E}(Y)= \beta_0 + \beta_1 x_1 + \beta_2 x_2 + \beta_3 x_3 + \beta_{12} x_1 x_2 + \beta_{13} x_1 x_3 + \beta_{23} x_2 x_3 + \beta_{11} x_1^2 +\beta_{22} x_2^2 +\beta_{33} x_3^2$$

& estimating the coefficients $\beta$ by ordinary least squares.

I fitted the model: the estimated mean response has a stationary point, a maximum, of $\mathrm{E}(Y)=102$ at $x_1=0.62, x_2=-0.11, x_3=0.13$. (You can check for yourself by differentiating the equation for the response with respect to each predictor, setting each derivative to zero (all slopes are zero at a stationary point), & solving the resulting simultaneous equations.) Contour plots at slices through the stationary point are a good way of visualizing the fitted model: response surface contour plots

The $95\%$ confidence interval for $\mathrm{E}(Y)$ at the maximum is $(89,114)$, & the $95\%$ prediction interval for $Y$ at the maximum is $(68,136)$. These are rather wide compared to the range of the response across the whole design. Indeed the residual standard deviation is $14$ (just look at the spread of responses over your centre points). I don't know the context of your experiment, but in many situations this would be cause for concern—are there other factors significantly contributing to process variability that you haven't taken into account?

$\endgroup$
  • $\begingroup$ Thanks for a very elaborate and thorough answer! Really appreciated. I will try to have thisin mind. The exercise I have mandates that the exercise is to be completed using MiniTab software. Yur answer certainly gave me a higher insight into fitted models and the actual mathematic behind MiniTab, which I fathom is fairly similar. $\endgroup$ – Cenderze Nov 3 '13 at 19:00
  • $\begingroup$ You're welcome. And to be clear, if you understand exactly what you want to do statistically, & just want to know the Minitab commands to do it, it's off-topic for CV (& more likely to get a quick response elsewhere, I think); but if you've any doubts at all about the statistics please do ask here - questions about experimental design & analysis are very welcome. Your question was fine for CV; just not very clear for people who don't use Minitab. $\endgroup$ – Scortchi - Reinstate Monica Nov 3 '13 at 19:17

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.