You are on page 1of 7

TAPAJYOTI GHOSH

CHBE 5779

HOMEWORK #9

Problem 6.7:
a) Solution:
= 0.05

First we test the model with all the possible effects.


Here the factors have been taken as categorical variables
because we have non numeric variables. Every factor has
only two levels. Thus this is 2^k factorial design. It is
used for screening factors.
From the whole model test, we find that atleast one of the
factors is significant. So we go to effects test. The effects
shows us which factors are significant( p < alpha) and
which are not significant. (p > alpha). Now we determine
how to get to am adequate model by removing the factors
which are not significant.. However, we should always
remember that we should remove one factor at a time.
And refit it everytime rather than removing all the not
significant effects at once.
So we start by moving the higher order effects. First
removing the ternary effect of C*D*A. And redoing the
analysis.

From the whole model test, we find that atleast one of


the factors is significant. So we go to effects test. The
effects shows us which factors are significant( p <
alpha) and which are not significant. (p > alpha). Now
we determine how to get to an adequate model by
removing the factors which are not significant.
So for the next analysis we remove the effect of
B*C*D and redo the entire analysis.

From the whole model test, we find that atleast one of


the factors is significant. So we go to effects test. The
effects shows us which factors are significant( p <
alpha) and which are not significant. (p > alpha). Now
we determine how to get to an adequate model by
removing the factors which are not significant.
So for the next analysis we remove the effect of B*D
and redo the entire analysis. The other two ternary
effects are now significant so we keep them in our
model.

From the whole model test, we find that


atleast one of the factors is significant. So we go to
effects test. The effects shows us which factors are
significant( p < alpha) and which are not
significant. (p > alpha). Now we determine how to
get to an adequate model by removing the factors
which are not significant.
So for the next analysis we remove the effect of
B*C and redo the entire analysis

From the whole model test, we find


that atleast one of the factors is
significant. So we go to effects
test. The effects shows us which
factors are significant( p < alpha)
and which are not significant. (p >
alpha). Now we determine how to
get to an adequate model by
removing the factors which are not
significant.
So for the next analysis we remove
the effect of B*C and redo the
entire analysis

From the whole model test, we find that


atleast one of the factors is significant. So we go to
effects test. The effects shows us which factors are
significant( p < alpha) and which are not
significant. (p > alpha). Now we determine how to
get to an adequate model by removing the factors
which are not significant.
So for the next analysis we remove the effect of
A*C and redo the entire analysis

From the whole model test, we find that


atleast one of the factors is significant. So we go to
effects test. The effects shows us which factors are
significant( p < alpha) and which are not
significant. (p > alpha). Now we determine how to
get to an adequate model by removing the factors
which are not significant.
So for the next analysis we remove the effect of
C*D and redo the entire analysis.

From the whole model test, we find that


atleast one of the factors is significant. So we go to
effects test. The effects shows us which factors are
significant( p < alpha) and which are not
significant. (p > alpha). Now we determine how to
get to an adequate model by removing the factors
which are not significant.
Thus finally we get the best model which is the
most adequate. I believe this model is correct and
adequate fro the given data. I decide to keep the
main effect of B even thought it is not significant
as it is a main effect and we should include it in the
model. We might have decided to remove it too.
Thus now as we have the model, we can proceed
for residual analysis.

Oh No! OUTLIER. So we had an outlier in the experiment. I might have to do my entire analysis over
again. But I decide not to. Why? Because the number of outliers are really less. Only one. And we have
32 runs. So probably, or model would be robust enough such that even if we remove this outlier from
our analysis, out model which we have already derived will still work fine. And also, and we have two
replicates for a data point. So it is difficult to say which one is the outlier. I think the number of runs is
not adequate. We should have more runs to decide whether it is bad data or bad model.

Yes! My model works without the outlier data point


too. So it is robust and I have probably done right.
Hence, even if I had included in the process of
developing the model, the model was quite robust and
the presence of just a single outlier ensured that I
didnt stray towards a very wrong direction. Hence, if I
had redone the entire analysis, I would have probably
got the same model.

Ah! So Finally I reach residual analysis. Very nice. Normaly distributed data. Goodness of fit test is
showing that is actually normally distributed. No outliers. Graph all good. Yup.Good to Go.

Plot of residuals vs predicted values shows equal variance at all levels. Hence our assumption was
correct. Thus our model is adequate.

A quick power analysis of the main effect shows that the main effect of B would might have been
rendered significant if we had done nearly 284 runs. TOO many runs! No need. Forget B. Power is low
though. High chance of type II error..hmmmm.

You might also like