Insignificant priors from the Pivot Design experiment
Posted: Wed Apr 15, 2015 8:33 pmHi,
I have generated a pivot design experiment and conducted a pilot survey in order to find the prior values of parameters. Following are the details of the survey design:
Design Type: Pivot Design
Model command used in Ngene: ;eff = fish(mnl,d) --> (Eventually I wish to use rppanel for estimation. However, I received a feedback through forum to use MNL for the pilot stage).
No of alternatives: 3
No of attributes per alternative: 4
No. of levels per attribute: 5
No. of blocks: 2
Scenarios per block: 10
Valid responses received: 15
Total rows of SP data: 150
I fit a MNL model on these 150 rows of data and getting the following estimates:
Name Coefficient p-value
TTIME -0.146 0.00
SNGO -0.0163 0.39
SNTTIME -0.0121 0.56
VRC -0.949 0.00
The sign of all these estimates is fine. The issues are, The estimates SNGO and SNTTIME are insignificant (even at 30%) & coefficient of SNGO is too low.
I checked another post on the forum with the subject "insignificant priors from the pilot data" and found it helpful.
--> Is it okay to say that the deviations from the expected outcome might be because of the small sample size & I should try to get more responses.
--> Is any intervention required to alter the experimental design or the format of the survey?
--> Prof. Rose mentioned something about "preferences in the population are close to zero, but not zero". Does that apply to my SNGO estimate too? Currently, I used the attrbute levels as {-.5, -.25, 0, +.25, +.5}. Will I be able to capture the preference of SNGO by expanding this interval beyond {-.5, +.5} ??
I'll be grateful if someone can assist me in clarifying the doubts.
Thanks in advance.
Neeraj
I have generated a pivot design experiment and conducted a pilot survey in order to find the prior values of parameters. Following are the details of the survey design:
Design Type: Pivot Design
Model command used in Ngene: ;eff = fish(mnl,d) --> (Eventually I wish to use rppanel for estimation. However, I received a feedback through forum to use MNL for the pilot stage).
No of alternatives: 3
No of attributes per alternative: 4
No. of levels per attribute: 5
No. of blocks: 2
Scenarios per block: 10
Valid responses received: 15
Total rows of SP data: 150
I fit a MNL model on these 150 rows of data and getting the following estimates:
Name Coefficient p-value
TTIME -0.146 0.00
SNGO -0.0163 0.39
SNTTIME -0.0121 0.56
VRC -0.949 0.00
The sign of all these estimates is fine. The issues are, The estimates SNGO and SNTTIME are insignificant (even at 30%) & coefficient of SNGO is too low.
I checked another post on the forum with the subject "insignificant priors from the pilot data" and found it helpful.
--> Is it okay to say that the deviations from the expected outcome might be because of the small sample size & I should try to get more responses.
--> Is any intervention required to alter the experimental design or the format of the survey?
--> Prof. Rose mentioned something about "preferences in the population are close to zero, but not zero". Does that apply to my SNGO estimate too? Currently, I used the attrbute levels as {-.5, -.25, 0, +.25, +.5}. Will I be able to capture the preference of SNGO by expanding this interval beyond {-.5, +.5} ??
I'll be grateful if someone can assist me in clarifying the doubts.
Thanks in advance.
Neeraj