Discriminant analysis: Fisher’s scores (like SPSS)Discriminant analysis: Fisher’s scores (like SPSS)
Statistics and AnalyticsRe: Discriminant analysis: Fisher’s scores (like SPSS) - Status changed to: InvestigatingRe: Discriminant analysis: Fisher’s scores (like SPSS)Hi Sam,
Here is more information, if helpful: https://www.ibm.com/docs/en/spss-statistics/beta?topic=analysis-discriminant-statistics
I am running a simple LDA on mycluster with all 6 basis variables (orange...blueberry).
Here is an example:
id orange grapefruit lemon strawberry raspberry blueberry mycluster
1.00 89.00 81.00 85.00 55.00 70.00 56.00 2
2.00 90.00 83.00 94.00 59.00 42.00 61.00 2
3.00 95.00 79.00 87.00 61.00 58.00 49.00 2
4.00 56.00 41.00 62.00 88.00 84.00 83.00 1
5.00 71.00 56.00 61.00 92.00 86.00 85.00 1
6.00 66.00 68.00 60.00 83.00 90.00 89.00 1
7.00 68.00 62.00 45.00 79.00 98.00 91.00 1
8.00 82.00 81.00 85.00 55.00 70.00 56.00 2
9.00 90.00 88.00 94.00 59.00 42.00 61.00 2
10.00 95.00 79.00 84.00 61.00 58.00 49.00 2
11.00 56.00 41.00 62.00 93.00 84.00 83.00 1
12.00 71.00 56.00 61.00 92.00 84.00 85.00 1
13.00 66.00 68.00 60.00 83.00 90.00 91.00 1
14.00 68.00 62.00 45.00 79.00 95.00 91.00 1
Desired LDA Output ("mycluster 1 2" are headers for the 3 columns in the table below):
Classification Function Coefficients
mycluster 1 2
orange 7.318 15.466
grapefruit -6.189 -10.719
lemon .488 12.336
strawberry 42.241 13.117
raspberry 17.201 14.129
blueberry 45.019 30.221
(Constant) -4625.059 -2422.181
Fisher's linear discriminant functions
Re: Discriminant analysis: Fisher’s scores (like SPSS) - Status changed to: Needs InfoTo help us with understanding the issue and need more fully, would you be able to provide an example data table or more full analysis using your desired approach?
Re: Discriminant analysis: Fisher’s scores (like SPSS) - Status changed to: Acknowledged@reese_sj thank you for your suggestion. We have captured your request and will take it under consideration.
Re: Discriminant analysis: Fisher’s scores (like SPSS)Detail:
Occasionally, R&D creates product category segments from a large number of questions using cluster analysis. The next step is to use stepwise linear discriminant analysis (LDA) to reduce the number of questions needed to accurately predict the respondent segment. From the output, we used to provide a set of "Fisher's Scores" (using the SPSS language) from the LDA model to research suppliers so that they could create an algorithm to classify respondents. (Let n=number of respondents, k=number of questions in the reduced LDA model, and m=number of segments). Using linear algebra, this required multiplying the n x k respondent data matrix times the k x m Fisher Score matrix (and summing over k plus the intercept) to calculate the LDA score for each respondent by segment. The segment with the highest score was assigned the respondent. Could you produce a similar Fisher's Score matrix (k x m, plus the intercept) for the m segments as part of your output?
Re: Discriminant analysis: Fisher’s scores (like SPSS)Hi again,
I have also seen these scores called, "Fisher's classification function coefficients" if it helps to track them down.
Here's my motivation --
Once a useful cluster analysis has been performed on survey respondents' data, the researchers would like a way to predict the cluster of new respondents in future surveys using just a few items from the original list. So, we turn to Stepwise LDA to create a prediction model. However, I would like to have an easy set of "Fisher's scores" (i.e. sets of coefficients plus the intercept --- one set per cluster) that I can embed in my next survey to screen/qualify respondents depending on their predicted cluster. Basically, I just need to multiply the data times the coefficients and then add the intercept -- for each cluster and then compare the scores to predict the cluster. Alternatively, I can take these "SUMPRODUCT" scores (using Excel function vernacular) and exponentiate them and then express them as a ratio vs. the sum of the exponentials to calculate the probability of belonging to the cluster.
Finally, it would be really helpful to add Leave One Out or Cross-validation to your stepwise selection process to avoid overfitting the model (for one reason). Moreover, the goal of the variable selection is to find the ABSOLUTE MINIMUM number of items in the model that are needed to predict cluster membership, since the researcher does not wish to waste survey space asking a bunch of items (survey questions) just to predict a cluster.
Re: Discriminant analysis: Fisher’s scores (like SPSS) - Status changed to: InvestigatingThank you for providing this additional information @hengehold_da1. We are researching.
Re: Discriminant analysis: Fisher’s scores (like SPSS) - Status changed to: AcknowledgedRe: Discriminant analysis: Fisher’s scores (like SPSS)Hi JMP friends,
Any news to share on this idea? Have you been able to find how to compute Fisher classification scores? If not, I believe a colleague of mine may have found an R package that can calculate them (in addition to SPSS). Let me know if you need a lead. (We are trying to make the method accessible to all of our (P&G) JMP users.).
Thanks!
