Cannonical correlation

39
Canonical Correlation

description

 

Transcript of Cannonical correlation

Page 1: Cannonical correlation

Canonical Correlation

Page 2: Cannonical correlation

Introduction

 If we have two sets of variables, x1,...., xn and y1,….., ym, and there are correlations among the variables, then canonical correlation analysis will enable us to find linear combinations of the x's and the y's which have maximum correlation with each other.

Canonical correlation begin with the observed values of two sets of variables relating to the same set of areas, and a theory or hypothesis that suggests that the two are interrelated.

The overriding concern is with the structural relationship between the two sets of data as a whole, rather than the associations between individual variables

Page 3: Cannonical correlation

Canonical correlation is the most general form of correlation.

Multiple regression analysis is a more specific case in which one of the sets of data contains only one variable, while product moment correlation is the most specific case in that both sets of data contain only one variable.

Canonical correlation analysis is not related to factor/principal components analysis despite certain conceptual and terminological similarities. Canonical correlation analysis is used to investigate the inter-correlation between two sets of variables, whereas factor/principal components analysis identifies the patterns of relationship within one set of data.

Page 4: Cannonical correlation

Difficulties in Canonical Correlation

Canonical correlation is not the easiest of techniques to follow, though the problems of comprehension are conceptual rather than mathematical.

Unlike multiple regression and principal components analysis, we cannot provide a graphic device to illustrate even the simplest form. For with canonical correlation analysis we are dealing with two sets of data. Even the most elementary example must, therefore, have at least two variables on each side and so we require 2 + 2 = 4 dimensions. Tied as we are, however, to a three dimensional world, a true understanding of the technique in the conventional cognitive/visual sense of the term, is beyond our grasp.

Page 5: Cannonical correlation

Conceptual Overview

Data Inputi. The size of the matrices : There is no requirement in

canonical analysis that there must be the same number of variables (columns) in each matrix, though there must be the same number of areas (rows). (There must of course be more than one variable in each set otherwise we would be dealing with multiple regression analysis)

ii. The order of the matrices : Neither set of data is given priority in the analysis so it does not matter which we term the criteria and which the predictors. Unlike simple linear regression there is no concept of a 'dependent' set or an 'independent' set. But in practice the smaller set is always taken second as this simplifies the calculation enormously

Page 6: Cannonical correlation

Advantages

Useful and powerful technique for exploring the relationships among multiple dependent and independent variables. Results obtained from a canonical analysis should suggest answers to questions concerning the number of ways in which the two sets of multiple variables are related, the strengths of the relationships.

Multiple regressions are used for many-to-one relationships, canonical correlation is used for many-to-many relationships.

Canonical Correlation- More than one such linear correlation

relating the two sets of variables, with each

such correlation representing a different

dimension by which the independent set of

variables is related to the dependent set.

Page 7: Cannonical correlation

Interpretability:

Although mathematically elegant, canonical solutions are often un-

interpretable. Furthermore, the rotation of canonical variates to

improve interpretability is not a common practice in research, even

though it is commonplace to do this for factor analysis and principle

components analysis.

Linear relationship:

Another problem using canonical correlation for research is that

the algorithm used emphasizes the linear relationship between

two sets of variables. If the relationship between variables is not

linear, then using a canonical correlation for the analysis may

miss some or most of the relationship between variables.

Page 8: Cannonical correlation

The Canonical ProblemLatent Roots and weights Canonical ScoresResults and Interpretation

i. Latent Roots ii. Canonical Weights iii. Canonical Scores

Page 9: Cannonical correlation

Mathematical Model

The partitioned intercorrelation matrix

where R11 is the matrix of intercorrelations among the p criteria variablesR22 is the matrix of intercorrelations among the q predictor variablesR12 is the matrix of intercorrelations of the p criteria with the q predictorsR21 is the transpose of R12

Page 10: Cannonical correlation

The Canonical Equation

i. The product matrix

Page 11: Cannonical correlation

ii. The canonical roots

• The significance of the roots:Wilk’s Lambda (ᴧ) :

Bartlett’s chi squared:

Page 12: Cannonical correlation

• The canonical vectors Weights B for the predictor variables are given by :

Weights A for the criteria variables are given by :

Page 13: Cannonical correlation

The canonical scoresThe scores Sa for the criteria are given by

Sa = Zp A

The scores Sb for the predictors are given by

Sb = Zq B

where Zp and Zq are the standardized raw data

Page 14: Cannonical correlation

RESEARCHERS-A. O. UNEGBU &JAMES J. ADEFILA

Canonical correlation analysis-promotion bias scoring detector

(a case study of American university of Nigeria(AUN))

`

Page 15: Cannonical correlation

Introduction

Problem: AUN bids to keep with her value statement i.e. highest standards of integrity, transparency and academic honest.

Solution: Appraise & select Faculties for promotion based on various promotion committees’ scores.

Issues : Dwindling funding, need for a bias free selection

technique,

Page 16: Cannonical correlation

Research Hypotheses

H01 : CCA cannot detect bias scoring for any of the candidates from any of the named committees with 90% confidence level.

H02: CCA cannot detect significantly whether or not score-weights of each of the Promotion Assessors have over bearing influence on the

promotability of candidates.H03: CCA cannot at 90% level of certainity

discriminate between candidates that have earned promotion scores and those that could

not from various promotion committees of the university.

Page 17: Cannonical correlation

Research objectives

To test the efficacy of Canonical Correlation Analysis as a relevant statistical tool for adaption in bias free promotion score processing and promotion bias scoring detector so as to ensure fairness, integrity, transparency and academic honest in analysis of applicants’ score and in reaching Faculties’ promotion decision.

Page 18: Cannonical correlation

Steps of the Research

1) Data collection2) Manual computations3) SPSS analysis4) Test the Hypothesis

Page 19: Cannonical correlation

AUN promotion procedure

Weights:Dean of the School 7.5%

School Promotion Committee 7.5%

The Academic Vice President 10%

External Assessor/Reviewer 10%

University Wide Promotion Committee 15%

The Senate Committee 20%

President of the University 30%

Total 100%

The benchmark for promotion is securing a weighted average score should be more than 65%age.

Page 20: Cannonical correlation

Each of the Committee’s point allocation will be based on the below criteria

Teaching Effectiveness 40 %

Scholarship, research & creative works 40 %

Service to the University & to Community

20 %

Page 21: Cannonical correlation

Supporting documents for Teaching Effectiveness

Peer evaluation Student evaluation Course Syllabi Record of participation in teaching seminars,

workshops, etcContributions to the development of new

academic programsFaculty awards for excellence in teaching

Page 22: Cannonical correlation

Scholarship, Research and Creative Works

Terminal degrees/Professional qualificationsAt least Five publications, three of which shall

be journal articlesComputer Software and Program developmentCreative work in the areas of advertising,

public relations, layout design, photography and graphics, visual arts etc.

Page 23: Cannonical correlation

Service to the University, Profession and Community

Membership/leadership in departmental, school-wide or university-wide committees

Planning or participation in workshops, conferences, seminars .

Evidence of participation in mentoring or career counseling of students.

Membership in Civil Society organizationsEvidence of service as external assessor or external examiner on examination committees

Page 24: Cannonical correlation

Raw Scores of Candidates

Page 25: Cannonical correlation

Processed scores of the Candidates

Page 26: Cannonical correlation

Scores of Promotable and Non-promotable Candidates

Page 27: Cannonical correlation

Data Input

The data input view containing the three groups of assessors and individual assessors

Page 28: Cannonical correlation

SPSS Results

Analyze ⇒General Linear Model⇒Multivariate

SPSS classified candidates into two groups of promotable and non promotable of 5 and 9 respectively.

The result leads to the rejection of Null hypothesis Ho3 which states that Canonical Correlation Analysis cannot with 90% confidence level discriminate between promotable and non promotable candidates

Page 29: Cannonical correlation
Page 30: Cannonical correlation

Multivariate Test

The Multivariate tests indicate the effect of scores of the group and individual assessors both on status determination and bias impact on such status. The figure shows that the computed values and critical table values differences are very insignificant.

Candidate’s status determination resulting from scores across the assessors and those that might result from bias scoring are very insignificant(Wilk’s lambda value =0.041)

There is no between-status differences in the scores between assessors of both group and individuals

Rejection of Null hypothesis (Ho1) which states that Canonical Correlation Analysis cannot detect bias

Page 31: Cannonical correlation
Page 32: Cannonical correlation

The results of the table show that the scores of each assessor had a significant effect on the determination of each Candidate Status as the significance is 0.135.

Page 33: Cannonical correlation

Test for homogeneity of variance

Overbearing score weight influence test hypothesis is aimed at detecting across the individual assessors’ mark allocations and weights assigned to each.

In this test, the assessors having low significance value mean that there is homogeneity of variance.

Page 34: Cannonical correlation
Page 35: Cannonical correlation

This Leads to rejection of null hypothesis (Ho2) which states that Canonical Correlation Analysis cannot detect significantly whether or not score-weights of each of the promotion assessors has overbearing influence on the promotability of candidates.

Page 36: Cannonical correlation

Shortcomings and limitations of the process

Procedures that maximize correlation between canonical variate pairs do not necessarily lead to solutions that make logical sense. it is the canonical variates that are actually being interpreted and they are interpreted in pairs. a variate is interpreted by considering the pattern of variables that are highly correlated (loaded) with it. variables in one set of the solution can be very sensitive to the identity of the variables in the other set.

The pairings of canonical variates must be independent of all other pairs.

Page 37: Cannonical correlation

Conclusion from research analysis:

From Table it can be seen that the order of promotable rankings but application of Canonical Correlation Analysis results produced different ranking of candidates.

Rejection of Null Hypothesis(H03):The results as shown in tables indicate the Canonical Correlation Analysis status discriminatory ability of grouping Candidates into promotable and Non-promotable status. The result leads to the rejection of Null hypothesis Ho3 which states that Canonical Correlation Analysis cannot with 90% confidence level discriminate between promotable and nonpromotable candidates based on their earned scores.

Page 38: Cannonical correlation

Continued………….

Rejection of Null Hypothesis(Ho1):Pillar’s trace of 0.041, Wilk’s Lambda of 0.041, Hotelling’s trace of 0.041 and Roy’s Largest Root of 0.041 - all of them showed that p<0.05, it means that there is no between-status differences in the scores between assessors of both group and individuals, thereby leading to the rejection of Null hypothesis (Ho1) which states that Canonical Correlation Analysis cannot detect bias.

Rejection of Null Hypothesis(Ho2):For Group Assessors - Internal Assessors with p=0.096, External Academic Assessors with p=0.526 and The President’s Assessment with p=0.0001, shows that except that of the President, the weight assigned to scores of other two are group assessors are insignificant- lead us to reject the Null hypothesis (Ho2) which states that Canonical Correlation Analysis cannot detect significantly whether or not score-weights of each of the promotion assessors has overbearing influence on the promotability of candidates.

Page 39: Cannonical correlation

Thank You !!