glb.algebraic: Find the greatest lower bound to reliability.
In frenchja/psych: Procedures for Psychological, Psychometric, and Personality Research

Description Usage Arguments Details Value Author(s) References See Also Examples

The greatest lower bound solves the “educational testing problem". That is, what is the reliability of a test? (See guttman for a discussion of the problem). Although there are many estimates of a test reliability (Guttman, 1945) most underestimate the true reliability of a test.

For a given covariance matrix of items, C, the function finds the greatest lower bound to reliability of the total score using the csdp function from the Rcsdp package.

1	glb.algebraic(Cov, LoBounds = NULL, UpBounds = NULL)

`Cov`	A p * p covariance matrix. Positive definiteness is not checked.
`LoBounds`	A vector L = (l1 ... lp) of length p with lower bounds to the diagonal elements x_i. The default l=(0, . . . , 0) does not imply any constraint, because positive semidefiniteness of the matrix C0 + Diag(x) implies 0 ≤ xi.
`UpBounds`	A vector u =(u1, . . . , up) of length p with upper bounds to the diagonal elements xi. The default is u = v.

If C is a p * p-covariance matrix, v = diag(C) its diagonal (i. e. the vector of variances v_i = c_{ii}), C0 = C - Diag(v) is the covariance matrix with 0s substituted in the diagonal and x = the vector (x1, . . . , xp) the educational testing problem is (see e. g., Al-Homidan 2008)

(Sum i = 1 to p xi) -> min

s.t.

C0 + Diag(x) >= 0

(i.e. positive semidefinite) and xi ≤ vi, i = 1 ..., p. This is the same as minimizing the trace of the symmetric matrix

C0 + Diag(x)

s. t. C0 + Diag(x) is positive semidefinite and xi ≤ vi.

The greatest lower bound to reliability is

(sum cij (i \ne j) + sum xi )/ sum cij

Additionally, function glb.algebraic allows the user to change the upper bounds xi ≤ vi to xi ≤ ui and add lower bounds li ≤ xi.

The greatest lower bound to reliability is applicable for tests with non-homogeneous items. It gives a sharp lower bound to the reliability of the total test score.

Caution: Though glb.algebraic gives exact lower bounds for exact covariance matrices, the estimates from empirical matrices may be strongly biased upwards for small and medium sample sizes.

glb.algebraic is wrapper for a call to function csdp of package Rcsdp (see its documentation).

If Cov is the covariance matrix of subtests/items with known lower bounds, rel, to their reliabilities (e. g. Cronbachs α), LoBounds can be used to improve the lower bound to reliability by setting LoBounds <- rel*diag(Cov).

Changing UpBounds can be used to relax constraints xi ≤ vi or to fix xi-values by setting LoBounds[i] < -z; UpBounds[i] <- z.

`glb`	The algebraic greatest lower bound
`solution`	The vector x of the solution of the semidefinite program. These are the elements on the diagonal of C.
`status`	Status of the solution. See documentation of csdp in package Rcsdp. If status is 2 or greater or equal than 4, no glb and solution is returned. If status is not 0, a warning message is generated.
`Call`	The calling string

Andreas Moltner
Center of Excellence for Assessment in Medicine/Baden-Wurttemberg
University of Heidelberg

William Revelle
Department of Psychology
Northwestern University Evanston, Illiniois
http://personality-project.org/revelle.html

Al-Homidan S (2008). Semidefinite programming for the educational testing problem. Central European Journal of Operations Research, 16:239-249.

Bentler PM (1972) A lower-bound method for the dimension-free measurement of internal consistency. Soc Sci Res 1:343-357.

Fletcher R (1981) A nonlinear programming problem in statistics (educational testing). SIAM J Sci Stat Comput 2:257-267.

Shapiro A, ten Berge JMF (2000). The asymptotic bias of minimum trace factor analysis, with applications to the greatest lower bound to reliability. Psychometrika, 65:413-425.

ten Berge, Socan G (2004). The greatest bound to reliability of a test and the hypothesis of unidimensionality. Psychometrika, 69:613-625.

For an alternative estimate of the greatest lower bound, see glb.fa. For multiple estimates of reliablity, see guttman

Cv<-matrix(c(215, 64, 33, 22,
              64, 97, 57, 25,
              33, 57,103, 36,
              22, 25, 36, 77),ncol=4)

Cv                    # covariance matrix of a test with 4 subtests
Cr<-cov2cor(Cv)       # Correlation matrix of tests
if(!require(Rcsdp)) {print("Rcsdp must be installed to find the glb.algebraic")} else {
 glb.algebraic(Cv)     # glb of total score
glb.algebraic(Cr)      # glb of sum of standardized scores

 w<-c(1,2,2,1)         # glb of weighted total score
 glb.algebraic(diag(w) %*% Cv %*% diag(w))  
alphas <- c(0.8,0,0,0) # Internal consistency of first test is known

glb.algebraic(Cv,LoBounds=alphas*diag(Cv))

                      # Fix all diagonal elements to 1 but the first:

lb <- glb.algebraic(Cr,LoBounds=c(0,1,1,1),UpBounds=c(1,1,1,1))
lb$solution[1]        # should be the same as the squared mult. corr.
smc(Cr)[1] 
}