Description

Given a data.frame or matrix of n items and N observations and a list of the direction to score them (a keys.list with k keys) find the sum scores or average scores for each person and each scale. In addition, report Cronbach's alpha, Guttman's Lambda 6, the average r, the scale intercorrelations, and the item by scale correlations (raw and corrected for item overlap). Replace missing values with the item median or mean if desired. Items may be keyed 1 (score it), -1 ) (reverse score it), or 0 (do not score it). Negatively keyed items will be reverse scored. Although prior versions used a keys matrix, it is now recommended to just use a list of scoring keys. See make.keys for a convenient way to make the keys file. If the input is a square matrix, then it is assumed that the input is a covariance or correlation matix and scores are not found, but the item statistics are reported. (Similar functionality to cluster.cor). response.frequencies reports the frequency of item endorsements fore each response category for polytomous or multiple choice items. scoreFast and scoreVeryFast just find sums/mean scores and do not report reliabilities. Much faster for large data sets.

Usage

Arguments

Details

The process of finding sum or average scores for a set of scales given a larger set of items is a typical problem in applied psychometrics and in psychometric research. Although the structure of scales can be determined from the item intercorrelations, to find scale means, variances, and do further analyses, it is typical to find scores based upon the sum or the average item score. For some strange reason, personality scale scores are typically given as totals, but attitude scores as averages. The default for scoreItems is the average as it would seem to make more sense to report scale scores in the metric of the item.

When scoring more than one scale, it is convenient to have a list of the items on each scale and the direction to score the items. This may be converted to a keys.matrix using make.keys or may be entered as a keys.list directly.

Various estimates of scale reliability include “Cronbach's alpha", Guttman's Lambda 6, and the average interitem correlation. For k = number of items in a scale, and av.r = average correlation between items in the scale, alpha = k * av.r/(1+ (k-1)*av.r). Thus, alpha is an increasing function of test length as well as the test homeogeneity.

Surprisingly, more than a century after Spearman (1904) introduced the concept of reliability to psychologists, there are still multiple approaches for measuring it. Although very popular, Cronbach's α (1951) underestimates the reliability of a test and over estimates the first factor saturation.

alpha (Cronbach, 1951) is the same as Guttman's lambda3 (Guttman, 1945) and may be found by

Lambda 3 = (n)/(n-1)(1-tr(Vx)/(Vx) = (n)/(n-1)(Vx-tr(Vx)/Vx = alpha

Perhaps because it is so easy to calculate and is available in most commercial programs, alpha is without doubt the most frequently reported measure of internal consistency reliability. Alpha is the mean of all possible spit half reliabilities (corrected for test length). For a unifactorial test, it is a reasonable estimate of the first factor saturation, although if the test has any microstructure (i.e., if it is “lumpy") coefficients beta (Revelle, 1979; see ICLUST) and omega_hierchical (see omega) (McDonald, 1999; Revelle and Zinbarg, 2009) are more appropriate estimates of the general factor saturation. omega_total (see omega) is a better estimate of the reliability of the total test.

Guttman's Lambda 6 (G6) considers the amount of variance in each item that can be accounted for the linear regression of all of the other items (the squared multiple correlation or smc), or more precisely, the variance of the errors, e_j^2, and is

lamada 6 = 1 - sum(e^2)/Vx = 1-sum(1-r^2(smc))/Vx.

The squared multiple correlation is a lower bound for the item communality and as the number of items increases, becomes a better estimate.

G6 is also sensitive to lumpyness in the test and should not be taken as a measure of unifactorial structure. For lumpy tests, it will be greater than alpha. For tests with equal item loadings, alpha > G6, but if the loadings are unequal or if there is a general factor, G6 > alpha. Although it is normal when scoring just a single scale to calculate G6 from just those items within the scale, logically it is appropriate to estimate an item reliability from all items available. This is done here and is labeled as G6* to identify the subtle difference.

Alpha and G6* are both positive functions of the number of items in a test as well as the average intercorrelation of the items in the test. When calculated from the item variances and total test variance, as is done here, raw alpha is sensitive to differences in the item variances. Standardized alpha is based upon the correlations rather than the covariances. alpha is a generalization of an earlier estimate of reliability for tests with dichotomous items developed by Kuder and Richardson, known as KR20, and a shortcut approximation, KR21. (See Revelle, in prep; Revelle and Condon, in press.).

A useful index is the ratio of reliable variance to unreliable variance and is known as the Signal/Noise ratio. This is just

s/n = n r/(1-nr)

(Cronbach and Gleser, 1964; Revelle and Condon (in press)).

Standard errors for unstandardized alpha are reported using the formula from Duhachek and Iacobucci (2005).

More complete reliability analyses of a single scale can be done using the omega function which finds omega_hierchical and omega_total based upon a hierarchical factor analysis. Alternative estimates of the Greatest Lower Bound for the reliability are found in the guttman function.

Alpha is a poor estimate of the general factor saturation of a test (see Revelle and Zinbarg, 2009; Zinbarg et al., 2005) for it can seriously overestimate the size of a general factor, and a better but not perfect estimate of total test reliability because it underestimates total reliability. None the less, it is a common statistic to report. In general, the use of alpha should be discouraged and the use of more appropriate estimates (omega_hierchical and omega_total) should be encouraged.

Correlations between scales are attenuated by a lack of reliability. Correcting correlations for reliability (by dividing by the square roots of the reliabilities of each scale) sometimes help show structure. This is done in the scale intercorrelation matrix with raw correlations below the diagonal and unattenuated correlation above the diagonal.

There are several alternative ways to treat missing values. By default, missing values are replaced with the corresponding median value for that item. Means can be used instead (impute="mean"), or subjects with missing data can just be dropped (missing = FALSE). For data with a great deal of missingness, yet another option is to just find the average of the available responses (impute="none"). This is useful for findings means for scales for the SAPA project (see https://www.sapa-project.org/) where most scales are estimated from random sub samples of the items from the scale. In this case, the alpha reliabilities are seriously overinflated because they are based upon the total number of items in each scale. The "alpha observed" values are based upon the average number of items answered in each scale using the standard form for alpha a function of inter-item correlation and number of items.

Using the impute="none" option as well as asking for totals (totals="TRUE") will be done, although a warning will be issued because scores will now reflect the number of items responded to much more than the actual pattern of responses.

The number of missing responses for each person for each scale is reported in the missing object. One possibility is to drop scores just for those scales with missing responses. This may be done adding the code:

scores$scores[scores$missing >0] <- NA

This is shown in the last example.

Note that the default for scoreItems is to impute missing items with their median, but the default for scoreFAst is to not impute but must return the scale scores based upon the mean or total value for the items scored.

By default, scoreItems will drop those items with no variance. This changes the reliability calculations (number of items is reduced), and it mean that those items are not used in finding scores. Using the delete=FALSE option, these variables will not be dropped and their scores will be found. Various warnings are issued about the SMC not being correct. What do you you expect when there is no variance for an item?

scoreItems can be applied to correlation matrices to find just the reliability statistics. This will be done automatically if the items matrix is symmetric.

scoreFast just finds the scores (with or without imputation) and does not report other statistics. It is much faster! scoreVeryFast is even more stripped down, no imputation, just scores based upon the observed data. No statistics.

Value

Note

It is important to recognize in the case of massively missing data (e.g., data from a Synthetic Aperture Personality Assessment (https://www.sapa-project.org/) or the International Cognitive Ability Resources (https://icar-project.org)) study where perhaps only 10-50% of the items per scale are given to any one subject)) that the number of items per scale, and hence standardized alpha, is not the nominal value and hence alpha of the observed scales will be overestimated. For this case (impute="none"), an additional alpha (alpha.ob) is reported.

More importantly in this case of massively missing data, there is a difference between the correlations of the composite scales based upon the correlations of the items and the correlations of the scored scales based upon the observed data. That is, the cor object will have correlations as if all items had been given, while the correlation of the scores object will reflect the actual correlation of the scores. For SAPA data, it is recommended to use the cor object. Confidence of these correlations may be found using the cor.ci function.

Further note that the inter-scale correlations are based upon the correlations of scales formed from the covariance matrix of the items. This will differ from the correlation of scales based upon the correlation of the items. Thus, although scoreItems will produce reliabilities and intercorrelations from either the raw data or from a correlation matrix, these values will differ slightly. In addition, with a great deal of missing data, the scale intercorrelations will differ from the correlations of the scores produced, for the latter will be attenuated.

An alternative to classical test theory scoring is to use scoreIrt to find score estimates based upon Item Response Theory. This is particularly useful in the case of SAPA data which tend to be massively missing. It is also useful to find scores based upon polytomous items following a factor analysis of the polychoric correlation matrix (see irt.fa). However, remember that this only makes sense if the items are unidimensional. That is to say, if forming item composites from (e.g., bestScales), that are empirically derived, they will necessarily have a clear factor structure and the IRT based scoring does not make sense.

When reverse scoring items from a set where items differ in their possible minima or maxima, it is important to specify the min and max values. Items are reversed by subtracting them from max + min. Thus, if items range from 1 to 6, items are reversed by subtracting them from 7. But, if the data set includes other variables, (say an id field) that far exceeds the item min or max, then the max id will incorrectly be used to reverse key. min and max can either be single values, or vectors for all items. Compare two examples of scoring the bfi data set.

If scales are formed with overlapping items, then the correlations of the scales will be seriously inflated. scoreOverlap will adjust the correlations for this overlap.

Yet another possibility for scoring large data sets is to ignore all the reliability calculations and just find the scores. This may be done using scoreFast or scoreVeryFast. These two functions just find mean scores (scoreVeryFast without imputation) or will do imputation if desired scoreFast. For 200K cases on 1000 variables with 11 scales, scoreVeryFast took 4.7 seconds on a Mac PowerBook with a 2.8GHZ Intel I7 and scoreFast took 23.2 seconds. scoreIrt.1pl for the same problem took xxx with options("mc.cores"=1) (not parallel processing) and 1259 seconds with options("mc.cores"=NULL) (implying 2 cores) and with four cores was very slow (probably too much parallel processing).

Yet one more possibility is to find scores based upon a matrix of weights (e.g. zero order correlations, beta weights, or factor weights.) In this case, scores are simply the product of the weights times the (standardized) items. If using coefficients from a regression analysis (lm), a column of 1's is added to the data and labeled "(Intercept)" and this is used in the calculation. This is done by scoreWtd.

Author(s)

William Revelle

References

Cronbach, L.J. and Gleser G.C. (1964)The signal/noise ratio in the comparison of reliability coefficients. Educational and Psychological Measurement, 24 (3) 467-480.

Duhachek, A. and Iacobucci, D. (2004). Alpha's standard error (ase): An accurate and precise confidence interval estimate. Journal of Applied Psychology, 89(5):792-808.

McDonald, R. P. (1999). Test theory: A unified treatment. L. Erlbaum Associates, Mahwah, N.J.

Revelle, W. (in preparation) An introduction to psychometric theory with applications in R. https://personality-project.org/r/book/

Revelle, W. and Condon, D.C. Reliability. In Irwing, P., Booth, T. and Hughes, D. (Eds). the Wiley-Blackwell Handbook of Psychometric Testing (in press).

Revelle W. and R.E. Zinbarg. (2009) Coefficients alpha, beta, omega and the glb: comments on Sijtsma. Psychometrika, 74(1):145-154.

Zinbarg, R. E., Revelle, W., Yovel, I. and Li, W. (2005) Cronbach's alpha, Revelle's beta, and McDonald's omega h, Their relations with each other and two alternative conceptualizations of reliability, Psychometrika, 70, 123-133.

See Also

make.keys for a convenient way to create the keys file, score.multiple.choice for multiple choice items,
alpha, correct.cor, cluster.cor , cluster.loadings, omega, guttman for item/scale analysis.

If scales are formed from overlapping sets of items, their correlations will be inflated. This is corrected for when using the scoreOverlap function which, although it will not produce scores, will report scale intercorrelations corrected for item overlap.

In addition, the irt.fa function provides an alternative way of examining the structure of a test and emphasizes item response theory approaches to the information returned by each item and the total test. Associated with these IRT parameters is the scoreIrt function for finding IRT based scores as well as irt.responses to show response curves for the alternatives in a multiple choice test.

scoreIrt will find both IRT based estimates as well as average item response scores. These latter correlate perfectly with those found by scoreItems. If using a keys matrix, the score.irt results are based upon the item difficulties with the assumption that all items are equally discriminating (effectively a Rasch model). These scores are probably most useful in the case of massively missing data because they can take into account the item difficulties.

scoreIrt.1pl finds the item difficulty parameters and then applies a 1 parameter (Rasch like) model. It chooses items based upon a keys.list.

Examples