You are on page 1of 10

Intelligence 32 (2004) 499 508

A test of Spearmans bLaw of Diminishing ReturnsQ in two large samples of Danish military draftees
Peter Hartmanna,*, Thomas W. Teasdaleb
a

PNE centre, Institute of Psychology, University of Aarhus, Denmark b Department of Psychology, University of Copenhagen, Denmark

Received 15 February 2004; received in revised form 18 June 2004; accepted 25 June 2004

Abstract Spearmans bLaw of Diminishing ReturnsQ (SLODR) predicts that g saturation for cognitive tests will be lower at high ability levels than at low ability levels. This hypothesis was tested in two large samples of Danish military draftees (n=33,833 and n=25,020). The subjects were representative samples of the young adult male population and 95% were aged 1819. Both samples were tested with a group intelligence test comprised of four subtests which was used to form high and low ability groups within each sample. The g saturation of the battery was estimated within each group using the average intercorrelation and the total explained variance of the first principal component (PC1) and first factor from a principal axis factoring (PAF1). Two different cutoff points were used to illustrate the effect of difference in standard deviation (S.D.) for subtest and total ability score. For both samples and for both cutoff points, SLODR could not be confirmed. In fact, support was found for the opposite effect, namely, a higher g saturation for the high ability group as compared to the low ability group. Furthermore, SLODR may be dependent upon the cutoffs used to define ability groups and also the number and overall g loading of the cognitive tests used. D 2004 Elsevier Inc. All rights reserved.
Keywords: Law of Diminishing Returns, g saturation, Cognitive tests

1. Introduction Early this century, Spearman (1927) proposed a hypothesis that later would be known as Spearmans bLaw of Diminishing ReturnsQ (SLODR). The hypothesis is twofold and states first that the g
* Corresponding author. Tel.: +45 86 16 66 46. E-mail address: Peterhartmann@stofanet.dk (P. Hartmann). 0160-2896/$ - see front matter D 2004 Elsevier Inc. All rights reserved. doi:10.1016/j.intell.2004.06.009

500

P. Hartmann, T.W. Teasdale / Intelligence 32 (2004) 499508

saturation of cognitive ability tests decreases as a function of ability, meaning that for a cognitive ability test, the proportion of the variance that can be attributed to g is higher for low ability (e.g., low g score) subjects, as compared to high ability (e.g., high g score) subjects [later SLODR(ability)]. The SLODR also states that the g saturation of cognitive ability tests decrease as a function of age, meaning that for a cognitive ability test, the proportion of the variance that can be attributed to g is higher for younger subjects, as compared to older subjects tested on the same cognitive ability tests [later SLODR(age)]. The purpose of the present study is solely to investigate SLODR(ability) in two large groups whose high degree of homogeneity for age makes it impossible for us to test SLODR(age). The SLODR(ability) has been much debated and researched throughout the last century, but without reaching a firm conclusion. A recent review by Hartmann and Nyborg (2004) has highlighted some of the important methodological considerations in the study of SLODR and also has reviewed much of the relevant published literature. This review found moderate support for SLODR(ability). However, close inspection revealed that, within an ability span of about 1 1/2 standard deviation (S.D.) values, only approximately 2/3 of the studies showed a tendency towards SLODR(ability), and only about half of these were statistically significant (Hartmann & Nyborg, 2004). These findings indicate that when looking at an average normal distributed population, the differences in g saturation between the low and high ability subjects would be rather small, thereby indicating that the differences in g saturation between ability groups are more of academic interest than of practical interest in regards to the utility of the IQ test. As a corollary of this, one would expect an even bigger difference when comparing extremely low and extremely high ability subjects, however, these finding would be of little practical interest when dealing with bordinaryQ subjects. It is therefore of more interest both academic and practical to look at ability groups differing in ability but at the same time representing a large proportion of the normal population. We predict SLODR to be confirmed.

2. Methods 2.1. Subjects There has been conscription in Denmark continuously since the Second World War and all Danish males, at age 18 or shortly thereafter, are required to appear before a draft board assessing suitability for military service. About 9095% appear in person, the remainder being exempted on the grounds of documented medical conditions. The data for this study comprise all young men appearing before the draft board in 1988 (n=33,833) and in 1998 (n=25,020). The decline in numbers between the two years is due to a falling birth rate in Denmark during the 1970s. The two samples have been described in detail elsewhere (Teasdale & Owen, 1986, 1989, 2000). 2.2. The cognitive ability test The draft board assessment includes the administration of a group test of general cognitive ability, Bbrge Priens Prbve (BPP) which has been unchanged since 1957 and has been strongly inspired by Ravens Matrices. The BPP consists of four subtests:

P. Hartmann, T.W. Teasdale / Intelligence 32 (2004) 499508

501

Letter matrices: 19 items, time limit 15 min. This is based on the principles of Ravens Progressive Matrices, but without a multiple-choice response. Verbal analogies: 24 items, time limit 5 min. This is similar to Millers Analogies Test but it involves deducing analogous relationships and is not a test of vocabulary. Number sequences: 17 items, time limit 15 min. Through the use of simple arithmetic operations the subjects must produce the fifth number following a given series of four numbers. Geometric figures: 18 items, 10 min. The task in this test is to indicate which and how many of five given simple figures are represented in a gestalt. Each item is one gestalt. The items within each subtest are ordering according to difficulty level and the time limit is announced in advance. The subjects are tested in groups of 2040. Each subtest is scored as the number of correct responses. For purposes of the draft board, a total score is calculated from the simple addition of the scores from the four subtests. The BPP test is described in further detail elsewhere (Teasdale & Owen, 1994). The correlation between the BPP total score and the full-scale score of Danish Translation of the WAIS is reported elsewhere to be r=.82 (Mortensen, Reinisch, & Teasdale, 1990). 2.3. Construction of groups It was decided to divide the samples into only two ability groups (e.g., high and low) based on their BBP total score. This method has recently been employed by Jensen (2003) and was selected because we believed it to be the most appropriate for our data. Alternatively, we could have chosen to construct more groups in order to get a more continuous estimate of SLODR. However, according to Jensen, this could result in attenuation of the correlations and even artificial negative correlation between subtests, since a high score on one subtest would necessarily result in a lower score on another subtest. In order to avoid the spurious correlations when dividing the sample, one can choose a single subtest as selection tests and then exclude this from further analysis. However, a single subtest usually is a poorer estimate than the total score. Furthermore, since we only had four subtests, then we could not afford to exclude one of these. Jensen points out that the construction of two groups based on the total score is preferable because (1) all subtests are included, and (2) the potential attenuation of the correlations is minimal. Based on Jensen, and the fact that we could not afford to lose a subtest, the methods employed by Jensen was also used in the present study. Because age differences among the subjects were minimal, no consideration was taken of this variable. For all analyses, we used the obtained raw scores from the four subtests or total raw scores. From that we also calculated the factor scores using the principal axis factoring (PAF), and obtained a correlation between PAF1 and BPP total of .99 for both samples. As pointed out by some authors (e.g., Abad, Colom, Juan-Espinosa, & Garcia, 2003; Deary, Egan, Gibson, & Austin, 1996; Jensen, 1998; Legree, Pifer, & Grafton, 1996; see Hartmann & Nyborg, 2004, for a review on this), one should avoid differences in S.D. because these could affect the intercorrelation between the variables, and thereby artificially inflate the estimated g saturation. These authors argue that to avoid inflating estimates of g saturation, groups should be constructed with equal S.D. values on the dependent selection test. The use of the S.D. for the selection test is properly used as a substitute for the subtest S.D. values, based on the assumptions that if the S.D. for selection is equal, then the subtests S.D. values are equal as well. This assumptions is in

502

P. Hartmann, T.W. Teasdale / Intelligence 32 (2004) 499508

generally sound but not necessarily perfectly valid. We argue (also see Hartmann & Nyborg, 2004) that the S.D. for the selection test is of less importance as compared to the actual subtests variables that enter into the factor analysis. So one should keep the S.D.s for the subtest variables equal instead of the S.D. on the selection test. This method is not novel and has as far as we know been used in regards to SLODR since 1935 (Garrett, Bryan, & Perl, 1935) and most recently by Jensen (2003). However, despite the knowledge of this, we have not found studies specifically addressing this difference in methodology and the possible effect of this difference (also see Hartmann & Nyborg, 2004). We have therefore decided to use both methods in order to investigate SLODR(ability), first by keeping the S.D. for the selection test constant and secondly by keeping the S.D. values for the subtest variables constant. In order to ensure equal S.D. for the selection test across ability groups, we chose different cutoff points on the total BPP score and compared the S.D. for the selection test in the high-scoring and lowscoring groups in order to obtain the best fit. For both the 1988 and 1998 samples, this resulted in a cutoff point at a BPP total score of 38. Subjects scoring 38 or above were thus considered high ability for both samples, whereas subjects scoring below 38 were considered low ability. As seen from Table 1, the S.D.s for the total BPP (the selection test) are highly similar across the two ability groups in both samples. Furthermore, there is a marked difference in the size of the two ability groups across both samples pointing to a skewed distribution of the BPP total score. However, we see no reason for this size difference to introduce confounders. We did not see any reason to test the differences in S.D. across the ability groups for significance using the F variance ratio test, because of the large sample sizes, namely, a higher g saturation for the high ability group as compared to the low ability group even small and negligible differences would become highly significant. As is also apparent, the lower ability groups have slightly higher S.D.s for the four subtests and the factor scores (described below). In order to ensure equal S.D.s for the subtest variables across ability groups, we selected different cutoff points on the total BPP score and compared the S.D.s for the subtest variables. This resulted in a cutoff point at a BPP total score of 35 as shown in Table 2. Subjects scoring 35 or above were considered
Table 1 Approximately equal S.D. for total BPP score Ability groups Letter matrices Verbal analogies Number series Geometric figures BPP total Factor scores (PAF) No. of subjects

1988 Mean score Low ability 7.43 (2.28) 8.28 (3.04) 6.37 (2.55) 7.47 (2.56) 29.54 (6.56) 1.10 (0.59) 10,396 BPP b38 High ability 11.12 (1.78) 14.04 (3.01) 11.05 (2.09) 11.22 (2.72) 47.43 (6.32) 0.48 (0.54) 23,437 BPP z38 1998 Mean score Low ability 7.55 (2.22) 8.28 (3.08) 6.34 (2.53) 7.77 (2.50) 29.94 (6.50) 1.19 (0.61) 6628 BPP b38 High ability 11.13 (1.76) 14.06 (2.96) 11.04 (2.09) 11.58 (2.77) 47.82 (6.31) 0.43 (0.55) 18,392 BPP z38 S.D. is reported in brackets.

P. Hartmann, T.W. Teasdale / Intelligence 32 (2004) 499508 Table 2 Approximately equal S.D. for the four subtests scores Ability groups Letter matrices Verbal analogies Number series Geometric figures BPP total

503

Factor Number of scores (PAF) subjects

1988 Mean score Low ability 6.89 (2.23) 7.55 (2.99) 5.71 (2.36) 7.01 (2.57) 27.16 (6.10) 1.31 (0.55) 7614 BPP b35 High ability 10.89 (1.90) 13.64 (3.15) 10.75 (2.27) 10.95 (2.77) 46.23 (6.93) 0.38 (0.59) 26,219 BPP z35 1998 Mean score Low ability 6.96 (2.18) 7.46 (3.05) 5.63 (2.38) 7.26 (2.50) 27.31 (6.08) 1.43 (0.57) 4643 BPP b35 High ability 10.92 (1.86) 13.69 (3.11) 10.75 (2.28) 11.33 (2.82) 46.68 (6.94) 0.33 (0.61) 20,377 BPP z35 S.D. is reported in the brackets.

high ability for both samples, whereas subjects scoring below 35 were considered low ability. As with the results obtained from Table 1, we did not test the differences in S.D. values for statistical significance. As seen in Table 2, the S.D.s for the four subtests are highly similar, however, when looking at the S.D. for the BPP total, the high ability groups has a higher S.D., as is also the case when looking at the factor scores (described below). Based on these two different cutoff points between high and low ability groups, we constructed 22 ability groups for each sample and later comparing them pairwise. As for both cutoff points, the difference in ability level for the subtests, BPP total, and factor scores were highly significant, thereby ensuring a difference in total and subtest ability across the low and high ability groups. After constructing the groups, the next step was the analysis of the data. 2.4. Method for investigation the g saturation In order to investigate differences in g saturation across groups differing in ability, we factor-analyzed the four subtests in the low and high ability groups separately using both a principal component analysis (PC) and PAF. This was done for two reasons. First, with the use of the eigenvalues of the first principal component, PC1, it is possible to calculate the average intercorrelation of correlation matrix using the Kaiser formula (Kaiser, 1968). The average intercorrelation of the variables (g)=((k1)/( p1)), where k is the eigenvalue of the PC1 and p is the number of variables. Secondly, with the use of the total explained variance of the first principle axis factor, PAF1, we get a better estimate of the g saturation of the test as compared to the PC1 (Jensen & Weng, 1994) due to the introduction of less specific variance into the estimate as seen by the PC1. After obtaining the average intercorrelation of the matrix and the total explained variance of the PC1 and the PAF1, we tested these for significance. In regards to the average intercorrelations, we treated these as simple correlations using Fishers z to compare the correlations and determine significance, as proposed by Escorial, Juan-Espinosa, Garcia, Rebollo, and Colom (2003) as an alternative to what Hartmann and Nyborg (2004) labeled bthe LynnCooper MethodQ (Lynn & Cooper, 1993, 1994). We

504

P. Hartmann, T.W. Teasdale / Intelligence 32 (2004) 499508

also compared the total explained variance of the PC1 and PAF1 by using the F variance ratio test for significance. We thus employed two ways of testing for significance for two cutoff points for two different samples.

3. Results Tables 3 and 4 show the results of the PAF and PCA of the four subgroups factor analyzed separately and the total sample for both the 1988 and 1998 samples. All of the congruence coefficients between the respective lowhigh ability subgroups were above .95, as were the congruence coefficients for the total sample (in both 1988 and 1998) and the respective four subgroups. This indicates that the factors measured are indeed highly similar. The reported factor loadings were found through the PAF, but the PC1 eigenvalues and total explained variances have also been include in the tables. As Tables 3 and 4 show, there is a tendency for the g saturation to increase as a function of ability, and not to decrease as predicted by SLODR. This tendency proved to be significant in every single case when comparing the total explained variance of the PC1, PAF1, and the average intercorrelation calculated from the eigenvalue of the PC1 with the use of the Kaiser formula. In the majority of cases, the significance level reached Pb.00001, the exceptions being two cases where it was Pb.0001 and one single case at Pb.001. The last three cases were all for the division at BPP 38 finding the least tendency for a change in the g saturation as a function of ability. As also seen from Tables 3 and 4, the difference in g saturation is larger when comparing PAF1 as opposed to PC1. Furthermore, the difference in g saturation is also larger for the second cutoff point at BPP 35 (similar S.D. for the subtests) as opposed to BBP 38 (similar S.D. for total score/selection test). Oddly enough, the results for the two samples (1988 and 1998) were highly similar for the total sample and the 22 ability subgroups. These results therefore contradict SLODR by finding significant support for the contrary, although the absolute difference in g saturation differs markedly depending on the specific cutoff point chosen.

Table 3 Factor analysis of the 1988 sample for low and high ability groups Sample anno 1988 Principal axis factoring n Letter matrices Verbal analogies Number series Geometric figures Total explained variance of the PAF1 Eigenvalue of the PC1 Total explained variance of the PC1 Average intercorrelation All 33,833 .77 .77 .78 .61 54.2% 2.61 65.3% .54 Low ability BPP b38 10,396 .59 .44 .47 .28 20.8% 1.59 39.8% .20 High ability BPP z38 23,437 .51 .56 .54 .35 24.4% 1.72 42.9% .23 Low ability BPP b35 7614 .55 .36 .39 .26 16.3% 1.45 36.4% .15 High ability BPP z35 26,219 .57 .60 .58 .40 29.7% 1.87 46.8% .29

All congruence coefficients between any of the ability subgroups and the total sample were above .95, as were the congruence coefficients between the respective low versus high groups.

P. Hartmann, T.W. Teasdale / Intelligence 32 (2004) 499508 Table 4 Factor analysis of the 1998 sample for Low and High ability groups Sample anno 1998 Principal axis factoring n Letter matrices Verbal analogies Number series Geometric figures Total explained variance of the PAF1 Eigenvalue of the PC1 Total explained variance of the PC1 Average intercorrelation All 25,020 .78 .76 .77 .60 53.1% 2.58 64.5% .53 Low ability BPP b38 6628 .65 .39 .48 .27 21.8% 1.60 39.9% .20 High ability BPP z38 18,392 .56 .54 .51 .36 24.8% 1.73 43.2% .24 Low ability BPP b35 4643 .61 .31 .42 .22 17.5% 1.46 36.5% .15

505

High ability BPP z35 20,377 .61 .59 .58 .41 30.3% 1.89 47.3% .30

All congruence coefficients between any of the ability subgroups and the total sample were .95 or greater, as were the congruence coefficients between the respective low versus high groups.

As a further analysis, we also compared the total 1988 and 1998 samples. The two samples were highly similar in total score (1988, M=41.93, S.D.=10.44; 1998, M=43.08, S.D.=10.14; pb0.000) and also for the subtest scores and their S.D., although the differences were significant due to the large sample sizes. The congruence coefficients for the two factors were 0.99 and there was no significant difference in the total explained variance of the PAF1 and the PC1 or average inter-correlation, although the differences in PAF1 came close to significance (slightly above p=0.05). This indicates no significant difference in ability level across the two samples and also no changes in g saturation due to a minimal ability difference (although in the direction predicted by SLODR). The analysis reaching a near significant result is interpreted as an artefact of the large sample sizes.

4. Discussion The present study finds no support for SLODR in any of the samples or in any of the subdivision of the two samples. Rather, the results indicate a statistically significant increase in g saturation as a function of ability. However, although reaching significance, the specific cutoff point between low and high ability groups has highly affected the absolute difference in g saturation. The subgroups with the cutoff point at BPP 38 had similar S.D. values for the total score (and thereby the selection test) across the ability groups but higher S.D. values for the subtest variables for the low ability group as compared to the high ability groups. The results here showed small difference in total explained variance for the PAF1 and PC1 (ca. 34%) and in average intercorrelations (ca. .04). In subgroups with the cutoff point of BPP 35, S.D.s were similar for the four subtests across ability groups, but were higher for total score for the high ability group. These results on the other hand showed large differences in total explained variance of the PC1 and the PAF1 (ca. 1015%) and the average intercorrelation (ca. .1.15). Thus, our results show marked differences in the estimated difference in g saturation across ability groups as a function of cutoff point. This cannot be explained in terms of the difference in ability between the ability groups, since the b35Q groups have a smaller difference in ability as compared to the b38Q groups. Intuitively, one should

506

P. Hartmann, T.W. Teasdale / Intelligence 32 (2004) 499508

expect to find a larger effect of SLODR when the ability difference was large, however, this was not the case here. Another possible explanation is the difference in S.D. across total score and subtest scores as a function of cutoff point. When consulting Tables 14, it becomes apparent that if differences in total score S.D. confound the estimated g saturation, the cutoff point of BPP 35 inflates the differences in g saturation because the high ability groups have higher S.D. for total score. On the other hand, if differences in the subtest S.D. values confound the estimated g saturation, then the BPP cutoff point of 38 deflates the difference in g saturation because the lower ability groups have higher S.D. values for the four subtests scores. As argued previously, we believe that the former cutoff point of BPP 35 is the more accurate because of the control for S.D. for the bactiveQ variables entering into the correlation matrix for the factor analysis. We therefore deem the results obtained by the former cutoff point at BPP 35 to be more valid. If this assumption is accepted, then the present study strongly supports moderate differences in g saturation as a function of ability in the opposite direction as predicted by Spearman. If this assumption is not endorsed however, the present study finds support for a small increase in g saturation as a function of ability. In either case, the bLaw of Diminishing ReturnsQ as proposed by Spearman is definitely not supported in the present study. The reason for the discrepancy between the present findings and the results from the recent review (Hartmann & Nyborg, 2004) is unknown. What the present findings show is that with increasing level of general cognitive ability, the g saturation increased, thereby indicating that a single subtest is a better predictor of g for high ability subjects than for low ability subjects. If one would accept the present findings as valid, what could be the explanation of these findings? Differences in education can be ruled out since the subjects have received approximately the same amount of education, before considering even whether education could affect the g saturation (see Hartmann & Nyborg, 2004 for a comment on this). The two samples only consisted of young males and one could therefore argue that SLODR would still hold for female subjects and/or older subjects. However, there is no theoretical explanation for this in regards to the hypothesis, and neither is there empirical evidence pointing to the importance gender or age as a moderator of SLODR(ability) (Hartmann & Nyborg, 2004). A more likely explanation is the number and type of subtests in the battery. With only four subtests the estimation of the g factor, the total explained variance and the average intercorrelations are less accurate than could be hoped for. This point was emphasized and elaborated on by Jensen (personal communication). He has suggested that the presence of other significant non-g factors is necessary for the SLODR to be present. Since this present battery only supported a one-factor solution and consisted of four highly g loaded and non-g factor free tests, the high ability subjects might not have had the opportunity to show their more differentiated ability pattern due to the lack of other cognitive factors besides g in the tests. Following this line of reasoning, one can imagine three ways in which SLODR can occur. (1) Increased differentiation among Stratum I and II factors. The differentiation of ability is uniform in the entire hierarchy of the cognitive structure. This would result in lower correlations between Stratum I factors, Stratum II factors, and within each cluster of Stratum I factors that have their salient loading on the same Stratum II factor. (2) Increased differentiation primarily among Stratum II (and thereby for Stratum I as well), but less so within each cluster of Stratum I factor with their salient loading on the same Stratum II factors. This would result in decreasing correlations among Stratum II factor and therefore also among Stratum I factor, but not necessarily within each cluster of Stratum I factor loading on the same Stratum II factor. In this case, differentiation is at the Stratum II level. (3) The reverse of 2,

P. Hartmann, T.W. Teasdale / Intelligence 32 (2004) 499508

507

where the primarily differentiation occurs within each cluster making the Stratum II factor less prominent for the cluster of Stratum I factors but affecting the correlations among Stratum II factors less. Depending on how SLODR works, one would need a battery where the subtests load on non-g factors as well as g. In the first case, the subtests would need to have loadings on Stratum I and II factors in order to support SLODR. In the second case, only subtests loadings on Stratum II factor are important, and therefore the selection of subtests within gross different cognitive categories is important. In the third case, Stratum I factor loadings become important to ensure SLODR, and similar subtests within the same cognitive domain could be satisfactory. The findings of this study highlight the importance of several methodological considerations when studying SLODR. Firstly, the marked difference between the estimated g saturation across low and high ability groups for different cutoff points accentuates the importance of determining a common standard for the control of S.D. variation, either at the variable level or at the total score/factor score level, because different methods yield different results. Secondly, the number and the nature of the ability tests, and obviously many diverse tests included in the analysis, are preferable, because missing diversity among tests could result in lack of opportunity for the high ability subjects to be able to express themselves and thereby show their differentiated ability structure. But since the inclusion of many diverse tests is not always possible, one should perhaps consider at what level (I and II, I or II) this differentiation is primarily taking place.

5. Conclusion The present study, based on two samples of Danish military draftees from 1988 and 1998 totaling over 50,000 young male subjects, found no support for SLODR. The study did find support for an increase in the g saturation as a function of ability (i.e., the opposite of that predicted by Spearmans hypothesis). This increase was highly significant, but the absolute difference (and thereby also the degree of significance) varied as a function of the cutoff point for the low and high ability groups. The largest differences were found when the S.D. values for the four subtest variables were held constant as opposed to the total score, as compared to holding the S.D. for the total score and allowing variance in the S.D. for the four subtests variables. We believe that the reason for this is that the latter method is less accurate and for the present study, this resulted in a deflation of the difference in g saturation across ability groups. The present study emphasizes the need for a common method for controlling for S.D. when investigating SLODR, and more studies aimed at investigating SLODR in order to determine the accuracy and relevancy of this hypothesis.

References
Abad, F. J., Colom, R., Juan-Espinosa, M., & Garcia, L. F. (2003). Intelligence differentiation in adult samples. Intelligence, 31, 157 166. Deary, I. J., Egan, V., Gibson, G. J., & Austin, E. J. (1996). Intelligence and the differentiation hypothesis. Intelligence, 23, 105 132. Escorial, S., Juan-Espinosa, M., Garcia, L. F., Rebollo, I., & Colom, R. (2003). Does g variance change in adulthood? Testing the age de-differentiation hypothesis across sex. Personality and Individual Differences, 34, 1525 1532.

508

P. Hartmann, T.W. Teasdale / Intelligence 32 (2004) 499508

Garrett, H. E., Bryan, A. I., & Perl, R. E. (1935). The age factor in mental organization. Archives of Psychology, 176, 1 31. Hartmann, P., & Nyborg, H. (2004). Spearmans bLaw of Diminishing ReturnsQ: A critical eye on a century of methods, results, and current standing on the theory. Unpublished work. Jensen, A. R. (1998). The g factor. Westport, CT7 Prager. Jensen, A. R. (2003). Regularities in Spearmans law of diminishing returns. Intelligence, 31, 95 105. Jensen, A. R., & Weng, L. J. (1994). What is a good g? Intelligence, 18, 231 258. Kaiser, H. F. (1968). A measure of the average intercorrelation. Educational and Psychological Measurement, 28, 245 247. Legree, P. J., Pifer, M. E., & Grafton, F. C. (1996). Correlations among cognitive abilities are lower for higher ability groups. Intelligence, 23, 45 57. Lynn, R., & Cooper, C. (1993). A secular decline in Spearmans g in France. Learning and Individual Differences, 5, 43 48. Lynn, R., & Cooper, C. (1994). A secular decline in Spearmans g in Japan. Current Psychology, 13, 3 7. Mortensen, E. L., Reinisch, J. M., & Teasdale, T. W. (1990). Intelligence as measured by the WAIS and a Military Draft Board Group Test. Scandinavian Journal of Psychology, 31, 315 318. Spearman, C. (1927). The abilities of man: Their nature and measurement. London7 Macmillan. Teasdale, T. W., & Owen, D. R. (1986). The influence of paternal social class on intelligence and educational level in male adoptees and non-adoptees. British Journal of Educational Psychology, 56, 12. Teasdale, T. W., & Owen, D. R. (1989). Continuing secular increases in intelligence and stable prevalence of high intelligence levels. Intelligence, 13, 262. Teasdale, T. W., & Owen, D. R. (1994). Thirty-year secular trends in the cognitive abilities of Danish male school-leavers at a high educational level. Scandinavian Journal of Psychology, 35, 328 335. Teasdale, T. W., & Owen, D. R. (2000). Forty-year secular trends in cognitive abilities. Intelligence, 20, 115 120.

You might also like