- Email: [email protected]
Body perception index: benefits, pitfalls, ideas
Journal of Psychosomatic Research, Vol. 44, Nos. 3/4, pp. 457–464, 1998 Copyright 1998 Elsevier Science Inc. All rights reserved. 0022-3999/98 $19.00 1 .00
S0022-3999(97)00142-6
BODY PERCEPTION INDEX: BENEFITS, PITFALLS, IDEAS MONIQUE A. M. SMEETS,* FILIP SMIT,† GEERT E. M. PANHUYSEN* and J. DAVID INGLEBY* (Received 9 July 1996; accepted 12 March 1997) Abstract—Estimates of body size are often expressed as a ratio of actual size [body perception index or BPI5(estimated size/actual size)3100%]. In this article, we examine the possibility that overestimation of body size in patients with anorexia nervosa, as measured by the BPI, is due to their smaller body size rather than to their being anorexic. Using 50 mean body sizes derived from seven studies we investigated whether the error of estimation is a constant proportion of the body size to be estimated, as the use of the BPI assumes. A negative linear relation between BPI and actual body size was found, confirming that smaller size is associated with greater overestimation. However, although both groups showed a strong tendency to overestimate smaller sizes, anorexic subjects showed even greater overestimation than controls. Hence, overestimation of body size in AN can only partially be accounted for by the smaller body size of anorexic patients. Recommendations for future use of the BPI are put forward. 1998 Elsevier Science Inc. Keywords: estimation.
Anorexia nervosa; Body image; Body perception index; Body size estimation; Visual size
INTRODUCTION
On average, patients with anorexia nervosa (AN) estimate their body size to be greater than it actually is. This was recently established in a statistical meta-analysis of 33 studies comparing body size estimates from patients with anorexia nervosa (AN) to controls without an eating disorder [1]. This finding implies that these patients have a visual mental image of their own body as fatter than it really is, which is part of the more general disturbance of body image—a diagnostic marker of AN [2]. An alternative explanation for the overestimation of body size by patients with AN, however, might be that it is an artifact of their being thin rather than a consequence of their being anorexic. According to this explanation, the habit of displaying the body size estimate as a fraction of actual size puts anorexic subjects— who, by definition, have smaller actual sizes than normal-weight controls—at a disadvantage. The ratio of estimated size to actual size (3100%) is known as the body perception index (BPI; see Slade and Russell [3]). The BPI is typically used in so-called body part studies, which are directed at estimation of actual body widths.
* Department of Psychonomics, Utrecht University, Utrecht, The Netherlands. † The Netherlands Institute of Mental Health and Addiction, Utrecht, The Netherlands. Address correspondence to: Monique A. M. Smeets, Pennington Biomedical Research Center, 6400 Perkins Road, Baton Rouge, LA 70808. Tel: (504) 763-3111; Fax: (504) 763-3022; E-mail: smeetsma@ mhs.pbrc.edu
457
458
M. A. M. SMEETS et al.
Using the BPI, Ben-Tovim et al. [4] made the curious discovery that mean BPI was lower in anorectics than in controls. This, in combination with the fact that the controls in their study had smaller body widths than the AN patients, prompted them to investigate the relationship between BPI and actual width. They found a significant tendency for BPI to increase with actual width: the smaller the body part, the greater its percentage overestimation. When this tendency was taken into account, no statistical difference emerged between the anorexic and control group. Thompson [5] subsequently showed that a significant difference between anorexic subjects and controls, apparent when the BPI was used in the traditional manner, became insignificant when the actual measures of the controls were used as the denominator for the anorexic group in the absence of data on their own premorbid body sizes. On the basis of the results obtained in a follow-up study to further investigate this phenomenon, Ben-Tovim and Crisp [6] claimed that estimates are random in relation to actual body sizes. Consequently, the smaller body sizes that are likely to be found in individuals with AN will increase the apparent size of perceptual error, as a function of dividing a random number by a relatively smaller number. The above findings urge a more cautious approach to applying the BPI in populations differing in weight and size. On the basis of the conclusion from the meta-analysis that anorexic patients, on average, overestimate their body sizes as compared to normal controls, we now investigate whether anorectics overestimate because they are anorexic, or merely because of their smaller actual sizes. On the basis of the results, we wish to formulate specific recommendations for future applications of body size estimation methods. It is appropriate first to highlight several assumptions that seem to have been made in applying the BPI, starting from the broader perspective of classical psychophysics. We will restate Ben-Tovim and Crisp’s [6] claim that people with AN show stronger body size overestimation than the average-sized controls they tend to be compared to because of smaller actual body size, starting from Weber’s law. Weber’s law has generally been used in psychophysics to express the relation between the perceived and actual value of a stimulus, and seems to apply to this case. If person A’s BPI is higher than person B’s, this implies either that A’s error of estimation in terms of centimeters is greater than B’s, or that A’s actual body size is smaller.1 Comparing BPIs with each other when the denominators (i.e., actual body sizes) are unequal would seem to be justified by the additional assumption that smaller actual sizes are accompanied by smaller errors of estimation—or vice versa, that greater actual sizes may lead to proportionately greater errors. The assumption that error/ actual size5k where k is a constant, is equivalent to Weber’s law that dS5kS or dS/ S5k: that is, the standard intensity of a stimulus (S) must be increased by a constant fraction (dS) to be distinguished (see Baird and Noma [8]). Correspondingly, in body size estimation studies an increase in body size is assumed to be accompanied by a proportional increase in estimation error. Hence, in the present study we first investigate whether, in accordance with We1 If person A’s body actual size is smaller than person B’s, his or her BPI may be higher not only when A’s estimation error is larger or equal to B’s, but even if it is slightly smaller. (For a computational example the reader is referred to Ben-Tovim and Crisp [6], p. 107).
Body perception index
459
ber’s law, the BPI is constant over a range of body sizes, or whether there is an (artifactual) correlation with actual size. If, as Ben-Tovim and Crisp [6] put it, body size estimates are randomly related to actual sizes, the function relating BPI to actual size will be a negative linear one rather than a constant. Second, we investigate whether the difference in body size estimates between anorexic patients and controls remains after controlling for the influence of actual size on BPI. This study focuses on size estimation data cumulated over studies, using the method of statistical meta-analysis.
METHOD Sample of studies The sample of studies used for this investigation was derived from the sample used in Smeet et al. [1]2. For this meta-analysis, studies carried out between 1973 and 1993 were collected which compared anorexic patients to normal controls on visual body size estimates. Of the original sample of 33 studies, 13 made use of body part methods. Six studies had to be excluded from this subsample because actual body width measures, which are essential to our analysis, were not mentioned. For an overview of the remaining studies the reader is referred to Table I. It should be noted that the usual statistic employed in meta-analysis is the effect size, which quantifies the difference in score between groups. In the present study we are not interested in a measure that combines the scores obtained by the groups. Rather, we investigate the relation between BPI and actual size for each of the groups separately. This implies that the effect size (d) was not suitable here. In this study, raw mean scores (BPI’s) were used instead. This strategy is usually discouraged in favor of using standardized measures such as d, because scales tend to vary when using different measurement techniques. However, because, in the present analyses, the sample was comprised of studies all using the same technique, namely visual size estimation, we feel that analyzing raw-score means, under the assumption of equality of sds, is acceptable (see also Hunter and Schmidt [9]). In total, 50 BPIs with corresponding actual sizes could be inferred from seven studies, which amounts to 25 per group. Each BPI corresponded to the mean for either of the groups.
Analyses All analyses involve ordinary least squares linear regression and were carried out using the LIMDEP program [10]. Extracting 50 BPIs from seven studies yields observations which are not independent of each other. To take account of this nonindependence, dummy variables were entered into the analyses to distinguish each study that provided the BPI. To test the influence of actual size, group membership (anorexic vs. control), and study membership on BPI, the following model was fitted: BPIi5a1b1 Gi1b2 Acti1Skj51b3Sij1ui
(1)
where BPIi5body perception index for i51, 2 . . . N group means; a5intercept (constant); b5regression weight (slope); Gi5group (05control, 15anorexia nervosa); Acti5actual size; Skj51Sij5set of study dummies (to account for dependency of BPIs within studies for j51, 2 . . . K studies); ui5 unique term. The first research question (concerning the relationship between BPI and actual size) involves testing the statistical significance of the regression weight b associated with Acti. Testing the second research
2 The primary purpose of the study was to evaluate whether the main finding of the meta-analysis reported in Smeets et al. [1]—that individuals with AN overestimate their body sizes as compared to controls—could be explained by the fact that the AN subjects had smaller actual body widths. Therefore, a subsample that provided the necessary data to perform the present analyses was taken from the original 33 studies, published between 1973 and 1993. However, because some time had passed since 1993, it was decided to perform another literature search to see if any appropriate studies had been published since then. Only two studies were found in which an AN sample was compared to a normal control sample, using the BPI to express the resulting estimates. However, because these studies did not provide the actual mean sizes for the body widths used in the estimation, these studies could not be included in the regression. We believe that the recent dearth of studies should be attributed to the nonpopularity of body size estimation methods, partly due to the artifact discussed here.
460
M. A. M. SMEETS et al.
Table I.—An overview of actual measures of body sizes and related body perception index (BPI) scores Anorexia nervosa Controls
Body part
Actual
BPI
Actual
BPI
Slade and Russell (1973)
face chest waist hips
11.63 23.32 19.84 27.18
157.6 134.2 146.6 126.8
10.9 25.58 21.77 29.46
94.7 95 100.2 96.6
Button et al. (1977)
face chest waist hips
11.94 25.75 21.24 29.78
115.2 115.5 119 102.5
12.57 28.24 24.66 34.46
131.1 121.6 113.5 107.1
Ben-Tovim et al. (1979a)
face waist
12.6 22.3
139.9 145
12.2 22.0
167.1 167.4
Casper et al. (1979)
face chest waist hips
11.32 22.69 18.64 27.71
130.8 114.1 124.6 103.2
11.39 24.91 22.23 29.22
122.4 108.9 113.7 107.7
Norris (1984)
head waist hips thigh
14.2 19.6 28.5 11.2
126.5 118 98 129.5
14.3 23.5 32.3 15.5
115.5 100 91.5 98
Proctor and Morley (1986)
face bust waist hips
11.2 24.3 20.9 28.3
153.6 118 140.1 117.4
11.4 26.1 23 32
132.0 103.9 113.9 97.5
Hartley (1989)
chest waist hips
27.13 24.15 30.51
153.24 156 147.98
32.50 29.20 36.40
123.92 122.72 113.88
Study
Note: actual measures are displayed in centimeters. a Actual measures and BPIs estimated from scatter diagrams. question (concerning the difference between groups) involves doing the same for the regression weight associated with Gi (see footnote 3).
RESULTS
The regression of BPI on Act and G yielded the following equation [F(8, 41)510.83, p,0.001], which accounted for 68% of variance: BPIi5138.50112.05G21.21Act For the regression coefficients associated with the control dummy variables corresponding to studies the reader is referred to Table II. The negative sign of the coefficient associated with Act(21.21; t524.38, p,0.001) demonstrates that the smaller the body width, the greater the relative overestima3 To account for the possibility that there is an interaction of the variables Act and G, a model was first tested that included the interaction vector G*Act (see Pedhazur [11]). Testing of the regression weight b associated with G*Act revealed no significance ( p50.80). Hence, because group membership and actual size do not interact, the interaction term can be omitted from the regression equation, yielding Model 1.
461
Body perception index
Table II.—Regression equations for the regression of BPI on actual size and group Variables Predictors Group Actual size Control variables S1 S2 S3 S4 S5 S6 Intercept R R2
B
seB
p
12.05 21.21
3.55 0.28
0.001 0.000
138.08 169.64 134.98 127.57 142.72 166.40 138.50
8.29 8.18 7.71 7.47 7.96 10.11 7.76
0.82 0.68
Note: B is the regression weight; se is the standard error. Group: 0 5 control; 1 5 anorexia nervosa. Control variables S1 to S6 are the dummy variables associated with the studies.
tion. Hence, the hypothesis that the slope of the regression lines is zero, which is implicit in the general use of the BPI, was rejected. Moreover, from the positive sign of the coefficient associated with G (12.05; t53.40; p,0.010) it was concluded that there is a real difference between the anorexic and control group in BPI, with anorectics (G51) showing stronger overestimation than controls (G50). Having determined that regression slopes do not differ (see footnote 2)—hence must be parallel—but that intercepts were different, separate regression functions were obtained with identical coefficients for Act but different intercepts. The Group effect (anorexic vs.control), captured by the group dummies, indicates that anorectics overestimate their body size by 12.05 BPI units as compared to normal subjects. The regression functions for the AN and control group, respectively, are shown in Fig. 1. DISCUSSION
Body size estimation methods, which became increasingly popular in the 1970s and 1980s for assessing the size at which people image their bodies, and especially for the disturbances in this ability in eating-disordered patients, have suffered a decline in research interest in recent years (see Hsu and Sobkiewicz [12]). The main reasons for this decline seem to lie in the apparent inconsistency4 of research findings and methodological problems. In the present study we have revealed one such problem, which, in our view, can be adequately dealt with in future research. Researchers have implicitly assumed that people make smaller absolute errors when estimating smaller body sizes and that, as a consequence, the proportion of the size of the error to the actual size is constant. If, contrary to this assumption, the size of the error does not decrease in proportion to actual body size, BPIs will 4 The results obtained with body size estimation methods are not as inconsistent as they seem (see Smeets et al. [1]).
462
M. A. M. SMEETS et al.
Fig. 1. Regressions of body perception index (BPI) on actual size after testing of regression functions for equality: BPIcon5138.521.21Act and BPIAN5150.5521.21Act, where Act is actual size. Solid line: An; dotted line: control.
increase as actual size decreases. Should this be the case, then overestimation of body size by patients with anorexia nervosa could be a result of being thin rather than of being anorexic. Contrary to the assumption of constancy, the BPI was found to be higher for smaller body sizes, implying that smaller body sizes were overestimated to a greater extent. Thus, and in agreement with Ben-Tovim and Crisp [6], it was concluded that the BPI transformation introduces a bias into the results when it is used to represent the outcome in body part studies. Although both anorectics and normal controls share the tendency to overestimate thinner bodies, the BPI scores for anorexic patients were even larger than those obtained by normal controls. Hence, overestimation of body size by anorexic patients cannot be entirely explained by the tendency to overestimate thinner actual sizes: on average, there is still a relative overestimation of 12.05% for anorexic patients. In view of the tendency for smaller body sizes to be overestimated to a greater extent than larger sizes, we recommend that variability in actual body sizes should be taken into account in future research. If the emphasis is not on eating-disordered populations, subjects can be matched on actual size prior to testing, as recommended by Thompson [13]. Alternatively, groups can be equated statistically by covariance procedures, which allow initial differences in body size across groups to be removed as a confounding variable [14]. Another possibility is to perform the regression analyses that were conducted here. In this way, the relation between BPI and estimated size can be easily examined. Between-group testing then comprises testing of the regression coefficients and intercepts belonging to the regression equations. It should be noted that the problem discussed in this study specifically for the BPI,
Body perception index
463
might also apply to studies using so-called whole body techniques, in which subjects estimate the size of their body as a whole by adjusting the size of a video image of their own body on a television screen. Also, in this kind of set-up, errors will be percentually larger when expressed against smaller body widths. Unfortunately, it is very hard to evaluate this artifact because such studies never provide actual body sizes or widths of the stimulus picture. However, if the present findings can be extended to the whole body procedures, we might anticipate that a difference between the AN and control groups will remain after controlling for this tendency. It is suggested to use thin controls without an eating disorder in future studies, to minimize differences in body size widths shown in the video images. The above recommendations deal on a post hoc basis with the unexpected finding of a negative relation between BPI and actual size, but do not offer any insight into this curious relation. The finding that smaller body parts are overestimated to a greater extent certainly runs counter to our intuition that smaller sizes can be estimated with smaller error. We will put forward an explanation to account for this finding, according to which the tendency to overestimate smaller body sizes to a greater extent than larger sizes has to do with the size of the image spontaneously constructed during size estimation. This explanation originated from analyzing the cognitive information processing that goes on during a body size estimation task (see Smeets [15]). To be able to estimate the width of the body at designated parts, people have to consult a visual mental image. One of the properties of imagery is that the image can be constructed at any size, while preserving the underlying proportions [16, 17]. In body part methods the estimated size is compared to the actual size, thereby assuming the image is formed at real size and that subjects “read off” the corresponding sizes of that particular image. However, they may go about this task in different ways. It may well be the case that subjects generate an image for each of the body parts separately. Smaller body parts, such as the head, may then be generated at a larger size to be able to better inspect the image for the requested size property. Alternatively, subjects may construct an image of the body as a whole, but may “zoomin” on the image to better assess the body width of interest. Hence, if smaller body sizes are imaged at greater sizes or zoomed-in on, it can be predicted that the estimated sizes will show greater discrepancies in the direction of overestimation than is the case for larger body parts, which would account for the relation found. Such an artifact might be resolved by carefully instructing the subjects how to construct the image (see Kosslyn [18]), and how to transfer the width of the body part as inferred from the image as a whole to the size estimation apparatus. Under such circumstances we would expect the negative relation to disappear. If this procedural modification is combined with the statistical corrections suggested earlier, we would see no problem in further application of the BPI, nor in further application of body size estimation methods. Acknowledgments—Preparation of this manuscript was supported by the Niels Stensen Foundation.
REFERENCES 1. Smeets MAM, Smit F, Panhuysen GEM, Ingleby JD. The influence of methodological differences on the outcome of body size estimation studies in anorexia nervosa Brit J Clin Psychol 1997;36:263–277.
464
M. A. M. SMEETS et al.
2. American Psychiatric Association. Diagnostic and statistical manual of mental disorders, 4th ed. Washington, DC: APA 1994. 3. Slade PD, Russell GFM. Awareness of body dimensions in anorexia nervosa: cross-sectional and longitudinal studies. Psychol Med 1973;3:188–199. 4. Ben-Tovim DI, Whitehead J, Crisp AH. A controlled study of the perception of body width in anorexia nervosa. J Psychosom Res 1979;23:267–272. 5. Thompson JK. Body size distortion in anorexia nervosa: Reanalysis and reconceptualization. Int J Eat Disord 1987;6:379–384. 6. Ben-Tovim DI, Crisp AH. The reliability of estimates of body width and their relationship to current measured body size among anorexic and normal subjects. Psychol Med 1984;14:843–846. 7. Shontz FC. Perceptual and cognitive aspects of body experience. New York: Academic Press 1969. 8. Baird JC, Noma E. Fundamentals of scaling and psychophysics. New York: Wiley 1978. 9. Hunter JE, Schmidt FL. Methods of meta-analysis: correcting error and bias in research findings. Newbury Park, California: Sage 1990. 10. LIMDEP, version 7.0 [computer software]. Bellport, New York: Econometric Software. 11. Pedhazur EJ. Multiple regression in behavioral research: explanation and prediction, 2nd ed. New York: Holt, Rinehart and Winston 1982. 12. Hsu LKG, Sobkiewicz TA. Body image disturbance: time to abandon the concept for eating disorders? Int J Eat Disord 1991;10:15–30. 13. Thompson JK. Body image disturbance: assessment and treatment. New York: Pergamon 1990. 14. Williamson DA, Davis CJ, Goreczny AJ, Blouin DC. Body-image disturbances in bulimia nervosa: influences of actual body size. J Abnorm Psychol 1989;98:97–99. 15. Smeets MAM. The rise and fall of body size estimation research in anorexia nervosa: a review and reconceptualization. Eur Eat Disord Rev 1979;5:75–95. 16. Kosslyn SM. Image and mind. Cambridge, Massachusetts: Harvard University Press 1980. 17. Kossyln SM. Image and brain: the resolution of the imagery debate. Cambridge, Massachusetts: MIT Press 1994. 18. Kosslyn SM. Measuring the visual angle of the mind’s eye. Cogn Psychol 1978;10:356–389.
S0022-3999(97)00142-6
BODY PERCEPTION INDEX: BENEFITS, PITFALLS, IDEAS MONIQUE A. M. SMEETS,* FILIP SMIT,† GEERT E. M. PANHUYSEN* and J. DAVID INGLEBY* (Received 9 July 1996; accepted 12 March 1997) Abstract—Estimates of body size are often expressed as a ratio of actual size [body perception index or BPI5(estimated size/actual size)3100%]. In this article, we examine the possibility that overestimation of body size in patients with anorexia nervosa, as measured by the BPI, is due to their smaller body size rather than to their being anorexic. Using 50 mean body sizes derived from seven studies we investigated whether the error of estimation is a constant proportion of the body size to be estimated, as the use of the BPI assumes. A negative linear relation between BPI and actual body size was found, confirming that smaller size is associated with greater overestimation. However, although both groups showed a strong tendency to overestimate smaller sizes, anorexic subjects showed even greater overestimation than controls. Hence, overestimation of body size in AN can only partially be accounted for by the smaller body size of anorexic patients. Recommendations for future use of the BPI are put forward. 1998 Elsevier Science Inc. Keywords: estimation.
Anorexia nervosa; Body image; Body perception index; Body size estimation; Visual size
INTRODUCTION
On average, patients with anorexia nervosa (AN) estimate their body size to be greater than it actually is. This was recently established in a statistical meta-analysis of 33 studies comparing body size estimates from patients with anorexia nervosa (AN) to controls without an eating disorder [1]. This finding implies that these patients have a visual mental image of their own body as fatter than it really is, which is part of the more general disturbance of body image—a diagnostic marker of AN [2]. An alternative explanation for the overestimation of body size by patients with AN, however, might be that it is an artifact of their being thin rather than a consequence of their being anorexic. According to this explanation, the habit of displaying the body size estimate as a fraction of actual size puts anorexic subjects— who, by definition, have smaller actual sizes than normal-weight controls—at a disadvantage. The ratio of estimated size to actual size (3100%) is known as the body perception index (BPI; see Slade and Russell [3]). The BPI is typically used in so-called body part studies, which are directed at estimation of actual body widths.
* Department of Psychonomics, Utrecht University, Utrecht, The Netherlands. † The Netherlands Institute of Mental Health and Addiction, Utrecht, The Netherlands. Address correspondence to: Monique A. M. Smeets, Pennington Biomedical Research Center, 6400 Perkins Road, Baton Rouge, LA 70808. Tel: (504) 763-3111; Fax: (504) 763-3022; E-mail: smeetsma@ mhs.pbrc.edu
457
458
M. A. M. SMEETS et al.
Using the BPI, Ben-Tovim et al. [4] made the curious discovery that mean BPI was lower in anorectics than in controls. This, in combination with the fact that the controls in their study had smaller body widths than the AN patients, prompted them to investigate the relationship between BPI and actual width. They found a significant tendency for BPI to increase with actual width: the smaller the body part, the greater its percentage overestimation. When this tendency was taken into account, no statistical difference emerged between the anorexic and control group. Thompson [5] subsequently showed that a significant difference between anorexic subjects and controls, apparent when the BPI was used in the traditional manner, became insignificant when the actual measures of the controls were used as the denominator for the anorexic group in the absence of data on their own premorbid body sizes. On the basis of the results obtained in a follow-up study to further investigate this phenomenon, Ben-Tovim and Crisp [6] claimed that estimates are random in relation to actual body sizes. Consequently, the smaller body sizes that are likely to be found in individuals with AN will increase the apparent size of perceptual error, as a function of dividing a random number by a relatively smaller number. The above findings urge a more cautious approach to applying the BPI in populations differing in weight and size. On the basis of the conclusion from the meta-analysis that anorexic patients, on average, overestimate their body sizes as compared to normal controls, we now investigate whether anorectics overestimate because they are anorexic, or merely because of their smaller actual sizes. On the basis of the results, we wish to formulate specific recommendations for future applications of body size estimation methods. It is appropriate first to highlight several assumptions that seem to have been made in applying the BPI, starting from the broader perspective of classical psychophysics. We will restate Ben-Tovim and Crisp’s [6] claim that people with AN show stronger body size overestimation than the average-sized controls they tend to be compared to because of smaller actual body size, starting from Weber’s law. Weber’s law has generally been used in psychophysics to express the relation between the perceived and actual value of a stimulus, and seems to apply to this case. If person A’s BPI is higher than person B’s, this implies either that A’s error of estimation in terms of centimeters is greater than B’s, or that A’s actual body size is smaller.1 Comparing BPIs with each other when the denominators (i.e., actual body sizes) are unequal would seem to be justified by the additional assumption that smaller actual sizes are accompanied by smaller errors of estimation—or vice versa, that greater actual sizes may lead to proportionately greater errors. The assumption that error/ actual size5k where k is a constant, is equivalent to Weber’s law that dS5kS or dS/ S5k: that is, the standard intensity of a stimulus (S) must be increased by a constant fraction (dS) to be distinguished (see Baird and Noma [8]). Correspondingly, in body size estimation studies an increase in body size is assumed to be accompanied by a proportional increase in estimation error. Hence, in the present study we first investigate whether, in accordance with We1 If person A’s body actual size is smaller than person B’s, his or her BPI may be higher not only when A’s estimation error is larger or equal to B’s, but even if it is slightly smaller. (For a computational example the reader is referred to Ben-Tovim and Crisp [6], p. 107).
Body perception index
459
ber’s law, the BPI is constant over a range of body sizes, or whether there is an (artifactual) correlation with actual size. If, as Ben-Tovim and Crisp [6] put it, body size estimates are randomly related to actual sizes, the function relating BPI to actual size will be a negative linear one rather than a constant. Second, we investigate whether the difference in body size estimates between anorexic patients and controls remains after controlling for the influence of actual size on BPI. This study focuses on size estimation data cumulated over studies, using the method of statistical meta-analysis.
METHOD Sample of studies The sample of studies used for this investigation was derived from the sample used in Smeet et al. [1]2. For this meta-analysis, studies carried out between 1973 and 1993 were collected which compared anorexic patients to normal controls on visual body size estimates. Of the original sample of 33 studies, 13 made use of body part methods. Six studies had to be excluded from this subsample because actual body width measures, which are essential to our analysis, were not mentioned. For an overview of the remaining studies the reader is referred to Table I. It should be noted that the usual statistic employed in meta-analysis is the effect size, which quantifies the difference in score between groups. In the present study we are not interested in a measure that combines the scores obtained by the groups. Rather, we investigate the relation between BPI and actual size for each of the groups separately. This implies that the effect size (d) was not suitable here. In this study, raw mean scores (BPI’s) were used instead. This strategy is usually discouraged in favor of using standardized measures such as d, because scales tend to vary when using different measurement techniques. However, because, in the present analyses, the sample was comprised of studies all using the same technique, namely visual size estimation, we feel that analyzing raw-score means, under the assumption of equality of sds, is acceptable (see also Hunter and Schmidt [9]). In total, 50 BPIs with corresponding actual sizes could be inferred from seven studies, which amounts to 25 per group. Each BPI corresponded to the mean for either of the groups.
Analyses All analyses involve ordinary least squares linear regression and were carried out using the LIMDEP program [10]. Extracting 50 BPIs from seven studies yields observations which are not independent of each other. To take account of this nonindependence, dummy variables were entered into the analyses to distinguish each study that provided the BPI. To test the influence of actual size, group membership (anorexic vs. control), and study membership on BPI, the following model was fitted: BPIi5a1b1 Gi1b2 Acti1Skj51b3Sij1ui
(1)
where BPIi5body perception index for i51, 2 . . . N group means; a5intercept (constant); b5regression weight (slope); Gi5group (05control, 15anorexia nervosa); Acti5actual size; Skj51Sij5set of study dummies (to account for dependency of BPIs within studies for j51, 2 . . . K studies); ui5 unique term. The first research question (concerning the relationship between BPI and actual size) involves testing the statistical significance of the regression weight b associated with Acti. Testing the second research
2 The primary purpose of the study was to evaluate whether the main finding of the meta-analysis reported in Smeets et al. [1]—that individuals with AN overestimate their body sizes as compared to controls—could be explained by the fact that the AN subjects had smaller actual body widths. Therefore, a subsample that provided the necessary data to perform the present analyses was taken from the original 33 studies, published between 1973 and 1993. However, because some time had passed since 1993, it was decided to perform another literature search to see if any appropriate studies had been published since then. Only two studies were found in which an AN sample was compared to a normal control sample, using the BPI to express the resulting estimates. However, because these studies did not provide the actual mean sizes for the body widths used in the estimation, these studies could not be included in the regression. We believe that the recent dearth of studies should be attributed to the nonpopularity of body size estimation methods, partly due to the artifact discussed here.
460
M. A. M. SMEETS et al.
Table I.—An overview of actual measures of body sizes and related body perception index (BPI) scores Anorexia nervosa Controls
Body part
Actual
BPI
Actual
BPI
Slade and Russell (1973)
face chest waist hips
11.63 23.32 19.84 27.18
157.6 134.2 146.6 126.8
10.9 25.58 21.77 29.46
94.7 95 100.2 96.6
Button et al. (1977)
face chest waist hips
11.94 25.75 21.24 29.78
115.2 115.5 119 102.5
12.57 28.24 24.66 34.46
131.1 121.6 113.5 107.1
Ben-Tovim et al. (1979a)
face waist
12.6 22.3
139.9 145
12.2 22.0
167.1 167.4
Casper et al. (1979)
face chest waist hips
11.32 22.69 18.64 27.71
130.8 114.1 124.6 103.2
11.39 24.91 22.23 29.22
122.4 108.9 113.7 107.7
Norris (1984)
head waist hips thigh
14.2 19.6 28.5 11.2
126.5 118 98 129.5
14.3 23.5 32.3 15.5
115.5 100 91.5 98
Proctor and Morley (1986)
face bust waist hips
11.2 24.3 20.9 28.3
153.6 118 140.1 117.4
11.4 26.1 23 32
132.0 103.9 113.9 97.5
Hartley (1989)
chest waist hips
27.13 24.15 30.51
153.24 156 147.98
32.50 29.20 36.40
123.92 122.72 113.88
Study
Note: actual measures are displayed in centimeters. a Actual measures and BPIs estimated from scatter diagrams. question (concerning the difference between groups) involves doing the same for the regression weight associated with Gi (see footnote 3).
RESULTS
The regression of BPI on Act and G yielded the following equation [F(8, 41)510.83, p,0.001], which accounted for 68% of variance: BPIi5138.50112.05G21.21Act For the regression coefficients associated with the control dummy variables corresponding to studies the reader is referred to Table II. The negative sign of the coefficient associated with Act(21.21; t524.38, p,0.001) demonstrates that the smaller the body width, the greater the relative overestima3 To account for the possibility that there is an interaction of the variables Act and G, a model was first tested that included the interaction vector G*Act (see Pedhazur [11]). Testing of the regression weight b associated with G*Act revealed no significance ( p50.80). Hence, because group membership and actual size do not interact, the interaction term can be omitted from the regression equation, yielding Model 1.
461
Body perception index
Table II.—Regression equations for the regression of BPI on actual size and group Variables Predictors Group Actual size Control variables S1 S2 S3 S4 S5 S6 Intercept R R2
B
seB
p
12.05 21.21
3.55 0.28
0.001 0.000
138.08 169.64 134.98 127.57 142.72 166.40 138.50
8.29 8.18 7.71 7.47 7.96 10.11 7.76
0.82 0.68
Note: B is the regression weight; se is the standard error. Group: 0 5 control; 1 5 anorexia nervosa. Control variables S1 to S6 are the dummy variables associated with the studies.
tion. Hence, the hypothesis that the slope of the regression lines is zero, which is implicit in the general use of the BPI, was rejected. Moreover, from the positive sign of the coefficient associated with G (12.05; t53.40; p,0.010) it was concluded that there is a real difference between the anorexic and control group in BPI, with anorectics (G51) showing stronger overestimation than controls (G50). Having determined that regression slopes do not differ (see footnote 2)—hence must be parallel—but that intercepts were different, separate regression functions were obtained with identical coefficients for Act but different intercepts. The Group effect (anorexic vs.control), captured by the group dummies, indicates that anorectics overestimate their body size by 12.05 BPI units as compared to normal subjects. The regression functions for the AN and control group, respectively, are shown in Fig. 1. DISCUSSION
Body size estimation methods, which became increasingly popular in the 1970s and 1980s for assessing the size at which people image their bodies, and especially for the disturbances in this ability in eating-disordered patients, have suffered a decline in research interest in recent years (see Hsu and Sobkiewicz [12]). The main reasons for this decline seem to lie in the apparent inconsistency4 of research findings and methodological problems. In the present study we have revealed one such problem, which, in our view, can be adequately dealt with in future research. Researchers have implicitly assumed that people make smaller absolute errors when estimating smaller body sizes and that, as a consequence, the proportion of the size of the error to the actual size is constant. If, contrary to this assumption, the size of the error does not decrease in proportion to actual body size, BPIs will 4 The results obtained with body size estimation methods are not as inconsistent as they seem (see Smeets et al. [1]).
462
M. A. M. SMEETS et al.
Fig. 1. Regressions of body perception index (BPI) on actual size after testing of regression functions for equality: BPIcon5138.521.21Act and BPIAN5150.5521.21Act, where Act is actual size. Solid line: An; dotted line: control.
increase as actual size decreases. Should this be the case, then overestimation of body size by patients with anorexia nervosa could be a result of being thin rather than of being anorexic. Contrary to the assumption of constancy, the BPI was found to be higher for smaller body sizes, implying that smaller body sizes were overestimated to a greater extent. Thus, and in agreement with Ben-Tovim and Crisp [6], it was concluded that the BPI transformation introduces a bias into the results when it is used to represent the outcome in body part studies. Although both anorectics and normal controls share the tendency to overestimate thinner bodies, the BPI scores for anorexic patients were even larger than those obtained by normal controls. Hence, overestimation of body size by anorexic patients cannot be entirely explained by the tendency to overestimate thinner actual sizes: on average, there is still a relative overestimation of 12.05% for anorexic patients. In view of the tendency for smaller body sizes to be overestimated to a greater extent than larger sizes, we recommend that variability in actual body sizes should be taken into account in future research. If the emphasis is not on eating-disordered populations, subjects can be matched on actual size prior to testing, as recommended by Thompson [13]. Alternatively, groups can be equated statistically by covariance procedures, which allow initial differences in body size across groups to be removed as a confounding variable [14]. Another possibility is to perform the regression analyses that were conducted here. In this way, the relation between BPI and estimated size can be easily examined. Between-group testing then comprises testing of the regression coefficients and intercepts belonging to the regression equations. It should be noted that the problem discussed in this study specifically for the BPI,
Body perception index
463
might also apply to studies using so-called whole body techniques, in which subjects estimate the size of their body as a whole by adjusting the size of a video image of their own body on a television screen. Also, in this kind of set-up, errors will be percentually larger when expressed against smaller body widths. Unfortunately, it is very hard to evaluate this artifact because such studies never provide actual body sizes or widths of the stimulus picture. However, if the present findings can be extended to the whole body procedures, we might anticipate that a difference between the AN and control groups will remain after controlling for this tendency. It is suggested to use thin controls without an eating disorder in future studies, to minimize differences in body size widths shown in the video images. The above recommendations deal on a post hoc basis with the unexpected finding of a negative relation between BPI and actual size, but do not offer any insight into this curious relation. The finding that smaller body parts are overestimated to a greater extent certainly runs counter to our intuition that smaller sizes can be estimated with smaller error. We will put forward an explanation to account for this finding, according to which the tendency to overestimate smaller body sizes to a greater extent than larger sizes has to do with the size of the image spontaneously constructed during size estimation. This explanation originated from analyzing the cognitive information processing that goes on during a body size estimation task (see Smeets [15]). To be able to estimate the width of the body at designated parts, people have to consult a visual mental image. One of the properties of imagery is that the image can be constructed at any size, while preserving the underlying proportions [16, 17]. In body part methods the estimated size is compared to the actual size, thereby assuming the image is formed at real size and that subjects “read off” the corresponding sizes of that particular image. However, they may go about this task in different ways. It may well be the case that subjects generate an image for each of the body parts separately. Smaller body parts, such as the head, may then be generated at a larger size to be able to better inspect the image for the requested size property. Alternatively, subjects may construct an image of the body as a whole, but may “zoomin” on the image to better assess the body width of interest. Hence, if smaller body sizes are imaged at greater sizes or zoomed-in on, it can be predicted that the estimated sizes will show greater discrepancies in the direction of overestimation than is the case for larger body parts, which would account for the relation found. Such an artifact might be resolved by carefully instructing the subjects how to construct the image (see Kosslyn [18]), and how to transfer the width of the body part as inferred from the image as a whole to the size estimation apparatus. Under such circumstances we would expect the negative relation to disappear. If this procedural modification is combined with the statistical corrections suggested earlier, we would see no problem in further application of the BPI, nor in further application of body size estimation methods. Acknowledgments—Preparation of this manuscript was supported by the Niels Stensen Foundation.
REFERENCES 1. Smeets MAM, Smit F, Panhuysen GEM, Ingleby JD. The influence of methodological differences on the outcome of body size estimation studies in anorexia nervosa Brit J Clin Psychol 1997;36:263–277.
464
M. A. M. SMEETS et al.
2. American Psychiatric Association. Diagnostic and statistical manual of mental disorders, 4th ed. Washington, DC: APA 1994. 3. Slade PD, Russell GFM. Awareness of body dimensions in anorexia nervosa: cross-sectional and longitudinal studies. Psychol Med 1973;3:188–199. 4. Ben-Tovim DI, Whitehead J, Crisp AH. A controlled study of the perception of body width in anorexia nervosa. J Psychosom Res 1979;23:267–272. 5. Thompson JK. Body size distortion in anorexia nervosa: Reanalysis and reconceptualization. Int J Eat Disord 1987;6:379–384. 6. Ben-Tovim DI, Crisp AH. The reliability of estimates of body width and their relationship to current measured body size among anorexic and normal subjects. Psychol Med 1984;14:843–846. 7. Shontz FC. Perceptual and cognitive aspects of body experience. New York: Academic Press 1969. 8. Baird JC, Noma E. Fundamentals of scaling and psychophysics. New York: Wiley 1978. 9. Hunter JE, Schmidt FL. Methods of meta-analysis: correcting error and bias in research findings. Newbury Park, California: Sage 1990. 10. LIMDEP, version 7.0 [computer software]. Bellport, New York: Econometric Software. 11. Pedhazur EJ. Multiple regression in behavioral research: explanation and prediction, 2nd ed. New York: Holt, Rinehart and Winston 1982. 12. Hsu LKG, Sobkiewicz TA. Body image disturbance: time to abandon the concept for eating disorders? Int J Eat Disord 1991;10:15–30. 13. Thompson JK. Body image disturbance: assessment and treatment. New York: Pergamon 1990. 14. Williamson DA, Davis CJ, Goreczny AJ, Blouin DC. Body-image disturbances in bulimia nervosa: influences of actual body size. J Abnorm Psychol 1989;98:97–99. 15. Smeets MAM. The rise and fall of body size estimation research in anorexia nervosa: a review and reconceptualization. Eur Eat Disord Rev 1979;5:75–95. 16. Kosslyn SM. Image and mind. Cambridge, Massachusetts: Harvard University Press 1980. 17. Kossyln SM. Image and brain: the resolution of the imagery debate. Cambridge, Massachusetts: MIT Press 1994. 18. Kosslyn SM. Measuring the visual angle of the mind’s eye. Cogn Psychol 1978;10:356–389.