Reconstruction of body height from the skeleton: Testing a dozen different methods for consistency of their results

Abstract

There are a number of methods of physical anthropology available to reconstruct living stature from skeletal remains. Some methods use dimensions of just a few bones, together with regression equations (mathematical, see Table 1: 1-7), while other methods require the whole skeleton and simply add the heights of specific skeletal components (anatomical, see Table 1: 8-11). This study investigates the consistency that mathematical and anatomical methods can provide in the determination of stature from skeletal remains. A significant difference was found between average heights of the same 20 individuals determined from seven mathematical and four anatomical methods (paired t-test, p < 0.001, df = 19). Mathematical methods provided taller height estimates than anatomical methods; the average difference was 47 mm. A repeated measures ANOVA indicated significant differences in the heights determined by all methods (p < 0.0001). Analysis of variance indicated significant differences in the heights determined by various mathematical methods (p < 0.03), whereas there were no significant differences in the heights amongst various anatomical methods (p < 0.77). When simple proportions of the length of the long bones to stature are used for reconstruction (see Table 1: 12), a bias is shown by mathematical methods to overestimate statures of short individuals and underestimate statures of taller individuals. To reduce this bias of linear regressions, we suggest that alternate methods, such as reduced major axis or organic correlation, should be employed (see Table 1: 13-15).

Stature estimation equations for South Asian skeletons based on DXA scans of contemporary adults

“…Comparing the equations produced using OLS, RMA, and robust regression models, there was no clear indication that one performed better than the others in terms of the errors associated with the regression models, or the results from applying them to Nat's () dataset. The advantages and disadvantages of the various regression models have been discussed previously (Hens et al, ; Konigsberg, Hens, Jantz, & Jungers, ; Pablos et al, ; Ruff et al, ; Sierp & Henneberg, ; Sjøvold, ; Smith, ), and while OLS is criticized for underestimating tall statures and overestimating short statures, the RMA and robust equations did not perform noticeably better in this respect. In fact the RMA equations gave more significant differences between measured and estimated statures in Nat's () dataset, although the correlations between stature and prediction errors were not significant for the RMA humerus and ulna equations, indicating that unlike most of our other equations, the pattern of overestimated statures for shorter individuals and underestimated statures for taller individuals was not a problem.…”

“…There has been considerable debate over the most appropriate way to derive stature estimation equations from long bone lengths (Hens et al, ; Konigsberg et al, ; Pablos et al, ; Ruff et al, ; Sierp & Henneberg, ; Sjøvold, ; Smith, ), and there appears as yet to be no consensus. A known limitation of the widely‐used ordinary least squares (OLS) regression is its tendency to overestimate statures of smaller individuals and underestimate those of taller individuals because of the nature of the line fitting process (Sjøvold, ).…”

“…A known limitation of the widely‐used ordinary least squares (OLS) regression is its tendency to overestimate statures of smaller individuals and underestimate those of taller individuals because of the nature of the line fitting process (Sjøvold, ). Reduced major axis (RMA) regression reduces this effect and so has been advocated by some, and is considered to be more reliable when extrapolation beyond the range of the original dataset may be necessary and when both the predictor and outcome are measured with error (Auerbach & Ruff, ; Ruff, Scott, & Liu, ; Ruff et al, ; Sierp & Henneberg, ; Sjøvold, ; Smith, ). RMA regression minimizes the deviation from the fitted line in both the x and y variables, rather than just the y variable as in OLS regression, and thus results in a symmetrical line (i.e., it makes no difference which variable is placed on which axis) (Sjøvold, ; Smith, ).…”