Application of DSP2 for biological sex estimation in a Spanish sample: analysis based on sex and side

Current results show that, whereas the sexing % decreased progressively when the number of variables or their discrimination power was reduced, the accuracy % remained high irrespective of the combination of employed variables. A comprehensive comparison to previously reported similar literature is shown in Table 8. Overall, results from all ten, or the first eight variables are almost identical across the table, so it appears that when the first eight variables are available for analysis, incorporating the remaining two (SIS and VEAC) does not render better outcomes. As advocated by Murail et al. and Bruzek et al. [10, 11], the last two features are useful within degraded and/or fragmentary contexts, where other more accurate variables are non-viable. In such scenarios, these variables with relatively lower accuracy can aid in identification by helping achieve the minimum four-variable requirement mandated by the DSP software.

Sexing % values obtained with all ten variables in the present study were 94.44% and 93.55% for the right and left halves, respectively, which is slightly lower than the 97.40% reported by Chapman et al. [18], the maximum value found in the literature, and higher than the 85.43% reported by Salles Machado et al. [17], the minimum value achieved across referenced studies (Table 8). Regarding % accuracy, the present study achieved 97.65 and 98.85 (right and left, respectively), within the range from the minimum 86.20% [24] to the maximum, 100.0% [11, 12, 14, 18‐20]. A comparative analysis (Table 8) indicates that general % accuracy results are consistently high in literature, with most of the authors reporting values higher than 95% with only two exceptions: a Brazilian miscegenated population [17] and a very small sample from South Korea [24]. In fact, several authors found no cases of misclassification during their sex estimation by DSP [11, 12, 14, 18‐20]. In the current study, this sort of success rate was achieved only in females.

The present results of % accuracy with the best 4 variables were 97.67% and 97.83% for right and left, respectively. According to the previous studies (Table 8), the reported results ranged from 80% [24] to 100% [10‐12, 15, 18, 20]. However, the lowest value (80%) came from an already mentioned small sample from Korea comprising of only 29 individuals. The second lowest result generated with a larger sample size (n = 103) was reported as 90.29% [17]. In this specific case [17], the results for % accuracy should be taken with caution since some reported values are unexpected and difficult to justify; i.e. they rendered 88.34%, 90.29% and 93.20% using 10 variables, best 4 variables and worst 4 variables, respectively. Continuing with best 4 variables results, all other studies yielded accuracy percentages higher than 90%. However, outcomes of % sexing are quite different: while the present study’s ones were 87.76% and 89.32% for right and left sides, respectively, previously literature results ranged from 71.09% [16] to 92.8% for a specific European subsample [10].

Results of the worst 4 variables achieved good accuracy % outcomes in the present study: 94.55% and 94.74% for the right and left sides, respectively. Regarding previous studies, if the Korean small sample study is not considered due to its smaller sample size, the results ranged between 90.48% [18] and 100% [10‐12]. In the case of the % sexing, the scenario is different. The minimum outcome was reported in a British subsample published by the original authors [11] as 32.79%, while the maximum was also found within the original publication [10] as 76%. Murail and colleagues’ original publication reported a sexing % of 40.23% in the worldwide sample (n = 2040) but, surprisingly, higher outcomes were reported in its different subsamples (even when 7 out of 9 of them were testing reference models on geographically different target samples), ranging from 63 to 76%. The rest of the revised literature, including the updating of DSP as DSP2 [11], reported % sexing between 32.79% [11] and 61.0% [15]. A scientific explanation for Murail’s differential findings is currently lacking.

One of the drawbacks of DSP compared with other discriminant analysis is the high number of undetermined cases that the software generates, especially when a reduced number of variables are considered. These results arise of the original authors decision about employing a 0.95 threshold instead of the 0.50 value. This settlement, however, guarantees very high levels of accuracy, with 100% or close values with a combination of the most discriminant variables. Interestingly, even taking into account just the four variables with relatively lower discrimination power (SS, SA, SIS and VEAC), accuracy percentages were lower than 94.0% in just 3 out of 32 revised literature outcomes, including the Korean one with a sample size of 29 individuals. Nevertheless, as other authors have suggested [20], the poorer results with the worst 4 combination of variables (or with a reduced number of variables) need to be interpreted with caution and complemented with non-metric methodologies.

In the present study, the posterior probability threshold was decreased from 95 to 85% to assess its potential impact on % sexing and % accuracy. This resulted in an increase in the % sexing in all cases in exchange of reduced % accuracy in most of the cases. Since the current female outcomes were comparably better than males, most of the changes were associated with sexing in males: improvement by increasing the sexing rate but gaining imprecision by incorporating errors. While this reduced accuracy still has acceptable values, the reduction in posterior probability is not recommended within future forensic practice from a reliability point of view, as a decrease in the number of indeterminate bones does not necessarily make up for decreased reliability and accuracy. Given the medico-legal and forensic contexts of interest here, an unreliable sex estimation approach can be more problematic than an unapplicable one [57, 58]. Furthermore, the reduction in posterior probability here, by and large, also resulted in a decrease in % accuracy, negating the need for this reduction. Moreover, even with a 95% posterior probability, % accuracy obtained in the present study is high (even with the worst four variables), demonstrating an excellent applicability for the DSP2 approach when it comes to sex estimation.

In relation to misclassification cases, some documentary or human error should not be completely ruled out in the osteological collection itself, since worldwide researchers and curators work with them all the time and misplacing skeletal elements or typographical errors could be contributing factors [18, 59].

Most of the authors who have analyzed DSP performance in diverse populations have used the left side throughout to be consistent [13, 15, 17, 18, 20]. However, some researchers used left and right hip bones in their studies, combining both sides without analyzing for potential differences [14, 19]. Whereas other studies only partially examined them [12, 16]. The present study has found higher values in all left cases compared to the right-side results, apart from ISMM, with opposite results, and a quite identical mean value in SIS between sides. However, these differences were only statistically significant in DCOX, IIMT, ISMM and SA. This finding could not be corroborated as no similar results have been reported previously, warranting further investigation into the causality of this anomaly. In fact, other authors concluded that the DSP values from the right and left coxal bones were comparable [12], even suggesting that single measurements may be substituted in cases of non availability of both halves of the pelvis. This absence of asymmetry agrees with previous results from a CT-based investigation which utilised different pelvic measurements [60].

Overall, current results did not indicate any clear side-based results regarding the % sexing in different number of variables sets. However, left side achieved better % accuracy compared to the right in all cases (Table 6). In females, similar results were found. No differences were displayed in most of the cases of % accuracy in females as left and right sides both achieved 100%. In males, on the other hand, right side displayed higher % sexing in all cases except for the best 4 variables, while left side achieved better % accuracy than right side with one exception: the worst 4 variables case. The only previous investigation to compare those results was from Rodrigo de Oliveira Lopes et al. [16], who reported accuracy comparison for the best 4 variables (PUM, SPU, DCOX, IIMT). These authors found higher % sexing and % accuracy in the left side in females and in the right side in males, i.e., contrasting findings in comparison to the present study (Table 6). Accuracy % results in females were not comparable because both left and right sides achieved 100% in the present study.

Neither PUM nor SA mean metric values were found to have significant differences between sexes in the left side coinciding with Salles Machado et al. [17]. In the right side, PUM alone did not exhibit significant sex differences, a result unreported in the literature. Other authors describe non-significant differences between males and females in SA [11, 13‐15, 19] and in the combination of PUM, SCOX and SA [12]. These PUM results highly contrast with the fact that this variable has the highest discrimination power across the analyzed ten variables according to the DSP creators [10], data which contrasts with Kranioti et al., who described the four most discriminant variables as ISMM, SPU, DCOX and VEAC in a Greek sample [13].

Tables 9 and 10 display the comparison of sex-specific percentages of sexing and accuracy depending on the number of DSP variables used across scientific literature. In the female comparative analysis (Table 9), the present study achieved 100% in most of the cases of % sexing in the first three sets (10, 8 and best 4 variables), exceeding the reported outcomes in other samples. Regarding % accuracy, no misclassification cases were reported here, agreeing with the ones rendered in Mexican [14], French [12, 19] and Brazilian [15] samples. However, the low % sexing exhibited in French (41.66%) compared to the rest of analyzed samples using the best four variables, ranging from 85.2 to 100%, is hard to explain, and is more similar to results from the set of the worst 4 variables. Since some of the values reported by Quatrehomme et al. [12] are internally inconsistent within their own tables, these results could have been caused by formatting typos and should be compared with caution. In the case of the worst 4 variables, present results are slightly lower than other studies [11, 12, 17]. It is remarkable that the high percentage achieved in this case by de Almeida and colleagues [15] in a Brazilian sample (81.2%), is more similar to the results of 10, 8 or best 4 variables in other geographical regions.

On the other hand, male-specific analysis is showed in Table 10. Regarding the 10-variable and 4-best cases, where more references were reported, current study % sexing are lower than the previous publications, except in Mexican [14] and Brazilian populations [16, 17]. For % accuracy, similar overall lower accuracy was achieved in the current study when compared with other publications with the exception of the Brazilian population [16, 17]. As mentioned previously, Salles Machado et al. [17] results are anomalous and should be considered cautiously; e.g. % accuracy using the worst 4 variables (100%) is higher than the value resulted from using 10 variables (90.57%). With regards to the worst 4 variables results, present % sexing results are higher than Northeastern Brazilian [15] and the worldwide sample utilised by Bruzek and colleagues [11] but lower than the outcomes achieved for France [12] and Southeast Brazil populations [17]. It is worthy to highlight the high percentage achieved by Salles Machado et al. [17] in a Brazilian sample (79.24%), comparable to the results of 10, 8 or best 4 variables in other geographical regions. The existence of an outlier outcome had been found and previously commented in Brazilian females, but in a different sample [15]. These differences could be explained by the great geographical extent and human diversity of this country, likely existing intrapopulation differences between Northeastern [15] and Southeast [17] Brazilian samples. To conclude the comparison, the current % accuracy results for the worst 4 variables in males were lower than all the revised literature ones [11, 12, 15, 17].

Comparing sexes across populations, in the case of the performance of the DSP methodology, both the sexing and accuracy results were better for females than males in most of the cases in the current sample. Higher % sexing, with less undetermined individuals in females, has already been reported in Mexicans [14], Danish [20], Brazilians [16] and the worldwide sample offered by Bruzek et al. [11]. Contrasting results were found for a French population [12]. Furthermore, outcomes in favor of females or males depending on the number of used variables, with no clear sex-specific trend, were found in French [19], Brazilian [15, 17], and Belgians [22]. According to % accuracy, on the other hand, the female predominance compared to males found here agreed with some Brazilian outcomes [15, 16] and were also reported in pre-Columbian mummies [22]. Besides, Bruzek et al. (worldwide sample) found more accurate results in males in the first two subsamples (10 and 8 variables) and coinciding or similar outcomes with the best and worst 4 variables subsamples, respectively [11]. No differences were found in % accuracy between sexes in Mexicans [14], and French [12, 19], where the systematic shared results between sexes were 100%. Contrasting with current findings, Gonzalez et al. examined sexual dimorphism in the great sciatic notch and ischiopubic complex in a Portuguese sample [61], concluding females are misclassified more often than males.

Current DSP variables values were within the ranges described by the worldwide sample used by the original authors [10, 11] with the exception of VEAC, where the current data was higher than the worldwide-based maximum threshold reported by Bruzek et al. [11]. Some research [62] stated that skeletal sexually dimorphic characters show inter and intra-population variability and other authors stated that pelvic-based sex determination may be impacted by population differences and sample variability [56, 63]. However, other researchers advocate for the opposite, specifically regarding the pelvis [64]. In keeping with this, Bruzek and colleagues [11] stated that the pelvis shows a similar pattern of sexual dimorphism across diverse geographical areas which appeared approximately 100–150 ky ago in early modern humans, defended by previous literature [65‐69]. Thus, they offered a worldwide database (software DSP2) of pelvic measurements to any anthropologist who needs to sex skeletal remains, regardless of their geographical origin. This global reference not only comes from samples around the world but also from different ethnical groups (Zulu, Soto, Afrikaner, African American, and European American) and temporal periods (from 18th to late 20th centuries), enhancing their potential applicability in miscellaneous populations. At this respect, it is worthy to stress that when reference models were used to determine the sex of geographically different target populations [10], accuracy results from 97.4 to 100% were displayed (Table 8). However, some large geographical areas were not considered in the reference sample used to calculate the posterior probabilities on DSP2, such as Central or South America. In addition, Africa and Asia were scarcely represented, with just Thai and South African samples analyzed. Although posterior tests on Mexican [14] and Brazilian [15, 16] samples reported comparable percentages of sexing and accuracy, application to miscegenated Brazilian-identified sample [17] yielded lower values compared to published literature (Table 8). Furthermore, similar results arose when DSP2 and population-specific formulae were applied to a Greece sample [13]. However, Kranioti et al. recommended population-specific formulations whenever possible to maximize the outcomes. This existing inter-population variation was also suggested by the application of DSP on Pre-Columbian mummies [22], which garnered lower measurements for the DSP variables when compared to the original reference ranges [11].