The Impact of Identity and Population History on Population Affinity Analysis in New Mexico Using Cranial Macromorphoscopic Data

Abstract

Background/Objectives: This paper presents cranial MMS data for an American Indian (AI) sample from New Mexico. We compare these data to other population reference samples to understand biological distance and classification patterns. Methods: Cranial MMS data was collected from CT scans of AI individuals (n = 839) in the New Mexico Decedent Image Database. We used 12 traits following a published protocol for CT data and excluded nasal bone contour. The AI sample is compared to other samples of African American or Black, Asian, Hispanic, and White individuals to assess biological distance and classification accuracy. Results: Biological distance analysis indicated similarity among the AI, Hispanic, and Black samples, but dissimilarity among the Asian and White samples to the AI sample. Classification accuracy was low for the overall model, with Hispanic and Black individuals frequently misclassifying as AI. Conclusions: As is true everywhere, ideas about identity are complex in New Mexico. AI and Hispanic individuals in NM self-ascribe to one or both social groups. We found that biological data cannot meaningfully differentiate between these social groups, as crania are morphologically similar when examining cranial MMS traits. These results, coupled with New Mexican identity and genetics, contributes to the difficulty in group attribution in forensic casework. Nevertheless, the reference data for AI presented here can provide more robust statistics that support the likelihood of AI and Hispanic affinity in population affinity analysis. We underscore the need for an understanding of regional population history and structure and reference samples while assessing population affinity in forensic casework.

1. Introduction

When human remains are discovered, forensic anthropologists are often called upon to estimate demographic variables due to their expertise in skeletal morphology and human variation. Collectively, this suite of demographic variables, which includes age, sex, stature, and population affinity, is called a biological profile, and it is one of the first steps toward identification. Here we focus on the estimation of population affinity, which is a measure of similarity between an unknown set of skeletal remains and reference population groups based on genetic or morphological variation within statistical models [1,2]. Human variation is shaped by environmental and cultural factors impacting the physical body both on a macro and micro evolutionary scale [3,4] and grouping labels should be defined according to research and/or societal parameters [2,5]. Current methods using population affinity modeling are statistically robust, with most models using morphological and metric variables from various cranial and postcranial elements (see [6,7]). However, models are only as good as the reference data available to create and test them [8]. Reference samples are often very small and, in many instances, absent, for American Indian (AI; we use the term “American Indian” to align with U.S. Census terminology, except when referring to authors who use alternate terms. “Native American” and “Indigenous Person” are considered synonyms for “American Indian”.) and other historically understudied populations [8,9], making population affinity estimates of these groups in forensic anthropology difficult and prone to inaccuracies [8].

The lack of research on contemporary samples is partly due to the absence of modern AI data in reference samples. The dearth of reference data is related to cultural ideas about internment, and mistrust in science, which stems from abuse of AI groups by anthropologists and researchers in the past [10,11]. While literature assessing aspects of the biological profile using AI archaeological data is available [12,13,14], studies examining the efficacy of biological profile methods using modern AI populations are limited [15,16]. As archaeological AI samples may not reflect variation in modern individuals that become part of medicolegal investigations, it is important to develop and test methods using appropriate reference samples.

AI, Asian, and Hispanic (we use the term ‘Hispanic’ to align with U.S. Census terminology, but acknowledge this term homogenizes culturally and linguistically different groups of people) groups share a proportion of genetic ancestry due to a deep shared population history [17,18,19]. In population affinity analysis in forensic anthropology, these groups are often hard to differentiate, despite vastly different population, cultural, and recent migratory histories [20]. However, research into variation among these groups shows geographic patterning among samples tested [12,21,22]. Considering the crisis of the Missing and Murdered Indigenous People (MMIP) in Canada and the U.S., and the humanitarian crisis of missing migrants at the Mexico–U.S. border, cranial macromorphoscopic (MMS) traits used in affinity analysis to distinguish among AI and Hispanic individuals could positively impact forensic casework, if they are useful in estimating affinity for members of these groups. This is particularly salient in New Mexico (NM), which is at the nexus of these two crises. Several cities in the state are identified as having statistically significant higher rates of missing AI females compared to expected numbers based on population demographics [23]. Additionally, NM borders Mexico and has recently experienced an uptick in clandestine migration and associated migrant mortality [24,25]. This, along with the state’s unique population, is reflected in statistics from the New Mexico Office of the Medical Investigator (NM OMI). In 2023, casework at the OMI included 32% White Hispanic individuals, 8% American Indian, 33% White Non-Hispanic, 16% White Unknown ethnicity, 2% Black, <1% Asian/Pacific Islander, and 10% Unknown race and ethnicity [26]. The OMI reports and all related investigative data follow the U.S. Census, meaning Hispanic is defined as an ethnicity, not a race. These data roughly mirror the state’s demographic of 80% White, which includes Hispanic, 10% AI, 2% Black, 0.8% Asian and 0.09% Pacific Islander [27]. Approximately 2.3% of Hispanics identify as AI, 0.7% as Black, and 0.3% as Asian [28].

While census level data employ biologically inconsequential ‘race’ categories on the U.S. population at large, local definitions about group membership may be forensically meaningful if they can be recognized in patterns of biological variation to provide information about socially relevant groups [29]. These considerations suggest research on a local level to determine labeling systems that are appropriate and useful for forensic casework and estimation of biological profile parameters [29]. This is especially true in NM where ideas about identity are complex and specific to New Mexicans. Within the socially defined Hispanic group, there are genetic contributions from Mexican, Spanish, Anglo (a term used locally to refer to non-Spanish Europeans and European Americans), and Indigenous AI groups in varying proportions among at least seven sub-groups, which are also socially defined [30]. Many New Mexicans socially ascribe themselves and each other to either AI or Hispanic (though sometimes to both) [30,31]. Group membership in New Mexican Hispanic subgroups can have biological consequences for members, including differential mortality [32]. In a medico-legal setting, the disconnect between self-identity, social race, and census-assigned groups has created challenges in the accurate assignment of race and ethnicity, an issue documented in a comparison of next of kin and investigator assigned race and ethnicity from the NM OMI [33]. In general, the number of Hispanic individuals is underestimated due to incorrect descriptions of decedents as non-Hispanic by death investigators whereas next of kin would have described them as Hispanic [33]. Further, the number of AI reported missing, and the number of unidentified individuals estimated to be AI do not match in national databases, such that there are more missing than unidentified [34]. It is likely that some AIs are likely categorized as Hispanic due to shared genetic ancestry and culture [35,36]. These errors can keep individuals unidentified for decades, prolonging uncertainty about missing people for families. Because these ascriptions of social group membership are included in descriptions of missing persons and limit the pool of potential identifications for unknown human remains, it is valuable to determine whether forensic anthropologists can estimate social group membership of unidentified individuals from phenotypic data.

The unique demographics and population history of NM, coupled with the intersection of these two national crises in the state, require improved population affinity estimates to improve overall identifications. The goal of this research is to examine cranial MMS traits to better understand trait presentation in contemporary AI and Hispanic groups and improve population affinity estimation in forensic casework. A notable benefit of this project is the contribution of cranial MMS reference data for modern AIs for inclusion in the Macromorphoscopic Databank (MaMD) [37], addressing a sample gap and allowing for future research into craniofacial variation.

While our focus here is on improving methods used in forensic casework, we employ a biocultural approach. We incorporate biological distance analysis and socio-cultural ideas about identity together to examine craniofacial variation among different census groups in NM. We use a classification framework, which is common practice in forensics, and use a biocultural approach to explain some of the issues in our results. Our approach provides insight into the complex population structure of NM and can explain some of the challenges with classification modeling in the region.

1.1. The Missing and Murdered Indigenous People Crisis on a National Level

The MMIP crisis, in which AI individuals experience violence at a much-elevated rate compared to other social race groups, is widespread throughout Canada and the U.S. [38,39,40]. The Bureau of Indian Affairs [41] states that rates of murder, rape, and violent crimes against AI are higher than the national averages for other population groups and are drastically underreported in AI communities. This crisis is the result of multi-faceted, institutionalized issues relating to social and criminal justice, racial/ethnic discrimination, and colonization through violence to AI populations in the U.S. occurring from the country’s inception to the present day [23,38,39,42]. Grassroots efforts have led to legislation at the federal and state level (in 14 states) to address MMIP; however, agencies indicate under-reporting and under-collection of data regarding missing Indigenous people [38]. Although this crisis is present at a national level, there are varying degrees of legislative efforts, governmental responses, and barriers to overcome when examining the MMIP crisis at a state level.

Many state efforts towards identifying unknown human remains use The National Missing and Unidentified Persons System (NamUs), a database containing information on both missing persons and unidentified persons, which can be accessed by medico-legal institutions, the public, and law enforcement for searches [43]. Information housed in NamUs can contribute to personal identification, case resolution, and research on topics related to missing and unidentified individuals [43]. One study using NamUs data explored differences between missing and unidentified women using race as a variable (the term ‘race’ refers to U.S. governmental classifications for race/ethnicity, reflecting current social and cultural constructions of identity). They found the number of missing persons cases of American Indian/Alaska Natives were proportional to other groups; however, these women were 135% more likely to remain unidentified than women of other census groups [39]. These long-term unidentified cases can have ramifications for living family members searching for their loved ones. Ambiguous loss is stressful for family and community members as there is no proof of finality (see [44]). One potential reason for a lack of unidentified remains classified as AI could stem from the failure of forensic anthropological methodology to correctly determine unidentified remains as coming from this group [45].

1.2. The Missing and Murdered Indigenous Peoples Crisis in New Mexico

NM has the fifth largest Indigenous population in the U.S., and it has the highest number of missing and murdered Indigenous women and girls [46]. State agencies, such as the New Mexico Department of Justice and the New Mexico Indian Affairs Department, have assembled various task forces and committees to increase awareness of MMIP issues and compile resources for families of the missing or those wanting to help their community. In 2019, the Missing and Murdered Indigenous Persons Advisory Council Act was established to create an MMIP State Response Plan. In 2022, state legislation created Missing in New Mexico Day so families can connect with law enforcement and receive support [47]. As of 2024, the state launched the Missing or Murdered Indigenous People portal where missing persons can be reported or searched for [48]. There has been limited research that examines this crisis, violence against AI groups, or statistical trends within NM, but what does exist has been undertaken by local task forces to raise awareness and expose gaps in the justice system [46].

Reports from the NM OMI corroborate the findings of state task forces as they relate to the death of and violence against AI communities. The 2022 report from the NM OMI states that 31.42% of all cases investigated by the office involve individuals of Hispanic (n = 2217) and AI (n = 770) racial and ethnic identities [49]; figures that increased slightly for 2023 [26]. Decedents from these groups had manners of death classified as ‘homicide’ and ‘undetermined’ at a rate nearly double that of White, non-Hispanic decedents [50]. There is a high caseload involving decedents from AI and Hispanic groups in the state and many missing and unidentified across the U.S. according to NamUs statistics. In a deeper examination of NamUs data, Joseph [23] found the number of missing persons cases for Indigenous women proportionally higher for several counties served by the NM OMI when compared to other places in the U.S. with large Indigenous populations. These counties include Navajo County in Arizona (AZ), and San Juan, McKinley, and Bernalillo Counties in NM. Cities in these districts, like Farmington, Gallup, and Albuquerque, had significantly higher numbers of missing Indigenous Persons than expected given the general population demographic information [23]. In our experience at the NM OMI, when following standardized protocol and using FORDISC 3.1 [51] to estimate population affinity of unknown skeletal remains, analyses of cases suspected to be AI, due to contextual information gathered during death investigation, often provide unclear results with poor statistics to support classifications. This results in more conservative approaches that list multiple groups (i.e., Hispanic and/or AI) for population affinity assessment.

1.3. Cranial Macromophoscopic Trait Data and Population Affinity

Cranial MMS traits are morphological variables in the craniofacial skeleton that correspond with soft tissue differences in the living [52,53]. This type of data has been used to assess affinity for decades with little regard for statistical or evolutionary frameworks (e.g., [54]). Within the last ten years, cranial MMS data have been standardized and accompanied by robust statistics [52,55]. Critiques of this approach liken the use of cranial MMS data in forensic anthropology to typological, non-statistical, trait-list approaches for estimating geographic origin [56], arguing that research into evolutionary mechanisms and global patterning of cranial MMS data outside of the forensic sphere is lacking. However, recent studies that model the relationship between phenotypic and genomic data find that several cranial MMS traits are tied to neutral genomic variation and can be reliably used in biological distance studies [57,58]. Plemons [4] further examined the selective forces of cranial MMS expression via climate variables (temperature and humidity), finding that cooler climates increase selection pressures for specific traits situating cranial MMS data within an evolutionary framework. Biological distance research also shows that cranial MMS data capture craniofacial variation resulting in similar group patterning as craniometric data from the same individuals [59].

Unlike metric approaches, cranial MMS data can be assessed visually and relatively quickly with a low level of observer error, within a standardized framework on fragmentary or partial remains [60,61,62]. This is especially important in forensic contexts as skeletal remains may be damaged due to trauma and/or taphonomy. Several studies have used cranial MMS data to understand patterns in biological distance and secular change; however, most have focused on samples using Black and White data from the U.S., with limited data from groups considered Hispanic [63,64,65,66,67,68,69]. Importantly, cranial MMS variables can differentiate between groups on a smaller, regional scale [59]. Models using cranial MMS data from AIs have not been tested as modern skeletal reference samples for this group were not available until recently. The MaMD contains the only cranial MMS trait data for AI groups, but these data come exclusively from the Smithsonian Institution’s Office of Repatriation [37], and comprise archaeological skeletons not suitable for contemporary, forensic case comparisons. The absence of modern AI data in population affinity modeling compounds the MMIP crisis by complicating forensic anthropologists’ ability to accurately estimate the biological profile of potentially unidentified AI remains.

Here, we present cranial MMS data for a modern AI sample from NM and explore biological distance and classification patterns with this sample and other contemporary samples included in forensic casework in the U.S. We aim to understand: (1) the relationships among AI and comparative samples using cranial MMS data within biological distance analysis, and (2) whether cranial MMS data can differentiate between AI and other comparative reference samples representing U.S. demographic groups (i.e., U.S. Black, U.S. White, U.S. Asian, New Mexico Hispanic).

2. Materials and Methods

2.1. Ethics Statement

The data used in this project come from forensic casework and include individuals affiliated with Pueblos and Tribes throughout NM and AZ. Data included here are deidentified and derive from deceased individuals and so do not qualify for human subjects review. However, due to the inclusion of deceased AI individuals, one of the authors (HJHE) established the University of New Mexico Missing and Murdered American Indian Advisory Group (UNM MMAIAG) at the outset of the larger project the current works was part of, to foster open dialogue regarding research and development of methods for identifying unknown decedents from AI groups. The UNM MMAIAG consists of scholars and members of Tribal communities from nearby Pueblos and Nations in NM and AZ who work for forensic and law enforcement agencies, universities, and nongovernmental organizations or are involved in public health or the search for missing Indigenous community members. Researchers involved in this project consulted with the UNM MMAIAG to discuss the research project and its initial results. Following discussions with the UNM MMAIAG, data sharing was acceptable given that it aligned with the overall goal of improving the identification of AI individuals in forensic contexts.

2.2. Samples

Data for this project comes from postmortem computed tomography (PMCT) scans housed in the New Mexico Decedent Image Database (NMDID; https://nmdid.unm.edu/). The NMDID is a large databank (N = ~15,000) of computed tomography scans collected from decedents during postmortem examination at the NM OMI [50]. A subset of AI (n = 839), Hispanic (n = 404), Asian American (n = 95), Black or African American (n = 281), and White or European American individuals (n = 47) from the NMDID were selected for analysis. We adopted population group labels from the NMDID, which used terms from the U.S. Census. Here, the term Asian American groups individuals described as Asian Indian, Chinese, Filipino, Japanese, Korean, Other Asian, Pacific Islanders, Samoan, Native Hawaiian, and Vietnamese. All decedents come from NM or AZ and Tribal affiliation (Puebloan and/or Athabaskan) is available for some of the AI sample. AI individuals were pooled in a single sample. We recognize that pooling the AI sample may obscure within-group variation; however, we opted to pool the AI sample as a first step; future analyses will examine variation within this sample. Any individual labeled as ‘White, Hispanic’ by the NM OMI were grouped in the Hispanic sample here. Individuals are 18 to 90 years old and had manners of death listed as ‘natural’ or ‘accidental,’ to minimize potential perimortem trauma or medical intervention to the cranium and ensure the visibility of all cranial MMS. We supplemented our data with samples from the MaMD [37]. These include continental Asian groups Chinese (n = 58), Japanese (n = 15), and Thai (n = 157) samples broadly grouped as Asian (n = 230), along with American Black (n = 38) and American White (n = 274) samples to meet a minimum threshold of 318 individuals per group, indicated by a power analysis. We pooled sex in our analysis following previous research design. Samples are presented in

Table 1. Sample demographic.

Sample	Population	Males	Females	Unknown	Total
NMDID
	Black or African American	198	72	0	270
	American Indian or AI	598	241	0	839
	White or European American	32	15	0	47
	Asian American	71	24	0	95
	Hispanic	305	99	0	404
MaMD
	American Black	40	9	0	49
	American White	162	112	0	274
	Chinese	53	5	0	58
	Japanese	10	5	0	15
	Thai	101	54	2	155

along with the labeling system used in analysis.

Data were collected for 12 cranial MMS traits (

Table 2. Cranial MMS traits.

Trait	Abbreviation
Anterior nasal spine	ANS
Inferior nasal aperture	INA
Interorbital breadth	IOB
Malar tubercle	MT
Nasal aperture shape	NAS
Nasal aperture width	NAW
Nasal bone shape	NBS
Nasal overgrowth	NO
Orbital shape	OBS
Postbregmatic depression	PBD
Posterior zygomatic tubercle	PZT
Palate shape	PS

; see [53]) from the NMDID samples following a protocol for collecting cranial MMS on CT data [70]. The CT protocol omits sutural-based cranial MMS traits, as these traits are not reliably observable on bone renderings and includes additional comments specific to CT data [70]. We further omitted collecting nasal bone contour (NBC) as a tool for emulating the contour gauge does not exist in medical imaging software, which is required for collecting the trait on dry bone [60].

Analysis was separated into three phases: initial data analysis (IDA), biological distance analysis, and classification. IDA aims to identify characteristics of and patterns in the dataset with and without grouping labels (see [71]). Biological distance analysis assesses similarities and differences among datasets according to the grouping variable (e.g., population) and is used to understand group relatedness. Finally, classification models are applied to evaluate how well skeletal variables can discriminate among the groups. These models are then used to assess model performance and provide information about their accuracy. Statistical analyses were performed in the open-source computational software, R (v. 4.4.2) [72].

Summary statistics provide information about the dataset and allow users to examine variability. Frequency data for each trait and character state in the AI sample are calculated using the ‘psych’ package and presented as an introduction to cranial MMS data in a modern AI sample. As the AI sample was initially the largest NMDID sample available, we first performed a power analysis to determine the appropriate comparative sample sizes using the ‘pwr’ package. We set the parameters for a medium effect (0.8) at the 0.05 significance level. Data from the MaMD were added when necessary to achieve the appropriate number of individuals. Second, all data are assessed for completeness. Missing data can occur for several reasons, including antemortem and perimortem trauma [59,73]. In bone renderings derived from postmortem imaging, objects from medical intervention may obscure the visibility of traits. Patterns of missing data are visualized using the ‘nanier’ package [74]. Imputation is a necessary step for Multiple Correspondence Analysis (MCA) analysis. Data were imputed using the ‘mice’ package following the Multiple Iterative Chained Equations (MICE) method [75]. Imputation is appropriate if the dataset contains less than 50% missing for each variable [73]. We used the predictive mean matching approach where a random observation is selected from the pool of observed values over n-iterations and n-imputations. After all iterations and imputations are performed, a consensus value approach assigns the missing value. Next, a polychoric correlation test in the ‘polychor’ package was employed to assess inter-trait correlations among the cranial MMS variables in the AI sample (see [69]). Possible outcomes include a positive correlation, a negative correlation, or no correlation, and results are visualized as a correlation plot. We employed MCA to identify patterns and relationships with respect to the grouping variable ‘population’. This type of analysis is a generalization of principal component analysis for categorical data [76], and we used the ‘factoextra’ and ‘FactoMineR’ packages to visualize results.

The second portion of this research aims to understand group relatedness using underlying morphological skeletal variables [77]. To assess this, we employed distance analysis using the Rank Estimator of Grade Differences (RED) statistic in the ‘RED’ package in R. Originally developed to assess dental nonmetric traits, RED is well suited for application to cranial MMS variables, by avoiding common issues like data reduction and compression that may diminish important biological relationships [78]. RED was used on the pooled, non-imputed dataset. Results are provided as a distance matrix and a hierarchical dendrogram to illustrate relationships among the samples.

To assess classification accuracy of cranial MMS data among all samples, we used a machine learning approach, artificial neural network (aNN) analysis. This method functions by introducing several variables that move through a series of nodes or layers within a ‘black box’ to arrive at an outcome based on patterns present in the dataset [55]. In aNN, random weights are assigned to each variable, and several models are generated over multiple iterations with the model that best fits the data selected. We used a randomized 30/70 training/test sample approach for cross validation using the ‘nnet’ package on the imputed group samples. We downsampled the training data to match the smallest training sample size to remove any potential bias toward the largest (AI) sample. Classification results are reported as contingency tables among population samples tested. Variable importance was assessed to understand the weighted connections among nodes in the model [79]. Variables were plotted by importance using the ‘caret’ package.

3. Results

3.1. Initial Data Analysis

Frequency data for each trait is presented in

Table 3. Frequency Data for the AI Sample.

Trait	Character State						Total
	0 1	1	2	3	4	5
ANS	—	328	396	91	—	—	815
INA	—	10	82	281	341	117	831
PZT	107	381	267	77	—	—	832
PBD	619	201	—	—	—	—	820
NO	176	647	—	—	—	—	823
NAS	—	405	183	240	—	—	828
NBS	—	85	510	164	52	—	811
NAW	—	105	578	143	—	—	826
MT	156	479	181	19	—	—	835
IOB	—	409	385	35	—	—	829
OBS	—	552	145	134	—	—	831
PS	—	206	91	179	44	—	520

¹ To note, scales for each character state begin at different numbers (see [52,53]). Dashes are used to denote character state scores that are not valid options for that trait.

for the AI sample only. Patterns are present in higher frequencies for certain character states; however, all character states are present for each trait in the AI sample.

Missing data is visualized by population label and trait in Figure 1. All samples have the highest amount of missing data for PS. Other missing values correspond to traits in the nasal area (NAS, NO, NBS, and ANS), but these variables are missing at a much lower percentage. An imputed dataset was used for the remainder of the IDA.

Polychoric correlations for all variables in all samples are shown in Figure 2. There are negative correlations between several nasal derived and midfacial cranial MMS traits like NO and IOB; INA and MT; NO and OBS; INA and IOB; NAW and ANS, indicating a reverse relationship in expression. While the level of correlations appears low, around 0 with slight positive or negative values, there were several traits with significant correlations (marked with *, **, or ***) in the figure.

The MCA indicates that traits in the nasal region contribute the most to variation in the sample. In dimension one NBS, NAW, IOB, INA and NAS drive variation, while NAS, NAW, and IOB are influential in dimension two (Figure 3).

When separating the variables into individual character states and population labels, a more nuanced picture of patterned variation emerges (Figure 4). Dimension 1 separates the White and Asian samples from the Hispanic, Black, and AI samples, while dimension 2 separates the White and Asian samples. The White sample exhibits a more projecting nasal spine (ANS 3), a narrower nasal aperture (NAW 1), and more rhombic-shaped orbits (OBS 3). Individuals in the Asian sample tend to have more projecting malar tubercules and more sloping inferior nasal apertures. Hispanic, Black, and AI samples exhibit overlap indicating individuals in these samples share similar characteristics like circular shaped orbits (OBS 1), the absence of a sloped INA (3,4,5), less pronounced MTs (0,1,2), and intermediate nasal aperture widths (NAW 2). In general, individuals in the AI sample have a more bowed nasal aperture shape (NAS 3), a bony overgrowth of the inferior end of the nasal bones (NO 1), the absence of an anterior nasal spine (ANS 1), a wider nasal aperture width (NAW 3), more triangular-shaped nasal bones (NBS 4), and an elliptical palate shape with more constricted third molars (PS 4).

3.2. Biological Distance Analysis

A non-imputed dataset was used for the RED analyses. A similarity/dissimilarity matrix is presented in Figure 5 along with a heatmap of distances for each sample. A darker color indicates more dissimilarity while the lighter color indicates more similarity among the samples. The AI and Asian samples are the most distant from each other followed by the Hispanic and Asian samples. The most similar samples with the smallest distance measures are the AI and Hispanic samples, then the Hispanic and Black samples, then the Asian and White samples. The most distant samples are the Asian and AI, then White and AI, then Hispanic and White. A hierarchical dendrogram shows the relationship of the samples (Figure 5). The dendrogram splits the samples into two main clusters. The first cluster includes White and Asian samples, while the second cluster comprises the Black, AI, and Hispanic samples. The second cluster splits again, separating the Black sample from the AI and Hispanic.

3.3. Classification

Table 4. Train/Test Samples.

	AI	Asian	Black	Hispanic	White
Train	224	224	224	224	224
Test	251	97	95	121	96

provides the sample size for the train/test dataset for aNN modeling.

Classification results from the training sample on the aNN model are listed in

Table 5. Classification Results from the Training Sample.

Predicted	Actual
	AI	Asian	Black	Hispanic	White	%CCR
AI	149 1	26	72	89	24	66.5%
Asian	7	134	23	17	4	59.8%
Black	28	40	93	21	21	41.5%
Hispanic	16	20	17	70	9	31.3%
White	24	4	19	27	166	74.1%
					Total:	55.4%

¹ Italicized values indicate correct classifications.

. The total correct classification accuracy is 55%, which is better than chance (20%), but lower than classification modeling accuracy for population affinity in the literature (i.e., FORDISC and MaMD Analytical). The highest classification rates were for the White, AI, and Asian groups, ranging from 60% to 74% correct classification. The Hispanic group had the lowest classification rate. The most misclassifications for the AI sample were with the Black then White sample. While individuals from the Hispanic and Black samples most frequently misclassified as AI.

Classification results from the test sample on the trained model are listed in

Table 6. Classification Results from Test Sample.

Predicted	Actual
	AI	Asian	Black	Hispanic	White	%CCR
AI	123 1	9	35	43	10	55.9%
Asian	21	47	10	7	6	51.6%
Black	55	23	29	16	12	21.5%
Hispanic	34	8	10	39	8	39.4%
White	18	10	11	16	60	52.2%
					Total:	45.2%

¹ Italicized values indicate correct classifications.

. The total correct classification accuracy is 45%, which is, again, better than chance (20%). Patterns here are similar to those illustrated by the confusion matrix from the training data. The AI, Asian, and White samples had the highest classification rates from 52% to 56% correct classification. The highest misclassification rates for AI were with the Hispanic then Black samples. The Black sample had a correct classification rate just above random. In all groups except for AI, the highest number of misclassifications were into the AI sample. More Blacks misclassified as AI then were correctly classified.

Variable importance is presented in Figure 6. The cranial MMS variables are ordered in descending order from most to least importance. The most important variables are nasal derived traits: NAW, NAS, NBS, NO, ANS, then INA, IOB, OBS.

4. Discussion

This study collected cranial MMS data from PMCT scans of modern AI individuals from NM. Our goals were to assess biological distance and classification among the AI sample with comparative samples encountered in casework and to understand if these data could be useful for differentiating between AI and other socially labeled population groups. Additionally, we provided our data to the MaMD for use in the MaMD Analytical tool and to make them available to other researchers.

4.1. Data Analysis

While PMCT data have many advantages, such as access to large, diverse datasets and the ability to collect data from skeletal models without handling physical human remains [80], one disadvantage is the potential for trauma or medical devices to obscure skeletal features on 3D renderings. Here, we filtered our sample by excluding individuals where the manner of death was recorded as ‘homicide’ or ‘suicide’ to minimize potential cranial trauma; however, because of the nature of the NMDID sample, several individuals had objects like intubation tubes in place while scanned. Additionally, metal, like that used in dental restorations, causes artifacts that can obscure features on skeletal 3D renderings. This created a pattern of missing data for traits near the hard palate oral cavity. In this study, the variable with the most missingness was PS, which was missing in higher frequency for the AI and Hispanic samples. Other explanations for the high amount of missing data for PS in the NMDID sample are related to oral health in these two groups [81]. PS was a trait added in later versions of MaMD and therefore this trait information is absent from data in populations previously collected, which may explain its absence in some of the MaMD supplemented samples. Despite high levels of missingness, our sample sizes for PS in AI and Hispanic were large (n > 250), so it is unlikely to have significantly impacted our results. In the past, an elliptical PS was considered a hallmark of Native American affinity [82]. This trait was included in our analysis and statistically tested here but did not indicate significance in sample separation.

In the MCA, midfacial and nasal traits were the driver for sample separation. However, only 11% of the variation is captured by the first two principal components, which is relatively low, indicating that no single dimension strongly separates the data. The MCA and RED analyses indicate that the AI sample is most dissimilar to the White and Asian samples and most like the Hispanic sample, followed by the Black sample. This is visually presented in the MCA biplot with overlap among the AI, Hispanic, and Black samples and the similar trait expressions among the midfacial and nasal variables for all three groups. Classification results indicate that the AI, Asian, and White samples show a sufficient degree of distinction and classified effectively, but the overlap of the AI sample with the Hispanic and Black samples in the MCA analysis is confirmed by low classification accuracy for the latter groups. Additionally, a large portion of the Black and Hispanic samples misclassified as AI.

4.2. Population Structure and Identity in New Mexico

Given the population structure and complex population history of NM and the Southwest, the misclassification of AI and Hispanic groups was an expected outcome. Genetics studies have highlighted the overlap in ancestry between these groups, which aligns with this finding [83]. New Mexican individuals who identify as Hispanic often have high levels of AI admixture, and those who self-identify as Mexican or Mexican American had higher levels of AI admixture than those who self-identify as half-White, half-Hispanic or Spanish [36]. This aligns with other findings that New Mexican Hispanics who emphasize their Mexican heritage and those who emphasize their Spanish heritage are genetically different on the level of AI and European genetic ancestry [30]. Specifically, New Mexicans related to recent immigrants from Mexico identified differently from those who have lived in NM for generations and have greater ties to their Spanish heritage [31,84]. This complicated relationship between race and ethnicity is not limited to groups in NM [85]. The complexity and heterogeneity within the Hispanic group alone continues to be addressed in forensic research [21,59,86,87,88,89,90,91], its necessity reinforced here by the strong overlap of the Hispanic, AI, and Black samples.

The more unexpected results are among the AI, Black, and Asian samples. The similarity of the AI sample to the Black sample and the high misclassification of Black Americans as AI was surprising. Based on the history of enslaved Africans arrival to the Americas, we expected dissimilarity between the AI sample and Black sample. Bryc and colleagues [83] find higher levels of AI genetic material in individuals who self-identify as Black in the American Southwest than other parts of the country. However, in a study by Klimentidis and colleagues [36], self-identified Hispanic and AI individuals had low West African admixture. Further genetic analyses of populations from NM highlight the overall low African ancestry present in New Mexicans who self-identify as Hispanic or Latino [30].

Despite MCA and RED analyses showing a strong dissimilarity between AI and Asian groups, a high number of Asian individuals were also misclassified as AI. These conflicting results may be related to a secular change, recent changes in population structure not previously accounted for, or the inclusion of MaMD samples broadly grouped as ‘Asian’ that are genetically, phenotypically, and temporally distinct from American populations [22]. There is a general understanding that ancient populations migrated from Asia into the Americas and that AI populations have genetic similarities to Asian populations [17,92,93,94]. While current census data list New Mexican Asian individuals with heritage from both Northern and Southern regions in Asia, refugee groups from Southeast Asia (Laos, Cambodia, and Vietnam) resettled in Albuquerque in the 1990s [95]. Recent, contemporary Asian groups that share ancestry with AIs may be better represented by Northern Asian samples rather than Southeast Asian samples, like Thai [96,97]. This assumption is supported with morphological dental data, as George and Pilloud [98] found Indigenous groups in the Americas to be more similar to Northern Asian groups (e.g., Chinese, Honshu Japanese). Therefore, sampling may account for the dissimilarity seen between AI and Asian groups in our modeling. The focus of the paper was on AI and Hispanic samples due to the challenges of discerning among these groups in forensic casework. While we did not explore patterns of misclassification of the Asian samples here, we agree that future research is needed to address these results.

4.3. Forensic Casework and Reference Data

Results presented here could reflect a population structure that is unique to NM, implying that regional reference samples are more appropriate for regional casework, rather than using broad U.S. samples as homogenous reference data [29]. It is also possible that due to the unique history of population mixing in NM, population structure may not be preserved in skeletal data. One of our goals was to determine whether cranial MMS data could help differentiate AI and Hispanic groups in NM samples. Within a classification framework with multiple groups, correct classification rates were low. However, when we tested a model with only AI and Hispanic samples, we achieved a higher classification rate of 62%. This indicates cranial MMS variables are discerning patterns of biological variation within the local population structure, just not strongly alone. This underscores the difficulty of confidently distinguishing these two groups in our sample. Our findings suggest that only using cranial MMS data may not be enough for strong classification accuracies; however, there is value in their application. One approach could be to combine cranial MMS data with other skeletal data types (i.e., craniometric, dental nonmetric). Despite this, a modern AI reference sample in the MaMD will still prove valuable as it can provide more robust statistical reporting for forensic case reports and be used in other research into forensic methods. As it currently stands, our population affinity estimations in NM include “Hispanic or AI” due to poor statistical results and the understanding that since these groups are not well represented in reference populations, classification results are weak. By filling in reference data gaps, even if reports still include “Hispanic or AI,” statistics more strongly support the estimation. Additionally, previous studies have demonstrated that multiple skeletal data types complement each other and provide a more comprehensive picture of variation and population structure [57], we aim to combine cranial MMS data used here with craniometric data we have already collected to assess biological distance and classification accuracies. As a final point, we followed previous research design, pooling age and sex categories in our samples, under the assumption sex and age do not impact cranial morphology and subsequently MMS traits. However, future research should examine population affinity in relation to sex and age to understand the interaction of all three variables on biological profile estimates.

Readers unfamiliar with day-to-day forensic anthropological casework might wonder whether there is a point to conducting this research in the genomic era, when DNA analyses alone can determine an unidentified person’s sex and estimate their biogeographic ancestry, hair and eye color, and facial features [99]. However, forensic anthropology practitioners will understand that, at least for the present, such DNA analyses are slow, expensive, and beyond the scope for most cases. If it can be reliable, morphological analyses of skeletal remains will continue to be core to casework for the foreseeable future. Whether the tool applied to the estimation of population affinity is DNA, MMS, or any other source of information, two factors are of utmost importance. First, the reference sample to which unknown individuals are compared must reflect the population from which the individual may derive [100] and second, the nomenclature, sampling structure, and statistics used must be based in historical and ethnographic research into the populations concerned. This second factor is where morphologists and geneticists both consistently fall short.

5. Conclusions

We collected cranial MMS data for a modern AI sample from NM, and then we used the sample to assess biological distance and classification accuracies with other reference samples included in casework in the state. Our results indicate that cranial MMS data alone do not provide enough resolution to differentiate between AI and Hispanic groups in NM. However, if paired with other skeletal data types, these data could be useful in population affinity estimation in forensic casework. This is a good first step in examining craniofacial variation in modern AIs in the region. Given the similar gene pool of AI and Hispanic groups in the SW, the population structure of NM may be quite unique and require a different approach for population affinity estimates. We support the employment of an anthropological perspective in forensic anthropology and population affinity assessment. We stress the importance of understanding population structure and population history in an area in which forensic anthropologists practice and being aware of the limits of methodology when applied to the local population. The sample collected here is available in the MaMD and could be useful for distinguishing among modern AI and non-AI groups in other parts of the U.S.