We thank Haas and Kelly (Reference Randall and Kelly2026) for their comment on our article (Buchanan, Hamilton, and Walker Reference Buchanan, Hamilton and Walker2026) and for identifying potential problems with dart and arrow classification in the Great Basin. In their comment, Haas and Kelly (Reference Randall and Kelly2026) use a logistic model that they claim outperforms the model we presented for North America; however, their model is specifically for the Great Basin. There are myriad ways to construct statistical models, but Haas and Kelly (Reference Randall and Kelly2026) do not provide evidence that their model outperforms ours. In fact, their model performs worse according to our cross-validation testing. Moreover, they claim that our model does not account for normality or collinearity in our data. As for normality assumptions, the residuals of our model are well behaved, as we noted in our original article. Although Haas and Kelly log-transform width, this does not improve cross-validation performance and is unnecessary because the assumption of normality applies to the residuals of the model, not the predictor variables. As for collinearity, the variance inflation factors for our model are not high, ranging from 1.1 to 1.5. In addition, our focus is less on the interpretation of individual predictor variables and more on improving predictive performance, as measured by cross-validation, which, as we stated, is improved with the variables we have chosen. However, despite these misunderstandings, we agree with Haas and Kelly and think that a regional classification model for the Great Basin should be developed. Here, we improve on Haas and Kelly’s (Reference Randall and Kelly2026) modeling approach by offering a new model for the Great Basin. It is one that incorporates archaeological and ethnographic data, a necessity for creating a Great Basin–specific classification model. Our new model uses a large dataset that includes training and test specimens, which allows the performance of our model, and subsequent revisions, to be evaluated.
Classifying stone dart and arrow points in the Great Basin is a particularly difficult problem because of the predominance of surface assemblages and undated sites. In our new classification for the Great Basin presented here, we rely on types—as do Haas and Kelly (Reference Randall and Kelly2026). However, we emphasize that we do not think that any given type is necessarily congruent with functional changes in weapon systems. Moreover, we use types here out of necessity to place point specimens in chronological sequence, and we note that our Bayesian models provide uncertainty measures for the classification of each specimen independent of type. The use of types entails risk from poorly defined types, the inclusion of misclassified specimens, and the use of types with poor chronological control (see Schroedl Reference Schroedl2024, Reference Schroedl2025; Smith and Rosencrance Reference Smith and Rosencrance2025); however, here we are interested in broad trends that allow for these uncertainties.
Based on our recently published study of directly dated atlatl and bow specimens in western North America (Buchanan, Hamilton, et al. Reference Buchanan, Hamilton, Eren and Walker2026), we hypothesize that the transition from atlatl to bow occurred around 1400 cal BP, which agrees with Haas and Kelly’s (Reference Randall and Kelly2026) estimate for the Great Basin. To build our model, we use one type that is considerably earlier and one type that is later than the hypothesized transition to the bow in the Great Basin. Buchanan, Hamilton, et alia’s (Reference Buchanan, Hamilton, Eren and Walker2026) dataset included one hafted stone dart point from the Great Basin. The Gatecliff Split Stem point (see Smith et al. Reference Smith, Barker, Hattori, Raymond and Goebel2013) was from Kramer Cave in Nevada and radiocarbon dated to 3,830 ± 110, about 4,230 calibrated years ago (Hattori Reference Hattori1982). Therefore, we reasoned that we could assign typed Gatecliff Split Stem points to the dart category. The generally accepted date range for the Gatecliff Split Stem type in the Great Basin is about 5,000 to 3,250 years ago (Smith et al. Reference Smith, Barker, Hattori, Raymond and Goebel2013; Thomas Reference Thomas1981, Reference Thomas2024). Similarly, we reasoned that the hafted Desert Side-Notched arrow points from the Powell Collection justify the inclusion of additional specimens of this type into the training model as arrows. The Desert Side-Notched point type first appears around 1,000 years ago in Oregon and is dated between 900 and 200 years ago in different regions of the Great Basin (Smith et al. Reference Smith, Barker, Hattori, Raymond and Goebel2013). By including a sample of both Gatecliff Split Stem and Desert Side-Notched archaeological specimens as known dart and arrow points, respectively, into our model, we run the risk of using a sample of dart and arrow points in our training data that are not definitively known as either darts or arrows, but this can be further tested as more data become available.
We obtained additional data to train and test our new model from Thomas’s (Reference Thomas2020, Reference Thomas2024) database of Great Basin points, which is compiled from various research projects. This database includes more than 7,000 measured and typed points. Excluding incomplete specimens, we created a subset of 3,951 complete points from 62 named types, including 360 Desert Side-Notched and 268 Gatecliff Split Stem points. From these, we sampled specimens to be included in the training model, with the remaining points to be included as test points. Our sampling procedure allowed us to create a balanced dataset of darts and arrows; consequently, we set the classification cutoff set at 0.5. However, we should note that the down-sampling procedure used by Haas and Kelly (Reference Randall and Kelly2026) has the same effect as the procedure used by Buchanan, Hamilton, and Walker (Reference Buchanan, Hamilton and Walker2026), which adjusts the decision cutoff using an iterative approach that balances sensitivity and specificity. Here, using a larger dataset, we can generate a balanced sample of 100 darts and 100 arrows. To do this, we randomly sampled 23 Desert Side-Notched arrow points and 99 Gatecliff Split Stem darts from the Thomas (Reference Thomas2020, Reference Thomas2024) database and added these to the known sample of 77 Desert Side-Notched arrows and a single Gatecliff Split Stem dart point.
Using cross-validation as our model selection technique, we compared the effectiveness of a classification model with the new Great Basin sample in addition to the overall North American dataset used by Buchanan, Hamilton, and Walker (Reference Buchanan, Hamilton and Walker2026) to a classification model using only the Great Basin regional sample of 100 darts and 100 arrows. We consistently found that the exclusively Great Basin sample exhibits a distinct regional signature that outperforms the model that includes all North American data. For this reason, we focus only on the Great Basin regional sample in our subsequent model development.
We used the same classification modeling procedures outlined by Buchanan, Hamilton, and Walker (Reference Buchanan, Hamilton and Walker2026), with slight modifications to fit the new data. We created a balanced dataset and therefore set the classification cutoff to 0.5. Next, we sought the optimal set of variables that maximized cross-validated Area Under the Curve (AUC) using a logistic regression model. The large dataset of measured points from Thomas (Reference Thomas2020, Reference Thomas2024) includes length, width, basal width, neck width, distal shoulder angle, proximal shoulder angle, and thickness for complete or nearly complete points. Weight was not reported for most of the sample and was therefore not included. We also considered all two-way (area measurements) and three-way (volume) interactions, but we found that these did not to contribute to the cross-validated performance. To find the best model, we used 10-fold cross-validation, each with five repeats, in the caret package of R (version 4.5.2). Width and thickness were found to yield the highest cross-validated AUC of 0.936. A Generalized Additive Model that allows for the subtle nonlinearities that emerge in the relationships improves the cross-validated AUC to 0.951. The addition of length, neck width, and basal width were found to improve the model slightly, but given their U-shaped relationships, we considered their addition as overfitting.
Our final Bayesian model for prediction is a logistic regression of dart versus arrow with spline smoothed terms for width and thickness. This model was implemented in R with the brms package (Bürkner Reference Bürkner2017, Reference Bürkner2018). Following our previous article, and the Generalized Additive Model results above, we found a slight nonlinear decision boundary between darts and arrows. The nonlinearities are subtle, so we used low basis dimensions (k = 3) for the spline smooths in the model. The model passes a posterior predictive check, and the residuals are well behaved. The mean of the Bayesian r 2 distribution is 0.61. The nonlinear gradient defining the probability assignment of darts and arrows is shown in Figure 1. Similar to the North American dart-arrow classification model of Buchanan, Hamilton, and Walker (Reference Buchanan, Hamilton and Walker2026), we make data and code available, and the new model is available to use through an online application (https://github.com/RobertSWalker/darts_v_arrows_gb).
Figure 1.The best variables for classifying darts versus arrows in the Great Basin are width and thickness. Our model with smoothed splines produces a slightly nonlinear decision boundary (curved line, cutoff = 0.5). (Color online)
Figure 1 Long description
A scatter plot with the horizontal axis labeled Width (millimeter) and tick labels at 10, 15, 20, 25 and 30. The vertical axis is labeled Thickness (millimeter) with tick labels at 0.0, 2.5, 5.0 and 7.5. A slightly curved diagonal line runs from the upper left area of the plotting region toward the lower right area. Two point categories are shown in the legend: arrow and dart. Points labeled arrow are concentrated at lower widths and lower thicknesses, with many points around width 10 to 16 and thickness about 1.5 to 4.5. Points labeled dart are concentrated at higher widths and higher thicknesses, with many points around width 18 to 28 and thickness about 4.5 to 7.5. Several points from both categories appear in the middle region around width 15 to 20 and thickness about 3.5 to 5.5. A vertical color scale bar appears to the right of the plot with the word dart at the top and arrow at the bottom.
To test the new model, we analyzed 3,829 of typed Great Basin points with associated date ranges as our test sample. We focus on six series: Desert (n = 658), Rosegate (n = 680), Elko (n = 608), Gatecliff (n = 778), Great Basin Stemmed (n = 595), and Humboldt (n = 240). We focus on these series because they have large samples and generally agreed-upon date ranges that allow us to estimate changes in dart and arrow proportions over time. We use the date ranges for these point series primarily from Thomas (Reference Thomas1981, Reference Schroedl2024) and Smith et alia (Reference Smith, Barker, Hattori, Raymond and Goebel2013).
In Figure 2, the expected probabilities of being a dart derived from our Great Basin model are plotted by the median age of the point series, with associated age ranges shown with horizontal error bars. The series we predicted based on time range to be arrows, the Desert Series, is mostly arrows (97%). The series we predicted to be darts, Gatecliff and Great Basin Stemmed, include a majority of points classified as darts (91% and 99%). The Humboldt Series also has mostly dart predictions (71%). The Rosegate (78% arrows) and Elko Series (89% darts), both having date ranges that span the proposed first appearance of the bow and arrow, have distributions that are skewed toward arrows for Rosegate and skewed toward darts for Elko.
Figure 2.Expected probabilities for Great Basin points from archaeological series. Swarm plot of individual points is shown along with the maximum age range for each series using horizontal error bars. The dashed vertical line is at 1400 cal BP, the postulated first appearance of the bow and arrow in western North America based on directly dated specimens. (Color online)
Figure 2 Long description
Predicted probability of dart A swarm plot with multiple series shown as vertical swarms of points and horizontal error bars. The x-axis is labeled Calibrated years BP, ranging from 0 to 11000, with labeled ticks at 0, 1000, 2000, 3000, 4000, 5000, 6000, 7000, 8000, 9000, 10000 and 11000. The y-axis is labeled Predicted probability of dart, ranging from 0.00 to 1.00, with labeled ticks at 0.00, 0.25, 0.50, 0.75 and 1.00. A dashed vertical line is drawn at x equals 1400. A horizontal line is drawn at y equals 0.50. Legend on the right is titled Series and lists: Desert, Elko, Gatecliff, Great Basin, Humboldt, Rosegate. Desert: A small cluster of points near y close to 0.00 at x near 0. A horizontal error bar spans roughly from x near 0 to x near 1000 at y near 0.00. Rosegate: A vertical swarm centered near x around 1200 to 1600, with points spanning from y near 0.00 up to y near 1.00. A horizontal error bar is centered near x around 1400 with endpoints roughly near x around 900 and x around 2000 at y near 0.25. Gatecliff: A vertical swarm centered near x around 2500 to 3000, with most points near y close to 1.00 and a narrow spread downward. A horizontal error bar near y close to 1.00 spans roughly from x around 2000 to x around 3500. Great Basin: A vertical swarm centered near x around 3500 to 4200, with most points near y close to 1.00 and a narrow spread downward. A horizontal error bar near y close to 1.00 spans roughly from x around 3000 to x around 5000. Elko: A vertical swarm centered near x around 3800 to 4500, with points concentrated near y close to 1.00. A horizontal error bar near y close to 0.75 spans roughly from x around 2500 to x around 6000. Humboldt: A vertical swarm centered near x around 9000, with points concentrated near y close to 1.00 and a dotted vertical line extending downward from near y close to 1.00 toward y around 0.60. A horizontal error bar near y close to 1.00 spans roughly from x around 7000 to x around 11000.
Our new Great Basin regional model identified width and thickness as the most important variables to distinguish between dart and arrow points. This contrasts with the overall North American model, which found length and width to be the best variables (Buchanan, Hamilton, and Walker Reference Buchanan, Hamilton and Walker2026). For the Great Basin sample, the inclusion of length slightly improves the model; however, the relationship is U-shaped, suggesting overfit of this variable. This means that the probability of a point being a dart increases for relatively longer and shorter specimens within the width and thickness tolerances for darts. This could be the result of original raw material or blank size variation, the impact of resharpening, or typological issues. Interestingly—and a topic for future research—this is not the case for the overall North American model of dart and arrow classification that identifies nonlinear point length and width as the important variables in discriminating between darts and arrows (Buchanan, Hamilton, and Walker Reference Buchanan, Hamilton and Walker2026). We recommend, when classifying points found outside of the Great Basin (or Southwest), that researchers use a nonregional overall model such as the one presented in Buchanan, Hamilton, and Walker (Reference Buchanan, Hamilton and Walker2026). We now have three models for classifying darts and arrows—the Great Basin, Southwest, and North America—and we hope that with more data, additional regional models can be constructed in the future. Although our models are more complicated than simply using cutoffs of length, width, or thickness (depending on the region), they perform better. We fully agree with Thomas’s (Reference Thomas1980:344) sentiment that “given a choice, archaeologists must reject statistics whenever simpler ones will work as well,” but our more complicated (but easy-to-use) models outperform these other options. The goal in classification modeling is performance—to find the best model, not the simplest model.
You have
Access
We thank Haas and Kelly (Reference Randall and Kelly2026) for their comment on our article (Buchanan, Hamilton, and Walker Reference Buchanan, Hamilton and Walker2026) and for identifying potential problems with dart and arrow classification in the Great Basin. In their comment, Haas and Kelly (Reference Randall and Kelly2026) use a logistic model that they claim outperforms the model we presented for North America; however, their model is specifically for the Great Basin. There are myriad ways to construct statistical models, but Haas and Kelly (Reference Randall and Kelly2026) do not provide evidence that their model outperforms ours. In fact, their model performs worse according to our cross-validation testing. Moreover, they claim that our model does not account for normality or collinearity in our data. As for normality assumptions, the residuals of our model are well behaved, as we noted in our original article. Although Haas and Kelly log-transform width, this does not improve cross-validation performance and is unnecessary because the assumption of normality applies to the residuals of the model, not the predictor variables. As for collinearity, the variance inflation factors for our model are not high, ranging from 1.1 to 1.5. In addition, our focus is less on the interpretation of individual predictor variables and more on improving predictive performance, as measured by cross-validation, which, as we stated, is improved with the variables we have chosen. However, despite these misunderstandings, we agree with Haas and Kelly and think that a regional classification model for the Great Basin should be developed. Here, we improve on Haas and Kelly’s (Reference Randall and Kelly2026) modeling approach by offering a new model for the Great Basin. It is one that incorporates archaeological and ethnographic data, a necessity for creating a Great Basin–specific classification model. Our new model uses a large dataset that includes training and test specimens, which allows the performance of our model, and subsequent revisions, to be evaluated.
Classifying stone dart and arrow points in the Great Basin is a particularly difficult problem because of the predominance of surface assemblages and undated sites. In our new classification for the Great Basin presented here, we rely on types—as do Haas and Kelly (Reference Randall and Kelly2026). However, we emphasize that we do not think that any given type is necessarily congruent with functional changes in weapon systems. Moreover, we use types here out of necessity to place point specimens in chronological sequence, and we note that our Bayesian models provide uncertainty measures for the classification of each specimen independent of type. The use of types entails risk from poorly defined types, the inclusion of misclassified specimens, and the use of types with poor chronological control (see Schroedl Reference Schroedl2024, Reference Schroedl2025; Smith and Rosencrance Reference Smith and Rosencrance2025); however, here we are interested in broad trends that allow for these uncertainties.
Based on our recently published study of directly dated atlatl and bow specimens in western North America (Buchanan, Hamilton, et al. Reference Buchanan, Hamilton, Eren and Walker2026), we hypothesize that the transition from atlatl to bow occurred around 1400 cal BP, which agrees with Haas and Kelly’s (Reference Randall and Kelly2026) estimate for the Great Basin. To build our model, we use one type that is considerably earlier and one type that is later than the hypothesized transition to the bow in the Great Basin. Buchanan, Hamilton, et alia’s (Reference Buchanan, Hamilton, Eren and Walker2026) dataset included one hafted stone dart point from the Great Basin. The Gatecliff Split Stem point (see Smith et al. Reference Smith, Barker, Hattori, Raymond and Goebel2013) was from Kramer Cave in Nevada and radiocarbon dated to 3,830 ± 110, about 4,230 calibrated years ago (Hattori Reference Hattori1982). Therefore, we reasoned that we could assign typed Gatecliff Split Stem points to the dart category. The generally accepted date range for the Gatecliff Split Stem type in the Great Basin is about 5,000 to 3,250 years ago (Smith et al. Reference Smith, Barker, Hattori, Raymond and Goebel2013; Thomas Reference Thomas1981, Reference Thomas2024). Similarly, we reasoned that the hafted Desert Side-Notched arrow points from the Powell Collection justify the inclusion of additional specimens of this type into the training model as arrows. The Desert Side-Notched point type first appears around 1,000 years ago in Oregon and is dated between 900 and 200 years ago in different regions of the Great Basin (Smith et al. Reference Smith, Barker, Hattori, Raymond and Goebel2013). By including a sample of both Gatecliff Split Stem and Desert Side-Notched archaeological specimens as known dart and arrow points, respectively, into our model, we run the risk of using a sample of dart and arrow points in our training data that are not definitively known as either darts or arrows, but this can be further tested as more data become available.
We obtained additional data to train and test our new model from Thomas’s (Reference Thomas2020, Reference Thomas2024) database of Great Basin points, which is compiled from various research projects. This database includes more than 7,000 measured and typed points. Excluding incomplete specimens, we created a subset of 3,951 complete points from 62 named types, including 360 Desert Side-Notched and 268 Gatecliff Split Stem points. From these, we sampled specimens to be included in the training model, with the remaining points to be included as test points. Our sampling procedure allowed us to create a balanced dataset of darts and arrows; consequently, we set the classification cutoff set at 0.5. However, we should note that the down-sampling procedure used by Haas and Kelly (Reference Randall and Kelly2026) has the same effect as the procedure used by Buchanan, Hamilton, and Walker (Reference Buchanan, Hamilton and Walker2026), which adjusts the decision cutoff using an iterative approach that balances sensitivity and specificity. Here, using a larger dataset, we can generate a balanced sample of 100 darts and 100 arrows. To do this, we randomly sampled 23 Desert Side-Notched arrow points and 99 Gatecliff Split Stem darts from the Thomas (Reference Thomas2020, Reference Thomas2024) database and added these to the known sample of 77 Desert Side-Notched arrows and a single Gatecliff Split Stem dart point.
Using cross-validation as our model selection technique, we compared the effectiveness of a classification model with the new Great Basin sample in addition to the overall North American dataset used by Buchanan, Hamilton, and Walker (Reference Buchanan, Hamilton and Walker2026) to a classification model using only the Great Basin regional sample of 100 darts and 100 arrows. We consistently found that the exclusively Great Basin sample exhibits a distinct regional signature that outperforms the model that includes all North American data. For this reason, we focus only on the Great Basin regional sample in our subsequent model development.
We used the same classification modeling procedures outlined by Buchanan, Hamilton, and Walker (Reference Buchanan, Hamilton and Walker2026), with slight modifications to fit the new data. We created a balanced dataset and therefore set the classification cutoff to 0.5. Next, we sought the optimal set of variables that maximized cross-validated Area Under the Curve (AUC) using a logistic regression model. The large dataset of measured points from Thomas (Reference Thomas2020, Reference Thomas2024) includes length, width, basal width, neck width, distal shoulder angle, proximal shoulder angle, and thickness for complete or nearly complete points. Weight was not reported for most of the sample and was therefore not included. We also considered all two-way (area measurements) and three-way (volume) interactions, but we found that these did not to contribute to the cross-validated performance. To find the best model, we used 10-fold cross-validation, each with five repeats, in the caret package of R (version 4.5.2). Width and thickness were found to yield the highest cross-validated AUC of 0.936. A Generalized Additive Model that allows for the subtle nonlinearities that emerge in the relationships improves the cross-validated AUC to 0.951. The addition of length, neck width, and basal width were found to improve the model slightly, but given their U-shaped relationships, we considered their addition as overfitting.
Our final Bayesian model for prediction is a logistic regression of dart versus arrow with spline smoothed terms for width and thickness. This model was implemented in R with the brms package (Bürkner Reference Bürkner2017, Reference Bürkner2018). Following our previous article, and the Generalized Additive Model results above, we found a slight nonlinear decision boundary between darts and arrows. The nonlinearities are subtle, so we used low basis dimensions (k = 3) for the spline smooths in the model. The model passes a posterior predictive check, and the residuals are well behaved. The mean of the Bayesian r 2 distribution is 0.61. The nonlinear gradient defining the probability assignment of darts and arrows is shown in Figure 1. Similar to the North American dart-arrow classification model of Buchanan, Hamilton, and Walker (Reference Buchanan, Hamilton and Walker2026), we make data and code available, and the new model is available to use through an online application (https://github.com/RobertSWalker/darts_v_arrows_gb).
The best variables for classifying darts versus arrows in the Great Basin are width and thickness. Our model with smoothed splines produces a slightly nonlinear decision boundary (curved line, cutoff = 0.5). (Color online)
Figure 1 Long description
A scatter plot with the horizontal axis labeled Width (millimeter) and tick labels at 10, 15, 20, 25 and 30. The vertical axis is labeled Thickness (millimeter) with tick labels at 0.0, 2.5, 5.0 and 7.5. A slightly curved diagonal line runs from the upper left area of the plotting region toward the lower right area. Two point categories are shown in the legend: arrow and dart. Points labeled arrow are concentrated at lower widths and lower thicknesses, with many points around width 10 to 16 and thickness about 1.5 to 4.5. Points labeled dart are concentrated at higher widths and higher thicknesses, with many points around width 18 to 28 and thickness about 4.5 to 7.5. Several points from both categories appear in the middle region around width 15 to 20 and thickness about 3.5 to 5.5. A vertical color scale bar appears to the right of the plot with the word dart at the top and arrow at the bottom.
To test the new model, we analyzed 3,829 of typed Great Basin points with associated date ranges as our test sample. We focus on six series: Desert (n = 658), Rosegate (n = 680), Elko (n = 608), Gatecliff (n = 778), Great Basin Stemmed (n = 595), and Humboldt (n = 240). We focus on these series because they have large samples and generally agreed-upon date ranges that allow us to estimate changes in dart and arrow proportions over time. We use the date ranges for these point series primarily from Thomas (Reference Thomas1981, Reference Schroedl2024) and Smith et alia (Reference Smith, Barker, Hattori, Raymond and Goebel2013).
In Figure 2, the expected probabilities of being a dart derived from our Great Basin model are plotted by the median age of the point series, with associated age ranges shown with horizontal error bars. The series we predicted based on time range to be arrows, the Desert Series, is mostly arrows (97%). The series we predicted to be darts, Gatecliff and Great Basin Stemmed, include a majority of points classified as darts (91% and 99%). The Humboldt Series also has mostly dart predictions (71%). The Rosegate (78% arrows) and Elko Series (89% darts), both having date ranges that span the proposed first appearance of the bow and arrow, have distributions that are skewed toward arrows for Rosegate and skewed toward darts for Elko.
Expected probabilities for Great Basin points from archaeological series. Swarm plot of individual points is shown along with the maximum age range for each series using horizontal error bars. The dashed vertical line is at 1400 cal BP, the postulated first appearance of the bow and arrow in western North America based on directly dated specimens. (Color online)
Figure 2 Long description
Predicted probability of dart A swarm plot with multiple series shown as vertical swarms of points and horizontal error bars. The x-axis is labeled Calibrated years BP, ranging from 0 to 11000, with labeled ticks at 0, 1000, 2000, 3000, 4000, 5000, 6000, 7000, 8000, 9000, 10000 and 11000. The y-axis is labeled Predicted probability of dart, ranging from 0.00 to 1.00, with labeled ticks at 0.00, 0.25, 0.50, 0.75 and 1.00. A dashed vertical line is drawn at x equals 1400. A horizontal line is drawn at y equals 0.50. Legend on the right is titled Series and lists: Desert, Elko, Gatecliff, Great Basin, Humboldt, Rosegate. Desert: A small cluster of points near y close to 0.00 at x near 0. A horizontal error bar spans roughly from x near 0 to x near 1000 at y near 0.00. Rosegate: A vertical swarm centered near x around 1200 to 1600, with points spanning from y near 0.00 up to y near 1.00. A horizontal error bar is centered near x around 1400 with endpoints roughly near x around 900 and x around 2000 at y near 0.25. Gatecliff: A vertical swarm centered near x around 2500 to 3000, with most points near y close to 1.00 and a narrow spread downward. A horizontal error bar near y close to 1.00 spans roughly from x around 2000 to x around 3500. Great Basin: A vertical swarm centered near x around 3500 to 4200, with most points near y close to 1.00 and a narrow spread downward. A horizontal error bar near y close to 1.00 spans roughly from x around 3000 to x around 5000. Elko: A vertical swarm centered near x around 3800 to 4500, with points concentrated near y close to 1.00. A horizontal error bar near y close to 0.75 spans roughly from x around 2500 to x around 6000. Humboldt: A vertical swarm centered near x around 9000, with points concentrated near y close to 1.00 and a dotted vertical line extending downward from near y close to 1.00 toward y around 0.60. A horizontal error bar near y close to 1.00 spans roughly from x around 7000 to x around 11000.
Our new Great Basin regional model identified width and thickness as the most important variables to distinguish between dart and arrow points. This contrasts with the overall North American model, which found length and width to be the best variables (Buchanan, Hamilton, and Walker Reference Buchanan, Hamilton and Walker2026). For the Great Basin sample, the inclusion of length slightly improves the model; however, the relationship is U-shaped, suggesting overfit of this variable. This means that the probability of a point being a dart increases for relatively longer and shorter specimens within the width and thickness tolerances for darts. This could be the result of original raw material or blank size variation, the impact of resharpening, or typological issues. Interestingly—and a topic for future research—this is not the case for the overall North American model of dart and arrow classification that identifies nonlinear point length and width as the important variables in discriminating between darts and arrows (Buchanan, Hamilton, and Walker Reference Buchanan, Hamilton and Walker2026). We recommend, when classifying points found outside of the Great Basin (or Southwest), that researchers use a nonregional overall model such as the one presented in Buchanan, Hamilton, and Walker (Reference Buchanan, Hamilton and Walker2026). We now have three models for classifying darts and arrows—the Great Basin, Southwest, and North America—and we hope that with more data, additional regional models can be constructed in the future. Although our models are more complicated than simply using cutoffs of length, width, or thickness (depending on the region), they perform better. We fully agree with Thomas’s (Reference Thomas1980:344) sentiment that “given a choice, archaeologists must reject statistics whenever simpler ones will work as well,” but our more complicated (but easy-to-use) models outperform these other options. The goal in classification modeling is performance—to find the best model, not the simplest model.
Acknowledgments
No permit was required for this research.
Funding Statement
This research received no specific grant funding from any funding agency or from commercial or not-for-profit sectors.
Data Availability Statement
All data and code used in this article are available at https://github.com/RobertSWalker/darts_v_arrows_gb .
Competing Interests
The authors declare none.