The authors thank the commenters for their interest. They are pleased to note that the authors’ study (Yoganandan et al., 2016) addressed some of the concerns documented in the paper by the commenters (McMurry and Poplin, 2015), although the original aim of the authors’ study was to improve and provide a more robust and generalized process to derive injury risk curves than the currently existing ISO method (ISO, 2014). In any PMHS experimental data, as with most real-world outputs, the true data-generating distribution will always be unknown (Hastie et al., 2001; Vapnik, 1998). The authors believe that assuming a particular distributional form holds for all types of impact biomechanics data, ignoring all other competing or equivalent functional forms without investigation or validation, is quite a stretch. It is also worth noting and well known that simulations have their limitations (Hastie et al., 2001). For example, simulated data for survival times require assumption of a probability distribution function (which could be Weibull), assumption of a censoring distribution (which could be something quite arbitrary, such as a Uniform distribution), and assumption of nature of dependence of the censoring on the survival distribution [often one assumes censoring independent of the survival distribution] (Therneau and Grambsch, 2000). Thus, simulated survival data sets need not reflect the actual survival time generating distribution after imposition of censoring, nor should it be expected that the assumptions for a simulated scenario mimic all possible real world conditions (Therneau and Grambsch, 2000). The AIC is a well-accepted metric for choosing an appropriate model in usual regression analysis and for parametric survival models (Harrell, 2001; Hurvich and Chih0Ling, 1989; Klein and Moeschberger, 2003). As with most statistical procedures, the efficiency of the AIC improves with increase in sample size. However, the AIC is still a principled way of choosing a particular functional form, rather than guessing an ad-hoc single candidate form for observed data. It is also worth noting that many candidate probability distribution functions have similar risk profiles, noted by the commenters in their work. Biomechanical considerations and the experts/experimenter's perspectives are definitely important considerations in the determination of the risk curve. As stated in the authors’ paper (Yoganandan et al., 2016), the judgment of the “experimental group and subject matter expert” should help decide overly-influential/outlier observations and also the final metric depending on the application, e.g., injury assessment risk curve for a dummy. The
[1]
B. Turnbull.
The Empirical Distribution Function with Arbitrarily Grouped, Censored, and Truncated Data
,
1976
.
[2]
Tze Leung Lai,et al.
For censored and truncated data
,
1999
.
[3]
Lucila Ohno-Machado,et al.
The use of receiver operating characteristic curves in biomedical informatics
,
2005,
J. Biomed. Informatics.
[4]
Robert Tibshirani,et al.
The Elements of Statistical Learning: Data Mining, Inference, and Prediction, 2nd Edition
,
2001,
Springer Series in Statistics.
[5]
Timothy L McMurry,et al.
Statistical Considerations in the Development of Injury Risk Functions
,
2015,
Traffic injury prevention.
[6]
Clifford M. Hurvich,et al.
Regression and time series model selection in small samples
,
1989
.
[7]
J. Klein,et al.
Survival Analysis: Techniques for Censored and Truncated Data
,
1997
.
[8]
Stanley Lemeshow,et al.
Applied Logistic Regression, Second Edition
,
1989
.
[9]
Rolf H Eppinger,et al.
Development of Side Impact Thoracic Injury Criteria and Their Application to the Modified ES-2 Dummy with Rib Extensions (ES-2re).
,
2003,
Stapp car crash journal.
[10]
P. Grambsch,et al.
Modeling Survival Data: Extending the Cox Model
,
2000
.
[11]
F Scott Gayzik,et al.
Deriving injury risk curves using survival analysis from biomechanical experiments.
,
2016,
Journal of biomechanics.
[12]
John M. Cavanaugh,et al.
Thorax Injury Biomechanics
,
2015
.