Age Estimation

A confidence interval tells how confident you are that the mean of your data will fall in between the interval, while a prediction interval shows the confidence that a future prediction will fall between those confidence intervals. The given values for both of these intervals would be 95%, you are 95% confident that data will fall between the values and 95% confident that a future prediction will too. 

Some of the problems with age estimation currently are the fact that we don’t actually have any methods aside from the new transition analyses method, every other “method” used are just descriptive statistics for the reference sample. These so called standard deviation methods are based on the average age in which certain traits of development or degeneration are present or absent in a population, which is why these methods are actually just descriptive statistics of the sample that the method analyzed. These standard deviation methods also assume a normal distribution, which is not reflective of the actual population. They are unable to take into consideration any noise or error, and are unable to make predictions about the ability of the model to predict a future unknown. 

These commonly used adult age indicators are very weakly correlated to chronological age. In addition, when statistics are used in these so-called methods, they are often linear regression models which are not appropriate for age estimation. These regression models tend to overestimate the age of younger individuals and underestimate the age of older individuals, in addition to contributing to age mimicry in which the age distribution of your unknown will tend to mimic the age distribution of the reference sample. In addition, these approaches produce prediction intervals that are often many decades wide if not terminal and open-ended for older adults, which although technically accurate, are overall uninformative. 

However, these predictive regression models are able to provide error ranges along with prediction intervals, meaning that we can know how well the model performs in estimating a future unknown. These methods for model testing include bootstrapping, k-nearest neighbor and leave-one-out cross validation.

Today we still see a dependency on stage based approaches, which is when a certain anatomical component is scored in its entirety. This includes the Suchey-Brooks pubic symphysis, the Lovejoy auricular surface, and Hartnett’s sternal rib ends. These methods are poor because they only use a single indicator and they assume a regularity of the aging process which doesn’t take into account the considerable variability, such as occupation or health. 

Component based models follow the same vein as phase based approaches, however instead of scoring the entire anatomic unit, the unit is broken down into components which are then scored individually. Component based approaches include McKern and Stewart pubic symphysis and Langley. Once again, as these are still focusing on one anatomical unit, they are still dependent on a single age indicator. 

Another problem with commonly used age estimation methods is the inability to combine scores from multiple different traits of different skeletal regions into a meaningful age estimation. To combat this, the overlap method is used. In this method the mean age from multiple different scoring systems is obtained, and the age range for which these means “overlap” the most is the age range chosen by the practitioner. The problem with this method is the subjectivity of this age interval, as it is simply chosen by the practitioner with no statistical backing, uncertainty, or confidence interval. 

There is a push to move age estimations towards multivariable analyses using either a bayesian or machine learning approach. These approaches have been shown to improve accuracy both from a performance metric standpoint and an individual standpoint. These models can also take into account the bias, that is the over or underestimation of an age, and contribute to greater precision or narrowness of age intervals. The overall conclusion being reached in newer research is that no single skeletal indicator is capable of producing accurate and efficient age estimates across the entire human age span. These multivariable analyses and approaches are needed to truly improve the standards of age estimation methods within the field. 

One such new method is transition analysis. Transition analyses are statistical algorithms that combine information from multiple probability curves into a single predictive interval based on conditional independence. 

Transition analysis is also a method used in age estimation, this method is based on the concept of transition, which is the age at which a certain trait is equally likely to be present or absent. In this method, a suite of skeletal traits are looked at and scored ordinally. You receive an output that is the maximum likelihood point estimate in which the age interval can be calculated from the probability distance for age-at-death. This method is represented by logarithmic (s-shaped) curves that are as wide as the age range. Therefore, a narrow nerve represents a narrower age range and vice versa. The steepness of these curves also gives information about how quickly the transition period takes, with a steeper curve being a quicker transition versus a less steep curve indicating that transition occurs slowly. A steep curve is more informative, as it gives us this narrower age range and tells us that if that trait is in transition, it is indicative of that narrow age range. With a less steep curve, the curve spans a broader area and thus is more uninformative in telling us that the transition phase can occur across a number of different ages. Most developmental traits have steeper curves while degenerative traits tend to be more broad. 

Another new and useful method different from transition analysis is based in bayesian statistics. In this approach, we have an initial hypothesis (prior), and data is collected that either supports or disproves this hypothesis. In Bayesian statistics, the outcome we want is a probability distribution representing the age of the unknown decedent. We arrive at this posterior probability using age indicators of the skeleton (known data), multiplied by the previous knowledge on the age of forensic decedents in the United States, divided by the total frequency of that age indicator being present across all age ranges. An uninformative prior in this example would be every single individual in the United States having an equal probability of entering the forensic record, whereas an informative prior would be that subadults are more likely to enter the forensic record at specific age periods. It is important to remember with these prior distributions that depending on the context in which you are estimating age, the prior will change. A forensic case should not be based on a prior obtained from archaeological data and vice versa. 

Finally, one last method is the DRNNAGE method that uses neural networks, or biologically inspired computational models that enable learning. This method allows for the combination of data from multiple traits across the skeleton in addition to being able to handle complex data that is either missing or noisy. 

What are some of these new traits that are being looked at?? These new traits include textures, shapes and combinations of developing and deteriorating skeletal elements. For example, Dr. Getz, a practicing forensic anthropologist for the Oklahoma’s medical examiner’s program said that she looks to the proximal humerus, as the lesser tubercle is bubbly in appearance in teenagers, sharper in the 30s, and becomes osteoarthritic in the 60s. She also looks at the femur for things like lipping of the fovea capitis or the texture of the greater trochanteric fossa. You can also look at degenerative aspects of the spine, however it has been discovered that certain pathological conditions such as DISH are not good age indicators and in fact obscure age indicators for the rest of the skeleton. Instead, DISH is pathognomonic of other disease processes. 

In age estimation, variables tend to be correlated. When you consider variables as dependent on each other, you get a shallower but wider prediction distribution that is less precise but more accurate. When you assume conditional independence, you get a distribution that is overly optimistic and overly precise, which can lead to lower accuracy.