*7.2. Assumptions and Limitations*

The key limitation of this work is the small size of the data set. There is limited statistical power associated with small data sets. Of note here is the just statistically significant result of the odds of a fatality based on aircraft age. A larger dataset would help to either confirm of disprove this. Similarly, the data set size has influenced the trend in percentage of maintenance accidents over time, with significant 'noise' in the last three years.

It could also be argued that the lack of information about incidents, which are far more common than accidents, is a limitation. It should, however, be noted that accidents result in significant damage or injury, even hull loss or death, unlike incidents. As such, research in accidents is arguably more important in aviation safety.

The use of uniform expected distributions for some of the categorical variables is also a limitation. For the maintenance issue and the systems/components involved the goal is to simply assess if one of these is statistically speaking more likely than the others. In contrast, for the operator (business model) and operation (type of service), these would benefit from a non-uniform expected distribution. With time and e ffort, these codes could be created for all 1277 accidents in the ICAO o fficial dataset. The additional insight gained from this could be useful or limited.

It would be ideal to analyze the dataset looking for covariances between the categorical variables considered. This would ideally help identify latent classes in the data (combinations of variable values that are more likely to occur together, and hence present a greater safety risk, e.g., an Ilyushin cargo aircraft in the Middle East). However, the limited number of accidents with maintenance contributions means that performing cross tabulations for these would result in such small counts that the associated Fisher's exact test would likely yield no statistically significant results.
