U. S. Department of Transportation
Download 2.66 Mb.
|
- Navigate this page:
- Merging Disparate Datasets
- Event-Specific Data
- Weather Data
- Summary Statistics
Merged Data SetThe above datasets were aggregated through a variety of processes that resulted in one overall dataset. The processes used to combine datasets fall into two major categories: event-specific data and more general data. The event-specific data are contained in the Runway Incursion database and the two ATQA databases. The more general information constitutes the airport characteristics, and operations data. The weather data required special treatment before they could be combined with the Runway Incursion database.
General InformationMatching the more general data to the Runway Incursion database was simple. Using the incident location (airport code), date, and time the general data could be easily matched. There is no need to differentiate between multiple incidents at the same airport for certain variables, such as number of runways at an airport.30 Thus, adding these variables to the underlying Runway Incursion database was simple. Event-Specific DataConversely, for event-specific data, there is a need to distinguish between multiple incidents at the same place and time. For the ATQA OE data, this was accomplished using a unique event identifier. Approximately 249 records in the Runway Incursion database did not have matching records in the ATQA OE dataset. The process for combining the Runway Incursion database with the ATQA PD data was more complicated. The ATQA PD database did not contain a unique record identifier that matched any identifier in the Runway Incursion database. A sequential matching procedure was employed to pair records from the ATQA PD database with the Runway Incursion database. The first step involved matching records that were unique by date and location. That is, records in each database that were the only one at that airport on that date were considered to be matches. A spot check of those matches indicates that they describe the same incident (e.g., aircraft involved, type of incident). The second step in the sequential match involved hand pairing records that were not already matched. Records were considered matches if they were identical on an increasingly looser set of criteria. For example, the exact times of the incidents were compared. If this did not result in a match, a comparison of information such as the aircraft involved and the hour of the incident followed. This process resulted in 4,193 records that were in both databases and 1,547 records only in the RI database. Weather DataAs mentioned previously, the weather data is reported hourly, representing point estimates of the conditions at that time. The Runway Incursion database contains the time of the event down to the minute. Because weather data did not necessarily align with the timing of the incursion event, a way to interpolate the weather at the time of the event was developed. Two methods were developed: one for variables that change continuously (like temperature) and one for variables that change discretely (such as precipitation). The method for continuous variables relied on linear interpolation. The two weather readings on either side of the incident were used as the basis for the interpolation. The method for variables that changed discretely relied on picking the observation closest to the time of the incident. The weather readings occur roughly hourly so the closest reading is, in general, less than 30 minutes away. This method was used for the variables including the weather code (indicating precipitation, fog, smoke, haze, etc.). The remainder of the variables (temperature, cloud cover, etc.) were all subject to the linear interpolation method. The combination of these two methods provided a set of data that could be matched exactly to the Runway Incursion database, making the matching trivial after the interpolation steps.
Before examining specific sets of variables, some general characteristics of the merged dataset are worth presenting. It is important to keep these facts in mind when examining specific variables, as the context in terms of the larger dataset is important. As mentioned previously, incursion events are categorized along two major axes: incident severity and incident type. Table 1 presents the cross tabulation of these two categories and the results of Pearson’s Chi-Squared test (Chi-Squared for short). Additionally, Table 2 presents the expected frequency. The expected frequencies represent the hypothetical distribution of observations across the two categories if the two variables were unrelated. That is, the expected distribution holds the row totals constant but divides observations proportionally among the columns. Deviation from that expected distribution is taken as indication that the rows and columns are not unrelated. For more information please see Appendix C.1. Table – Observed Incident Type Distribution by Severity
Table – Expected Incident Type Distribution by Severity
[A] policy intervention directed at reducing PD incursions would do less to reduce category A incursions than an intervention targeted at OE incursions. The first thing worth noticing is the frequencies across the various cells. Interestingly, across the time period covered by our sample, PD incidents occur about twice as often as V/PD incidents and four times as often as OE incidents. The predominance of PD incidents is also true for the different severity categories. The overall frequency is not the only metric of importance, however. Note that OE incidents are the least frequent overall but are the second most frequent for categories A, B, and C. In fact, category A OE incidents occur approximately four times as often as category A PD incidents (giving OEs the highest rate of category A incidents). Thus, while overall frequency is interesting, it is also important to understand the relative frequency of each category. For example, a policy intervention directed at reducing PD incursions (as they are the most common) would do less to reduce category A incursions than an intervention targeted at OE incursions. The difference between relative frequency and overall frequency raises the need to test for differences in the two. This is where a Chi-Squared test can be useful.32 As reported in Table 1, the Chi-Squared statistic is extraordinarily high and associated with a p-value of approximately zero. This indicates that the distribution of incursion severity is not uniform across the different incident types. Because this is a joint test, it is unable to distinguish which categories are over or under represented. That is, this test indicates that there is some relationship between incident type and severity, but cannot shed light on what that relationship might be. A cursory look at the observed and expected numbers reveals that OE incidents appear to be over represented in categories A, B, and C while being underrepresented in category D incursions. The opposite is true for PD and V/PD incidents, which are underrepresented in categories A, B, and C and overrepresented in category D. This pattern may be the result of one or more underlying processes. Firstly, the increased severity among OE incidents might merely be a function of the nature of OE incidents; in other words, OE incidents are naturally more dangerous. An alternative explanation is that controllers have been successfully trained to avoid category D incidents.33 If controllers were trained to avoid category D incidents (i.e., relatively minor incidents) the remaining incidents would be the more severe incidents. Under this scenario, the rate of OE category A, B, and C incursions is natural, but the rate of OE category D incursions is artificially low. This would be consistent with the observations in Table 1. Yet a third possibility is that controller actions are always double-checked by the pilot. That is, each command given by a controller must be enacted by a pilot. That pilot has the ability to error check those commands and perhaps forestall the least dangerous situations (such as turning onto a closed runway). This is in contrast to pilots who are able to take actions without someone double-checking them, such as rolling over a hold short line or turning onto a runway without contacting the tower. Future Research
While some of the causes suggested above might be more or less likely, it is important to note that there may be multiple explanations. The results presented in Table 1 indicate that additional research is required to understand the true nature of the relationship between incident type and incident severity. Results presented later in this paper may help focus research on why OE incidents may be more severe than other incident types. Table 1 indicated that there is a relationship between severity and incident type. Table 3 further explores this focusing on OE events. The results presented in Table 3 represent the impact of an incident being categorized as OE on severity. As with all regression results, it is important to note that these results represent correlation rather than causation. Table – Logit Estimate of Impact on Severity, OE Incident
Table 3 presents the results of the logit output in terms of odds ratios.34 As described in Table 3 the odds of a severe incident are approximately 3.4 times as high for OE events as for non-OE events. This is in accordance with the results seen in Table 1, but is a more precise measure of how much more likely OEs are to be severe. Table 4 presents the same information, but restricted to only conflict events. Here the alternative to “severe” is category C rather than both categories C and D. The effect of being an OE still persists, though in reduced magnitude. Table – Logit Estimate of Impact on Severity, OE Incident, Conflict Only
The pattern of incursions across regions is also informative. Table 5 and Table 6 present the breakdown of incident type by region while Table 7 and Table 8 present the breakdown of incident severity by region. While the above results presented in Table 1 indicate that there is a relationship between incident type and severity, it is difficult to control for such relationships in a two-way table. Table – Observed Incident Type Distribution by Region
Table – Expected Incident Type Distribution by Region
Table – Observed Severity Distribution by Region
Table – Expected Severity Distribution by Region
[A]ny policy intervention will have differing impacts across regions The most striking feature of these tables is the Chi-Squared statistics rather than the individual cells. The test statistics indicate that the distribution by region is not uniform for incident type or incident severity. This is not surprising given the result of the relationship between severity and incident type noted above. There are likely a variety of causes of this discrepancy, such as varying traffic patterns between regions and the prevalence of general aviation in each region. The overarching point is that any policy intervention will have differing impacts across regions.
| |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||