Quantifying the Risk of Accidents and Serious Incidents Due to Maintenance in General Aviation

1. Introduction

General Aviation (GA) moves a negligible amount of passengers per year compared to airline transport operators, but it accounts for 71.6% of all accidents and serious incidents reported to the United States’ National Transportation Safety Board (NTSB) between 2008 and 2024 (see Table 1). (For brevity, we will call event any accident or serious incident that is reported to the NTSB, per definitions in ICAO Annex 13 or, equivalently, 49 CFR §830.2 in the US Federal regulations.) Of those events, 66.0% were imputed to aircraft-related causes, as either the primary factor or a co-factor, suggesting that mechanical factors are a prime concern. The cumulative loss of life, health, and property associated with these events suggests that GA safety deserves more attention and systematic analysis than it has received so far.

Maintenance plays a critical factor in aircraft-related events. A 2002 FAA study[1] found that at least 7.05% of all GA accidents that occurred between 1988 and 1997 were maintenance-related. Furthermore, the authors

believe that 7.05% is a conservative estimate of GA maintenance-related accidents during this period. [...] there are probably many more maintenance-related accident reports in the NTSB Database that were not included in this study because they were not designated with a maintenance code.

Two known, major causes of failure after maintenance are: (1) early mortality of newly installed components, and (2) Maintenance-Induced Failures (MIFs), i.e., failures whose leading cause is improperly performed maintenance. In GA, a large fraction of all maintenance is performed as a result of scheduled inspections. All GA planes undergo annual inspections. Then, planes involved in training or rental are typically also inspected every 100 hours in service. During an inspection, not only faulty systems are impacted, being replaced or repaired, but healthy systems are touched as well, as necessary to determine their health. As a result, systems that will later be deemed healthy are disassembled and re-assembled. This exposes those systems to human error. More specifically, there is a non-zero risk that, while repairing faulty systems, mechanics accidentally degrade healthy ones, thus causing early failures after the inspection. This is just one example of MIF, but numerous articles have examined the diverse incarnations of MIF [1,4,6,7,10].

Maintenance risks are a popular discussion topic among GA pilots, owners and mechanics, with some explicitly voicing the concern that airplanes would be more dangerous to fly during the first hours in service after maintenance, compared to their baseline risk level (i.e., after accumulating more service time without incident). But in these discussions, beliefs are typically only supported by anecdotal evidence and personal experience. Some mechanics[2] even propose a maintenance minimalism approach: avoid maintenance except when strictly necessary, with the intent of reducing MIF risks. Despite deserving quantitative analysis, the matter has not been explored in prior literature. We focus on that creating that analysis.

In our approach, we isolate the sample of NTSB events that occurred between 2008 and 2024, involving airplanes operating under Part 91 regulations, in which the aircraft was a causal factor and human performance was not, and study the two metrics Airframe Time in Service since Last Inspection (ATSLI), and Engine Time in Service since Last Inspection (ETSLI). We identify the distribution of events with respect to ATSLI and ETSLI values. Finally, we estimate the hazard rates associated with strictly aircraft-caused events.

Paper organization. The remainder of this paper is organized as follows: Section 3 describes the public datasets we consume; Section 4 describes the statistical methods we employ to correct sampling bias and estimate hazard functions; Section 5 presents the results; and Section 6 draws our conclusions.

Novelty claims. To the best of our knowledge, this paper is the first to focus on General Aviation while investigating a link between maintenance and accidents. It is also the first to propose employing strictly human-caused events as a proxy for the over-representation of low times in service, to correct for sampling bias.

2. Related Literature

Boyd[3] offers a broad analysis of the safety record of General Aviation (GA) operations, discussing accident risk factors and possible mitigations, but does not cover maintenance-induced failures. Hieminga and Turkoglu[4] examine maintenance-associated risks and propose a taxonomy to classify them, but they analyzed commercial air transport only. Insley and Turkoglu[6] also limited their review and analysis of aircraft maintenance-related accidents and serious incidents to commercial air transport aviation. Ewertowski et al.[5] discuss risk management in GA, but do not address maintenance-related risk. Tyagi et al.[7] offer a comprehensive review of all literature on safety management and hazard in the aircraft maintenance industry, but it excludes GA. Janovec and Mojžišová[8] explore in depth the root causes of maintenance-related accidents in civil aviation (once again, excluding GA) and propose countermeasures, but do not focus on quantitative risk estimation. Kiyak[9] models the impact of preventive maintenance on total cost (including reliability cost and maintenance cost), but does not estimate risk quantitatively in any area of flight operations. Wild[10] performs an in-depth statistical examination of maintenance-related accidents that occurred in civil aviation as reported in the Aviation Safety Network dataset, but he does not cover GA, and the research questions he explores do not include a relationship between risk and time in service after maintenance.

3. Data Sources

3.1. NTSB Events

Our research uses the public NTSB Aviation Database1, in the format that the NTSB adopted in 2008. This is a relational database comprising 20 tables in Microsoft Access MDB format. Four tables are of interest to our exploration: "events", "aircraft", "findings", and "engines". At the time of our analysis, the database includes accidents and serious incidents that occurred between January 2008 and November 2024.

Our analysis focuses exclusively on GA airplanes (Part 91). The aircraft categories we exclude (rotorcraft, powered lift, lighter than air, etc.) account for a minor fraction of the dataset. We also exclude space launch operations and unmanned craft events, which are also negligible. Events can have a record of ATSLI, ETSLI, both, or neither. We restrict our analysis to events for which either ATSLI or ETSLI is tracked. The cardinality of these sets are in Table 2.

Causal Factors. The NTSB offers causal findings for each event. Causal factors are crucial to the statistical analysis we conduct in this paper. Most events have more than one causal factor. Our approach distinguishes two major causal factors: aircraft and human.

• We call an event human-related and aircraft-related if any of the NTSB findings associated with that event are in the aircraft and in the human category, respectively.
• We call an event strictly human-caused if at least one of its findings is in the human category and no findings are in the aircraft category.
• We call an event strictly aircraft-caused if at least one of its findings is in the aircraft category and no findings are in the human category.

A majority of events are caused by factors in both categories. Table 2 offers a breakdown of events by causal factor, and the charts in Figure 1 visualize how events distribute over ATSLI.

3.2. FAA General Aviation and Part 135 Activity Surveys

We approach post-maintenance reliability by comparing the sample of adverse events caused by the aircraft alone against the sample of those caused by human error alone. In this way we can estimate the proportional hazard ($\psi \left(t\right)$) due to recent maintenance. This choice is justified in detail in Section 3.4.

In the last section of this paper, we will report the absolute hazard rates (${\lambda }_1\left(t\right)$) for strictly aircraft-caused events, expressed in events per Million hours (pMh) flown and, equivalently, Mean Times Between Failures (MTBF) expressed in hours in service. We derive those as the product: ${\lambda }_1\left(t\right)=\psi \left(t\right){\lambda }_0$, where ${\lambda }_0$ is the baseline hazard rate of strictly aircraft-caused events over all GA hours flown.

In turn, we estimate the ${\lambda }_0$ baselines using data from Table 2.1 of the General Aviation and Part 135 Activity Surveys published by the United States’ Federal Aviation Administration (FAA)2. We use survey years 2018 through 2023, inclusively. Note that the surveys lump together Part 91 and Part 135 statistics, while, in the context of this paper, we focus on Part 91 flights only. In spite of this difference, the FAA surveys are still the best resource available to us to estimate ${\lambda }_0$. We compute ${\lambda }_0$ as the ratio between 2018-2023 Part 91+135 event counts we extracted from the NTSB dataset and the total hours flown reported in the FAA survey. The results are in Table 3. Since the survey distinguishes between single- and twin-engine aircraft, we compute distinct values of ${\lambda }_0$ for single- and twin-engine aircraft, and for generic piston aircraft regardless of number of engines. In the table, we report baseline rates for both strictly aircraft-caused events and aircraft-related events, for comparison.

3.3. Metrics of Interest

Two metrics tracked in the NTSB dataset are crucial to our research question: Airframe Time in Service since Last Inspection (ATSLI) and Engine Time in Service since Last Inspection (ETSLI). By Time In Service (TIS), we never intend to mean wall-clock time elapsed since a specified occurrence, but rather a cumulative count of hours during which the aircraft was operated since that occurrence. Operators are required by law to record the TIS when maintenance is performed. The law affords operators some freedom in how to measure TIS, with most GA operators choosing tachometer time. The tachometer accumulates, per each real-time hour, a fraction of hours equal to the ratio between the RPMs the engine was developing, and a reference RPM value, typically maximum-continuous-power RPMs. For example, if the pilot operates the engine at 60% of reference RPMs for one wallclock hour, tachometer time will increase its count by 0.6. Tachometer time can even run faster than real time, if the pilot exceeds the reference RPM value.

At each event, the NTSB makes an effort to reconstruct ATSLI and ETSLI from the aircraft records and other investigations, with variable degrees of success. In the NTSB database, about one third of the events report ATSLI and/or ETSLI values (Table 2). We limit our analysis only to events for which ATSLI or ETSLI is recorded.

The amount and type of maintenance information that the NTSB has decided to track limits the scope of research questions that one can answer. For example, the dataset does not track inspections preceding the last one, which makes it impossible to test a hypothesis that accounts for a plane’s longer-term maintenance history.

3.4. Sampling Bias and Low-time-in-service Over-representation

The NTSB event sample suffers from a sampling bias that causes low ATSLI and ETSLI values to be over-represented. Specifically, most GA airplanes do not fly many hours between mandated inspections, so events with low ATSLI and ETSLI are disproportionately frequent.

An explanation of the causes follow. Annual inspections must occur at calendar intervals, regardless of time in service, and most GA planes do not see many hours in service in each 12-month interval, due to the constraints inherent in personal ownership. Further evidence for this claim is offered by the breakdown of NTSB events by purpose of flight (see Table 4), where 75.7% of all Part 91 events occurred in personal flights. By contrast, aircraft used for commercial passenger carriage are used as much as business allows. Moreover, airplanes used for instruction or rental are typically subject to 100-hour inspections in addition to annual inspection. As a result, these planes are virtually never flown at ATSLI or ETSLI values that exceed 100 hours.

Without bias correction, the data would suggest artificially low survival rates at high time-in-service values. The bias countermeasures we employ are described in the next section.

4. Methods

To correct the sampling bias present in the sample of strictly aircraft-caused events we plan to divide hazard rates, at any given time in service t, by a measure of the over-representation of all GA flying at that t. Ideally, that measure would be the statistical distribution of ATSLI and ETSLI values across all GA hours flown. Unfortunately, that measure is not available because neither the NTSB nor the FAA dataset tracks uneventful flights.

We propose using a proxy population as a measure of the over-representation. The proxies we choose are the ATSLI and ETSLI distributions of strictly human-caused events. This is a creative choice and deserves justification: it seems reasonable to assume that human performance is independent from aircraft- and engine- time in service. By this logic, strictly human-caused events should occur uniformly over all hours flown in a given aircraft and therefore be a reasonable proxy for how ATSLI and ETSLI distribute across all Part-91 hours flown.

We plot histograms of the two distributions of strictly aircraft- and strictly human-caused events, in red and green, respectively, in Figure 2 with 5-h wide bins. Note that bins are drawn semi-transparent: a majority of red bins surpassing the corresponding green ones in the bins in the 0–75h range suggest that airplanes suffer from increased early hazards, even after sampling bias correction.

5. Results

In Section 5.1 we begin by comparing the statistical distributions of service hours since last inspection between the samples of strictly human-caused events and strictly aircraft-caused events. In Section 5.2 we visually compare their empirical survival functions. In Section 5.3 we model the effect of maintenance with a time-invariant multiplicative factor of the baseline hazard function. In Section 5.4 we fit the hazard functions independently and produce a time-varying hazard proportion function. Finally, in Section 5.5 we estimate the absolute hazard functions of post-maintenance flight.

5.1. Hypothesis Testing

We now perform a two-sample Kolmogorov-Smirnov hypothesis test to get a quantitative measure that aircraft-caused event rates exceed human-caused events, which serve as a measure of over-representation. We perform this test for both ATSLI and ETSLI.

• Alternate hypothesis $H_1$: "Strictly aircraft-caused events occur at lower ATSLI/ETSLI values than strictly human-caused events." (i.e., "Recent maintenance raises failure rates".)
• Null hypothesis $H_0$: The two samples come from the same distributions. "Strictly aircraft-caused and strictly human-caused events are drawn from the same distribution over ATSLI/ETSLI." (i.e., "Recent maintenance does not affect failure rates".)

Test protocol: we choose $\alpha =0.05$ as a significance level; if p-value $<0.05$, we reject $H_0$ in favor of $H_1$; if p-value $\ge 0.05$, we are unable to reject $H_0$.

The results are presented in Table 5: in both tests, p-values are well below significance, so we reject the null hypothesis in both cases. Data supports the alternate hypothesis: aircraft-caused events occur at disproportionately lower time-in-service values, i.e., recent maintenance raises accident and incident rates.

5.2. Survival Function Comparison

Additional confirmation of the increased aircraft risk due to recent inspective maintenance after bias correction is obtained comparing the empirical survival functions3 of the two populations of interest, charted in Figure 3. Charted in red is the empirical survival function of strictly aircraft-caused events; charted in green is the empirical survival function of strictly human-caused events, which is our measure of over-representation. The left and right charts depict ATSLI and ETSLI, respectively; they show a consistent story.

Note that the survival rates we chart here are not computed over the entire population of aircraft flown in the period examined–that population is not tracked by the NTSB–but only over the population of airplanes that incurred an event at some point. For example, if the survival function assumes value 0.2 at 70h ATSLI, that means that 20% of the aircraft that experienced an event at some point still have not experienced it at 70h. Actual survival rates are much higher. We will present absolute failure rates at the end of this paper.

Discussion: In both charts, the strictly aircraft-caused survival function curves descend earlier than the strictly human-caused survival curves, with a gap between the two that extends from 0h until approximately 60h. This indicates that aircraft risk is heightened during that window, even after correcting for low-TIS over-representation.

5.3. Risk Quantification via Cox’s Proportional Hazards Model

We now estimate the increased aircraft risk due to recent inspective maintenance while correcting for sampling bias, by fitting our data to Cox’s Proportional Hazards (CPH) model. We do that for both ATSLI and ETSLI values, obtaining similar results. CPH is a model from survival analysis theory which assumes that a population suffers a baseline risk, represented by hazard function $h_0\left(t\right)$ and, in addition to that, when a particular treatment or an action is applied, the hazard function becomes $h_1\left(t\right)=\psi h_0\left(t\right)$. The proportional hazard $\psi$ is a time-invariant factor that multiplies the baseline hazard. (For definitions of terms like hazard function, reliability function, ECHF, etc., please see the Reliability and Survival Statistics Term Summary in the Appendix A.)

We apply Breslow’s method[11] of non-parametric CPH fitting to the two samples: strictly human-caused events, representing $h_0\left(t\right)$, and strictly aircraft-caused events, representing $h_1\left(t\right)$. We truncate both samples at the 200h cutoff point, past which the proportionality assumption no longer seems to hold (see Figure 4). We obtain the following estimates and confidence intervals for $\psi$:

Table 6. Results of Breslow’s method of non-parametric CPH fitting applied to both measures of time in service.

	ATSLI	ETSLI
Estimated $\psi$	1.2268	1.2353
95% confidence interval for $\psi$	[1.1037, 1.3637]	[1.0930, 1.3962]

Table 6. Results of Breslow’s method of non-parametric CPH fitting applied to both measures of time in service.

	ATSLI	ETSLI
Estimated $\psi$	1.2268	1.2353
95% confidence interval for $\psi$	[1.1037, 1.3637]	[1.0930, 1.3962]

Discussion: CPH fitting reveals that in the 0...200h range of time in service after inspection, aircraft causes are responsible for a 23% average increase in accidents and serious incidents. Results are similar for both airframe and engine times in service. Assuming that strictly human-caused events are not associated with maintenance, this would indicate that airplanes are indeed more dangerous immediately after maintenance.

Figure 4. Region of applicability of Cox’s Proportional Hazard (CPH) model. Visually, the CPH assumptions hold in the region where the Empirical Cumulative Hazard Functions (ECHF) for the two populations are parallel when charted in log-log space. Here, that region appears to be to be 0...200 h. The red curve is the strictly aircraft-caused events’ ECHF, and the green is the strictly human-caused events’ ECHF, our measure of over-representation. The left and right charts depict ATSLI and ETSLI, respectively.

Figure 4. Region of applicability of Cox’s Proportional Hazard (CPH) model. Visually, the CPH assumptions hold in the region where the Empirical Cumulative Hazard Functions (ECHF) for the two populations are parallel when charted in log-log space. Here, that region appears to be to be 0...200 h. The red curve is the strictly aircraft-caused events’ ECHF, and the green is the strictly human-caused events’ ECHF, our measure of over-representation. The left and right charts depict ATSLI and ETSLI, respectively.

Limitations: Because CPH assumes a time-invariant $\psi$, it does not model how aircraft risk changes over the first few hours in service. Nevertheless, it is a useful summary indicator of the risk gap between the two populations, approximated by constant hazard functions $h_0\left(t\right)={\lambda }_0$ and $h_1\left(t\right)={\lambda }_1$. The two CHFs corresponding to ${\lambda }_0$ and ${\lambda }_1$ appear as straight lines in log-log space: see the red and the green dashed lines in Figure 5, plotted against the empirical CHF of the actual distributions, drawn with discrete dots.

In the next section of this paper, we obtain a time-varying estimate of the proportional risk.

5.4. Proportional Hazard Function Estimation

We now take the final step toward deriving the time-dependent proportional Hazard Function $\psi \left(t\right)$ of strictly aircraft-caused events. With that, we can compute absolute event rates ${\lambda }_1\left(t\right)$ for every hour of ATSLI or ETSLI, which express how dangerous an airplane is during that hour in service.

To do that, we adopt a parametric approach:

• We fit the ATSLI distribution $h_1\left(t\right)$ of strictly aircraft-caused events. This distribution suffers from sampling bias as described earlier. The best fit is a Weibull mixture, of parameters ${\alpha }_1=32.1546,{\beta }_1=0.919913,{\alpha }_2=858.447,{\beta }_2=0.560298,w_1=0.934676$, depicted in Figure 6(a).
• We fit the ATSLI distribution $h_0\left(t\right)$ of strictly human-caused events. This distribution is our chosen measure of sampling bias. The best fit is Weibull mixture, of parameters ${\alpha }_1=42.0653,{\beta }_1=0.941214,{\alpha }_2=1762.77{\beta }_2=0.825542,w_1=0.955302$, depicted in Figure 6(b).
• We derive $\psi \left(t\right)=h_1\left(t\right)/h_0\left(t\right)$, the proportional hazard functions corrected for sampling bias. It is charted in Figure 7.

We choose to derive $\psi \left(t\right)$ as a ratio between two smooth, parameterized PDF functions.

Figure 7 shows that during the first hour, the hazard rate is 33.8% higher then baseline, then it’s 30.0% higher in the second hour, and it keeps decreasing, crossing the baseline level at 31h ATSLI. It reaches a minimum at 188h ATSLI, at which the relative risk is 56.4% of baseline. Then it starts increasing again, crossing baseline level once more at 327h and reaching a new peak at 478h ATSLI, at a risk level 37.1% above the baseline. In its first 500h, the curve resembles a typical "bathtub curve" hazard function as seen in numerous reliability applications[12], with an early mortality period, a useful life period, and a wear-out period.

We decide not to extend our ATSLI considerations to ETSLI because the quality of parametric fits we were able to achieve on ETSLI data (see Figure 6(c),(d)) appeared insufficient for use in further analysis.

5.5. Absolute Hazard Function Estimation

We now combine $\psi \left(t\right)$, the bias-corrected proportional hazard function of strictly airplane-caused events as a function of ATSLI (derived in the previous section) with ${\lambda }_0$, the time-invariant absolute baseline hazard rate we derived earlier, in Section 3.2. We use three distinct estimates for ${\lambda }_0$, one for single-engine planes, one for twin-engines, and one for generic piston engines.

That yields absolute risk rates ${\lambda }_1\left(t\right)=\psi \left(t\right){\lambda }_0$ that we tabulate in Table 7, Table 8, and Table 9 for the first 15h ATSLI. Note that we estimated $\psi \left(t\right)$ on strictly aircraft-caused events to exclude the influence of other factors, but it seems now sound to also apply $\psi \left(t\right)$ also to aircraft-related events, so the tables present ${\lambda }_1$ for both event classes.

An example of using the tables follows: in the first hour of service since inspection, hazard is increased by a factor of $1.338\times$: in Single-Engine airplanes, this leads to a rate of 11.533 strictly aircraft-caused events per million hours and 69.059 aircraft-caused events per Million hours. Equivalently, the expected amounts of hours flown before a strictly aircraft-caused or an aircraft-related event are 86,707 and 14,480 hours. These hazard rates are significantly higher than the baselines presented in Section 3.2.

6. Conclusions

We examined a commonly held belief that maintenance raises the risk of accidents and serious incidents in general aviation in the immediate service hours after an inspection. We subjected this belief to quantitative scrutiny, and found it to be substantiated by data.

Specifically, we examined the distribution of time-in-service-since-last-inspection metrics in failures occurring in the population of accidents and serious incidents reported to the NSTB in 2008-2024, affecting general aviation airplanes. We found this sample to be biased towards low time-in-service values, because most planes are flown only a few hours between mandated inspections. We studied the population of strictly aircraft-caused events using the strictly human-caused events as a measure of bias. We verified via Kolmogorov-Smirnov hypothesis testing that the two populations are statistically distinct. Via CPH model fitting, we estimated the average proportional hazards for the first hours in service. Finally, we fitted the two populations with Weibull distributions and obtained an analytic estimate of the proportional hazard as a function of time $\psi \left(t\right)$, which shows that in the first hour in service, hazard rates are 33.8% higher than baseline.

Our findings consistently substantiate the belief that inspective maintenance increases danger in the first hours of service after a plane returns from inspective maintenance. This belief, commonly held on an empirical and anecdotal basis, is now supported by statistical evidence. Pilot and owners of Part 91 airplanes are therefore justified in exercising increased caution when performing post-maintenance acceptance flights and flying those planes in the subsequent hours in service. Efforts to increase the standards of maintenance processes and materials will likely reduce these risks.