Validation of an Offshore Occupational Accident Frequency Prediction Model—A Practical Demonstration Using Case Studies Daryl Attwood, Faisal Khan, and Brian Veitch Faculty of Engineering & Applied Science, Memorial University, St John’s, NL, Canada A1B 3X5;
[email protected] (for correspondence) Published online 14 March 2006 in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/prs.10128 A model has been developed to predict the frequency and associated costs of occupational accidents in the offshore oil and gas industry. Model inputs include: (i) direct factors, such as quality of personal protective equipment; (ii) corporate factors, such as training program effectiveness; and (iii) external factors, such as royalty regime. Three applications of the model are described, two for projects in eastern Canada and one for the Gulf of Mexico drilling sector. Expert opinion is used to provide the required model input associated with the regions’ safety programs. Published accident data are used to calibrate the model and validate results. The model is shown to predict actual results well, especially considering the subjective nature of the activity. The model’s versatility is demonstrated through its application to different types of accident statistics and regions, and its use in generating performance measures for operators. © 2006 American Institute of Chemical Engineers Process Saf Prog 25: 160 –171, 2006 Keywords: occupational accidents, offshore accidents, slips, trips and falls, accident models, safety systems, safety culture Current address of D. Attwood is Lloyd’s Register EMEA Aberdeen, 25 Union Terrace, Aberdeen, UK, AB10 1NN. © 2006 American Institute of Chemical Engineers
160
June 2006
INTRODUCTION
Occupational accidents constitute an area of significant and continuing risk for the oil and gas industry. The statistical data [1] show that fatalities are more likely to be caused by occupational accidents than by more catastrophic events, such as explosions or air transport incidents. The situation is consistent with that observed in the general workplace, where it has been reported [2] that over a third of all major injuries reported each year result from a slip or trip, this being the single most common cause of injuries at work. While workplace safety is regulated under national legislative schemes, analysis is not as rigorous for occupational accidents as for major event hazards. In order to suggest a more quantitative approach to the occupational accident issue than currently exists, the authors have developed a model to predict occupational accident frequency. While occupational accident statistics are available in many regions where oil and gas activities are undertaken, they are, by definition, based on past performance on existing installations. The present model offers stakeholders on new projects (even in the same region), having unique safety practices and philosophies, a tool to predict accident frequency under their specific regime. Furthermore, operators are always interested in optimizing management decisions, including the choice of where to spend on safety. The model offers the capability to compare predicted safety improvements resulting from Process Safety Progress (Vol.25, No.2)
Figure 1. Model structure.
changes in various safety elements, for example, enhanced personal protective equipment compared with improved corporate safety culture. The model’s development and structure have been described in detail elsewhere [3]. Its validation, using three cases, is described in this article. THE MODEL
The model’s basic premise is that while occupational accidents result from an unsatisfactory direct interaction between worker and workplace, workers’ behaviors are influenced by corporate culture, and their workplace environment and procedures are controlled corporately. Furthermore, corporate decisions and actions are, in turn, influenced by external elements. The model structure is shown in Figure 1 (for details, see [3, 4]). Essentially, the safety system is treated as a modified reliability network. Component reliability values determine overall system reliability, which is used to predict accident frequency. Quantitative inputs are required for: (i) direct factors, such as quality of personal protective equipment; (ii) corporate factors, such as training program effectiveness; and (iii) external factors, such as royalty regime. Model development was based on a review of reProcess Safety Progress (Vol.25, No.2)
lated literature, expert opinion, and reliability analysis concepts. The model uses quantitative data derived from a survey of safety experts to account for the differing relative importance of factors. The influences of external elements on corporate actions and of corporate actions on the direct accident process are also included in a quantitative manner, again benefiting from the expert opinion survey.
MODEL EXECUTION METHODOLOGY—CALIBRATION AND ACTUAL CASES
The accident frequency prediction process requires the model to be run in two distinct modes. First, a calibration run is executed, where known accident rates are used to determine base case component reliabilities. Second, the model is run in predictive mode following adjustment of the base case reliabilities. The degree of adjustment is determined using a quantified comparison of safety conditions in the specific and base cases, which requires expert input from safety personnel familiar with both situations. In many applications, the global average safety situation, having documented results and generally known conditions, is used as the base case. This section details the process
Published on behalf of the AIChE
DOI 10.1002/prs
June 2006
161
of running the model in both calibration (base case) and predictive (specific case) modes. Calibration The goal of the calibration process is to determine base case component reliabilities. Any situation where both safety results and operating conditions are known can be chosen as the base case. Because the subsequent predictive model run requires a comparison of specific and base cases, a convenient base case option is the average global offshore industry. Global average safety results are available, and in most cases experienced safety experts will be able to offer a reasonable comparison of any specific factor to the global average situation for that element. The remainder of the discussion is based on the assumption of global average conditions as the base case. The type of accident statistic used for calibration depends on which output statistic is desired. For example: • If a particular (total recordable, lost time) accident
rate in a region or industry sector is sought, then the corresponding global average value of that particular rate is used for calibration. • If the expected annual number of a specific kind of accident (total recordable, days away from work) on an installation having a given POB (persons on board) is required, then the global average rate of that type of accident is combined with the POB to determine accident numbers expected had the facility been operating under average safety conditions. An example of the latter type of calculation is presented here. Suppose the predicted number of annual recordable incidents on a platform is required. The global average value of this statistic (TRIR, or total recordable incident rate), 6.36 [1], is used for calibration. On a 70 POB installation operating 24 h per day, as most do, approximately 35 persons will be working at any given time, while their counterparts rest. The number of person-hours worked per year for the platform is then calculated as below. Person-hours worked ⫽ 共35 persons兲 ⫻ 共24 hours/day兲 ⫻ 共365.25 days/year兲 ⫽ 306,810 person-hours Based on the 2004 TRIR (6.36/1,000,000 hours), the expected number of installation specific annual recordable cases under average safety conditions is then calculated as follows. Expected number of recordable cases ⫽ 共306,810 person-hours兲 ⫻ 共6.36/1,000,000 person-hours兲 ⫽ 1.95 cases This figure is then used to calibrate the model for average safety conditions. An iterative process is used 162
June 2006
Published on behalf of the AIChE
to determine individual component inputs required for the model to have predicted this number of annual accidents. Software tools (goal seek function in Microsoft Excel) are available to make the exercise a quick and easy affair. Many sets of component reliabilities could produce the output required for calibration purposes. However, since model execution is based on a quantified comparison of specific and base cases, the absolute values of base case component reliabilities are not important. What is crucial, though, is the scale of the subsequent expert judgment based adjustments applied for the predictive run. Model execution requires a comparison of specific case to base case, not absolute component reliability values. Because of this, the individual component reliabilities assigned by the calibration process are identical to one another. Assigning different values would be an unnecessary complication. Specific Case Runs To predict accident frequency for a specific case, the model is run following adjustment of the base case component reliabilities in line with the safety environment of the installation or sector under study. The degree of component reliability adjustment is based on the opinion of experts familiar with both base (average global) and specific case safety conditions. The experts assign scores from 1 to 10 for each factor, representing the component’s specific case conditions, compared to global average, which is represented by a score of 5. Higher scores, in all cases, represent situations more favorable to safety results—for example, a high score on royalty regime corresponds to a situation where the government takes relatively less money in royalties, thereby leaving more free cash for operators to spend on everything, including safety measures. Likewise, a lower score (that is, less than 5) corresponds to a regime where more than average cash is taken by the government, leaving relatively less for safety spending. A high score on PPE would indicate that the specific case’s overall suite of safety equipment was considered superior to the global average. At first glance, it would seem reasonable to adjust component reliabilities in direct proportion to the experts’ assigned scores. Using this system, a score of 6 for a given component would result in the base case reliability being multiplied by 6/5, or 1.2, while a score of 10 would result in a doubling of base case reliability (10/5 ⫽ 2). However, other functions can be used to transform the expert panel’s subjective observations to factors used to adjust the base case reliabilities. The literature [5] describes the design of several alternative mechanisms for transforming subjective observations such as these to useful indices. For example, the use of power functions to generate indices for water quality based on pollutant variables is proposed. For the present application, results were seen to be improved by considering the “importance” of scores further away from the mean (5) to be greater than those for more centralized results, in effect magnifying extreme values’ effects beyond that applied when using a directly proportional approach. This is done by using a DOI 10.1002/prs
Process Safety Progress (Vol.25, No.2)
“power 2” function, in other words, making the changes (in component reliability in this case) proportional to the square of the ratio of the specific case to average case score (5). For example, an assigned score of 6 would produce a component reliability increase of (6/5)2 ⫽ 1.44 and a score of 10 would produce a component reliability increase of (10/5)2 ⫽ 4. Note that the difference between squared increases and the directly proportional approach is only 20% (1.44/1.2) for values relatively close to the mean (6), but the difference is 100% (4/2) at the extreme value (10). This process has the effect of making the “importance” directly proportional to the magnitude of the score. The use of powers greater than 2 makes the process overly sensitive to extreme values and is, therefore, to be avoided. In summary, the use of the “power 2” function provides the best fit with model formulation, calibration, and results. Model accuracy is described in the next section. Once the expert panel has assigned scores for the specific case, and adjusted component reliabilities have been determined using the method described above, accident frequency predictions can be made by running the model in one of three ways: 1. Direct layer component reliabilities can be input and system reliability and accident frequency calculated directly. 2. Corporate layer component reliabilities can be input and allowed to influence the direct layer values (the specific process is described in [3]). Following determination of the direct layer reliabilities, the calculation proceeds as in the previous method. 3. External layer reliabilities can be input, which can be allowed to determine the corporate values and, in turn, the direct values, facilitating the calculation as previously. In general, if specific case expert scores are known for all components, the model is run using all three methods, with the final prediction taken as the average of the three results. CASE STUDIES
Three case studies have been executed, as follows: 1. A comparison of the number of predicted and actual annual accidents on a Nova Scotia (NS), Canada, based installation 2. A comparison of the number of predicted and actual annual accidents on a Newfoundland (Nfld), Canada, based installation 3. A comparison of predicted and actual lost time incident (LTI) rate in the Gulf of Mexico, USA, drilling sector The general process for running the model was discussed above. The specific methodology for these case studies was as follows: 1. In cases 1 and 2, published global average accident rates (OGP) were combined with POB values to estimate how many accidents would have occurred on the installations had they been operating under average conditions. In case 3, the global average Process Safety Progress (Vol.25, No.2)
2.
3.
4. 5.
accident rate was directly available from the published data (IADC). Using the results obtained in Step 1, a calibration run was executed to calculate component reliabilities which would have produced the average conditions, thereby producing “base” values for the component reliabilities. Each component reliability was then adjusted according to the location-specific scoring assigned to each factor by the expert panel as they compared the specific situation to global average (using a 1–10 scale and assuming 5 ⫽ global average). The model, using the updated component reliabilities, then predicted accident occurrence numbers or rate for the specific situations. For cases one and two, published, region specific accident rates (CNSOPB, CNLOPB) were used to estimate the number of accidents likely to have been experienced on installations in the areas covered by the data. For case three, the region specific accident rate was directly available from published data [6]. These are taken to be the “actual” values against which the predictions generated in Step 4 were evaluated.
A few additional details on the process are noteworthy, as follows. Because of the subjective nature of the component scoring process, an analysis was conducted to study the sensitivity of output predictions to changes in individual component scorings. It was discovered that the greatest percentage change in prediction with a single step change (for example, from 7 to 8) in any one component was less than 3%. This means that if the panel erred (perhaps erred is too strong a word for this subjective activity) by a single digit in its scoring of a specific component, the effect on accident frequency prediction would be relatively small. Furthermore, since (i) the direction (that is, overrating versus underrating) of individual component scoring errors is expected to be equally divided and (ii) the prediction process relies on the scoring of multiple components, the effect of the errors is expected to cancel out. It was important for the panel to have an accurate general view of the overall situation, but a precise measure on each and every individual component is not essential. The opinionbased nature of the scoring process likely made such precision impossible in any event. Within the context of using expert opinion to produce a quantitative prediction of accident frequency, the expected level of output error associated with individual component scoring error is considered acceptable. Three separate methods of using the model to determine system reliability, and hence accident frequency, were discussed above. For these case studies, the experts’ opinions allowed the calculation to be performed using any of the methods. The final accident prediction (Step 4) was taken as the average of the results from the three runs.
Published on behalf of the AIChE
DOI 10.1002/prs
June 2006
163
Table 1. Accident rates (events per million hours).
Global offshore TRIR (OGP) Nfld offshore TRIR (CNOPB) NS installations TRIR (CNSOPB) Gulf of Mexico drilling LTIR (IADC) Global drilling LTIR (IADC)
2000 8.84 10.16 5.60
2001 6.85 9.49 3.35 3.35 3.09
2002 5.77 8.04 6.40 2.52 2.71
2003 4.87 11.45 3.35 2.57 1.81
2004 6.36 4.36 5.95 2.11 1.56
2005
1.96 1.92
Data Sources
• Engineer one—35 years oil and gas experience,
Accident Statistics
• Engineer two—10 years oil and gas experience,
the past 8 on Canadian oil and gas projects
Data for the calibration portion of the model application are publicly available [6 – 8]. For comparison of predicted to actual values, the ideal data would be installation specific accident frequencies, since they would offer an opportunity to directly compare predictions based on safety conditions with specific platform results. However, operators in general do not release this information and refused the request to do so for this research. In the absence of operator-supplied platform specific data, statistics published by the petroleum boards [7, 8] were used to estimate accident frequencies on installations in the respective regions. Listed below are the data sources that have been used to (i) calibrate the model for average conditions and (ii) evaluate the subsequent predictions. The pertinent data are shown in Table 1. • The International Association of Oil and Gas Pro-
ducers [1] Safety Performance Indicators, which provided global TRIR (total recordable incident rate) statistics. • The Canada-Nova Scotia Offshore Petroleum Board [8] and the Canada-Newfoundland and Labrador Offshore Petroleum Board [7] websites, which provided annual TRIR values for the Nova Scotia and Newfoundland offshore areas, respectively. The Nova Scotia data are split into installations, vessels, and aviation, but the Newfoundland data cover all offshore activity. • The International Association of Drilling Contractors [1] website, which provided LTIR (lost time incident rate) data for offshore drilling activities for several regions (for example, USA, Canada, Africa). Expert Opinion
Satisfactory model output accuracy requires a quality comparison of the specific case’s safety situation with global average conditions. This can be provided only by qualified safety professionals having both specific project or region experience, and a significant international offshore background. The present panel had both region-specific and general experience in safety design, project management, offshore surveying, and safety consultancy. The members averaged 18 years of oil and gas industry experience, ensuring that the group could draw from a sufficient depth of relevant knowledge. Details of the panel’s experience are listed below. 164
June 2006
Published on behalf of the AIChE
the past 5 on Canadian oil and gas projects • Project manager one—23 years oil and gas expe-
rience, most of the past 14 spent on Canadian and US oil and gas projects • Project manager two—18 years oil and gas experience, the past 3 spent on Canadian oil and gas projects • Project manager three—3 years oil and gas experience, all spent project managing Canadian and US oil and gas projects • Design specialist—25 years oil and gas experience, the past 8 partially spent appraising safety designs on Canadian and US oil and gas projects • Safety risk consultant—10 years oil and gas experience, participating in industry related safety research on a global basis. Results and Discussion A five step model execution process was described above. The discussion below makes reference to that process. Nova Scotia Production Installation
A Nova Scotia based 70 POB (persons on board) production installation was chosen as a case study for the model. The data in Table 2 will be discussed in this section. Step 1—Accidents under global average conditions The number of accidents expected on a 70 POB installation operating under global average conditions is calculated by combining annual average global accident rates (TRIR) available from the OGP database (Table 2, Row 1) with the POB. The assumption is made that, as is the norm on offshore oil and gas installations, 50% of workers are “on shift” while their opposite numbers rest. This produces the same numerical result as if 50% of the POB were working continuously. As an example, the expected number of accidents for the 2004 data is calculated as follows. Expected accidents ⫽ 6.36 accidents/1,000,000 manhours ⫻ 70 persons ⫻ 0.50 working ⫻ 24 hours/day ⫻ 365.25 days/year ⫽ 1.95 The results of this by year are presented in Table 2, Row 3. DOI 10.1002/prs
Process Safety Progress (Vol.25, No.2)
Table 2. Nova Scotia case study.
1. Global average TRIR 2. Nova Scotia TRIR 3. Number of accidents (based on global average TRIR) 4. Number of accidents predicted by model 5. Number of accidents (based on Nova Scotia TRIR) 6. % error
2000 8.84 5.60
2001 6.85 3.35
2002 5.77 6.40
2003 4.87 3.35
2004 6.36 5.95
2.71
2.10
1.77
1.49
1.95
2.00
2.26
1.65
1.32
1.03
1.50
1.55
1.72 32
1.03 61
1.96 ⫺33
1.03 0
1.83 ⫺18
1.51 3
Table 3. Expert opinion of Nova Scotia safety
environment compared to global average (global average ⫽ 5 on a 1–10 scale). Factor External factors Value placed on life Price of oil Shareholder pressure Royalty regime Corporate factors Safety culture Safety training Safety procedures Direct factors Attitude Motivation Lack of fatigue Coordination Fitness Knowledge Intelligence Safety design Weather Personal protective equipment
Expert Score 9 10 3 4 7 7 8 7 6 8 5 6 7 5 7 3 8
Step 2—Calibration run In order to set base component reliabilities, the model is run for each year with the results preset to the global average accident expectations listed in Table 2, Row 3. Step 3—Component reliability adjustment Table 3 shows the component scores assigned by the expert panel to the Nova Scotia region, as mentioned, on a scale of 1–10 compared to an industry average value of 5. These scores are used to adjust the base component reliabilities calculated in Step 2. Step 4 —Prediction The predicted numbers of accidents per year by year are shown in Table 2, Row 4. Step 5—Comparison of predictions with estimates of actual numbers of accidents Actual number of accidents expected on the platform can be estimated by repeating the calculation in Step 1, but, instead of using global average values, substituting the annual accident rates (TRIR) available from the CNSOPB (Table 2, Row 2). The results for Process Safety Progress (Vol.25, No.2)
Average 6.54 4.93
2000 –2004 are shown in Table 2, Row 5. The error between the predicted and actual results (that is, using Nova Scotia accident statistics) is shown in Table 2, Row 6. A graphical comparison of actual and predicted results for each year is shown in Figure 2. The results serve to validate the model. Some specific points follow. • The significance of the errors was evaluated using
the procedure described in Smith [9]. A null hypothesis, specifically, that the mean of the differences between predicted and actual results was not significantly different from zero, was tested. An analysis (t ⫽ -0.16) showed that the null hypothesis could not be rejected at the P ⫽ 0.2 confidence level. • It could be argued that results for any specific year have questionable reliability, and that five year rolling averages are more appropriate. The five year average number of accidents was predicted with a very small (3%) error. • The result for 2003 was excellent (0% error). • Trend matching on the basis of five data points may be of limited value, but it was interesting to note that with the exception of the 2001–2002 transition, the direction of year on year changes in actual accident frequency was matched by the predicted values. Newfoundland Production Installation
The model has been applied to a Newfoundland based 100 POB installation. The data in Table 4 will be discussed in this section. Step 1—Accidents under global average conditions The number of accidents expected on a 100 POB installation operating under global average conditions is calculated by combining the annual average global accident rates (TRIR) available from the OGP database (Table 4, Row 1) with the POB (and assuming that 50% of them are working continuously). As an example, the expected number of accidents for the 2004 data is calculated as follows. Expected accidents ⫽ 6.36 accidents/1,000,000 manhours ⫻ 100 persons ⫻ 0.50 working ⫻ 24 hours/day ⫻ 365.25 days/year ⫽ 2.79
Published on behalf of the AIChE
DOI 10.1002/prs
June 2006
165
Figure 2. Predicted versus actual results, Nova Scotia platform.
Table 4. Newfoundland case study.
1. Global average TRIR 2. Newfoundland TRIR 3. Number of accidents (based on global average TRIR) 4. Number of accidents predicted by model 5. Number of accidents (based on Newfoundland TRIR) 6. % error
2000 8.84 10.16
2001 6.85 9.49
2002 5.77 8.04
2003 4.87 11.45
2004 6.36 4.36
3.87
3.00
2.53
2.13
2.79
2.86
3.68
2.80
2.33
1.93
2.59
2.67
4.45 ⫺17
4.16 ⫺33
3.52 ⫺34
5.02 ⫺62
1.91 36
3.81 ⫺30
The results of this by year are presented in Table 4, Row 3. Step 2—Calibration runs The model is run for each year with the result set at the global average accident expectations shown in Table 4, Row 3. This allows the base component reliabilities to be set. Step 3—Component reliability adjustment Table 5 shows the scores assigned to model components by the expert panel for the Newfoundland region. Step 4 —Predictions The predicted numbers of accidents per year by year are shown in Table 4, Row 4. Step 5—Comparison of predictions with estimates of actual numbers of accidents Actual number of accidents expected on the platform can be determined by repeating the calculation in Step 1, but, instead of using global average values, substituting the annual accident rates (TRIR) available from the CNLOPB (Table 4, Row 2). The results for 2000 –2004 are shown in Table 4, Row 5. The error between the predicted and actual results (that is, using Newfoundland accident statistics) is shown in Table 4, Row 6. 166
June 2006
Published on behalf of the AIChE
Average 6.54 8.70
A graphical comparison of actual and predicted results for each year is shown in Figure 3. These results are less encouraging than those obtained in the Nova Scotia case study. In this case the “t” value was 1.90, which indicates that the null hypothesis (mean of the error is not significantly different from zero) can be rejected at the 0.2 confidence level. A possible explanation for this follows. The Newfoundland published (actual) TRIR results, compared in Figure 4 with global average values, were, from 2000 –2003, consistently and significantly worse than industry average. However, as discussed elsewhere [10, 11], safety experts’ views can offer an alternative (some say better) indicator of safety performance to the more commonly used accident statistics. In contrast to the relationship between Newfoundland and global performance implied by the 2000 –2003 TRIR statistics, the panel rated the Newfoundland offshore safety environment equal or superior to the average global situation in more than 86% of the elements considered. It could be argued that because the evaluation took place in mid-2005, it was most heavily influenced by the situation over the most recent few years (say, 2004 – DOI 10.1002/prs
Process Safety Progress (Vol.25, No.2)
Table 5. Expert opinion of Newfoundland safety
environment compared to global average (global average ⫽ 5 on a 1–10 scale). Factor External factors Value placed on life Price of oil Shareholder pressure Royalty regime Corporate factors Safety culture Safety training Safety procedures Direct factors Attitude Motivation Lack of fatigue Coordination Fitness Knowledge Intelligence Safety design Weather Personal protective equipment
Expert Score 9 10 3 4 8 7 9 6 7 8 5 6 8 5 7 1 9
2005). The panel’s view of the relationship between global and Newfoundland safety performance was consistent with that implied by the 2004 statistics (that is, Newfoundland results better than global). However, in the absence of an explanation for why Newfoundland safety performance would be significantly worse than global average from 2000 –2003, and then suddenly better in 2004, and based on the panel’s views, it is probably more likely that the Newfoundland offshore industry has in fact been performing better than global average over the entire 2000 –2004 period. This conclusion requires explanations of (i) why the relationship between global and Newfoundland published TRIR values from 2000 –2003 was opposite from the expert panel’s views and (ii) why this effect was not evident in the Nova Scotia statistics. Two possibilities are offered here. • The Newfoundland data included accidents on
supply boats, which are not usually included in oil and gas statistics. This made the Newfoundland data less applicable than the Nova Scotia data for the present exercise, which is concerned with activities on installations. The five-year average TRIR rate for vessels in Nova Scotia was 38% higher than for installations, so the inclusion of the supply boat data may have inflated the Newfoundland TRIR results. • The Newfoundland statistics may have suffered from a greater propensity to over-report occupational accidents than the Nova Scotia results. A possible explanation may be associated with the different union status of workers in the regions. Unlike the Nova Scotia sector, both of the Newfoundland projects’ workforces are unionized. A Process Safety Progress (Vol.25, No.2)
successful union certification vote for one project was held in late 2001 [12], and for the other in 2002 [13]. In situations where a struggle to unionize has recently been won, the workforce can sometimes be characterized by both an exaggerated and newfound perception of job security, and significant anger with the employer. In such situations, a healthy and appropriate willingness to report accidents can gradually turn into a desire to do so, resulting in trivial accidents finding their way into the statistics. Following a three year downward trend in the Newfoundland published accident rate from 2000 –2002, an upward spike occurred in 2003, in the first full year when both installations were operating with a unionized workforce. It is noted that the 2004 value could have been predicted by approximately continuing the 2000 –2002 downward trend. The panel’s view of safety performance in the Newfoundland offshore industry (relative to global average), contrasting as it does with the published TRIR values, explains why, with the exception of 2004, the model under-predicted the Newfoundland installation accident frequency. The model can be used as a diagnostic tool to investigate unexpected safety results. In this case, for example, a trial and error exercise was conducted to determine some component scorings necessary for the model to have accurately predicted the actual results. Figure 5 shows the comparison of predicted and actual values when external elements were scored as 3, instead of the values in Table 5. Under this scenario, predictions match actual values quite well for the years 2000 –2002. The 2003 prediction implies a continuation of the trend for the previous three years, which did not compare well with the rise experienced in the 2003 published data. If the 2003 and 2004 “actual cases” (5.02, 1.91) are both replaced by their average (3.47), thereby effectively removing the 2003–2004 fluctuations, the matching is improved even more (see circles in Figure 5). Gulf of Mexico Drilling Sector
In this case, the model has been used to predict lost time incident (LTI) rate in the Gulf of Mexico drilling sector. This application shows that the model can be used in different modes, in this case to predict accident rate in a given region rather than previously where it was used to predict accidents per year on a platform. The data in Table 6 will be discussed in this section. Step 1—Average global LTI rate Global average drilling LTI rates are shown in Table 6, Row 1. Step 2—Calibration run The model is run for each year with the result set at the global average accident rates shown in Table 6, Row 1, allowing the base component reliabilities to be set. Step 3—Component reliability adjustment Table 7 shows the scores assigned to the Gulf of Mexico offshore drilling sector by an expert panel.
Published on behalf of the AIChE
DOI 10.1002/prs
June 2006
167
Figure 3. Predicted versus actual results, Newfoundland production platform.
Figure 4. Global TRIR versus Newfoundland TRIR.
Step 4 —Prediction The results by year of predicted accident rate are shown in Table 6, Row 3. Step 5—Comparison of predictions with actual accident rate Annual Gulf of Mexico drilling sector LTI rates are available from the IADC website and are presented in Table 6, Row 2. The error between the predicted and actual results is shown in Table 6, Row 4. A graphical comparison of actual and predicted results for each year is shown in Figure 6. In all cases the error percentage was less than 27%, and in three of the five years studied it was less than 13%. The “t” value in this case was 1.08, which indicates that the null hypothesis (mean of the error is not significantly different from zero) cannot be rejected at the 0.2 confidence level. With the exception of the 2004 –2005 transition, the five year trend is matched, and the five year average predicted value is within 7% of the actual value. This case demonstrates the versatility of the model. The previous cases considered numbers of accidents 168
June 2006
Published on behalf of the AIChE
expected on production platforms in different regions, whereas this case considered accident rate in a region specific drilling sector. Use of Results to Predict Probability of Numbers of Accidents
The predictions of expected number of annual accidents can be used to give the operator an idea of the probability of experiencing specific numbers of accidents around the mean value. The Poisson distribution [14] represents the probability of an isolated event occurring a specified number of times in a given time interval when the average rate of occurrence is fixed. The assumption of constant failure rate can be made [3] for the present application. Assuming a constant failure rate (), then, the Poisson distribution proposes the following equation to calculate the probability of “x” occurrences in a unit time (1 year in this example). Px ⫽
x e ⫺ , x!
DOI 10.1002/prs
Process Safety Progress (Vol.25, No.2)
Figure 5. Actual versus predicted, Newfoundland platform, adjusted external element scorings.
Table 6. Gulf of Mexico case study.
1. 2. 3. 4.
Global average LTIR Gulf of Mexico LTIR Predicted LTIR % error
2001 3.09 3.35 3.19 ⫺5
2002 2.71 2.52 2.81 12
Table 7. Expert opinion of Gulf of Mexico safety
environment compared to global average (global average ⫽ 5 on a 1–10 scale). Factor External factors Value placed on life Price of oil Shareholder pressure Royalty regime Corporate factors Safety culture Safety training Safety procedures Direct factors Attitude Motivation Lack of fatigue Coordination Fitness Knowledge Intelligence Safety design Weather Personal protective equipment
Expert Score 4 7 1 6 3 5 4 4 4 5 5 4 4 5 4 8 5
2003 1.81 2.57 1.91 ⫺26
2004 1.56 2.11 1.66 ⫺21
2005 1.92 1.96 2.02 3
Average 2.22 2.50 2.32 ⫺7
predicts the expected number () of occupational accidents on a Nova Scotia installation for 1 year to be 1.50. Using the Poisson assumptions, the operator could expect his probability of having 0, 1, 2. . . . accidents to be as shown in Figure 7. The operator could conclude from this, for example, that assuming no safety related changes are made, his probability of having zero accidents would be 0.22, of having zero or one accident would be 0.55, of having two or fewer would be 0.80, and so on. Similar calculations could be made for other years, assuming the five year average value is applicable over the long term. An analysis such as this could be used by an operator to make reasonable safety challenges to its workforce or set key performance indicator (KPI) targets for itself or its contractors. The often stated goal of “zero accidents” is an extremely difficult one to achieve. Calculations such as these, however, would provide managers with more achievable, yet challenging, targets. Staff could be asked, for example, to improve on the number of accidents expected on a platform with a probability of, say, 60%. Or, contractors could earn scaled rewards based on beating the number of accidents expected with probability of 80%, 70%, and so on. CONCLUSIONS
where Px ⫽ probability of “x” occurrences, and ⫽ average or expected number of occurrences As reported in Table 2, Row 4, in 2004, the model Process Safety Progress (Vol.25, No.2)
The model has been validated against actual accident results. Results were most accurate in the Nova Scotia installation case study, followed by the Gulf of Mexico drilling sector lost time incident analysis. Results in the Newfoundland study were less accurate
Published on behalf of the AIChE
DOI 10.1002/prs
June 2006
169
Figure 6. Actual versus predicted, Gulf of Mexico drilling sector.
Figure 7. Probability of accidents, Nova Scotia production platform.
than the other two, but the published accident data in this case may be unreliable. It was demonstrated that the model can be used as a diagnostic tool to study unexpected safety results. For example, by adjusting input scores, we can simulate situations that would have been required to match actual results. If the theoretical input ratings are clearly at odds with reality, we may have possible grounds to question the reliability of the reported data. The versatility of the model has been demonstrated. It can be used to predict accident numbers on a single platform or accident rates in a specific sector. Following the validation described here, the model can be used by offshore oil and gas safety professionals in any of the following practical ways: • To predict occupational accident frequency under
any unique safety environment. • To observe improvements in results achievable 170
June 2006
Published on behalf of the AIChE
with changes in input conditions (either individual elements or groups), thereby facilitating optimal management decisions. These changes can result from either changes in asset status, or conscious adjustments in safety philosophy. • To set realistic safety targets.
LITERATURE CITED
1. International Association of Oil & Gas Producers (OGP), OGP Safety Performance Indicators 2004, Report Number 367, May 2005. 2. UK HSE (Health & Safety Executive), Preventing slips, trips and falls at work, UK HSE, Suffolk, United Kingdom, 1996. 3. D. Attwood, F. Khan, and B. Veitch, Can we predict occupational accident frequency? Process Safety DOI 10.1002/prs
Process Safety Progress (Vol.25, No.2)
4.
5. 6. 7. 8. 9.
and Environmental Protection, Trans IChemE, Part B, 84(B2), March 2005 pp 1–14. D. Attwood, F. Khan, and B. Veitch, Offshore Oil and Gas Occupational Accidents—What is Important?, Journal of Loss Prevention in Process Industries, In Press, Proof Corrected, available on line, December 20, 2005. W. Ott, Environmental indices: theory and practice, Ann Arbour Science Publishers, Ann Arbor, MI, 1978. International Association of Drilling Contractors (IADC), Website: www.iadc.org, 2005. Canada-Newfoundland and Labrador Offshore Petroleum Board (CNLOPB), Website: http://www. cnlopb.nl.ca/, 2005. Canada-Nova Scotia Offshore Petroleum Board (CNSOPB), Website: www.cnsopb.ns.ca, 2005. G.M. Smith, A simplified guide to statistics for psychology and education, Holt, Rinehart and Winston, New York, 1970.
Process Safety Progress (Vol.25, No.2)
10. R.C. Thompson, T.F. Hilton, and L.A. Witt, Where the safety rubber meets the shop floor: a confirmatory model of management influence on workplace safety, Journal of Safety Research 29 (1998), 15–24. 11. J.M. Tomas, J.L. Melia, and A. Oliver, A crossvalidation of a structural equation model of accidents: organisational and psychological variables as predictors of work safety, Work & Stress 13 (1999), 49 –58. 12. R. Hatfield, Extreme organising: a case study of Hibernia, JUST LABOUR, A Canadian Journal of Work and Society, York University Centre for Research on Work & Society, Volume 2, Spring, Toronto, Canada, 2003, pp. 14 –22. 13. Canadian Broadcasting Corporation (CBC), Terra Nova workers have a union, Website: www.cbc.ca, 16 April, 2003. 14. R. Billinton and R.N. Allan, Reliability evaluation of engineering systems: concepts and techniques, Plenum Press, New York, London, 1983.
Published on behalf of the AIChE
DOI 10.1002/prs
June 2006
171
