Objective To estimate the association between wearing a personal floatation device (PFD) and death by drowning among recreational boaters.
Design Matched cohort study analysis of Coast Guard data.
Setting United States.
Subjects Recreational boaters during 2000–2006.
Main outcome measures Risk ratio (RR) for drowning death comparing boaters wearing a PFD with boaters not wearing a PFD.
Results Approximately 4915 boater records from 1809 vessels may have been eligible for our study, but because of missing records and other problems, the analysis was restricted to 1597 boaters in 625 vessels with 878 drowning deaths. The adjusted RR was 0.51 (95% CI 0.35 to 0.74).
Conclusions If the estimated association is causal, wearing a PFD may potentially prevent one in two drowning deaths among recreational boaters. However, this estimate may be biased because many vessels had to be excluded from the analysis.
- Cohort studies
- conditional Poisson regression
- epidemiologic methods
- epidemiologic research design
- matched analysis
- missing data
- personal floatation device
- selection bias
Statistics from Altmetric.com
- Cohort studies
- conditional Poisson regression
- epidemiologic methods
- epidemiologic research design
- matched analysis
- missing data
- personal floatation device
- selection bias
Although personal flotation devices (PFDs), including life vests or jackets, are widely recommended for boaters, there are few estimates of their effectiveness.1 2 The United States Coast Guard collects data regarding deaths of recreational boaters. Our study goal was to use these data to estimate the association between wearing a PFD and drowning death among boaters in the USA. We planned to use a matched cohort design,3–5 comparing the outcomes of persons from the same boat who were involved in incidents such as capsizing or sinking which resulted in being in the water and at risk of drowning. Comparison of persons from the same vessel, who can be considered a matched set, has the advantage of comparing people at the same time of day, in the same water temperature, in the same water conditions, with the same distance to shore, and with the same proximity to help. Matched cohort methods have been used for many studies of traffic crashes,3 4 6 comparing persons in the same vehicle, to estimate the effectiveness of seat belts,7–11 air bags12 and motorcycle helmets.13
The Code of Federal Regulations requires that, in the event of death, injury or substantial property damage involving a recreational vessel, the operator or owner must provide details to government agencies such as police, marine patrols, game wardens or the Coast Guard itself.14 This information is entered into the Coast Guard's electronic Boating Accident Report Database. The Coast Guard provided us with electronic data for all 50 states, plus the District of Columbia, American Samoa, Guam, Northern Mariana Islands, Puerto Rico and the Virgin Islands, for the years 2000 through 2006. This study was approved by the Human Subjects Review Board of the University of Washington.
The electronic file had a record for each person in the vessel or event with these variables: vessel number, person number, state, date, vessel details (type, length), water conditions (calm or rough) and data for each person—age, sex, and whether they were the operator, wore a PFD, died or drowned. Each record had a narrative text; the longest was 4000 characters. The narrative sometimes had information about alcohol use, alcohol blood tests and swimming ability, but this information was not consistently collected.
US mortality data
To estimate the completeness of the Coast Guard data, we used US compressed mortality files15 to count all deaths related to watercraft using ICD-10 codes V90 through V94 for deaths that occurred during water transport.16 We omitted deaths involving merchant ships (V code fourth character 0.0) and passenger ships (fourth character 0.1), as they are not recreational vessels.
The study outcome: drowning death
We attributed death to drowning even if other factors contributed; for example, hypothermia might have contributed, but if drowning occurred, we considered this to be possibly preventable by wearing a PFD. If someone had a head injury and drowned, we coded that as a drowning outcome because a PFD might have kept the person on the surface and prevented the death.
Selection of records for analysis
We wished to compare people who entered the water from the same vessel. In a matched cohort analysis, no information is gained from matched sets (people from the same vessel) in which no one has the study outcome. Therefore, we selected the records of vessels with at least one death; a selected vessel could have more than one death, and in many vessels all the boaters died. We read the narratives for all selected vessels along with a listing of the variables and deliberately eliminated some records for the following reasons.
Death was due to trauma.
Death was due to another mechanism, such as carbon monoxide poisoning or obvious myocardial infarction.
The narrative indicated that only one person was in the water. These vessels were omitted, as we wished to compare two or more persons in the water.
People entered the water intentionally. Sometimes boaters left their vessel in order to swim or fix a motor. We do not expect all swimmers to wear PFDs, so we did not expect usual PFD wearing would prevent these deaths. Sometimes boaters left the vessel urgently to save a person or dog in the water; wearing a PFD might have been possible and helpful for these boaters. We eliminated vessels with persons who entered the water for pleasure or non-emergencies, but retained vessels of persons who dived overboard in a rescue attempt.
Sometimes the data included people who never entered the water or people, such as an owner, who were not in the vessel at all. When possible, we eliminated these excess records, but kept the eligible vessel and individual data in the analysis.
We then excluded some vessels for the following additional reasons.
The narrative revealed that there were more people in the water than there were records. For example, a boat submerged leaving four people in the water, but there were only two records. Often there were only records for the dead. A matched cohort analysis can produce biased estimates if some persons in each matched set (each vessel) are omitted. For example, an analysis that omitted many survivors who wore PFDs would tend to underestimate any protective effect of PFDs or even estimate that PFDs were harmful.
There was no narrative or we could not determine that death was due to drowning from the available narrative.
There were more records than boaters in the water and we could not determine which records were superfluous.
Other reasons—for example, some vessels were commercial, not recreational.
The exposure of interest: wearing a PFD
The data had variables for ‘PFD worn’ and ‘PFD used’. The meaning of PFD used, as distinct from wearing a PFD, was unclear. In the 9925 boater records from vessels with at least one death, PFD worn was missing for 52% and PFD used was missing for 84%. When the two variables were jointly known, they were both coded ‘no’ for 1557 records, both ‘yes’ for two records, PFD worn ‘no’ and used ‘yes’ for 55 records, and PFD worn ‘yes’ and used ‘no’ for two records. We defined the exposure of interest as wearing a PFD; holding on to a PFD or lying on top of a PFD were not considered to be wearing. If someone donned a PFD after entering the water, we coded them as wearing a PFD; only a few boaters did this.
We planned to adjust study estimates for both age and sex, but data were often missing for both these variables, as well as for whether or not a PFD was worn. We used the narratives to fill in some missing information for these three variables. We read the narratives for every record used in our analyses, and we corrected coded information if the narrative disagreed with the coding. Then we used multiple imputation17 18 to create 50 sets of data which were identical with respect to known information, but could differ on imputed values for missing information about age, sex and wearing of a PFD. We used the method of chained equations (regression switching), sampling imputed values from the posterior predictive distribution of the missing data.19–21 The multiple imputation model included variables for drowning, wearing a PFD, age, sex, age-squared and interaction terms between sex and age, sex and PFD wearing, and a term for the proportion of other occupants of the same vessel who were known to be wearing a PFD.
To compare the risk of drowning death of persons wearing a PFD with those not wearing one, we estimated risk ratios (RR) comparing the persons in each vessel. This was done using conditional Poisson regression.3 5 22 We used a robust variance estimator to account for the fact that the outcome was binary and therefore did not fit a Poisson distribution.22 23 RRs were adjusted for sex and age; for age we adjusted using quadratic splines with knots at ages 20, 40 and 60 years and the tails constrained to be linear.24 We generated RR estimates in each of the 50 sets of known and imputed data and combined the results by averaging the RRs on the log scale and creating CIs that accounted for variation within and between each set of data.
For 2000–2006, the Coast Guard provided electronic records for 104 683 persons in 46 234 vessels. Over the 6-year period there were 21% more boating-related deaths in Coast Guard data than in US mortality files (table 1).
There were 9925 person records for 4242 vessels with at least one death; 48% (4782/9925) of these boaters died and 2899 of the deaths were attributed to drowning. We reviewed the narratives of the 4242 vessels (9925 person records) with a death. We removed 1148 person records because a death was due to trauma, 214 because death was due to a mechanism other than trauma or drowning, 2981 because only one person was in the water, 82 because boaters entered the water intentionally, and 585 for people who were either not in the water or not even in the vessel. This left 4915 records from 1809 vessels with 1839 drowning deaths; this sample may have been close to the sample we wished to analyse. However, we had to remove an additional 1286 records because the number of records was less than the number of boaters in the water, 1906 because the narrative was unclear, 84 because there were more records than people in the water, and 42 for other reasons. The final analysis sample had 32% (1597/4915) of the records, 35% (625/1809) of the vessels, and 48% (878/1839) of the drowning deaths compared with approximate counts for the sample we wished to analyse.
In the final sample of 1597 boaters in 625 vessels, PFD wearing was missing for 34% of boaters, sex for 53%, and age for 19%. After filling in missing data using information from the narratives, PFD wearing was still missing for 357 (22%) of the boaters, sex for 462 (29%), and age for 283 (18%); 828 (52%) still had missing data for at least one of these three variables. The number of people in the water from each vessel ranged from two to nine. Many (317/878=36%) of the drownings were from a vessel with no survivors. A few (95) records were from vessels in which someone went overboard to rescue a person or pet, and 25 boaters left a disabled vessel in an attempt to swim to shore. Only 21% (336/1597) of the boaters wore a PFD (table 2). PFD wearers were more often female and less often aged 20–39 years compared with persons who did not wear a PFD. Among those who wore a PFD, 45% drowned, compared with 58% of boaters who did not wear a PFD.
Using the known and imputed data for analysis, and ignoring matching on vessel, the unadjusted RR for drowning among those wearing a PFD, compared with those who did not, was 0.45/0.58 (table 2)=0.78 (95% CI 0.68 to 0.91). When we accounted for matching on vessel, thereby comparing persons from the same vessel, the unadjusted RR was 0.50 (95% CI 0.34 to 0.71). After further adjustment for sex and age, the matched RR was 0.51 (95% CI 0.35 to 0.74). When we omitted the 120 boaters in 50 vessels with a person who entered the water to attempt a rescue or to swim to safety, the adjusted RR was 0.53 (95% CI 0.36 to 0.77). For comparison, when we analysed only the 201 vessels with complete data for their 497 boaters and accounted for matching, the adjusted RR was 0.49 (95% CI 0.31 to 0.78).
We estimated that wearing a PFD reduced the risk of drowning by 49% (95% CI 26% to 65%). However, for several reasons, this estimate may not be accurate.
Possible selection bias
We hoped to estimate the RR for PFD wearing using all US recreational vessels in which two or more persons entered the water and at least one drowned. There were 1809 vessels in the data that may have met our criteria. However, we had to remove most (1184/1809=65%) of these vessels from our analysis because we could not tell if they met our study criteria or because the number of person records was too small (boaters were missing from the data) or too large (some people in the data were not in the water or even in the vessel). Our adjusted RR estimate for PFD wearing was based on a selected subset of 625 vessels; our RR estimate may not be what we would have found if we could have analysed data from all vessels that met our study criteria. The direction and severity of any bias is unknown, but it could be substantial.25 26
Possible confounding bias
Because a matched analysis compares people from the same vessel, our RR estimates were not biased by factors that were shared by boating companions, such as water temperature, weather, visibility or distance to shore. We also adjusted for age and sex. However, we could not adjust for some variables that may be important. Alcohol use has been associated with the risk of drowning among boaters.27 If intoxicated boaters wear PFDs less often than other boaters, failure to adjust for alcohol use may exaggerate any benefit of wearing a PFD. Failure to adjust for swimming ability may also produce bias. Information about alcohol use and swimming ability was sometimes in the Coast Guard narratives, but it was not collected consistently and there were no variables for these factors.
Whether a PFD was worn was missing for 22%, sex for 29%, and age for 18% of boaters included in our final analysis. Analysis of records with known information only may be biased unless the data are missing completely at random. We used multiple imputation methods, which require the less stringent assumption that data were missing at random, conditional on the values of variables used for imputation.17 18 However, bias could still be present if missing values were strongly related to unmeasured characteristics of boaters.
Australian investigators selected boating mishaps with two or more people in the water and at least one drowning death.1 They estimated a crude OR for drowning of PFD wearers versus others: 0.47 (95% CI 0.25 to 0.88). The authors described their design as a case–control study in which cases and controls experienced the same ‘environmental conditions’. But in a case–control study matched on boat, each case of death should be matched to one or more surviving controls from the same boat. Furthermore, it is usually necessary to account for the matching in the analysis.28–31 If all occupants of a boat died, the Australian study could not match those cases to any controls. In our data, 36% of the dead were from a boat with no survivors; failure to match these cases to any living control could bias a case–control study. If the Australian data are treated as a matched cohort study, we can estimate3 4 29 30 (from their table 3) that the crude RR for drowning of PFD wearers compared with others was 0.64 (95% CI 0.42 to 0.97). This crude estimate may still be biased, as it ignores matching on boat and is not adjusted for any other variables.
A case–control study from China estimated that children aged 1–4 years who wore PFDs when playing near water were less likely to drown than other children: adjusted RR 0.43 (95% CI 0.22 to 0.71).2 It is unknown whether this estimate is relevant for adults or boaters.
Suggestions for the future
Statistical methods for estimating the effectiveness of PFDs comparing persons from the same vessel are available. The Coast Guard has a nation-wide system for collecting data about drowning of recreational boaters. These data have the potential to facilitate important injury prevention studies. Such studies would benefit from (1) improved completeness of records for all vessel occupants, (2) better identification of records for vessel occupants, (3) reduction in missing data, (4) addition of variables for alcohol use and swimming ability, and (5) a formal dictionary with definitions for each variable. A model for a reporting system is the Fatality Analysis Reporting System data collected by the National Highway Traffic Safety Administration since 1975.32 33 Improvements in data quality would allow more credible estimates of PFD efficacy in the future.
What is already known on this subject
Two studies have estimated the association between wearing of a personal floatation device (PFD) and death by drowning.
The estimates from one study are difficult to interpret, as the analysis did not fit the study design. The other study was limited to children aged 1–4 years who were playing near water.
What this study adds
We used data from the US Coast Guard for recreational boaters during 2000–2006 and estimated the association of PFD wearing with drowning death by comparing persons in the same boat.
Wearing a PFD appeared to prevent about one of two deaths that would have occurred otherwise in the vessels that could be analysed.
However, this estimate may be biased, as many vessels had to be excluded from the analysis because there were records for people who were not in the vessel, records were missing for boaters who were in the vessel, or the cause of death was unclear.
We thank the United States Coast Guard for providing us with the data for this study as well as information about these data. The paper is solely our responsibility and does not necessarily represent the official views of the United States Coast Guard. We thank Michael Terasaki who helped review the report narratives and code records for analysis.
Funding This work was supported by grant 5R49CE000197-05 from The Centers for Disease Control and Prevention, Atlanta, Georgia, USA. Its contents are solely the responsibility of the authors and do not necessarily represent the official views of The Centers for Disease Control and Prevention.
Competing interests None.
Ethics approval This study was conducted with the approval of the University of Washington, Seattle, Washington, USA.
Provenance and peer review Not commissioned; externally peer reviewed.
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.