Some like to claim that a lack of health insurance kills lots of people.
A widely reported study by Wilper, Wollhandler, Lasser, McCormick, Bor, and Himmelstein was published in the American Journal of Public Health (AJPH) in December 2009. The statistical analysis concludes that the uninsured have higher annual death rates than the insured. The higher death rate among the uninsured, according to the authors, is associated with nearly 44,789 annual deaths in the U.S. Co-author Dr. David Himmelstein of Harvard Medical School proclaims in an interview with CNN that for those without health insurance, "it means that you're at mortal risk." Many news outlets discussed this study, with reports of 45,000 deaths annually being linked to lack of health insurance. The study (a draft proof is available online at the Physicians for a National Health Program) used data from the Third National Health and Nutrition Examination Survey (NHANES III). This dataset is unique, as it contains interview, health examination, and insurance status records for nearly 34,000 participants, collected between 1988 and 1994. The NHANES III participants are matched by identifiable information to the National Death Index (NDI) by the National Center for Health Statistics (NCHS). This data allowed the researchers to follow these participants through time to identify variables which affect health. The researchers used the NCHS data through December 31, 2000, allowing for the study of participants for six to twelve years after their initial interview.
A total of 9,005 NHANES III participants were identified as uninsured and included in the Wilper, et al. study. The researchers were able to follow the participants for a total of 80,657 person-years, which equates to an average of 8.96 years per participant.
Of the 9,005 participants, the researchers identified through the NDI that 351 had died by December 31, 2000.
The NHANES III interviews identified 2,350 participants who were uninsured and 6,655 who were insured. The uninsured were 16.2% of the studied population, which is similar to other conclusions on the rate of the health insurance in the U.S. (Note: 2,350 / 9,005 will not equal 16.2%, since the researchers excluded many insured groups from the study, such as the elderly on Medicaid. The uninsured were overrepresented in the study.)
To calculate how many observed deaths occurred in each category (uninsured vs. insured), we need to look at Table 1 of the study. Since the researchers did not quantify the number of deaths in each group, we will need to back-calculate the status of observed deaths.
The study identifies that 17.2% of the 351 deceased were uninsured. This equates (0.172 X 351) to 60.4 uninsured persons having died in the study. Since 16.2% of the population was uninsured, we would have expected 56.9 persons (0.162 X 351) to have died if insurance status had no effect. The researchers should expect 56.9 uninsured persons to have died, but 60.4 uninsured persons actually died (a net increase of 3.5 deaths) during the nearly nine-year study.
Alternatively, the researchers state that 3.1% of all participants, and a higher percentage of the 2,350 uninsured participants (3.3%) died in the analysis. This equates (0.033 X 2350) to 77.6 uninsured persons having died in the study. We would have expected (0.031 X 2350) that 72.9 uninsured persons (a net increase of 4.7 deaths) would have died in the nine-year study if insurance status had no effect. (Note: this is higher than the previous 3.5 deaths, likely due to the over-sampling of the uninsured.)
We can conclude that the researchers observed 3.5 to 4.7 additional deaths in the uninsured participant population during the 8.96-year study. Don't ask how we have fractions of persons. Since the only whole number between the 3.5-4.7 interval is 4, let's take a leap of faith and say the researchers observed four additional deaths during the nearly nine-year study (or one death for every two years of the study). The research then models the population through sensitivity analyses to conclude that an additional 44,789 deaths per year may be associated with the lack of health insurance in the U.S.
This is similar to stating that since there were four observed deaths in the 8.96 year study, we can conclude that 401,309 (44,789 deaths/year X 8.96 years) deaths would occur in the uninsured American population in the same timeframe. There are just too few additional deaths observed in the study to make such a claim.
We may need to follow individual participants to determine the causes of their deaths and understand if health insurance could have made a difference. Full access to the NHANES III data on insurance status is not available to the public. However, the linked information for deaths and causes is available. From the data, we can obtain specific causes of death, such as participant #122, who died from an accidental fall (ICD-10 code 118, starting in column 25).
Insurance status has little to no impact on participants who die violently, as these persons would get emergency care, which is not dependent on heath insurance. If the four excess deaths we identified were from violent causes, there would be no underlying correlation between insurance status and early death.
We can identify forty persons who died from motor vehicle accidents (ICD-10 code 114) in the NHANES III study (participants #646, #1107, #1475, #2441, #2528, #3384, #3859, #3867, #4111, #5257, #9786, #10504, #12302, #12980, #13419, #13553, #16145, #16862, #18014, #18936, #19288, #19695, #19858, #33112, #33661, #36565, #39275, #40308, #42863, #43194, #43438, #46507, #46882, #47250, #47905, #49168, #49472, #51896, #52206 and #53476). The availability of insurance would have had little impact.
We can also identify eleven persons who died from suicide by firearm (ICD-10 code 125 -- participants #3947, #6138, #10655, #14336, #15860, #18222, #37902, #42061, #47057, #48163 and #48495). Health insurance would likely have had little impact on their deaths, either.
Statistical studies are onerous. The researchers of this study used the best available tools in very valid methods. Researchers must take into account that when looking at such small subsets of data (four additional deaths in nine years), claims of "mortal risk" become more difficult to ascertain regarding the population of the U.S. as a whole. Wilper, et al. should discount deaths from violent causes before making claims on the correlation of death and health insurance, especially when the net difference in deaths between the insured and uninsured groups is likely only four deaths during a nine-year study.