Statistical analysis
We used the oxygen dissociation curve and Kelman’s
equation to determine whether blood gas samples were
arterial or venous if this had not been recorded.26 We
compared the measured arterial oxygen concentrations
with the concentrations calculated from pulse oximetry
at the time of the blood gas sample. We took a difference greater than 5% between the calculated and measured oxygen concentrations to indicate that the blood
gas was a venous sample. For venous samples, we used
the formulations described by Ak to estimate arterial
values for pH, carbon dioxide, and bicarbonate.27 We
used non-paired Student’s t tests to compare arterial
and converted venous blood gases. As no significant
differences existed between arterial or converted
venous samples for pH, carbon dioxide, or bicarbonate, we combined them for subsequent analyses. We
could not convert venous oxygen concentrations, as
comparative arterial oxygen concentrations are not
provided in the conversion equations, and we thus
report only measured arterial oxygen concentrations.
The primary outcome was prehospital and in-hospital
mortality. We used log binomial regression to compare
the risk of death for patients in the two treatment arms.
We compared secondary outcomes—namely, length of
hospital stay and blood gas measurements (pH, arterial
carbon dioxide, bicarbonate, and arterial oxygen)—by
using Student’s t tests after transformation to remove
skewness. We combined ventilation categories (invasive and non-invasive), as the number of observations
was small, and used log binomial regression to compare
them. We based initial analyses on the intention to treat
principle and restricted a further set of per protocol analyses to patients who received treatment that complied
with the study protocol. We did all analyses for all study
patients and for the subgroup of patients with chronic
obstructive pulmonary disease confirmed by lung function tests. We adjusted standard errors to account for
any correlation between observations for the same