On 2014 Oct 22, Mark Jones commented:

The mortality effect reported in this study can be explained by time-dependent bias. Out of 1,859 included patients, 416 (25%) of 1,676 treated with NAI died compared to 76 (42%) of 183 of patients untreated. This provides a crude relative risk of 0.60 for NAI (40% reduction in mortality). However this crude analysis fails to take account of the time-dependent nature of NAI exposure. Another naïve way to compare groups is based on rate of events per person day in hospital. The median length of stay for NAI patients was 10 days compared to 6 days for untreated patients. If we assume mean length of stay is similar we have 1,676 x 10 = 16,760 person days in hospital at risk of death for patients treated with NAI, and event rate of 416 / 16,760 = 0.025 deaths per person day. In comparison we have 183 x 6 = 1,098 person days in hospital for untreated patients with event rate 76 / 1,098 = 0.069 thus mortality rate ratio = 0.025 / 0.069 = 0.36, an even larger protective effect of NAI on mortality (64%). However we need to take into account that patients received NAI hours or even days after hospital admission. The delay was a median of 1 day after admission as time from onset of symptoms to admission was a median of 3 days whereas time from onset to treatment was a median of 4 days. If we assume mean delay is also 1 day we need to subtract 16760 days from the person-days estimate of the NAI treatment group and add it to the person-days estimate of the untreated group. This now more accurately reflects the total number of person days at risk of death while untreated as well as total number of person days at risk of death after treatment with NAI. It also makes our estimates of deaths per person day: 416 / 15,084 = 0.028 after NAI treatment compared to 76 / 2,774 = 0.027 while untreated, with rate ratio 1.0, i.e. no difference between groups. Of course this is an approximate analysis and a better analysis would involve including NAI treatment as a time-dependent exposure in a hazards based survival model. It is concerning that the authors did not conduct an appropriate analysis and also that the journal review did not pick up this fatal flaw.

On 2014 Oct 22, Mark Jones commented:

The mortality effect reported in this study can be explained by time-dependent bias. Out of 1,859 included patients, 416 (25%) of 1,676 treated with NAI died compared to 76 (42%) of 183 of patients untreated. This provides a crude relative risk of 0.60 for NAI (40% reduction in mortality). However this crude analysis fails to take account of the time-dependent nature of NAI exposure. Another naïve way to compare groups is based on rate of events per person day in hospital. The median length of stay for NAI patients was 10 days compared to 6 days for untreated patients. If we assume mean length of stay is similar we have 1,676 x 10 = 16,760 person days in hospital at risk of death for patients treated with NAI, and event rate of 416 / 16,760 = 0.025 deaths per person day. In comparison we have 183 x 6 = 1,098 person days in hospital for untreated patients with event rate 76 / 1,098 = 0.069 thus mortality rate ratio = 0.025 / 0.069 = 0.36, an even larger protective effect of NAI on mortality (64%). However we need to take into account that patients received NAI hours or even days after hospital admission. The delay was a median of 1 day after admission as time from onset of symptoms to admission was a median of 3 days whereas time from onset to treatment was a median of 4 days. If we assume mean delay is also 1 day we need to subtract 16760 days from the person-days estimate of the NAI treatment group and add it to the person-days estimate of the untreated group. This now more accurately reflects the total number of person days at risk of death while untreated as well as total number of person days at risk of death after treatment with NAI. It also makes our estimates of deaths per person day: 416 / 15,084 = 0.028 after NAI treatment compared to 76 / 2,774 = 0.027 while untreated, with rate ratio 1.0, i.e. no difference between groups. Of course this is an approximate analysis and a better analysis would involve including NAI treatment as a time-dependent exposure in a hazards based survival model. It is concerning that the authors did not conduct an appropriate analysis and also that the journal review did not pick up this fatal flaw.