Table 1 Overview of bias concerns, consequences and adjustments in the PACT intervention study, Norway 2014–16

Survival / Lead-time bias

More patients (69%) in the intervention group than in the control group (19%) had Lead Days in the hospital before inclusion.

Mortality: Intervention patients must survive lead days to be referred and included in PACT. Survivors may be healthier and cause a survival bias.

Sum emergency inpatient bed-days: In controls, we count “inpatient days” from the first day of emergency admission. In PACT patients, we start counting from the time of referral to PACT, leaving out emergency Lead days before referral. If left unadjusted, this would bias comparisons towards a lower sum of emergency days in intervention patients.

Mortality risk analyses:

Restricted sub-group analyses to control-patients with a survival time equal to or greater than median Lead-days in the intervention group.

Sum emergency inpatient days:

We adjusted for Lead Days so that effect estimates are independent of prior lead days. We tried matching on lead-days, which would be the preferred avenue, but we could not find enough matching controls for this.

We restricted analyses of sum emergency bed-days to PACT patients with an index emergency hospitalisation so that both groups add emergency bed-days from their index episode to the 6-month outcome measure.

Indication bias

Referral to the PACT intervention is less likely for terminal patients, or patients they judge to be unsuitable for the intervention for other reasons. In the control group, providers are likely to refer all other patients, including terminal patients to emergency admissions who then become eligible to be controls. We have no data, on the judgements made by referring professionals in either group.

Adjustment for possible under-referral of terminal patients to the intervention: We used the Elixhauser death risk score and the modified (m)-PARR30 score for both matching and adjustment. The C-statistic was 0,74 and 0,71 for death within six months in a local hospital population for these two predictors respectively.

We made sub-group analyses restricted to control-patients who survived median Lead-days to exclude terminal controls who died in their first few days in the hospital.

We estimated crude mortality risk in intervention patients that the PACT-team excluded since these might include terminal patients.

Healthy selection bias:

69% of the intervention and 100% of the control patient index episodes were linked to an emergency admission. Intervention patients who had no index-episode emergency hospitalisation may be healthier than controls.

Sub-group analyses restricted to intervention patients whose index episode was an emergency admission.