We utilize data from the Norwegian Patient Registry (NPR). NPR contains complete patient-level observations (diagnosis, exact date and place for admission and discharges, the degree of urgency at arrival at the health institution) for all somatic public hospitals and private hospitals with contracts with regional health authorities. We link NPR with nationwide individual-level data from Statistics Norway (age, gender, immigration status, municipality and district of residence, date of death) using the encrypted version of a unique personal identification number issued to every resident of Norway at birth or upon first immigration.
Study sample and treatment variables
Our sample consists of individuals diagnosed with AMI for the first time in at least two years in 2010-2015, using data from 2008-2015. We compare health outcomes for patients diagnosed with AMI who are sent directly to a PCI-hospital with similar patients who are sent to a non-PCI-hospital (and may or may not be transferred to a PCI-hospital). To do this, we first identify all hospitalization spells that overlap with the admission where the patient is diagnosed with AMI as their primary diagnosis (hereafter called AMI spell). Specifically, an AMI spell includes all hospital admissions where there is no more than one day between last discharge and the next admission and where the patient is diagnosed with AMI in at least one of the admissions covered by the spell. We only include AMI spells where the first admission in the spell is registered as acute. A patient is defined as sent directly to a PCI-hospital if the first admission in the spell was to a hospital performing PCI, and directly to a non-PCI-hospital if the first hospital in the spell was a local hospital not performing PCI.
The instrument used is the historical municipal share of acute AMI patients sent directly to a PCI-hospital. For each AMI patient in our sample we calculate the instrument as the historic share of acute AMI patients sent directly to a PCI-hospital in the same municipality (and the 15 city districts in the municipality of Oslo) in the 365 days preceding the current AMI spell of the patient. Since the main sample consists of AMI patients from 2010-2015, we use data from 2009-2015 in this calculation. If there were fewer than ten AMI patients admitted in the municipality the previous year, the given individual is excluded from our sample. This was the case for 8 280 individuals, and leaves us with a sample of 53 773 individuals residing in 338 municipalities.
To compare patients sent directly to a PCI-hospital and directly to a hospital not performing PCI, we link to data on gender, age and date of death as well as municipality (district in Oslo) of residence at the beginning of the year of AMI spell.
Statistical analyses
First, we show that being sent directly to a PCI-hospital is not random by showing the share of AMI patients sent directly to a PCI-hospital in each municipality (using the sample of 53 773 patients). Second, to study the effect of initially being sent to a PCI-hospital on mortality we use several multivariate risk adjustment methods as well as the instrumental variable approach. In all models we control for observable characteristics such as age (dummy variable for yearly age groups), gender, and interactions between these, as well as time and municipality fixed effects. Importantly, the municipality fixed effects control for any time-invariant differences between the municipalities, such as absolute and relative distance to different hospitals.
The multivariate risk adjustment methods used to estimate the effect of being sent directly to a PCI-hospital on the probability of death, are linear probability (OLS) and logit models. The implemented model is illustrated by the OLS-equation (
1):
$$\begin{aligned} Y_{i} = \beta _1\ H _i +\beta _3\ X_i + u_i \end{aligned}$$
(1)
where
\(Y_i\) is a dummy variable equal 1 if the patient is dead within a set time period, say 1 month, after the start of the AMI spell (otherwise 0).
\(H_i\) takes the value 1 if individual i is initially sent to a PCI-hospital and 0 of the individual is initially sent to a non PCI- hospital.
\(X_i\) is a vector of observable characteristics for individual i, including age (dummy variable for yearly age groups), gender,interactions between these. We also include time and municipality fixed effects.
These models identify the effect of being sent directly to a PCI-hospital on a health outcome under the strong assumption that the variation in who are sent directly is (conditionally) uncorrelated with unobserved determinants of the health outcome. However, it is likely that the observable confounders will not capture all differences in health that affect the decision of whether or not to send a patient directly to a PCI-hospital. First, the treatment guidelines are different for different types of AMI. While patients with STEMI heart attacks have the most acute condition, they are more often diagnosed in the ambulance, and have higher probability of being sent directly to a PCI-hospital. Since STEMI patients also have higher in-hospital mortality, standard regression results would be expected to be biased towards finding worse health outcomes for those sent directly to a PCI-hospital. On the other hand, while evidence suggests that timely invasive management strategies primarily benefit elderly or high-risk patients, several studies have found that this intervention in practice is directed to lower-risk patients [
10]. Hence, estimates adjusting for observable characteristics are likely contaminated by omitted variable bias, though the direction of that bias is not clear.
To address concerns of omitted variable bias and endogeneity, we apply the instrumental variable (IV) model which can be illustrated as follows:
$$\begin{aligned} H_i = \pi _1 \ Z_i +\pi _3X_i +v_i \end{aligned}$$
(2)
$$\begin{aligned} Y_i =\gamma _1\ \hat{H}_i + \gamma _3X_i +w_i \end{aligned}$$
(3)
where the instrument
\(Z_i\) is historical municipal share of patients being sent directly to a PCI-hospital over the preceding year. Hence, since comparing patients with respect to the actual treatment received (sent directly to a PCI-hospital or not) may be biased by health characteristics of the specific patient, our instrumental variable analysis compares groups of patients that differ in the likelihood of being sent directly to PCI-hospital for reasons not related to the health condition of the specific patient.
H is a dummy variable set to 1 if patient
i is sent directly to a PCI-hospital, and 0 otherwise, and
\(\hat{H}\) is the predicted probability of being sent directly to a PCI-hospital. Note that the instrument takes on the same value only for patients in the same municipality whose AMI spell starts on the same day, and it can thus be different for patients in different municipalities or in the same municipality on different days.
The first-stage regression in Eq.
2 estimates to what degree the instrument affects the probability of being sent directly to a PCI-hospital. The second stage regression in Eq.
3 provides the main parameter of interest,
\(\gamma _1\), which captures the local average treatment effect (LATE) of being sent directly to a PCI-hospital, instead of first going to a non-PCI-hospital, for patients for whom the hospital they are sent to (PCI or not) shifts as a result of variation in the practice of the municipality (i.e. in the instrument
\(Z_i\)). Like in the multivariate risk adjustment models, we control for a vector of observable characteristics (
\(X_i\)) and time and municipality fixed effects. We estimate the LATE using two-stage least square (2SLS), and cluster on municipality level.
In order for the identification strategy to be valid, the independence assumption must hold, meaning that our instrument should be uncorrelated with individual patients’ potential outcomes. If this assumption holds, the intention to treat estimates (ITT) of the effect of the instrument on the outcome will estimate causal effects of being sent directly to a PCI-hospital vs. first to a non-PCI-hospital. This seems reasonable in our situation, as the share of preceding patients who are sent to a PCI-hospital is not likely to be determined by characteristics of future individual patients, which is also suggested empirically: Table A
1 shows that, as expected, observable characteristics of the patients are predictive of actually being sent directly to a PCI-hospital, but it also shows that the observable characteristics of the patients are not predictive of previous patients in the municipality being sent directly to a PCI-hospital (i.e. the instrument).
1
To identify the local average treatment effect (LATE), the so-called exclusion restriction must hold. It requires that the instrument only affects the outcome of interest through the treatment. Put differently, the health outcome (mortality) of a given patient should not be directly affected by the share of previous patients admitted directly to a PCI-hospital. Congestion, where the current patient is sent to a non-PCI-hospital because the previous patients were sent to a PCI-hospital, would lead to a violation of this assumption. To take this into account we do not include admissions at the same day as the given patient when we calculate the instrument (i.e., we calculate a leave-out mean). Also the instrument must be highly correlated with the probability of being sent directly to a PCI-hospital, which we can easily confirm empirically in the first stage in Eq.
2. The monotonicity assumption, requires that the instrument should (weakly) change the probability of treatment in the same direction. In our setting, we need to assume that a higher rate of patients being sent directly to a PCI-hospital (weakly) increases the likelihood of future patients being sent directly to a PCI-hospital. A possible violation of this assumption occurs if healthcare professionals experience that the health outcomes for patients sent directly are poor and therefore send fewer patients directly to a PCI-hospital in the future. However, we argue this is unlikely as emergency medical personnel often do not observe the health outcome of the patient.
If these assumptions hold the estimate of \(\gamma _1\) captures the local average treatment effect (LATE) of being sent directly to a PCI-hospital for the compliers. In our setting with municipality fixed effects, compliers are patients who are sent directly to a PCI-hospital because there is a change over time in the municipality’s inclination of sending patients directly to a PCI-hospital.
Analyses were performed by using STATA 16.