Data source
We performed this retrospective cohort study using the Diagnosis Procedure Combination database. The details for this nationwide inpatient database were described elsewhere [4]. The database includes hospitalizations for all factors across all ages and has about 1,000 participating hospitals ranging from academic to community hospitals (about 7 million patients every year). All 82 academic hospitals are obliged to participate in the database, while participation by community hospitals is voluntary. Data for about 7 million hospitalized patients of all ages are collected every year, equivalent to about 50% of the total acute-care hospitalizations in Japan. The database includes the following information: unique identifiers of hospitals; patient baseline characteristics; and diagnosis at admission, comorbidities at admission, and complications after admission recorded with text data in Japanese and International Classification of Diseases, Tenth Revision (ICD-10) codes. The database also contains information on medical procedures and treatments, including drug administrations, use of devices, and surgical and nonsurgical procedures, as well as length of stay, discharge status, and medical cost for hospitalization. Diagnoses at and after admission are clearly differentiated in the database. The attending physicians are obliged to record patient data for all diagnoses and comorbidities with reference to the medical records. Furthermore, because accurate reporting is linked with the payment system in Japan, the attending physicians and hospitals are required to accurately report the diagnoses and comorbidities. The present study was approved by the Institutional Review Board of The University of Tokyo (approval number: 3501-(3); 25 December 2017). The requirement for informed consent was waived because of the anonymous nature of the data.
Participants
We used the database to identify patients who were diagnosed with KD (ICD-10 code: M303) between July 2010 and March 2017. Patients whose first IVIG treatment was started within 5 days of hospitalization and who received at least 2 g/kg IVIG within 3 days of starting their first IVIG treatment were included in the study. We further checked the Japanese text describing the detailed diagnoses in each case to include atypical KD patients and exclude patients with a ‘suspected’ diagnosis of KD, inadequate treatment (<2 g/kg IVIG), and age >6 years. The included children with KD in the acute phase were divided into two groups depending on their initial treatment with high-Na IVIG (154 mEq/L Na) or low-Na IVIG (0.09–2.60 mEq/L Na). Patients who received both high-Na and low-Na IVIG were excluded.
Outcomes
The primary outcome was the occurrence of coronary artery abnormalities (CAAs) at the time of discharge in the two groups. CAAs were defined according to the following criteria: (i) recorded diagnosis of CAAs (ICD-10 code: I254); (ii) use of warfarin or clopidogrel; or (iii) cardiac catheterization. The secondary outcomes were IVIG resistance, length of stay, and medical cost. IVIG resistance was defined as use of IVIG at a total dose of ≥4.0 g/kg or a combination of infliximab, cyclosporine, or plasma exchange.
Covariates
The baseline characteristics were age, sex, weight, height, hospital days of illness at initial IVIG, type of hospital, complex chronic conditions [5], transportation by ambulance, activities of daily living at admission, Japan Coma Scale at admission, transportation from other hospital, fiscal year, additional treatment, and hospital volume. The Japan Coma Scale scores were categorized into two groups: alert and not alert. Japan Coma Scale assessment was previously shown to be well associated with Glasgow Coma Scale assessment [6]. Hospital volume was defined as the mean annual number of all KD patients at each hospital. We categorized the eligible patients into tertiles of hospital volume so that the numbers of patients in the groups were almost equal. Additional treatment was defined as: any steroid, infliximab, or cyclosporine, or plasma exchange beyond day 5 after starting the initial IVIG treatment. Because the initial IVIG treatment usually took 1–2 days and a further 1–2 days were required to assess the patient’s response, we chose 5 days for the definition of additional treatment.
Statistical analysis
Categorical variables are shown as number and percentage and were compared using Fisher’s exact test. Continuous variables are shown as mean and standard deviation (SD) or median and interquartile range (IQR). The Mann–Whitney U test was used to compare non-normally distributed variables between the two groups. To obtain meaningful interpretation of variables with skewed distributions, we calculated the median difference with 95% confidence interval (95% CI) using the Hodges–Lehmann estimator.
Propensity score-matched analysis
We conducted 1:1 propensity score-matched analyses to compare the outcomes between the two groups. For the propensity score matching, nearest-neighbor matching without replacement was performed. The caliper width was set at ≤0.2 of the pooled SD of the estimated propensity scores. We examined the covariate balance between the two groups before and after the propensity score matching using the absolute standardized difference. An absolute standardized difference of >10% was regarded as imbalanced [7].
The chi-square test was used to compare the proportions of CAAs and IVIG resistance between the two groups. We also estimated the risk differences and 95% CIs. The Mann–Whitney U test was used to compare length of stay and medical cost.
Instrumental variable analysis
Propensity score-matched analyses cannot eliminate the effects of unmeasured confounders such as laboratory data. To confirm our propensity score-matched analyses, we performed instrumental variable analyses. The key assumptions for instrumental variable analysis are that the instrumental variable: (i) is highly correlated with the treatment assignment, (ii) is not correlated with other confounders, and (iii) does not affect patient outcomes except through the treatment [8,9].
Instrumental variable analysis methods often utilize “physician prescribing preference” as the instrumental variable. This variable reflects the notion that a physician’s prescribing decision depends on not only the patient characteristics but also the physician’s preference for a specific medicine. This means that the physician’s preference is largely independent of the patient characteristics and outcomes and can therefore serve as an instrumental variable. One commonly used method to determine the physician’s preference is to employ the prescription for the last patient treated by the physician as the preference for the current patient [10].
In the present study, we used “prescription for the last patient” as the instrumental variable. Specifically, when the prescription for the last patient in the same institution was high-Na IVIG, the current patient was assumed to be more likely to receive high-Na IVIG. Conversely, when the prescription for the last patient in the same institution was low-Na IVIG, the current patient was assumed to be more likely to receive low-Na IVIG. The prescription for the last patient is assumed to be independent of the current patient’s characteristics and not directly related to the outcome. Therefore, this instrumental variable was considered to meet the above-described three key assumptions for an instrumental variable.
We used a two-stage residual inclusion method for both continuous and binary outcome variables [11,12]. In the first-stage model, we determined the row residual for each patient by calculating the difference between the model-predicted probability of receiving the treatment choice and the actual treatment received. The residuals were included as an additional covariate in the second-stage model. In the second-stage model, the association between treatment choice and outcome was estimated in an unbiased manner, after adjustment for covariates. We used a multivariable linear regression model for continuous outcome variables and a multivariable logistic regression model for binary outcome variables. All instrumental variable analyses were performed using robust standard errors. To assess the validity of the instrumental variable, we tested its association with our main predictor of actual treatment choice using the F-statistic (F-statistic >10 is considered to reflect a valid instrumental variable) [13]. We also determined that the instrumental variable was not associated with the patient background characteristics and outcomes. A two-sided p<0.05 was considered significant. All statistical analyses were conducted using Stata software version 16.1 (StataCorp LP, College Station, TX, USA).