This retrospective cohort study used patient data from the Diagnosis Procedure Combination database. As of November 2022, this nationwide database contained hospital administrative claims data and discharge abstracts of approximately 8,000,000 inpatients from more than 1200 acute care hospitals[22]. Participation in this database is compulsory for academic hospitals and voluntary for community hospitals. The database contains the following information: patient’s age, sex, body height, body mass index (BMI), smoking history, primary diagnosis, comorbidities at admission (International Classification of Diseases [ICD]-10 code), prescription information, tumor–node–metastasis stage of malignant tumor, length of hospital stay, discharge status, activities of daily living (ADL) following hospitalization at admission and discharge, and total hospitalization costs. The recorded diagnoses in the database were validated; for example, the specificity of lung cancer diagnosis was 96.7%, while the sensitivity was 50–80%. The specificity and sensitivity of the recorded procedures exceeded 90% [23, 24].

We identified inpatients who received cisplatin- or carboplatin-based chemotherapy for lung cancer (ICD-10 code: C34) and olanzapine 2.5 or 5 mg as CINV prophylaxis between January 2016 and March 2021. Prescription of olanzapine at chemotherapy initiation or before was regarded as prophylactic administration of olanzapine. We excluded patients (i) below 18 years old, (ii) with schizophrenia (F20, F22–25, F28, F29), (iii) with diabetes (F10–14) or who received treatment with insulin or oral hypoglycemic agents, or (iv) who received olanzapine for more than two days before chemotherapy or on day ≥ 5 after chemotherapy initiation, assuming that it could be for other purposes such as treatment for general anorexia, cachexia, and psychiatric symptoms [25, 26]. We divided the eligible patients into two groups: patients who were prescribed olanzapine 2.5 mg/day (the 2.5-mg olanzapine group) or olanzapine 5 mg/day (the 5-mg olanzapine group). The initial day of chemotherapy was defined as day 1.

The primary outcome was defined as additional antiemetic drug administration within the overall assessment period (days 2–5) and on each day (2, 3, 4, and 5). We considered the administration as a surrogate for CINV symptoms [9, 27, 28]. According to the antiemetics guidelines and common treatments in Japanese cancer hospitals, metoclopramide, domperidone, lorazepam, alprazolam, haloperidol, chlorpromazine, and prochlorperazine were considered additional antiemetic drugs [29, 30]. The secondary outcomes were dexamethasone use within days 2–5, length of hospital stay, and total hospitalization costs. One US dollar (USD) was equivalent to 110 Japanese yen.

This study was approved by the Institutional Review Board of the University of Tokyo (approval number 3501-5; May 19, 2021). The review board waived the requirement for patient-informed consent because of the use of anonymous data.

We examined the following patients’ characteristics: sex, age, BMI, smoking index (0/1–19/ ≥ 20 pack-years), Charlson comorbidity index [31], Parkinson’s disease, independence in ADL, and cancer stage. The cancer stage was made according to the TNM classification. Age was categorized into four groups: < 64, 65–74, 75–84, and ≥ 85 years. There were four BMI groups: underweight, < 18.5; normal weight, 18.5–24.9; overweight, 25.0–29.9; and obese, ≥ 30 kg/m2. Comorbidities were assessed using the Charlson comorbidity index and categorized into 2, 3, 4, or ≥ 5. We also categorized treatment history into radiotherapy, chemotherapy regimen (Supplementary Table 1), antiemetic regimen according to the guidelines of the American Society of Clinical Oncology and National Comprehensive Cancer Network (Supplementary Table 2), use of olanzapine-interacting drugs, number of chemotherapy cycles (one, two, three, or above three), support from a palliative care team, emergency admission, types of hospitals (teaching hospital or not), and fiscal year of admission. Regarding olanzapine-interacting drugs, we included corticosteroids (except dexamethasone), hypnotics (benzodiazepines, non-benzodiazepines, and other hypnotic drugs), barbiturates, selective serotonin reuptake inhibitors, serotonin-norepinephrine reuptake inhibitors, noradrenergic and specific serotonergic antidepressants, tricyclic antidepressants, tetracyclic antidepressants, other antidepressants, multi-acting receptor-targeted antipsychotic drugs (except for olanzapine 2.5 and 5 mg), serotonin-dopamine antagonists, dopamine receptor antagonists (phenothiazines, butyrophenones, and benzamides), anti-parkinsonian drugs (levodopa, levodopa-carbidopa, levodopa-benserazide, ropinirole, pramipexole, and rotigotine), carbamazepine, omeprazole, and rifampicin [16, 17, 32, 33]. We identified patients who received these medications within 7 days before chemotherapy initiation.

We performed a propensity score analysis to adjust for confounding by indication and to compare the outcomes between both groups. A propensity score analysis can effectively adjust for measured confounders and is used to balance patients’ backgrounds in a retrospective study [34]. We conducted propensity score matching at a 1:1 ratio. Propensity scores were calculated with a logistic regression model using the above-mentioned patient and treatment variables as covariates. Each patient who received olanzapine 2.5 mg was matched with a patient who received olanzapine 5 mg using the nearest-neighbor matching method without replacement. The caliper width was ≤ 0.2 of the pooled standard deviation of estimated logits of the propensity score. We calculated standardized differences to examine the balance in baseline covariates of patients between both groups. An absolute standardized difference below 10% denoted a negligible difference between both groups [35]. Continuous and categorical variables were compared using the t test and Chi square test, respectively.

We also performed the four subgroup analyses for the primary outcome. Olanzapine metabolism is reportedly strongly affected by sex, age, and smoking [16, 18,19,20]; therefore, we evaluated patients in the following subgroups: sex, age (< 65 or ≥ 65 years), and with and without a smoking history. Additionally, we evaluated the data of patients who received only a cisplatin regimen.

Finally, we conducted two sensitivity analyses to confirm the robustness of our results. First, we only included patients who were using three antiemetic drugs (a 5-HT3 receptor antagonist, an NK1 receptor antagonist, and dexamethasone). Second, we conducted an instrumental variable analysis to address unmeasured confounders. Facility proportion of annual 2.5 mg olanzapine usage was an instrumental variable because facility treatment proportion is the best-known instrumental variable type [36, 37]. We used a two-stage residual inclusion estimation framework with robust standard errors, using background patient and treatment characteristics as covariates [38]. An F-statistics of > 10 indicated that our instrumental variable was highly correlated with treatment using additional antiemetic drugs. The first stage was a generalized linear model adjusted for the patient and treatment backgrounds. This model was used to measure the association between the facility proportion of annual olanzapine 2.5 mg usage and our instrument. From this model, we determined the raw residual for each patient by calculating the difference between the model-predicted probability of facility proportion of annual olanzapine 2.5 mg usage and olanzapine 2.5 mg administration. The residuals were then included as an additional covariate in the second-stage model. In the second-stage binomial regression model adjusted for patient and treatment backgrounds used in the first stage and the residual calculated in the first stage, we estimated the association between olanzapine 2.5 or 5 mg administration and additional antiemetic drugs administration. We then calculated a risk difference for the primary outcome (i.e., additional antiemetic drug administration within the overall assessment period) in the olanzapine 2.5 mg group and compared it with that of the olanzapine 5 mg group. To demonstrate the quasi-randomization of treatment assignments by the instrumental variable, we described the patients’ characteristics according to the mean value of the instrumental variable. All analyses were performed using Stata/SE 17.0 software (StataCorp, College Station, TX, USA). Continuous variables are presented as median and interquartile range (IQR). Categorical variables are expressed as numbers and percentages. All reported p values are two-sided, and p < 0.05 was considered statically significant.