The study was carried out with the 2009 release of the SEER-Medicare linked database, which includes cases of lung cancer diagnosed up to 2005 and follow-up data to December 2007. The SEER registry keeps a national database that collects information on all incident cases of cancer in selected areas of the United States, covering nearly 26% of the US population.19
From the SEER-Medicare database we selected patients aged more than 65 years with a diagnosis of stages II-IIIA non-small cell lung cancer between 1992 and 2005 and who underwent surgical resection (lobectomy or pneumonectomy). Among these cases we excluded patients in healthcare maintenance organisations or those without part B Medicare insurance (coverage for outpatient care) for whom we were not able to ascertain comorbidities and use of chemotherapy.20
We also excluded patients who died during the perioperative period (within 30 days of surgery) or who were discharged to a nursing home after surgery, as they would have not been candidates for adjuvant chemotherapy. The final cohort consisted of 3324 patients with stages II-IIIA non-small cell lung cancer.
From the SEER-Medicare database we obtained sociodemographic information on characteristics such as age, sex, race or ethnicity, marital status, and estimated income. To evaluate the burden of comorbidities, we used the Deyo adaptation of the Charlson comorbidity index, applying lung cancer specific condition weights as described in the literature.21 22 23
From the SEER database we obtained data on tumour location, size, extension, involvement of lymph nodes, and histology. We classified histological subtypes into categories of adenocarcinoma, bronchioalveolar carcinoma, squamous cell carcinoma, large cell carcinoma, and other histological type.
We examined surgical treatment using information from the SEER-Medicare database.24
Using these data, we classified patients as having either a lobectomy or a pneumonectomy (SEER site specific surgical codes 30 to 70). From Medicare inpatient, outpatient, and physician files we identified those patients who experienced postoperative complications (extrapulmonary infections, cardiovascular complications, thromboembolic events, respiratory complications, reoperations, and transfusions) within 30 days of surgery.25 26 27
Use of radiation therapy was ascertained from the SEER database and Medicare claims. We classified patients as having received radiotherapy if they were coded in the SEER database as having received external beam radiation or if Medicare claims contained any code indicating use of radiation therapy within six months of cancer diagnosis. From Medicare files we identified patients treated with adjuvant chemotherapy. Using validated algorithms, we classified patients as being treated with chemotherapy if Medicare inpatient, outpatient, or physician claims contained any code indicating that the patient received platinum based chemotherapy within three months of surgery.28
Functional status is an important determinant of chemotherapy use. Although the SEER-Medicare database does not include information on patients’ functional status, all the study participants were eligible for surgery, which should have effectively excluded patients with poor functional status. To indirectly assess patients’ postoperative functional status, we used data from the Home Health Agency file to ascertain use of home health services such as home health aide, physiotherapy, speech therapy, occupational therapy, and medical social services. As beneficiaries must be homebound to be eligible for Medicare home services, we used this information as a proxy for poor functional status.
The primary study outcome, determined from Medicare data, was overall survival. Using information in the Medicare file we calculated survival times as the period from surgery to the date of death. We classified those surviving past 31 December 2007 (alive at the end of follow-up) as censored observations. The secondary study outcome was the rate of serious adverse events among older patients who did or did not receive adjuvant chemotherapy. Using a published algorithm, we defined serious adverse events as those requiring admission to hospital within 2-6 months of surgery (the usual period for occurrence of chemotherapy related adverse events). Serious adverse events were infection, fever, neutropenia, anaemia, thrombocytopenia, dehydration, nausea or emesis, acute renal dysfunction, and unspecified adverse events of systemic therapy.29
Additionally, we evaluated the number of patients who died within 12 weeks of initiation of chemotherapy (the typical duration of platinum based regimens) and number of patients with a diagnosis of neuropathy, a potential long term adverse event from chemotherapy, within two years of resection.
The χ2 test was used to evaluate differences in the distribution of baseline characteristics between patients who did or did not receive postoperative platinum based chemotherapy. We used propensity score methods to control for potential selection bias. The propensity score is a measure of the probability that a patient will receive adjuvant chemotherapy after resection on the basis of their baseline characteristics. We calculated propensity scores using a logistic model that included the patients’ sociodemographic characteristics, comorbidities, and cancer related factors (tumour location, size, involvement of lymph nodes, and grade). Additionally, we included dummy variables in the propensity score model indicating whether the patients had postoperative complications or received home services, as these patients were probably less likely to receive adjuvant chemotherapy. Once the model was fitted, we used regression analysis to evaluate whether baseline covariates were balanced across study groups after adjusting for the estimated propensity scores.
To compare survival of patients who did or did not receive adjuvant chemotherapy we used Cox regression analysis adjusting for propensity scores in three ways.30 31
Firstly, we included the propensity score as a continuous covariate in a Cox model comparing the survival of patients treated with and without chemotherapy. In a second approach we fitted a stratified Cox model according to fifths of propensity scores. Finally, we matched patients based on the propensity scores and compared the survival of patients treated with and without adjuvant chemotherapy using a marginal Cox model for correlated data.32
Adjuvant chemotherapy may be used in combination with postoperative radiotherapy. As radiotherapy is usually given concurrently or after chemotherapy this covariate was not included in the propensity score model. Thus we carried out secondary analysis adjusting for and stratifying the cohort by radiotherapy use to assess the effectiveness of chemotherapy among patients treated with and without postoperative radiotherapy. We also did secondary analyses to assess survival among patients treated with and without adjuvant chemotherapy separately for stage II and stage IIIA disease, and within strata according to the patients’ age at diagnosis (<70, 70-79, and >80 years). Finally, we repeated all the analyses controlling for year of diagnosis, to adjust for potential time trends in other aspects of lung cancer care.
The potential association of chemotherapy with increased survival may be confounded by the patients’ functional status, an important determinant of treatment. To evaluate this possibility we carried out a sensitivity analysis to test whether differences in functional status could account for the magnitude of the observed association of postoperative chemotherapy with survival.33
In our analysis we used published data on the prevalence of poor functional status (Eastern Co-operative Oncology Group functional status >2) among patients with stages II-IIIA non-small cell lung cancer and the relative hazard of death associated with poor functional status, to evaluate the robustness of our findings across different scenarios.34 35
We calculated the unadjusted odds for serious adverse events, with 95% confidence intervals, for patients receiving adjuvant chemotherapy. To estimate the odds of serious chemotherapy related adverse events among patients receiving chemotherapy compared with those not receiving chemotherapy we used logistic regression analysis after adjusting for propensity scores. All analyses were done with SAS software and using two tailed P values.