Stress Ulcer Prophylaxis and Postoperative Pneumonia Risk
Stress Ulcer Prophylaxis and Postoperative Pneumonia Risk
Study data came from the Premier Research Database, a hospital administrative database that contains information on approximately one sixth of all hospital admissions in the United States. The database contains a complete census of inpatients from approximately 500 hospitals from across the United States (numbers vary slightly by year). Preliminary analysis done by the Premier organization comparing characteristics of patients' and hospitals' for Premier hospitals with those from the National Hospital Discharge survey suggests that the profile of patients treated at Premier hospitals is similar to those treated nationally (Statement of Work Document, Premier Inc, 2012). It contains information for all patients treated at the included hospitals, independent of payer (Medicaid, Medicare, or commercial insurance). Premier collects data from member hospitals through its informatics products and then provides information back to hospitals for benchmarking purposes. Member hospitals pay Premier for these services. Member hospitals are primarily non-profit, non-governmental, community and teaching hospitals. The database includes charges for all drugs, procedures, and diagnostic tests during each hospital admission. It also includes patients' demographic characteristics and hospital characteristics, discharge diagnoses, and discharge status. Data are routinely audited, verified, and validated to ensure that the use of supplies and other hospital resources are within an acceptable range, but Premier does not directly verify that submitted data match individual patients' medical records. Several previous studies have used Premier data to study perioperative drug use and health outcomes.
The cohort consisted of patients aged 18 years or older who underwent the most common cardiac surgical procedure, coronary artery bypass graft surgery, identified by code 36.1, or any subcode thereof, from ICD-9 (international classification of diseases, 9th revision) between January 1, 2004 and December 31, 2010. For the main analysis, we included only those patients who had coronary artery bypass grafting on the third day of hospital admission or thereafter. We did this to obtain a preoperative period in which to measure patients' baseline comorbidities and other risk factors; these factors may affect both the choice of agent for stress ulcer prophylaxis and risk of pneumonia and thus confound the planned analysis. To ensure that we did not include in our analysis any patients who had pneumonia at the time of admission to the hospital, we excluded those who were exposed to systemic antibiotics before the day of surgery. We also excluded patients exposed to either proton pump inhibitors or H2 receptor antagonists before the day of surgery to isolate the effect of the type of stress ulcer prophylaxis given in the immediate postoperative period. We further limited our analysis to those patients who started either a proton pump inhibitor or an H2 receptor antagonist, but not both, on postoperative days 1, 2, or both and who survived to postoperative day 3 or beyond. A single dose of acid suppressive drug is commonly administered immediately before surgery to lessen the risk of aspiration on induction of anesthesia; as the focus of this study was the risk associated with stress ulcer prophylaxis in the postoperative period, we assigned exposure status independently of the type of acid suppression drug administered on the day of surgery.
We defined exposure to proton pump inhibitor by one or more charges on postoperative days 1 or 2 for omeprazole, lansoprazole, esomeprazole, pantoprazole, or rabeprazole. We defined exposure to H2 receptor antagonist by charges on postoperative days 1 or 2 for cimetidine, ranitidine, famotidine, or nizatidine.
The study outcome was the development of postoperative pneumonia during the index hospital admission and was assessed by the presence of the ICD-9 diagnosis codes of 481 to 486 or 507 (and any subcode thereof) among the reported discharge diagnoses. The exclusion of patients with systemic antibiotics before surgery would indicate that the pneumonia developed while the patient was in hospital, rather than being a condition present on admission. We also assessed two secondary outcomes: postoperative upper gastrointestinal hemorrhage and in-hospital death. We defined postoperative upper gastrointestinal hemorrhage on the basis of the presence of appropriate diagnosis and procedure codes. To ensure that any reported hemorrhage occurred postoperatively, we further required a charge for an esophagogastroduodenoscopy during the postoperative follow-up period. Esophagogastroduodenoscopies are routinely done for diagnostic and therapeutic purposes for patients with severe upper gastrointestinal bleeding.
We extracted five groups of potential confounders from the database: demographic characteristics, surgery characteristics, chronic comorbid conditions, markers of the severity of cardiac or coexisting conditions, and hospital characteristics. The demographic characteristics considered included age on admission, ethnicity (classified as white, black, or other), marital status (classified as married, single, or unknown), and sex. Surgery characteristics included type of admission (classified as emergency, elective, or unknown); number of bypass grafts (classified as one, two, three, or four or more); whether the patient received a mammary artery graft; whether the patient had a previous coronary artery bypass graft; whether the surgery was done using cardiopulmonary bypass; whether the patient received a transfusion of packed red blood cells; whether the patient received an intra-aortic balloon pump; whether the patient received an aortic, mitral, or tricuspid value repair or replacement or ventricular aneurysm repair concurrently with their coronary artery bypass graft; and season and year of admission.
We identified the presence of chronic comorbid conditions by using discharge diagnoses including chronic hypertension, liver disease, chronic obstructive pulmonary disease and allied conditions, malignancy, previous myocardial infarction, previous stroke, endocarditis, peripheral vascular disease, chronic renal disease, chronic hemostatic disorder, smoking (either past or current), alcohol/drug misuse or dependence, carotid artery stenosis, gastrointestinal reflux, and diabetes. Although these conditions were recorded as discharge diagnoses, all of these conditions would have been present before admission. We also calculated the Romano modification of the Charlson comorbidity index for each patient by using conditions identified by discharge diagnoses.
We evaluated markers of the severity of the patients' preoperative cardiac and coexisting disease with charges for drugs and procedures before the day of surgery. These included the use of drugs including ß blockers, calcium channel blockers, aspirin, Aggrenox or aspirin/dipyridamole, clopidogrel, statins, nitrates, fibrates, digoxin, insulin, oral anti-diabetes drugs, rhythm control drugs, warfarin, glycoprotein IIb/IIIa inhibitor, thiazide diuretics, loop diuretics, angiotensin converting enzyme inhibitors, angiotensin receptor blockers, aldosterone antagonists, direct renin inhibitors, systemic corticosteroids, inhaled steroids, leukotriene inhibitors, theophylline, inhaled ß2 agonists, probiotics, sucralfate, heparin, nitroglycerine, and inotropic support drugs. We also assessed charges for procedures that might indicate severity of cardiac and coexisting disease, including preoperative intensive care unit admission, cardiac resuscitation, oxygen use, telemetry, echocardiography, percutaneous angioplasty or stent, continuous positive airway pressure, mechanical ventilation, or dialysis. Finally, we determined the number of days the patient was in the hospital before surgery.
We also assessed the characteristics of the hospitals in which patients were treated. We recorded whether they were teaching hospitals and whether they were located in an urban or rural location. We estimated the annualized volume of coronary artery bypass graft surgeries performed by each hospital by dividing the total number of such surgeries for each hospital during the study time period by the number of years that each hospital performed one or more coronary artery bypass graft operations. We ranked hospitals in order of annualized volume and then categorized them into thirds of volume (high, medium, and low volume hospitals). We classified the region of the hospital as Midwest, Northeast, South, or West.
We determined baseline characteristics for the patients exposed to proton pump inhibitors and H2 receptor antagonists and summarized them as medians and interquartile ranges for continuous variables and counts and proportions for categorical variables. We determined the unadjusted risk and unadjusted risk ratios with 95% confidence intervals for postoperative pneumonia during the index hospital admission, comparing patients exposed to proton pump inhibitors with those exposed to H2 receptor antagonists during the defined exposure window. Because of the low event rate, we report the odds ratios from logistic regression analyses as relative risks.
Propensity Score Analyses. To account for potentially important baseline differences in those patients who received proton pump inhibitors and H2 receptor antagonists, we did two separate propensity score analyses. For both analyses, we first estimated a propensity score by using a logistic regression model in which exposure to a proton pump inhibitor (as opposed to an H2 receptor antagonist) was the dependent variable and was estimated on the basis of all demographic and surgery characteristics, chronic comorbid conditions, markers of the severity of the cardiac condition, markers of coexisting disease/disease severity, and hospital characteristics, as defined above. The model predicting exposure to a proton pump inhibitor had an area under the receiver operating curve of 0.70 (possible range 0.5 to 1.0).
In the primary propensity score analysis, we excluded all patients with a propensity score value that corresponded to the 2.5th centile or lower in the proton pump inhibitor exposed patients and to the 97.5th centile or higher in the H2 receptor antagonist exposed patients. This technique of trimming patients in the tails of the propensity score distribution excludes patients treated strongly contrary to expectation (for example, a patient who was strongly predicted to receive a proton pump inhibitor but received an H2 receptor antagonist instead) and thus reduces residual confounding. We calculated tenths of propensity score among the remaining patients (n=18,573; 87.6% of total). We then fitted a multivariable logistic regression model including indicator of tenth of propensity score as a covariate, from which we estimated a relative risk and 95% confidence interval for postoperative pneumonia associated with proton pump inhibitor exposure.
In the second propensity score analysis, we matched patients exposed to proton pump inhibitors to those exposed to H2 receptor antagonists in a fixed 1:1 ratio by using a nearest neighbor algorithm with a maximum matching distance of 0.05. Matching was restricted to patients undergoing coronary artery bypass grafting in the same year. After matching, the cohort included 15,074 patients (71.1% of original cohort). We again estimated a relative risk and 95% confidence interval for postoperative pneumonia.
Instrumental Variable Analysis. When hospitals show a strong consistency in whether they use either proton pump inhibitors or H2 receptor antagonists for prophylaxis, the decision to administer one or the other of these agents may be assumed to be made largely independently of individual patients' characteristics; in this scenario, a patient's receipt of either regimen is determined more strongly by the facility at which he or she is treated than by his or her specific risk factors. In such a circumstance, the hospital's preferred agent can act as an instrumental variable, thus setting the stage for a "natural experiment" that allows for an unbiased estimate of the risk of pneumonia, even in the presence of unmeasured confounding. This analysis assumes that patients' choice of hospital is made independently of the hospital's choice of acid suppressive drug for prophylaxis, and the hospital's use of proton pump inhibitors or H2 receptor antagonists is largely independent of markers for positive or negative outcomes in that facility. We note that these assumptions, although strong, are distinct from the assumption of no residual confounding that is needed for a causal interpretation of a "conventional" analysis. A consistent result from the instrumental variable analysis may serve as a useful confirmatory analysis for our primary conventional approach.
We analyzed 15,800 patients who received either a proton pump inhibitor or an H2 receptor antagonist and who were treated at hospitals that performed, on average, at least 12 coronary artery bypass graft procedures (average of one a month) during the year of surgery. We classified hospitals that administered proton pump inhibitors to 90% or more of their coronary artery bypass graft patients in a given year as hospitals with a preference for proton pump inhibitors and those that administered proton pump inhibitors to 10% or fewer of their coronary artery bypass graft patients in a given year as hospitals that prefer H2 receptor antagonists. Using the hospitals' preference as an instrumental variable, we computed differences in risk of postoperative pneumonia between the proton pump inhibitor and H2 receptor antagonist groups. We did this by using a two stage linear regression that, for robustness, also adjusted for the demographic and surgery characteristics, chronic comorbid conditions, markers of the severity of the cardiac condition, and markers of coexisting disease/disease severity. To account for any possible link between hospitals' choice of acid suppressive drug for prophylaxis and positive or negative outcomes in that hospital, we additionally adjusted for the hospitals' characteristics and procedure volume measurement noted above. Using Stata version 10 and the ivreg2 procedure, we estimated a risk difference and number needed to harm and their 95% confidence intervals.
Secondary Outcomes, Exposure Subgroup, and Sensitivity Analyses. We repeated each of the analyses described above to examine the secondary outcomes of postoperative gastrointestinal hemorrhage and in-hospital mortality. We also compared the risk of the primary outcome in patients exposed to the most commonly prescribed H2 receptor antagonist in our cohort, famotidine, and the most commonly prescribed proton pump inhibitor, pantoprazole, to test whether the main study findings applied specifically to the most commonly used agents in each class. We did a sensitivity analysis excluding those patients who were discordant with respect to the class of acid suppressive drug received on the day of surgery and during the postoperative period (that is, patients who received an H2 receptor antagonist on the day of surgery and a proton pump inhibitor during the exposure window, or vice versa).
Lastly, we tested the generalizability of our findings to all coronary artery bypass graft patients by examining the risk of pneumonia in patients who underwent surgery on hospital day 1 and hospital day 2. Although important in establishing generalizability, estimates from these analyses are more likely to be biased than are those obtained in the primary analysis. This is due to more limited ability to control for confounding bias (as less baseline information is available on patients before surgery) and greater likelihood of misclassification bias (as patients taking a proton pump inhibitor or an H2 receptor antagonist as outpatients may switch classes on admission, but their risk of pneumonia may still be affected by the drug class they took before admission, whereas in the primary analysis a two day washout period exists for anyone exposed before admission). The propensity score for patients having surgery on day 1 included information on demographic characteristics, surgery characteristics, chronic comorbid conditions, and hospital characteristics. For patients having surgery on hospital day 2, we performed all of the exclusions that were done in the main analysis and the propensity score model included all of the covariates used in the main analysis (although measured for a shorter period). Finally, to assess the potential effects of residual confounding, we determined the strength (confounder-outcome relative risk) of a hypothetical residual confounder whose existence would produce the estimate of the effect of proton pump inhibitors on the risk of pneumonia observed in the primary analysis. As the effect of confounders varies with relative prevalence, we assumed a prevalence of 5% among the unexposed patients and range of prevalences in the exposed patients.
Methods
Data Source
Study data came from the Premier Research Database, a hospital administrative database that contains information on approximately one sixth of all hospital admissions in the United States. The database contains a complete census of inpatients from approximately 500 hospitals from across the United States (numbers vary slightly by year). Preliminary analysis done by the Premier organization comparing characteristics of patients' and hospitals' for Premier hospitals with those from the National Hospital Discharge survey suggests that the profile of patients treated at Premier hospitals is similar to those treated nationally (Statement of Work Document, Premier Inc, 2012). It contains information for all patients treated at the included hospitals, independent of payer (Medicaid, Medicare, or commercial insurance). Premier collects data from member hospitals through its informatics products and then provides information back to hospitals for benchmarking purposes. Member hospitals pay Premier for these services. Member hospitals are primarily non-profit, non-governmental, community and teaching hospitals. The database includes charges for all drugs, procedures, and diagnostic tests during each hospital admission. It also includes patients' demographic characteristics and hospital characteristics, discharge diagnoses, and discharge status. Data are routinely audited, verified, and validated to ensure that the use of supplies and other hospital resources are within an acceptable range, but Premier does not directly verify that submitted data match individual patients' medical records. Several previous studies have used Premier data to study perioperative drug use and health outcomes.
Cohort
The cohort consisted of patients aged 18 years or older who underwent the most common cardiac surgical procedure, coronary artery bypass graft surgery, identified by code 36.1, or any subcode thereof, from ICD-9 (international classification of diseases, 9th revision) between January 1, 2004 and December 31, 2010. For the main analysis, we included only those patients who had coronary artery bypass grafting on the third day of hospital admission or thereafter. We did this to obtain a preoperative period in which to measure patients' baseline comorbidities and other risk factors; these factors may affect both the choice of agent for stress ulcer prophylaxis and risk of pneumonia and thus confound the planned analysis. To ensure that we did not include in our analysis any patients who had pneumonia at the time of admission to the hospital, we excluded those who were exposed to systemic antibiotics before the day of surgery. We also excluded patients exposed to either proton pump inhibitors or H2 receptor antagonists before the day of surgery to isolate the effect of the type of stress ulcer prophylaxis given in the immediate postoperative period. We further limited our analysis to those patients who started either a proton pump inhibitor or an H2 receptor antagonist, but not both, on postoperative days 1, 2, or both and who survived to postoperative day 3 or beyond. A single dose of acid suppressive drug is commonly administered immediately before surgery to lessen the risk of aspiration on induction of anesthesia; as the focus of this study was the risk associated with stress ulcer prophylaxis in the postoperative period, we assigned exposure status independently of the type of acid suppression drug administered on the day of surgery.
Classification of Drug Exposure and Study Outcome
We defined exposure to proton pump inhibitor by one or more charges on postoperative days 1 or 2 for omeprazole, lansoprazole, esomeprazole, pantoprazole, or rabeprazole. We defined exposure to H2 receptor antagonist by charges on postoperative days 1 or 2 for cimetidine, ranitidine, famotidine, or nizatidine.
The study outcome was the development of postoperative pneumonia during the index hospital admission and was assessed by the presence of the ICD-9 diagnosis codes of 481 to 486 or 507 (and any subcode thereof) among the reported discharge diagnoses. The exclusion of patients with systemic antibiotics before surgery would indicate that the pneumonia developed while the patient was in hospital, rather than being a condition present on admission. We also assessed two secondary outcomes: postoperative upper gastrointestinal hemorrhage and in-hospital death. We defined postoperative upper gastrointestinal hemorrhage on the basis of the presence of appropriate diagnosis and procedure codes. To ensure that any reported hemorrhage occurred postoperatively, we further required a charge for an esophagogastroduodenoscopy during the postoperative follow-up period. Esophagogastroduodenoscopies are routinely done for diagnostic and therapeutic purposes for patients with severe upper gastrointestinal bleeding.
Patient and Hospital Level Covariates
We extracted five groups of potential confounders from the database: demographic characteristics, surgery characteristics, chronic comorbid conditions, markers of the severity of cardiac or coexisting conditions, and hospital characteristics. The demographic characteristics considered included age on admission, ethnicity (classified as white, black, or other), marital status (classified as married, single, or unknown), and sex. Surgery characteristics included type of admission (classified as emergency, elective, or unknown); number of bypass grafts (classified as one, two, three, or four or more); whether the patient received a mammary artery graft; whether the patient had a previous coronary artery bypass graft; whether the surgery was done using cardiopulmonary bypass; whether the patient received a transfusion of packed red blood cells; whether the patient received an intra-aortic balloon pump; whether the patient received an aortic, mitral, or tricuspid value repair or replacement or ventricular aneurysm repair concurrently with their coronary artery bypass graft; and season and year of admission.
We identified the presence of chronic comorbid conditions by using discharge diagnoses including chronic hypertension, liver disease, chronic obstructive pulmonary disease and allied conditions, malignancy, previous myocardial infarction, previous stroke, endocarditis, peripheral vascular disease, chronic renal disease, chronic hemostatic disorder, smoking (either past or current), alcohol/drug misuse or dependence, carotid artery stenosis, gastrointestinal reflux, and diabetes. Although these conditions were recorded as discharge diagnoses, all of these conditions would have been present before admission. We also calculated the Romano modification of the Charlson comorbidity index for each patient by using conditions identified by discharge diagnoses.
We evaluated markers of the severity of the patients' preoperative cardiac and coexisting disease with charges for drugs and procedures before the day of surgery. These included the use of drugs including ß blockers, calcium channel blockers, aspirin, Aggrenox or aspirin/dipyridamole, clopidogrel, statins, nitrates, fibrates, digoxin, insulin, oral anti-diabetes drugs, rhythm control drugs, warfarin, glycoprotein IIb/IIIa inhibitor, thiazide diuretics, loop diuretics, angiotensin converting enzyme inhibitors, angiotensin receptor blockers, aldosterone antagonists, direct renin inhibitors, systemic corticosteroids, inhaled steroids, leukotriene inhibitors, theophylline, inhaled ß2 agonists, probiotics, sucralfate, heparin, nitroglycerine, and inotropic support drugs. We also assessed charges for procedures that might indicate severity of cardiac and coexisting disease, including preoperative intensive care unit admission, cardiac resuscitation, oxygen use, telemetry, echocardiography, percutaneous angioplasty or stent, continuous positive airway pressure, mechanical ventilation, or dialysis. Finally, we determined the number of days the patient was in the hospital before surgery.
We also assessed the characteristics of the hospitals in which patients were treated. We recorded whether they were teaching hospitals and whether they were located in an urban or rural location. We estimated the annualized volume of coronary artery bypass graft surgeries performed by each hospital by dividing the total number of such surgeries for each hospital during the study time period by the number of years that each hospital performed one or more coronary artery bypass graft operations. We ranked hospitals in order of annualized volume and then categorized them into thirds of volume (high, medium, and low volume hospitals). We classified the region of the hospital as Midwest, Northeast, South, or West.
Statistical Analysis
We determined baseline characteristics for the patients exposed to proton pump inhibitors and H2 receptor antagonists and summarized them as medians and interquartile ranges for continuous variables and counts and proportions for categorical variables. We determined the unadjusted risk and unadjusted risk ratios with 95% confidence intervals for postoperative pneumonia during the index hospital admission, comparing patients exposed to proton pump inhibitors with those exposed to H2 receptor antagonists during the defined exposure window. Because of the low event rate, we report the odds ratios from logistic regression analyses as relative risks.
Propensity Score Analyses. To account for potentially important baseline differences in those patients who received proton pump inhibitors and H2 receptor antagonists, we did two separate propensity score analyses. For both analyses, we first estimated a propensity score by using a logistic regression model in which exposure to a proton pump inhibitor (as opposed to an H2 receptor antagonist) was the dependent variable and was estimated on the basis of all demographic and surgery characteristics, chronic comorbid conditions, markers of the severity of the cardiac condition, markers of coexisting disease/disease severity, and hospital characteristics, as defined above. The model predicting exposure to a proton pump inhibitor had an area under the receiver operating curve of 0.70 (possible range 0.5 to 1.0).
In the primary propensity score analysis, we excluded all patients with a propensity score value that corresponded to the 2.5th centile or lower in the proton pump inhibitor exposed patients and to the 97.5th centile or higher in the H2 receptor antagonist exposed patients. This technique of trimming patients in the tails of the propensity score distribution excludes patients treated strongly contrary to expectation (for example, a patient who was strongly predicted to receive a proton pump inhibitor but received an H2 receptor antagonist instead) and thus reduces residual confounding. We calculated tenths of propensity score among the remaining patients (n=18,573; 87.6% of total). We then fitted a multivariable logistic regression model including indicator of tenth of propensity score as a covariate, from which we estimated a relative risk and 95% confidence interval for postoperative pneumonia associated with proton pump inhibitor exposure.
In the second propensity score analysis, we matched patients exposed to proton pump inhibitors to those exposed to H2 receptor antagonists in a fixed 1:1 ratio by using a nearest neighbor algorithm with a maximum matching distance of 0.05. Matching was restricted to patients undergoing coronary artery bypass grafting in the same year. After matching, the cohort included 15,074 patients (71.1% of original cohort). We again estimated a relative risk and 95% confidence interval for postoperative pneumonia.
Instrumental Variable Analysis. When hospitals show a strong consistency in whether they use either proton pump inhibitors or H2 receptor antagonists for prophylaxis, the decision to administer one or the other of these agents may be assumed to be made largely independently of individual patients' characteristics; in this scenario, a patient's receipt of either regimen is determined more strongly by the facility at which he or she is treated than by his or her specific risk factors. In such a circumstance, the hospital's preferred agent can act as an instrumental variable, thus setting the stage for a "natural experiment" that allows for an unbiased estimate of the risk of pneumonia, even in the presence of unmeasured confounding. This analysis assumes that patients' choice of hospital is made independently of the hospital's choice of acid suppressive drug for prophylaxis, and the hospital's use of proton pump inhibitors or H2 receptor antagonists is largely independent of markers for positive or negative outcomes in that facility. We note that these assumptions, although strong, are distinct from the assumption of no residual confounding that is needed for a causal interpretation of a "conventional" analysis. A consistent result from the instrumental variable analysis may serve as a useful confirmatory analysis for our primary conventional approach.
We analyzed 15,800 patients who received either a proton pump inhibitor or an H2 receptor antagonist and who were treated at hospitals that performed, on average, at least 12 coronary artery bypass graft procedures (average of one a month) during the year of surgery. We classified hospitals that administered proton pump inhibitors to 90% or more of their coronary artery bypass graft patients in a given year as hospitals with a preference for proton pump inhibitors and those that administered proton pump inhibitors to 10% or fewer of their coronary artery bypass graft patients in a given year as hospitals that prefer H2 receptor antagonists. Using the hospitals' preference as an instrumental variable, we computed differences in risk of postoperative pneumonia between the proton pump inhibitor and H2 receptor antagonist groups. We did this by using a two stage linear regression that, for robustness, also adjusted for the demographic and surgery characteristics, chronic comorbid conditions, markers of the severity of the cardiac condition, and markers of coexisting disease/disease severity. To account for any possible link between hospitals' choice of acid suppressive drug for prophylaxis and positive or negative outcomes in that hospital, we additionally adjusted for the hospitals' characteristics and procedure volume measurement noted above. Using Stata version 10 and the ivreg2 procedure, we estimated a risk difference and number needed to harm and their 95% confidence intervals.
Secondary Outcomes, Exposure Subgroup, and Sensitivity Analyses. We repeated each of the analyses described above to examine the secondary outcomes of postoperative gastrointestinal hemorrhage and in-hospital mortality. We also compared the risk of the primary outcome in patients exposed to the most commonly prescribed H2 receptor antagonist in our cohort, famotidine, and the most commonly prescribed proton pump inhibitor, pantoprazole, to test whether the main study findings applied specifically to the most commonly used agents in each class. We did a sensitivity analysis excluding those patients who were discordant with respect to the class of acid suppressive drug received on the day of surgery and during the postoperative period (that is, patients who received an H2 receptor antagonist on the day of surgery and a proton pump inhibitor during the exposure window, or vice versa).
Lastly, we tested the generalizability of our findings to all coronary artery bypass graft patients by examining the risk of pneumonia in patients who underwent surgery on hospital day 1 and hospital day 2. Although important in establishing generalizability, estimates from these analyses are more likely to be biased than are those obtained in the primary analysis. This is due to more limited ability to control for confounding bias (as less baseline information is available on patients before surgery) and greater likelihood of misclassification bias (as patients taking a proton pump inhibitor or an H2 receptor antagonist as outpatients may switch classes on admission, but their risk of pneumonia may still be affected by the drug class they took before admission, whereas in the primary analysis a two day washout period exists for anyone exposed before admission). The propensity score for patients having surgery on day 1 included information on demographic characteristics, surgery characteristics, chronic comorbid conditions, and hospital characteristics. For patients having surgery on hospital day 2, we performed all of the exclusions that were done in the main analysis and the propensity score model included all of the covariates used in the main analysis (although measured for a shorter period). Finally, to assess the potential effects of residual confounding, we determined the strength (confounder-outcome relative risk) of a hypothetical residual confounder whose existence would produce the estimate of the effect of proton pump inhibitors on the risk of pneumonia observed in the primary analysis. As the effect of confounders varies with relative prevalence, we assumed a prevalence of 5% among the unexposed patients and range of prevalences in the exposed patients.
Source...