Comparing Results from Cox Proportional Hazards Models using SUDAAN and SAS Survey Procedures to a Logistic Regression Model for Analysis of Influenza Vaccination Coverage

The National Immunization Survey-Flu (NIS-Flu) is an ongoing, national telephone survey of households with children in the United States used to measure influenza vaccination coverage. The data collected by NIS-Flu have similarities to data typically analyzed using survival analytic procedures. Estimates of vaccination coverage from the NIS-Flu survey are calculated using Kaplan-Meier survival analysis procedures to account for censoring of the data. However, multivariable models to examine socio-demographic characteristics associated with receipt of influenza vaccination using NIS-Flu data have typically been done using logistic regression rather than using survival analytic methods. The logistic regression approach ignores the time-to-event and censoring characteristics of the influenza data and assumes that censoring throughout the survey period occurs equally among the comparison groups of interest. If this assumption is untrue, measures of association for receipt of influenza vaccine could be biased. Another approach used to address the censoring issues in NIS-Flu data is to restrict the logistic regression analysis to interviews conducted at the end of the vaccination period (i.e., March-June) when it is unlikely that many respondents would be vaccinated after the time of interview. However, this approach ignores a large amount of data, results in a reduced precision of estimates, and potentially exacerbates recall bias. This project assessed the feasibility, methods, and advantages of using a Cox proportional hazards model as opposed to a logistic regression model using full NIS-Flu 2013-14 season data and a logistic regression model using end of vaccination period data, This project also compared the results of Cox proportional hazards model from SUDAAN SURVIVAL and from SAS SURVEYPHREG procedures. Despite the slight underestimate of the associations between vaccination status and demographic characteristics, the logistic model remains a reasonable alternative to the Cox proportional hazards model in analyzing the NIS-Flu data. The SAS SURVEYPHREG and the SUDAAN SURVIVAL produced nearly identical Cox proportional hazards model results. Conclusions drawn based on the results from logistic regression and either of the Cox proportional hazards models using full or post-vaccination period NIS-Flu data are comparable. INTRODUCTION The influenza vaccination data collected by the National Immunization Survey-Flu (NIS-Flu) has similarities to data typically analyzed using survival analytic procedures. Children for whom parents answer “No” to the question regarding whether their child had received an influenza vaccination are censored at the date of the telephone interview, indicating the child was not vaccinated by the time of interview. Children can, and some do, receive an influenza vaccination later in the influenza season after the NIS-Flu interview, especially for interviews occurring during the active vaccination months. However, the NIS-Flu does not follow-up with children to determine if vaccination occurred after interview. To account for this censoring in the data, estimates of vaccination coverage from the NIS-Flu are calculated using Kaplan-Meier survival analysis procedures (CDC, 2010a, 2010b, 2013b; Lu, Rodriguez-Lainz, O'Halloran, Greby, & Williams, 2014). However, multivariable models to examine sociodemographic characteristics associated with receipt of influenza vaccination using data from vaccination surveys have typically been done using logistic regression in SUDAAN with the RLOGIST procedure rather than using survival analytic methods (Lu et al., 2014). The logistic regression model, although familiar to most analysts and researchers, ignores the time-to-event and censoring characteristics of the influenza data. The approach also assumes that censoring throughout the survey period occurs equally among the comparison groups of interest. If this assumption is untrue, measures of association for receipt of influenza vaccine could be biased. Another approach that has been used for dealing with the censoring issues with influenza survey data is to restrict the logistic regression analysis to interviews conducted at the end of the vaccination period, when it is unlikely that many respondents would be vaccinated after the time of interview (Santibanez, 2012). However, this approach ignores a large amount of collected data, resulting in a reduced sample size and reduced precision of estimates, and potentially exacerbates recall bias since the interviews included in the analyses take place months after vaccination for many respondents. The SURVIVAL procedure in SUDAAN utilizes the Cox proportional hazards model with complex survey data (RTI, 2009). SAS has extended its software capabilities for handling complex survey data; after releasing SURVEYMEANS, SURVEYREG, SURVEYFREQ, and SURVEYLOGISTIC in version 8.0 and 9.0, in 2010, an experimental SURVEYPHREG procedure was released with the SAS/STAT 9.22 giving SAS the capability of performing Cox proportional hazards regressions for complex survey data (SAS, 2010). Cheng compared logistic