Regression modeling methods are commonly used to estimate influenza-associated mortality using covariates such as laboratory-confirmed influenza activity in the population as a proxy of influenza incidence.

We examined the choices of influenza proxies that can be used from influenza laboratory surveillance data and their impact on influenza-associated mortality estimates.

Semiparametric generalized additive models with a smoothing spline were applied on national mortality data from South Africa and influenza surveillance data as covariates to obtain influenza-associated mortality estimates from respiratory causes from 2009 to 2013. Proxies examined included alternative ways of expressing influenza laboratory surveillance data such as weekly or yearly proportion or rate of positive samples, using influenza subtypes, or total influenza data and expressing the data as influenza season-specific or across all seasons.

Based on model fit, weekly proportion and influenza subtype-specific proxy formulation provided the best fit. The choice of proxies used gave large differences to mortality estimates, but the 95% confidence interval of these estimates overlaps.

Regardless of proxy chosen, mortality estimates produced may be broadly consistent and not statistically significant for public health practice.