Purpose: The goal was to quantify effects of data collection and regression practices on estimation of inactivation model parameters using simulated isothermal experiments.
Methods: Using MATLAB, synthetic isothermal inactivation data were generated assuming a true response (log-linear or Weibull models with a Bigelow-type secondary model), with random error, taking into account data collection variables (subsamples, replications, come-up-time (CUT)). Monte Carlo simulations were used to generate data for 1000 synthetic studies, each encompassing a defined number of observations, subsamples, and replications. Log-linear and Weibull models were fit to the synthetic data, using nonlinear regression.
Results: For the log-linear model, regardless of instantaneous or non-zero CUT, there was negligible bias between estimated and true model parameters (<2%). For the Weibull model with instantaneous CUT, at least 2 subsamples or 4 replications were required to reduce relative bias below 5%. For the Weibull model and non-zero CUT, the estimated parameters were increasingly biased from the true model parameters as the model became more nonlinear. For example, the scale and shape parameters were different (P < 0.05) from the true values, with biases of 145 and 29%, respectively, when the shape parameter was 0.5. Overall, additional subsamples were more effective than additional observations or replications in reducing parameter variance and error.
Significance: When true bacterial response is log-linear, estimated model parameters are reliably accurate, and reproducibility is best improved with subsamples. When the true response is nonlinear, then typical isothermal experiments may not be adequate to accurately estimate inactivation parameters.