AIC and BIC are Information criteria methods used to assess model fit while penalizing the number of estimated parameters. As I understand, when performing model selection, the one with the lowest AIC or BIC is preferred. In a situation am working on, a model with the lowest AIC is not necessarily the one with the lowest BIC. Is there any reason to prefer one over the other?
Hi Paul,
AIC and BIC differ by the way they penalize the number of parameters of a model. More precisely, BIC criterion will induce a higher penalization for models with an intricate parametrization in comparison with AIC criterion.
Whether you decide to use rather AIC or BIC is all up to you as these criteria are slightly different but none is better than the other.
Best regards,
The reason to choose one relative fit measure over another is often based on a combination of statistical and substantive criteria. I often find disparate answers regarding relative model fit between the AIC and BIC in practice, and rely more heavily on the latter. However, I would report the values from both measures in a research paper (along with negative two times the log likelihood for each model). Then again, I deal primarily with multilevel models and complicated structural equation models (SEMs), and the BIC is used more frequently than the AIC (as well as sample size adjusted variants of these fit measures) in my field. Assuming your models are nested, you may also want to consider conducting a likelihood ratio test, if p-values are of importance to you.
As Geoffrey pointed out, the BIC penalizes more heavily for complex models. So the context certainly matters here.
Below link may helpful for you.
http://stats.stackexchange.com/questions/577/is-there-any-reason-to-prefer-the-aic-or-bic-over-the-other
When aic or bic increase does those model better than lower bic or aic ?
The BIC has heavier penalty term to penalize the model than AIC which always its value larger than the value of the AIC
I find this confusing at times. I need some help pls.
Between my original time series data and transformed time series, which am I performing arima on? which am i expected to forcast or predict? and what situation necessitates performing the arima, prediction and forecasting on any of the two?
I have notice when performing fitting to distribution using Python library Fitter (it uses scipy) I have a problem with AIC. I always get uniform distribution with the lowest AIC. For example if I create a random sample of normal distribution, the BIC and sumsquare error have the lowest values for normal or lognormal distributions. However, the AIC value for uniform distribution is the lowest. I guess is related to number of parameters of the uniform model but I haven’t find anyone mentioning this and I wanted to know if this happens to you as when you try to analyze fitting to distributions and include uniform distribution in the analysis.