Differences of GAMs and GLM I'm currently looking at a dataset and have to create a model to predict outcomes for whose values are non-negative.
I'm looking at using a GLM or GAM for my data though I'm unaware of what would work better. Are there any resources out there that could help me choose whether I should use a GAM or GLM?
 A: If your GAM includes penalised splines of the sort implemented in mgcv then you are really fitting a fancy GLM, and the penalties on the splines will shrink (or penalize) wiggly functions back to linear (on the link scale) functions should the relationship between $y$ and $x_j$ be linear. In other words, the wiggliness penalties on the smooths in a GAM could result in linear effects (on the link scale) being estimated, which is what you would have got if you'd fitted as a GLM.
The advantage of starting with the GAM is that you allow for the possibility that the effect of $x_j$ on $y$ is non-linear. The disadvantage is that because we are selecting smoothness parameters during fitting, the results of statistical tests based on the model fit are a little anti-conservative or need adjustment for extra uncertainty due to assuming that the smoothness parameters were known, when they aren't. There are also a few more things you need to concern yourself with (using a sufficiently large basis for each smooth, and concurvity [the GAM equivalent of collinearity, just non-linear] for example).
The advantage of the GLM is that it will estimate a linear effect (on the link scale) and if that is what theory in your system suggests then it is more direct to fit the GLM. The disadvantage is that unless you know the effect is linear, a GLM will only ever fit linear effects. A GAM can encompass both extremes.
