Residual plot for nonlinear regression

I have a couple of questions regarding performance of nonlinear regression models.

1. Are the residuals from a nonlinear regression model supposed to be randomly distributed too (as in linear regression)?
2. I am comparing two nonlinear regression models (non-nested). What model performance indicators can I use for the purpose? My understanding was to compare RSS, RSE, residual plots, autocorrelation and residual normality plots.
3. The models have different Y axes. So I tried standardizing variables and then re-estimated RSS, RSE. Is that the right approach? What other model performance measures should be estimated / plotted with standardized variables?
• What do you mean that the "models have different Y axes"? Are they different variables? – gung Jul 3 '14 at 14:31
• Yes, one set of comparison was for models having different response variable but same predictor. But I think it is not possible to do such a comparison. – Learner Jul 3 '14 at 15:03

1. Are the residuals from a nonlinear regression model supposed to be randomly distributed too (as in linear regression)?

Yes. The error term is the same and the residuals represent estimated errors. The fitting of the model will induce some structure among the residuals, and that won't be identical in form for a nonlinear model as it is for a linear model but taking a Taylor approximation to first order, the nonlinear model is approximated by a linear model, so even that structure will tend to be similar to that of a linear model near the optimum.

2. I am comparing two nonlinear regression models (non-nested). What model performance indicators can I use for the purpose? My understanding was to compare RSS, RSE, residual plots, autocorrelation and residual normality plots.

You can compare two different models fitted to the same data using those things, but what you should look at depends on what kind of performance matters for your present purpose.

Note that when you compare RSS, more parameters will tend to make the RSS smaller; you can't really compare them for models with different degrees of freedom. Similarly with MSE or RMSE. (You might better compare AIC or BIC instead, but I wouldn't use them as the basis for model selection - you need to properly account for out of sample error or you'll end up with the usual problems from doing model selection using the same data you measure model performance on)

3. The models have different Y axis.

If the response variable isn't the same, you can't really compare the models.

So I tried standardizing variables and then re-estimated RSS, RSE. Is that the right approach?

The right approach to achieve what exactly?

• Thanks @Glen_b! That answers the confusion I had. Actually, there are three models and two separate comparisons. One with different predictors but same response variable and second with same predictor but different response variable. I gather from your reply that the second scenario is not really possible. And so, there is no need of standardizing variables in the first place (last part of question). – Learner Jul 3 '14 at 15:00
• Some kinds of comparison may be possible, but it depends on what you're trying to do. It's difficult to compare apples with oranges, but some things can perhaps be compared. To be more meaningful than a fairly facile "this number is bigger than that one", it needs to be approached with some care. – Glen_b Jul 3 '14 at 15:02
• One last thing, regarding the model performance part: does the "residual analysis" include residual vs fitted plot as well as residual normality plot? – Learner Jul 3 '14 at 15:10
• Yes, the use of both will be essentially the same as they are in a linear model. – Glen_b Jul 3 '14 at 15:18