computing model deviance
5 posts
• Page 1 of 1
computing model deviance
by alqamy » Mon Aug 24, 2009 7:29 am
Hi I am new to PRESNCE. I have seen many papers using PRESENCE to estimate occupancy and the authors were reporting model deviance as the were favoring rival models over others. my question is" Is it good to use model deviance as criteria for model selection? and if so is it primary or secondary to AIC? and last how can I compute the model deviance? Regards
- alqamy
- Posts: 6
- Joined: Sun Aug 23, 2009 9:09 pm
- Location: Egypt
omputing model deviance
by jhines » Mon Aug 24, 2009 1:03 pm
Deviance is the difference in likelihood between the saturated model and the selected model. By adding parameters to the model, the deviance should be reduced (model fits better). So, basing model selection on deviance will result in choosing the model with the most parameters.
AIC imposes a penalty (2.0 for each parameter) for adding parameters. So, if two parameters are very similar and are estimated with a single parameter in one model, and two parameters in another model, the difference in deviances for the two models will be small (< 2.0) and you would choose the model with fewer parameters. If the actual parameters are very different, the difference in deviance will be larger than the penalty, and you will choose the model with more parameters.
If you're using MARK, deviance is computed for you and presented in the results table. If you're using PRESENCE, you can use the log-likelihood value in place of deviance.
Cheers,
Jim
AIC imposes a penalty (2.0 for each parameter) for adding parameters. So, if two parameters are very similar and are estimated with a single parameter in one model, and two parameters in another model, the difference in deviances for the two models will be small (< 2.0) and you would choose the model with fewer parameters. If the actual parameters are very different, the difference in deviance will be larger than the penalty, and you will choose the model with more parameters.
If you're using MARK, deviance is computed for you and presented in the results table. If you're using PRESENCE, you can use the log-likelihood value in place of deviance.
Cheers,
Jim
- jhines
- Posts: 632
- Joined: Fri May 16, 2003 9:24 am
- Location: Laurel, MD, USA
computing model deviance
by jhines » Mon Aug 24, 2009 1:09 pm
I should have mentioned that if the last post helps, you will need to read the chapter in the MARK book (4.3) dealing with model selection.
Cheers,
Jim
Cheers,
Jim
- jhines
- Posts: 632
- Joined: Fri May 16, 2003 9:24 am
- Location: Laurel, MD, USA
-2log(L) in MARK
by dhewitt » Mon Aug 24, 2009 1:15 pm
To see the log-likelihood for models in the Results Browser in MARK (in addition to the deviance), check the fourth option under File --> Preferences.
And I agree with Jim that you probably need to read a bit more on model selection as an inference framework. Three references to check out:
1. Model selection and multimodel inference. Burnham and Anderson (2002; second edition). Springer.
2. Model based inference in the life sciences. Anderson (2008). Springer.
3. Chapter 3 of Occupancy estimation and modeling. MacKenzie et al. (2006). Academic Press/Elsevier.
And I agree with Jim that you probably need to read a bit more on model selection as an inference framework. Three references to check out:
1. Model selection and multimodel inference. Burnham and Anderson (2002; second edition). Springer.
2. Model based inference in the life sciences. Anderson (2008). Springer.
3. Chapter 3 of Occupancy estimation and modeling. MacKenzie et al. (2006). Academic Press/Elsevier.
- dhewitt
- Posts: 150
- Joined: Tue Nov 06, 2007 12:35 pm
- Location: Fairhope, AL 36532
5 posts
• Page 1 of 1
Who is online
Users browsing this forum: No registered users and 2 guests