search

Philipp-Neubauer / FirstAssessment

Analysing the time to first assessment of fish stocks
0 stars 0 forks source link

other possible, non-influential predictors #16

Closed mcmelnychuk closed 7 years ago

mcmelnychuk commented 7 years ago

Just throwing this out there: what are your thoughts about including a couple additional numerical predictors in the analysis, such as:

James-Thorson commented 7 years ago

Trophic level is predicted to be strongly associated with max length under size structure theory, and I doubt it's well measure for all our stocks. Year of first development, hard to say, but might be with the acceleration rate.

If rather keep the existing analysis as is, unless you have a strong opinion Phil

On Nov 19, 2016 12:16 PM, "Michael Melnychuk" notifications@github.com wrote:

Just throwing this out there: what are your thoughts about including a couple additional numerical predictors in the analysis, such as:

  • trophic level
  • year of fishery development. I wouldn't expect these to have a strong influence, and there could even be some confounding with maximum body size, but I'm just wondering if there might be some value in showing what other variables are not influential, so as to highlight (by comparison) the variables that are (max landings, price).

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/Philipp-Neubauer/FirstAssessment/issues/16, or mute the thread https://github.com/notifications/unsubscribe-auth/AHnqTZRlocDbaFdXEaDzE06wpBVR8hSPks5q_1kugaJpZM4K3Wsg .

Philipp-Neubauer commented 7 years ago

Hi Mike;

I didn't include TL since i) its not as readily available as size info, so would need to use taxonomic imputation to get it for quite a few taxa and ii) its strongly correlated with size. Since the latter doesn't come out as a strong predictor, it may siluffice to reference the strong correlation of size with TL to suggest that TL would have a similarly small influence. I' d say its worth discussing in the context of Sethi et al....

I like the idea of having year of fishery dev in there - not sure if it'll work since it also enters in the time-to-assessment calculation. Doesn't cost much to give that a quick try...

On 20/11/2016 9:16 AM, "Michael Melnychuk" notifications@github.com wrote:

Just throwing this out there: what are your thoughts about including a couple additional numerical predictors in the analysis, such as:

  • trophic level
  • year of fishery development. I wouldn't expect these to have a strong influence, and there could even be some confounding with maximum body size, but I'm just wondering if there might be some value in showing what other variables are not influential, so as to highlight (by comparison) the variables that are (max landings, price).

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/Philipp-Neubauer/FirstAssessment/issues/16, or mute the thread https://github.com/notifications/unsubscribe-auth/ACJDC7gw8iZXJO13DG1JonSa5yV0S717ks5q_1kugaJpZM4K3Wsg .

mcmelnychuk commented 7 years ago

Sure, no problem at all leaving the analysis as is, especially if year of development is already part of the calculation. To clarify, is the "start time" of the time-to-event analysis the year of development (as Sethi et al use it, year when landings reached 25% of max historical landings), or is it the year of first recorded landings?

Philipp-Neubauer commented 7 years ago

I think the first year could be informative since it gives a more straightforward way of talking about assessment probabilities for old vs newly landed stocks. I'll give it a go to see what happens.

The development year is the year that species first appear in the landings DB.

On 20/11/2016 10:05 AM, "Michael Melnychuk" notifications@github.com wrote:

Sure, no problem at all leaving the analysis as is, especially if year of development is already part of the calculation. To clarify, is the "start time" of the time-to-event analysis the year of development (as Sethi et al use it, year when landings reached 25% of max historical landings), or is it the year of first recorded landings?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/Philipp-Neubauer/FirstAssessment/issues/16#issuecomment-261739586, or mute the thread https://github.com/notifications/unsubscribe-auth/ACJDCzSty_YAMgmTv_ddbreFGv9ZWv4cks5q_2SfgaJpZM4K3Wsg .

Philipp-Neubauer commented 7 years ago

A thought - I tried year of fishery development as a predictor and it is HIGHLY influential, much more so than anything else. But it also screws up the model and leads to unrealistic predicitons.

I think the effect is due to a "technology effect" - i.e., prior to some date, it was simply impossible to have a model that would fit our definition of assessment. So for all stocks that were first landed prior to that date, there is a mandatory waiting period...

Since the positive effect means the model assumes a linear increase in assessment probability with year of first landing, all recently landed but un-assessed stocks will have a near 1 probability of being assessed very soon - not very realistic I'd say for small stocks. So my feeling is that we'd need to have a non-linear effect OR have a binary variable for pre- and post "first ever stock assessment" instead of just a linear effect with year. I think the technology effect is probably something we need to account for in some way...

Hope this makes sense...

mcmelnychuk commented 7 years ago

Apologies if I opened up a can of worms with that suggestion. The other predictors we have thus far can legitimately be seen as independent, whereas this one is probably more correlative (when landings and price of some species pass some threshold, they begin to be recorded, and sometime thereafter they might be assessed). If inclusion of this predictor is screwing up the model and predictions, should we just omit it given that it's not really independent anyway? Either way I'm happy to defer to your judgement about changing it to include a technology effect. It looks like the first assessment was in 1960, so that only leaves 10 years between the censored start of landings data and the first assessment.

Mike

On 2016-11-21 12:42 PM, Philipp Neubauer wrote:

A thought - I tried year of fishery development as a predictor and it is HIGHLY influential, much more so than anything else. But it also screws up the model and leads to unrealistic predicitons.

I think the effect is due to a "technology effect" - i.e., prior to some date, it was simply impossible to have a model that would fit our definition of assessment. So for all stocks that were first landed prior to that date, there is a mandatory waiting period...

Since the positive effect means the model assumes a linear increase in assessment probability with year of first landing, all recently landed but un-assessed stocks will have a near 1 probability of being assessed very soon - not very realistic I'd say for small stocks. So my feeling is that we'd need to have a non-linear effect OR have a binary variable for pre- and post "first ever stock assessment" instead of just a linear effect with year. I think the technology effect is probably something we need to account for in some way...

Hope this makes sense...

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/Philipp-Neubauer/FirstAssessment/issues/16#issuecomment-262060956, or mute the thread https://github.com/notifications/unsubscribe-auth/AV_oVdKOEP9SoAZXsO3FkQQTCtA5yj5zks5rAgIagaJpZM4K3Wsg.

Philipp-Neubauer commented 7 years ago

Hmm, maybe the time-to-assessment should really be assessment_year - max(1960,first_landing) then...taht would settle it without having to have a predictor...

On Tue, Nov 22, 2016 at 10:16 AM, Michael Melnychuk < notifications@github.com> wrote:

Apologies if I opened up a can of worms with that suggestion. The other predictors we have thus far can legitimately be seen as independent, whereas this one is probably more correlative (when landings and price of some species pass some threshold, they begin to be recorded, and sometime thereafter they might be assessed). If inclusion of this predictor is screwing up the model and predictions, should we just omit it given that it's not really independent anyway? Either way I'm happy to defer to your judgement about changing it to include a technology effect. It looks like the first assessment was in 1960, so that only leaves 10 years between the censored start of landings data and the first assessment.

Mike

On 2016-11-21 12:42 PM, Philipp Neubauer wrote:

A thought - I tried year of fishery development as a predictor and it is HIGHLY influential, much more so than anything else. But it also screws up the model and leads to unrealistic predicitons.

I think the effect is due to a "technology effect" - i.e., prior to some date, it was simply impossible to have a model that would fit our definition of assessment. So for all stocks that were first landed prior to that date, there is a mandatory waiting period...

Since the positive effect means the model assumes a linear increase in assessment probability with year of first landing, all recently landed but un-assessed stocks will have a near 1 probability of being assessed very soon - not very realistic I'd say for small stocks. So my feeling is that we'd need to have a non-linear effect OR have a binary variable for pre- and post "first ever stock assessment" instead of just a linear effect with year. I think the technology effect is probably something we need to account for in some way...

Hope this makes sense...

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/Philipp-Neubauer/FirstAssessment/ issues/16#issuecomment-262060956, or mute the thread https://github.com/notifications/unsubscribe-auth/AV_ oVdKOEP9SoAZXsO3FkQQTCtA5yj5zks5rAgIagaJpZM4K3Wsg.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/Philipp-Neubauer/FirstAssessment/issues/16#issuecomment-262069545, or mute the thread https://github.com/notifications/unsubscribe-auth/ACJDC2kSpnEKr4qjfzsMeZgdr6k2-V8wks5rAgoZgaJpZM4K3Wsg .

Phil

Philipp-Neubauer commented 7 years ago

I just pushed an updated model version (need to sync my dropbox file, so it won't have updated for you guys yet). The updated version has

i) the time to assessment calculated as per my last comment. I think the technology effect was a legitimate concern, and by taking 1960 as the start year, I took the easy way to eliminate that concern.

ii) I added the interaction term of landings*price - at first only out of curiosity, but it seems that a) it leads to a positive effect for length (i.e. faster assessment for larger species), which I think is interesting, b) Is negative, suggesting the price matters most when landings are low, and c) it provides more clarity for some of the random effects (e.g., Gadidae are now found to have slightly higher than average assessment rates). Those are all pretty interesting results, I thought, so I thought it is probably worth keeping that term in...

On Tue, Nov 22, 2016 at 10:20 AM, Philipp Neubauer neubauer.phil@gmail.com wrote:

Hmm, maybe the time-to-assessment should really be assessment_year - max(1960,first_landing) then...taht would settle it without having to have a predictor...

On Tue, Nov 22, 2016 at 10:16 AM, Michael Melnychuk < notifications@github.com> wrote:

Apologies if I opened up a can of worms with that suggestion. The other predictors we have thus far can legitimately be seen as independent, whereas this one is probably more correlative (when landings and price of some species pass some threshold, they begin to be recorded, and sometime thereafter they might be assessed). If inclusion of this predictor is screwing up the model and predictions, should we just omit it given that it's not really independent anyway? Either way I'm happy to defer to your judgement about changing it to include a technology effect. It looks like the first assessment was in 1960, so that only leaves 10 years between the censored start of landings data and the first assessment.

Mike

On 2016-11-21 12:42 PM, Philipp Neubauer wrote:

A thought - I tried year of fishery development as a predictor and it is HIGHLY influential, much more so than anything else. But it also screws up the model and leads to unrealistic predicitons.

I think the effect is due to a "technology effect" - i.e., prior to some date, it was simply impossible to have a model that would fit our definition of assessment. So for all stocks that were first landed prior to that date, there is a mandatory waiting period...

Since the positive effect means the model assumes a linear increase in assessment probability with year of first landing, all recently landed but un-assessed stocks will have a near 1 probability of being assessed very soon - not very realistic I'd say for small stocks. So my feeling is that we'd need to have a non-linear effect OR have a binary variable for pre- and post "first ever stock assessment" instead of just a linear effect with year. I think the technology effect is probably something we need to account for in some way...

Hope this makes sense...

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/Philipp-Neubauer/FirstAssessment/issues/ 16#issuecomment-262060956, or mute the thread https://github.com/notifications/unsubscribe-auth/AV_oVdKOE P9SoAZXsO3FkQQTCtA5yj5zks5rAgIagaJpZM4K3Wsg.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/Philipp-Neubauer/FirstAssessment/issues/16#issuecomment-262069545, or mute the thread https://github.com/notifications/unsubscribe-auth/ACJDC2kSpnEKr4qjfzsMeZgdr6k2-V8wks5rAgoZgaJpZM4K3Wsg .

Phil

Phil

mcmelnychuk commented 7 years ago

thanks, and sounds good - looking forward to going through it again.

Mike

On 2016-11-23 1:19 AM, Philipp Neubauer wrote:

I just pushed an updated model version (need to sync my dropbox file, so it won't have updated for you guys yet). The updated version has

i) the time to assessment calculated as per my last comment. I think the technology effect was a legitimate concern, and by taking 1960 as the start year, I took the easy way to eliminate that concern.

ii) I added the interaction term of landings*price - at first only out of curiosity, but it seems that a) it leads to a positive effect for length (i.e. faster assessment for larger species), which I think is interesting, b) Is negative, suggesting the price matters most when landings are low, and c) it provides more clarity for some of the random effects (e.g., Gadidae are now found to have slightly higher than average assessment rates). Those are all pretty interesting results, I thought, so I thought it is probably worth keeping that term in...

On Tue, Nov 22, 2016 at 10:20 AM, Philipp Neubauer neubauer.phil@gmail.com wrote:

Hmm, maybe the time-to-assessment should really be assessment_year - max(1960,first_landing) then...taht would settle it without having to have a predictor...

On Tue, Nov 22, 2016 at 10:16 AM, Michael Melnychuk < notifications@github.com> wrote:

Apologies if I opened up a can of worms with that suggestion. The other predictors we have thus far can legitimately be seen as independent, whereas this one is probably more correlative (when landings and price of some species pass some threshold, they begin to be recorded, and sometime thereafter they might be assessed). If inclusion of this predictor is screwing up the model and predictions, should we just omit it given that it's not really independent anyway? Either way I'm happy to defer to your judgement about changing it to include a technology effect. It looks like the first assessment was in 1960, so that only leaves 10 years between the censored start of landings data and the first assessment.

Mike

On 2016-11-21 12:42 PM, Philipp Neubauer wrote:

A thought - I tried year of fishery development as a predictor and it is HIGHLY influential, much more so than anything else. But it also screws up the model and leads to unrealistic predicitons.

I think the effect is due to a "technology effect" - i.e., prior to some date, it was simply impossible to have a model that would fit our definition of assessment. So for all stocks that were first landed prior to that date, there is a mandatory waiting period...

Since the positive effect means the model assumes a linear increase in assessment probability with year of first landing, all recently landed but un-assessed stocks will have a near 1 probability of being assessed very soon - not very realistic I'd say for small stocks. So my feeling is that we'd need to have a non-linear effect OR have a binary variable for pre- and post "first ever stock assessment" instead of just a linear effect with year. I think the technology effect is probably something we need to account for in some way...

Hope this makes sense...

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/Philipp-Neubauer/FirstAssessment/issues/ 16#issuecomment-262060956, or mute the thread https://github.com/notifications/unsubscribe-auth/AV_oVdKOE P9SoAZXsO3FkQQTCtA5yj5zks5rAgIagaJpZM4K3Wsg.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub

https://github.com/Philipp-Neubauer/FirstAssessment/issues/16#issuecomment-262069545, or mute the thread

https://github.com/notifications/unsubscribe-auth/ACJDC2kSpnEKr4qjfzsMeZgdr6k2-V8wks5rAgoZgaJpZM4K3Wsg .

Phil

Phil

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/Philipp-Neubauer/FirstAssessment/issues/16#issuecomment-262464984, or mute the thread https://github.com/notifications/unsubscribe-auth/AV_oVVhkIw1phhmtNuXoqzoAvYn6Bc4sks5rBAUfgaJpZM4K3Wsg.

Philipp-Neubauer commented 7 years ago

My take on this as of today:

mcmelnychuk commented 7 years ago

Sounds good!

Mike