Webinar Short Course on
|
|
The US Geological Survey and Proteus Wildlife Research Consultants are offering a free introductory workshop, which is open to all who are interested.
The course has been completed, but we hope to offer it again.
Please contact Paul_Geissler@usgs.gov if you would like to join our emailing list for webinar course announcements.
The presence or absence of a species across a set of landscape units is a fundamental concept used widely in ecology (e.g., species range or distribution, epidemiology, habitat modeling, resource selection probability functions, as a monitoring metric, metapopulation studies, biodiversity and species co-occurrence). An important sampling issue, however, is that a species may not always be detected when present at a landscape unit. This will result in "false absences" causing parameter estimates to be biased if unaccounted for, possibly leading to misleading results and conclusions, even with moderate levels of imperfect detection.
This introductory workshop will cover many of the latest methods for modeling patterns and dynamics of species occurrence in a landscape while accounting for the imperfect detection of the species. Participants will be introduced to the basic methods of analysis with worked examples and a strong emphasis on study design issues. Due to limited time there will be no software demonstrations or class exercises. While primarily aimed at the beginner and intermediate level, more experienced researchers will also benefit from attending.
Darryl is also offering in-person workshops you may be interested in.
Because of travel restrictions, we will need to conduct the short course over the web. You will be able to view PowerPoints and demonstrations on your computer screen. You can listen either using your computer speakers or by calling a phone bridge long distance. Power point presentations, lecture notes, background information, and recordings of the sessions will be posted on the web. If you use the telephone or have a headset or microphone, you can ask questions orally. Otherwise, you can type in your questions.
Certificates of participation are available to those who participate. US Department of the Interior employees can receive credit through DOI Learn.
For more information, please contact Paul Geissler (Paul_Geissler@usgs.gov, 970-226-9482)
Please register for the webinar at https://www1.gotomeeting.com/register/895225144
You are encouraged to register, even if conflicts prevent you from participating in some or all of the sessions or if you are only interested in some sessions. You can participate in the sessions that you are interested in and that are at a convenient time. You can watch the recordings of other sessions at a more convenient time. Registering will allow us to provide you with updates and more information about the course. There is no charge.
Some people on military bases have had a problem with this link. If you have a problem, please call Paul at 970-226-9482, and he will register you over the phone.
Files are available at ftp://ftpext.usgs.gov/pub/cr/co/fort.collins/Geissler/SpeciesOccurrence. We will update the PowerPoints and Handouts after each presentation.
The FTP site will usually be updated before this web page. Recordings (.wmv) files for each session will be available a day after the presentation.
Internet Explorer works well. Download them to your hard disk before opening them by right clicking and selecting "save link as". Files on FTP server are not password protected. If asked, just leave the user
name and password blank. The files are too large to e-mail. PowerPoints and handouts will be updated shortly before the presentations.
| Hawaii | Alaska | Pacific | Mountain | Central | Eastern | UTC |
| 8:30-10:30 | 10:30-12:30 | 11:30-1:30 | 12:30-2:30 | 1:30-3:30 | 2:30-4:30 | 18:30-20:30 |
SESSION 2 - Tuesday, August 24
SESSION 3 - Wednesday, August 25
SESSION 4 - Thursday, August 26
SESSION 5 -Friday, August 27
• Darryl MacKenzie, Ph.D., Proteus Wildlife Research Consultants, Dunedin New Zealand, darryl@proteus.co.nz. Darryl is a young biometrician with a rapidly growing international reputation. His main area of expertise is in using occupancy models for monitoring and research.
• James Nichols, Ph.D, USGS Patuxent Wildlife Research Center, jnichols@usgs.gov, 301-497-5660. Jim is is a wildlife biologist and senior scientist at the USGS Patuxent Wildlife Research Center. His main interests are the dynamics and management of animal populations and communities.
ORGANIZER
• Paul Geissler, Ph.D., USGS Status and Trends of Biological Resources Program. Contact Paul if you have any questions or problems with the webinar.
Synthetic BookOther Publications:
DISCUSSION
The email discussion items will be posted here.
Jack: Is software of webinar for ubuntu linux?
Darryl: The underlying 'number crunching' code for MARK and PRESENCE are Fortran
and C respectively, though both use something else to create the GUI,
which is restricted to Windows. Apparently MARK will run fine using an
emulator like Wine
(http://www.phidot.org/software/mark/download/index.html) and I suspect
that PRESENCE would probably be ok too but have not tried it myself.
Andrew: Although I can see the files listed at the FTP site, most of the files neither open nor down-load, including the first two PowerPoint lessons. The FTP site responds by beginning the download, but then freezes, ending with errors, e.g. "Internet Explorer can not download ...<file name> ... The connection with the server was reset".
Paul: I suggest that you try a ftp client, which may be more robust to network issues than a browser. There are several open source (free) ftp clients, such as FileZilla (http://filezilla-project.org/), which I use.
Andrew: That worked. I tried the traditional guest FTP parameters and used Core FTP Lite.
***** Sent 8/30/10 *****
Paul: This morning, I emailed certificates of participation to those who attended three or more of the live webinars. If you attended fewer sessions, send me an email after you have watched the other recordings or viewed the handouts saying you have done so, and I will email you a certificate. If you watched in a group, and only one member of the group signed in, let me know who else watched in the group and I will send them certificates. I used the name on the webinar registration on the certificates. Please let me know if it is garbled or if you would like another version of your name.
***** Sent 8/26/10 *****
Ryan: I was unable to attend the webinar on Tuesday, so I downloaded to file from the ftp site. It seems that the audio went out after about one minute into the recording. I'm not sure if others are having this issue as well.
Paul: On Tuesday, I had a problem with my headset and could not hear anything. I had to reinstall the headset and could hear, BUT it apparently disconnected the recording although I did not know it. I am very sorry, but at this point there is nothing can do about it. You can read the PPT slides at ftp://ftpext.usgs.gov/pub/cr/co/fort.collins/Geissler/SpeciesOccurrence/
Tracey: I forwarded today’s reminder to my home email, should the link on that email be sufficient to access the webinar? Or should I provide my home email for you to send the reminder to?
Paul: The link will work on any computer connected to the internet. Allow a few minutes to download software.
Warren: Thank, Paul. I'm enjoying it very much. Quick question. Are there any archived webinars that I could access?
Paul: Yes, see http://www.fort.usgs.gov/brdscience/Courses.htm
Laura: Assume to work with a territorial species that often mark (defecate) in the same places. When I conduct repeated surveys (scat-surveys) to detect the species, I naturally tend to check in the places where I know the species mark. There is a tendency to detect the species where I have already detected it. How can I resolve this bias in detection probability?
Jim: This is where the so-called trap response model comes in handy. You use a dummy covariate for detection probability for each occasion following first detection.
Clara: Could you use the models of species occurrence on a multivoltine
insect for which you survey the adults only. My concern is that an
absence could mean either a) an extinction b) it was present but not
observed and c) the adults were not present but their other life
stages were. Would this last possible scenario (c) be taken into
account by the detection probability (b), or would you want to
incorporate another variable that would represent the probability that
the organism is an adult and therefore detectable?
Jim: My view is that if non-adults are completely invisible to your sampling methods, then you can say little about them. If non-adults can be detected, but just with a different probability than adults, you can use a multistate model (tomorrow lectures).
Diana: We have a project which is using Logistic Regression to predict the
occurence of certain invertebrates on the seafloor based on habitat and
other environmental variables. The data used to establish the
relationship between species occurence and these variables comes from
samples (video transects of the seafloor). The environmental variables
are mapped throughout a geographic region of interest (from which the
samples came), and from the Logistic Regression coefficients the
presence of certain invertebrate species are predicted given the mapped
environmental variables. We have not incorporated detectability into our
model - yet.
My question is: how would the models being presented in this webinar be
incorporated into what we are doing (i.e. predict species occurance
based on environmental factors). Would we replace our model with a
PRESENCE model altogether, or would a PRESENCE model be used to derive a
term to add our Logistic Regression? I am a beginner at this, so hope my
question makes sense.
Jim: You would replace LR analysis with PRESENCE analysis. PRESENCE essentially conducts logistic regression within a larger framework that permits nondetection. Key in you analysis would be a means of obtaining replicate detection data (e.g., different observers viewing the footage; perhaps spatial replication).
Jimmy: Due to time difference (its midnight here!!) I have not been able to follow the webinar directly, although I'm following up the discussions and ppt posted. I was just wondering if there would be any real time analysis in the webinars in future. Regards,
Paul: I think Darryl does software demonstrations and exercises in his in-person workshops (http://www.proteus.co.nz/workshops.html), but we do not do so during the webinars.
Paul: Darryl mentioned the possibility of using any two states not just presence absence, such as counts greater than a threshold. In that situation, one can have both false-positives as well as false-negatives. It would seem to me that it would be necessary to model these extra possibilities.
Jim: We have developed some models to deal with false positives (in review, new stuff). However, in this particular case, I believe that Darryl refers to situations in which if you detect large number of animals, you assume this to be the true state (you can’t see more animals than are really present). Ambiguity arises only with the state of fewer animals, as there may be more there and you simply fail to detect them.
Paul: Systematic and spatial balanced sampling (GRTS) is frequently used because they reduce the variance of estimates in the presence of environmental gradients, because the intraclass correlation (between points) is negative. However, systematic sampling has two major problems. (a) adding, deleting or replacing point negates the spatial balance, and (b) a good variance estimator is not available. GRTS sampling solves both problems, so one can add, delete and replace points with disturbing the spatial balance (and variance reducing properties) and it has an excellent local variance estimator. However, if an independent random sampling (IRS) variance estimator is used with GRTS sampling, the variance is overestimated, providing the variance appropriate for IRS. Even if IRS is used, GRTS is still advantageous because it gives the same variance as one would respect with IRS, but the estimates are more precise than they are estimated to be. Can these PAO estimates be used with GRTS sampling, taking advantage of the GRTS local variance estimator? My uninformed thought would be to use jackknife sampling over sites.
Jim: I hope that Darryl addresses this question as well. I don’t know as much about GRTS sampling as I should – just a basic idea. But the idea of embedding particular spatial sampling designs within an occupancy modeling framework is very reasonable. The key is always thinking about what sort of correlation structure you expect among sampling units for the parameters of interest. So what sort of spatial covariance among psi’s for different sampling units do you expect a GRTS design to induce?
Questions Log 2010_08_26
Q: I have a "presence" version from 2008, is there a newer one? Where to get? Thank you
A: I suspect so, check Jim Hines' website to be certain.
A: http://www.mbr-pwrc.usgs.gov/software.html
Q: Is spatially random sampling imperative for regional-scale studies, it approximately 5E6 systematically placed samples cover the entire study area?
A: No, there are many sampling designs and the selection of one should depend on the question you are asking. As noted on day 1, if you are focusing on habitat differences (2 types), you want most sampling on these 2 types and less intense sampling in the rest of the landscape.
Q: Where to find the 2nd session WMV file?
A: We lost the audio on the recording.
Q: Is spatially random sampling imperative for regional-scale studies, it approximately 5E6 systematically placed samples cover the entire study area?
A: I would suggest spatially balanced GRTS sampling. Unlike systematic sampling, it has an excellent variance estimator. A problem is that I do not know if a spatial balanced or spatial sample will work with PAO. Perhaps, Jim or Darryl knows.
Q: Is Darryl using probability of occupancy interchangeably with probability of detection
A: No, probability that a randomly selected sampling unit is occupied by species, vs. Probability of a detection on a unit that is occupied by species.
Q: In my case, grid cells are 100 m by 100m... Input data resolution ranges from 10 m cells to watersheds ~200 km^2, and the dependent variable is defined by a study using a 1/2 acre minimum mapping unit that conducts ground truthing only within wetlands, presumed to focus field sampling efforts where the species of concern is likely to occur. For this type of a correlation analysis that hopes to investigate invasive wetland plant distributions within the U.S. Great Lakes region with respect to environmental variables derived from publicly available data sets, could a systematic sampling approach be superior to a spatially random sampling approach?
A: Possibly. I would need more detail. Systematic sampling is reasonable for lots of objectives, but be aware of the conceptual difficulties with variance estimation. I make no claim that this is a big deal in practice, but you should be aware.
Q: I have a landbird sampling design that includes 30 bird routes with 12 points on each route. The points are 250 meters apart. Each route is surveyed once every 3 years. Is there an example out there that would give me an idea of 1) how to address the spatial autocorrelation and 2) breaking minutes at each point on a route into 1 minute sampling events so I can use the data in presence to address the detectability of a species on a route and in the sampling area (30 routes?
A: (1) If you use spatial stops as replicates, then look at: Hines, J.E., J.D. Nichols, J.A. Royle, D.I. MacKenzie, A.M. Gopalaswamy, N. S. Kumar, and K.U. Karanth. 2010. Tigers on trails: Occupancy modeling for cluster sampling. Ecol. Appl. 20:1456-1466. You can use time of detection approaches which make use of the so-called trap response occupancy models we spoke about the first day. For individual bird detection probability see: Farnsworth, G.L., K.H. Pollock, J.D. Nichols, T.R. Simons, J.E. Hines, and J.R. Sauer. 2002. A removal model for estimating detection probabilities from point count surveys. Auk 119:414-425. We would use this kind of thinking for occupancy (at how many sites did we first detect in 1st minute, second minute, etc.
Q: For a species like cheetahs who have a home range of 200 square Km, would you use more people for the study or less people and more time?
A: Again, specify question (not just study organism) before making this decision.
Q: What do you do if males and females have different size home ranges?
A: Depends on question. Some folks focus on sex with the smaller home range and use that as basis for selecting sample unit size.
Q: If a population is not restricted to a point of the habitat, is there reason to take independent sample units
A: Must specify question. For lots of questions, independent sample units are sensible, but if you can't do this, recall the auto-logistic models that deal with occupancy of neighbor sites.
Q: If you have say like a grid of some radiotracked animals, with each grid that can be "occupied" (say, where you have at least 1 radiolocation) or not occupied (no locations). It does make sense to apply these occupancy models to the grids of this example? (so within home range scale). The species is solitary and territorial
A: Don't understand question. If you are only concerned about occupancy or not by the radiomarked animals, then you may not need these models (you know what areas the animals do and don't use), but you must still specify question or quantity of interest. In this case you might use PRESENCE, setting p=1 (perfect detection).
Q: What does unequal probability sampling means?
A: Where the probability of selection can vary continuously. See Thompson, S.K. 2002, Sampling, Wiley
Q: Can current occupancy models accommodate more than 2 states? E.g. a 3 state problem: (1) absent, (2) present but not breeding, (3) present & breeding.
A: Yes, tomorrow lecture.
Q: So inference can only be made within the boundary which you're sampling? Like the BBS survey using point counts along a transect; they could only make inferences regarding the population within those transects
A: I think of BBS routes as representing a cigar-shaped area around the route. You draw inferences about the routes. If you have habitat covariates, then you may also draw inferences about occupancy of same-sized cigar-shaped areas without BBS routes (but for which habitat is known).
Q: Is NZ + or - 12 wrt GMT?
A: I think about +18 or 20 wrt EST.
Q: But what if you are in very remote areas, the cells of the grid that are occupied by totally inaccessible forest (for example) will have almost always zero probability of being selected. I mean, is it possible to "adjust" random sampling AND do the field work feasible?
A: If there are potential units for which sampling probability is 0, then they are not part of the population of units about which you can draw inferences.
Q: When there are different numbers of sampling units in a multiseason design, can the number of sampling units be included in the modeling procedure?
A: Yes, this is easy within PRESENCE and MARK.
Q: If initial selection yr. 1 of 15 sites using GRTS and then in yrs. 2-10 we use the same sites. Does this limit the ability to make inferences about other areas because the probability of sampling other areas is zero once have locked into fixed sampling locations?
A: If initial selection is random, for example, then inference about entire area is still reasonable. More important, you can make inferences about colonization and extinction by visiting the same sites through time. You cannot do this if you select different set of sites each year.
Q: How does one overcome spatial dependence when choosing sampling units? Especially when surveying larger species, sampling units may be larger in size and therefore often adjacent to each other, which increases the chances of spatial dependence.
A: OK, so here is a situation where you may just have to deal with dependence via autologistic occupancy modeling (occupancy at focal unit depends on occupancy of neighboring units).
Q: My point would be to compare results in the space use patterns obtained with presence and those obtained, for instance, by an utilization distribution (kernel density estimation). OS I would use occupancy as a "measure" of intensity of use. Don't know if it's clearer now.
A: If space use implies habitat relationships, then you will want to select a sample unit size (does not have to be as large as home range size) and then use occupancy modeling to address the degree to which one habitat type is preferred (used more) to another.
Q: ...so a season could be a series of years if we tried to use "current" (recent) species distribution in conjunction with past case studies to develop a predictive tool for managers and deciders
A: Bad news about treating multiple years as one season is the likelihood of violating the closure assumption. Can't estimate p well without assuming that places occupied at one survey in season are still occupied at other surveys during season. Except for random use case, this may be a problem.
Q: Going by the definition of season as given by Darryl, is one day also a season
A: Yes, 1 day can be a season, depending on question.
Q: Sometimes, if you have some constraints regarding the number of sites to survey, isn't it preferable to disperse sites almost homogenously in order to cover almost all the area of which one wants to know the occupancy pattern? Sometimes doing strictly random sampling, it can happen that several points are too close to one another, and it leaves some "holes" in the distribution of the population.
A: There are cases in which systematic sampling is smart. I agree.
Q: Can we still use PRESENCE when we're sampling several ponds, during reproduction season, and we detect new individuals (tadpoles) between samplings occasions as well as juveniles leaving the ponds? Does the closure assumption is met? Can we consider the reproduction season as one single season even with individuals coming in and out
A: You can use occupancy modeling for sure, with the caveat that you think hard about what sort of model you use (depending on timing of metamorphosis, desire to include multiple life stages, etc.)
Q: Can you have 2 repeat surveys? Will the estimate still be reliable?
A: 2 is LOTS better than 1, but >2 is better than 2.
Q: how can multiple survey plots within a larger sample unit give you a true estimate of p? Seems like you confound small-scale variation with detectability
A: p is now defined as Pr(species in plot | species in sample unit) * pr(detection | species in plot).So detection probability definition depends on the nature of the replication (space, time). But occupancy is estimated just fine.
Q: Could data from fishermen from a determined area be accounted as valid survey for these models
A: Sure, good idea.
Q: What is the difference in the interpretation of psi in these two different situations? 1-an array of sites where is not possible that an individual detected in one site can't be detected in other site; and 2-an array where an individual can be detected in more than one site?
A: Not much difference in meaning of occupancy: it is still the probability that a unit is used by the species.
Q: What is the citation for the avian flu paper to which Darryl referred?
A: Kendall and White J. Appl Ecology 2009, I think.
Q: With presence/absence surveys, can we stop sampling the units where we have already found the sp (cause we are sure they are there), and just keep on surveying the ones where we haven't seen them yet
A: You can use removal models with such data. Flexibility in modeling is reduced a bit, but this is a pretty good design, especially when visits to sites require a lot of effort.
Q: What do you mean by heterogeneity
A: I can't recall context, but heterogeneity in some parameter among sites basically means that each site is not exposed to the same probability (e.g., of occurrence). Probabilities differ among sites.
Q: How do you control detection heterogeneity and distinguish this from the actual underlying heterogeneity from other parameters?
A: Control: covariates, finite mixture models, abundance models. Heterogeneity in p has different effects on detection history data than heterogeneity in psi.
Q: Heterogeneity is not always something to avoid. Sometimes it can give information about the random distribution of an organism in space, correct?
A: Heterogeneity in occupancy probability associated with habitat, for example, is something that you would like to quantify with PRESENCE (which you can).
Q: is there any way to account for potential miss-identification of species, as can happen, for example, in interview surveys where interviewees confound species?
A: A lot of work on this is occurring now. One fix is to stress to folks the problems with miss-id as opposed to missed presence. If this does not work we must go to modeling and these are a current topic of work here.
Q: "When there are a limited number of potential sub- units that could be sampled, and the species is not in all of them, then sub- units should be sampled with replacement." can you clarify what you mean by limited # of potential sub units? I would think that any time you sub-sample there will be, by definition, a limited (finite) number of them
A: There are some cases where number of potential subunits is so large so that sampling with and without replacement is not likely to differ much (unlikely to sample same unit twice even with replacement)
Q: If I want to contrast occupancy in specific habitats within the area, is better to use habitat as covariate and see their influence or make a separated survey in each habitat type?
A: I'd use habitat as covariate, because you may be lucky enough to have similar detection probabilities (not occupancy) between the habitat types. Then you don’t have to estimate detection separately for each area and power of your test increases.
Q: if I can make a suggestion about the course. It would be great if you could make a break of a few minutes after one hour for different reasons.
A: Thanks - not sure how possible this is, but we'll see.
Q: For a species like cheetahs who have a home range of 200 square Km, would you use more people for the study or less people and more time?
A: I'm not quite sure what you mean. Perhaps the more recent slides have answered your question?
Q: if I want to do occupancy studies with mid and large size mammals in a forest, and I use baits for increasing probability of detection, what would be a minimum and maximum distance recommended between cameras traps
A: Primarily it depends on how you want to interpret occupancy. Certainly you don't want to have the lure ranges overlapping.
Q: Can you write down the reference Darryl just mentioned about sampling
A: Which reference? What slide? There's a list of references available from the ftp site Paul keeps referring to.
Q: Suppose, one has surveyed a series of sites twice during two different visits (but within a season) and time constraints or the fact that some sites are more remote make less expensive to conduct the third and four surveys during only one visit (the third). Is it ok, or once one has started doing repeated surveys, it is not possible to conduct multiple surveys at a site during a single visit? I am not sure I have been clear....
A: You could just combine the results of the 2 surveys in that final visit and use a covariate to allow those sites to have a different detection probability compared to those that really only had a single survey.
Q: so why s is different for each sampling design...if all of them at the end sum up 48.
A: 48 is the number of site*sampling occasions. Other way of thinking about it is that in each design there were 96 surveys in total.
Q: Re: optimal number of surveys per sample unit. What effects does missing data have on the optimum number and patterns in variance? I.e. if sample units surveyed max of 3 times, but large portion of site only 2x or 1x, then we assume variance increases? What other suggestions would you have for dealing with this case?
A: The variance/standard error will increase. The 'magic numbers' indicate the number of non-missing surveys for each unit.
Q: What happens if a species is very rare AND/OR very elusive, and one can obtain only a small number of presence records over all the repeated surveys? I mean, what is the minimum number of presence records one should have in order to run these occupancy models?
A: I'm not sure of an exact minimum or where things go from working ok - not working well. However, when you say elusive, if may that you mean low detection then I suggest you probably what to have more repeat surveys, possibly with fewer units. For example, if you only have 3 surveys and 50 sites, but only have 6 detections all at different sites, then things probably won't work very well and you may have been better to go to 25 units with 6 repeat surveys. But if all 6 of you detections were only at 2 sites, that would indicate p is actually higher so 3 repeat surveys may have been sufficient and your species is just very rare.
Q: Will we get a full list of citations mentioned in the slides
A: Should be available from the course ftp site.
Q: Any guidance for how to determine # of surveys for community type surveys such as point counts for birds, where detectability and occupancy vary by species?
A: If there's some species of specific interest you might gear the sampling towards those species, otherwise I think (but am not certain) that you may be best to go for some sort of average design.
Q: where can I find a reference to cite for conducting 3 surveys when p>0.5
A: MacKenzie et al. (2006) Occupancy estimation and modeling.
Q: If I'm surveying big mammals, and stay 5mins in every sample unit to verify that the animal isn't really there, am I making my "p" nearly to 1?
A: Depends how mobile the mammals are and your question of interest. If you come back the next day to a place where you detected them previously, and you didn't see them would you consider that to be a real or temporary absence? If it's a temporary absence, then p<1.
Q: I cannot access the ftp site since yesterday, there is always an error occurring, but i have succeeded the first day. Is there any chance you could email me the pdfs?
A: PDFS are large. I'll let Paul know people are having some problems. Have you tried refreshing the page?
Q: I have grouse survey data where only 1 site (lek) is surveyed once per season but the same site (lek) is surveyed over many years, maybe not each year but say surveyed 7 out of 10 years - could this count as a repeat visit
A: Likely not because you expect possible changes (absence of closure) between seasons, right?
Q: Will Jim be presenting any of today's material, or is it all on Darryl's plate again
A: You prefer his accent? ;-)
Q: How can I estimate detectability?
A: Do you have any pilot data?
Q: If initial selection yr 1 of 15 sites using GRTS and then in yrs 2-10 we use the same sites. Does this limit the ability to make inferences about other areas because the probability of sampling other areas is zero once have locked into a fixed sampling location? Which Table for the magic numbers? When talking about variance mentioned this - I must have missed earlier
A: The table on the slide previous to the one with the variance equation on it.
Q: How can I interpret the SE of occupancy?
A: It's a measure of how precise an estimate you have. This is a fairly general statistical question so if you're not sure you might need to do some more reading of some introductory statistical texts
Q: Jim said that when multiple plots are sampled within a transect and occupancy is calculated at the transect level that "p is now defined as Pr (species in plot | species in sample unit) * pr (detection | species in plot). So detection probability definition depends on the nature of the replication (space, time)." Is there an easy way to calculate both "Pr (species in plot | species in sample unit)" and "pr (detection | species in plot)" using PRESENCE or is it impossible to separate those
A: If you have another level of sampling, e.g. repeat surveys on the transect segment, you could tease them apart.
A: Example is Nichols, J.D., L.L. Bailey, A.F. O’Connell, Jr, N.W. Talancy, E.H. Campbell, E.H.C. Grant, A.T. Gilbert , E.M. Annand, T.P. Husband, and J.E. Hines. 2008. Multi-scale occupancy estimation and modeling using multiple detection methods. J. Appl. Ecol. 45:1321-1329.
Q: All the 5 lesson will be recorded and put on the web site? We'll need to listen them again and again
A: Yes, although it sounds like there was a problem with audio for day 2
Q: So, independently of the detection probability of any certain species (high prob vs. .los prob) is it safe to say that more and repeated surveys are preferred?
A: To a point, although once detectability is at a certain level further repeat surveys are a waste of effort (in terms of estimating occupancy)
Q: Can GENPRES factor in different numbers of staff available along with the other variables used when analyzing completing sampling designs?
A: Only if that determines the number of repeat surveys you can do in a given year
Q: Does anybody apply this kind of methods to large scales questions, like occupancy of species X at county level across long time periods? Assuming you have data on different collectors that surveyed the area randomly each year, makes sense to look for extinctions-colonization patterns at that scales
A: Sure. For example, some folks in India are using it to look at tiger distribution across large portions of India.
A: Karanth, K.K., J.D. Nichols, J.E. Hines, K.U. Karanth, and N.L. Christensen. 2009. Patterns and determinants of mammal species occurrence in India. J. Appl. Ecol. 46: 1189-1200.
*Karanth, K., J.D. Nichols, K.U. Karanth, J.E. Hines, and N. L. Christensen, Jr. 2010. The shrinking ark: patterns of large mammal extinctions in India. Proc. Royal Soc. London 277:1971-1979.
*Karanth, K.K., J.D. Nichols, and J.E. Hines. Occurrence and distribution of Indian primates. Biol. Conserv. (in press)
Q: There was an audio problem on day 2? I heard it all here, and will check with our IT guys if there's a backed up copy of streamed content.
A: Problem with the recording at Paul's end so if someone else could drag out a copy I'm sure Paul would really appreciate it.
Q: There was an audio problem on day 2? I heard it all here, and will check with our IT guys if there's a backed up copy of streamed content.
A: The problem was with the recording. I had a problem with my headset and did not realize it would affect the recording.
Q: Do you have the reference of the India tiger large scale distribution?
A: Jim could probably list them off the top of his head, but one of the lead guys is Ullas Karanth. Searching for him you should be able to find some.
A: see above.
Q: One can be constrained to historical sampling units. How do we correct for this?
A: Your scope of inference is really restricted them to only sites with known historic status if everywhere else has no chance of being surveyed.
A: You can make inferences about extinction. Occupancy of sites at which species was once present is essentially the complement of extinction (in what fraction is the species still extant).
Q: Would it be adequate to work with sample units of different sizes
A: Depends on question of interest. If working on patches, then it could be fine.
A; Often wise to use size as a covariate in that case.
Q: I am interested in persistence. I have grouse survey data where only 1 site (lek) is surveyed once per season but the same site (lek) is surveyed over many years, maybe not each year but say surveyed 7 out of 10 years - could this count as a repeat visit over a broad temporal scale?
A: On what scale do you want to talk about persistence? If annual, and p<1, then you may have a problem. If you want to talk about persistence over a longer timeframe, e.g. 5 years, you could combine the annual data and use them as repeat surveys.
Q: Could you explain the difference between occupancy and use, explained before
A: Occupancy = species is present for the entire season, use = species is present sometimes during the season
Q: Does Darryl's in person workshop cover more advanced topics? What is the main difference between this webinar and Darryl's in person course?
A: Greater time for explanations. Hands on exercise with software. Most of the theory we're covering here is the same in the other workshops. Much more time for asking questions. See http://www.proteus.co.nz/workshops.html
Q: can you give an example where panel-type designs are useful?
A: Sorry, I am usually focused on dynamics, so I am not a fan (JDN)
Q: So, when we are dealing with migratory species, we could only speak in terms of use
A: Depends on timeframes with which you're doing the sampling, and when you're doing that relative to the migration patterns of the species
Q: Are there similar occupancy modeling approaches for community level parameters, such as Species Richness? In other words, incorporating these detectability and occupancy parameters into community ecology questions
A: Yes, Jim will talk a bit about this tomorrow
Q: Is the equation (slide 23) published somewhere?
A: Yes, MacKenzie et al. (2006) book and MacKenzie and Royle (2004; I think) J. Animal Ecology
A: Darryl does not know his own papers; I think it is J. Applied Ecol.
Q: If I conduct scat surveys and I remove scats during the surveys, do I need to consider this interference with the system in the future surveys? If I leave scats, naturally I can find them in the future surveys and for this I remove them
A: Removing them is fine. If you do not, you may need to use a trap response model where detection probability depends on previous detections.
Q: Could Jim expand on what was meant by lateral thinking?
A: Ask Darryl, as I haven’t a clue (JDN).
Q: Regarding sample size: what Darryl is talking as well as the material covered in the book talks about number of sites and survey occasions to estimate occupancy. But what would be adequate guidelines when we want or need to include covariates in our models? At least for log. regression a rule of thumb is to secure at least 10 presence and 10 absences per parameter we want to model... is there any good guideline that we could apply here? (i.e. how many data points per parameter?)
A: In terms of number of units, I'd give some similar guidelines, but these would be absolute minimums. If you looked at standard errors I bet that SE's would still be fairly large.
***** Sent 8/26/10 *****
Questions Log 2010_08_25
Q: Paul, Do you know when the .wmv file for yesterday's session will be posted
A: I worked at home yesterday to baby sit a sick granddaughter (she is not very sick) and I left the recording at home. So it will not be available until tomorrow. Sorry. [Monday, Tuesday and Wednesday's recordings are on the FTP site now.] Be sure to refresh your browser to see the recently added files.
Q: Will you comment about the use of museum species records for modeling species across space? I am wondering if there is any acceptable way to re-sample a large set of occurrences that will allow these data to be used for modeling, even though it was collected in a biased way. Thoughts?
A: Unless you have a way to determine where they’ve looked for the species and not found it, I don’t think there’s much that can be done unfortunately. There may also be a consideration that museum specimens may have been collected over many years during which there may have been changes in the distribution which distorts exactly what a resulting analysis is representing.
Q: Paul: I had a meeting yesterday and missed the workshop. Is there any way that I can do some make-up
A: The handout is available now and the recording will be available tomorrow at ftp://ftpext.usgs.gov/pub/cr/co/fort.collins/Geissler/SpeciesOccurrence/
Q: I have not been able to find any of the .wmv files yet on the ftp site ftp://ftpext.usgs.gov/pub/cr/co/fort.collins/Geissler/SpeciesOccurrence/ Do I have the correct site
A: Please refresh your browser to see recent additions. Monday's recording is there, but Tuesday's will not be available until tomorrow.
Q: Do you know if the occupancy models are used in benthic shellfisheries (e.g. clams)? My concern is because fishers normally need to have a certain density to fish on that sea bed due to effort-capture trade off and occupancy models seems to be used mostly on a presence-absence scenario.
A: While we’ve talked about it terms of presence/absence, these models could be used any time you’re trying to categorize units into 1 of 2 states where there’s ambiguity associated with one of the observations of those states, for example. Instead of presence/absence it could be high abundance, or not; realizing that even if a site is ‘high abundance’ you won’t always see that evidence (i.e. you have the potential to occasionally observe low abundance). Possibly this type of thinking is useful to you?
Q: How do we access handouts? Are these .ppt slides?
A: Yes
Q: Can't load the handouts link... is there another site where can I find them.
A: Please try again. It may be a problem of several people trying to access them at the same time.
Q: You said number of surveys may vary between seasons. Can they vary within a season? So 001 and 0001 in the same season?
A: At a different unit? Sure.
Q: Is the mean (or 0 when standardized) the appropriate value to use for missing covariate data? How come you can use dashes for missing survey data but not missing covariate data?
A: You should use a dash to represent missing covariate data in PRESENCE. Note that it’s presumed if you have the corresponding detection/nondetection data, you also need to have the covariate value. If you have missing covariate value, but a valid detection/nondetection observation PRESENCE will give you a warning.
Q: What is a Markov process?
A: http://en.wikipedia.org/wiki/Markov_process
Q: Where can I get the recordings of the webinar?
A: ftp://ftpext.usgs.gov/pub/cr/co/fort.collins/Geissler/SpeciesOccurrence/
Q: Can I use this model complementary to ecological niche factor analysis (enfa)?
A: We think you can, provided you’ve got suitable data. In fact we think this type of approach is better because you can explicitly account for imperfect detection.
Q: The model predicts non-random species distributions as well random distributions
A: Not quite sure what you mean by a random distribution, sorry.
Q: Where can I find the schedule of new USGS webinars, as well as the recordings of the older webinars?
A: http://www.fort.usgs.gov/brdscience/Courses.htm
Q: What is the definition of p i,j .. Can u show it again?
A: pij is the probability of detecting the species in the jth survey of an occupied unit.
Q: Does each survey site have its own epsilon and gamma, and the transition probability matrix is then calculated from the average epsilon and gamma values over all the sites?
A: Each site is assumed to have the same value, or they differ according to a set of covariates. When they differ the transition probability matrix is calculated separately for each site.
Q: What kind of Markov models are the questions on slide 31 based on?
A: It’s a first order Markov process. It’s essentially a special case of a Hidden Markov Model.
Q: It blocks the entrance to me.
A: If the FTP site asks for a user name and password, just leave them blank. If not, what message are you getting?
Q: Server denied.
A: I do not know what is happening. You should be able to view ftp://ftpext.usgs.gov/pub/cr/co/fort.collins/Geissler/SpeciesOccurrence/ by entering the URL in your browser.
Q: Could you repeat the name of the Markov model for the explicit dynamics using the matrix algebra
A: Hidden Markov Model
Q: Can you include interaction term between two covariates in PRESENCE?
A: Yes
Q: What about data missing in space, but present in just a few years?
A: I’m not sure what you mean; can you please give an example?
Q: Can these models be used to study fish communities? So far I've only seen examples with terrestrial species
A: Sure, we know of a number of folks that are applying them in such a way.
Q: Wrt to missing data & using an implicit vs. explicit modeling approach for multi-season data --- does the pattern of missing data make a difference in which approach you would adopt? e.g. (1) sites added as study progresses, but followed faithfully once included (missing data clumped at start of study) vs. (2) missing data distributed more willy nilly throughout data set?
A: Explicit model not going to work well if you don’t have many instances of sites being surveyed in consecutive years; there’s little information on colonization and extinction in that case so implicitly modeling is likely to be more stable.
Q: I didn't get what p was, sorry.
A: Probability of detection
Q: I'm sorry I missed the meaning of epsilon on slide 41 in the first model; please repeat
A: Probability of local extinction: species is present at time t, but absent at time t+1
Q: ARE WE GOING TO HAVE ANY QUICK BREAK?
A: Sorry, no.
Q: Are epsioln and Gama complementary to each other, i.e. does epsilo+Gama=1
A: No
Q: If we think of patches of null data in xyz space, are some of these concepts applicable to dealing with modeling species distribution based on imagery gathered from fall, summer and spring over about one year. By patches of missing data, I mean primarily missing patches in input/independent controlling variables and 'vital rates'
A: Sorry, I’m not following you.
Q: Were fences installed in year 0 or year 1?
A: The skink data was collected prior to fence erection.
Q: Why in the last gamma of the equation there is a t and not a t+1?
A: It’s just the notation we use, denoting gamma and epsilon in terms of the first period. Recall these are between season event hence if you have 6 seasons of data you’re going to have 5 gamma and epsilons.
Q: I did transect to collect microhabitat data at all camera trap locations to assess the microhabitat effect on studied species. Is it makes sense for me to put these microhabitat variables on probability of occupancy (psi)?
A: If you’re using the camera-trap location as you ‘unit’ then sure, provided you think that’s biologically reasonable.
Q: I believe there's a lambda in biogeography that is the ratio of colonization to extinction. Is this related to the lambdas you've mentioned, or useful, or is it problematic?
A: I’m not familiar with the lambda you’re referring to so can’t comment on it’s usefulness or how it’s been used, but it’s probably a case where we’re using the same name for a different concept. I don’t think the two are directly comparable.
Q: As brought up on Monday, these models need data that are collected in the same way? Does "f" also include surveyor bias that would be found in data, such as BBS?
A: ‘f’ it was related to how many times the house finch was detected along a route in previous years. We’re using it to allow the probability of detecting HF to be different once it’s been detected at more than 10 stops along a route. Not sure what you mean by ‘data that are collected in the same way?’
Q: Considering the house finch example, if detection probability is different over time as a function of abundance, would that same idea apply to rare species or species that are declining in population numbers?[
A: Possibly, depending on how the sampling is being done and the question of interest.
Q: A nice map would come out of these values.
A: One day I’ll figure out how, and get the time, to do it! ;-)
Q: My audio's acting up. Sounds like Darryl's gargling water while speaking. Anyone else having these problems or is this just my connection?
A: I am not experiencing that problem.
Q: What does the 'work account has one new message' mean in the software created for this webinar (g2m_download.exe)?
A: I don't know. In what context did you get the message?
Q: what guides the distance choices when building models like this? Does it have anything to do with the flight range of the bird species?
A: In house finch example ‘distance’ related to distance of each BBS route (sample unit) from the release point. Flight range wasn’t considered.
Q: Also, are models like these utilized when dealing with invasive species to curtail their expansion
A: Not sure about ‘utilized’ but could be used to predict possible expansions and how the rate of that expansion might change if management could alter colonization or extinction rates.
Q: What is the fundamental difference between the results from slides 49-52 and 53-56?
A: 49-52 are about parameters that were estimated directly and can be obtained from the software output, 53-56 were derived values that required some secondary calculations.
Q: This might be a foolish question. Can you download and save the movie files that you have placed on the ftp website? It will not allow me this option on my computer, thus the seemingly silly question.
A: I am recording this session and will make the recording available on the FTP site later today. Participants cannot record using GoToWebinar software.
Q: He/she could pipe the stereo sound in to a second computer's microphone jack, and record it in real time with his own comments/questions.
A: I will forward your suggestion to the group, but I suspect that it would be easier to wait and download the audio/video recording.
Q: Are point counts along a transect replicates or pseudoreplicates?
A: Points along a transect have a spatial correlation, which must be accounted for in an analysis model. They would be pseudoreplicates if the correlation is not accounted for in the model.
Q: When/where is the full workshop?
A: See http://www.proteus.co.nz/workshops.html for Darryl's in-person workshops.
Q: I think I can see pictures appear instantly.
A: So can I, but it depends on the speed of your internet connection.
Q: Does PRESENCE use the same algorithms as MARK does for computing the likelihoods? I always have issues with partial convergence failure w/ MR data in MARK and was wondering if PRESENCE would be equally likely to give such problems, issues with data sparseness aside.
A: Not sure if they use the exact same algorithms, but certainly similar techniques.
Q: Is this on the Klamath National Forest?
A: Not sure, near Willow Creek.
Q: Darn it, undergrads are people too, and a black bear 'wandering by' a few meters away is pretty thrilling for people who aren't dead yet... No?
A: ;-)
A: Same for post-docs from New Zealand ;-)
Q: I also have the audio problem with Daryl gargling but it is very sporadically.
A: I did have a couple of sips of water at one point. ;-)
Q: Will you be doing demo of PRESENCE software application?
A: Sorry, no
A: I think Darryl will demo software in his in-person workshops - see http://www.proteus.co.nz/workshops.html
Q: Is the Status and Trends of Biological Resources program the only USGS unit that does these types of webinars? If there are others, how do you find out about them?
A: As far as I know S&T is the only program offering webinars.
Q: How about Fish and Wildlife, BLM, Forestry, etc., do they have these types of webinars available to the public.
A: I am not aware of them. Try googling webinar and your topic of interest.
Q: Would you say that model 1 is better than model 2 on slide 69 being that there is one extra parameter in model 2 and not much of a change in Delta AIC or would you model average both models?
A: The additional parameter is accounted for by AIC. If I was trying to get my most honest estimate of certain parameters then in this case I would use model averaging (over all 3 models)
Q: I have a doubt on slide 69: can we choose the first model based on w (0, 62)? Delta AIC of the second model was less than 2.
A: I’d argue that 2nd model has sufficient weight that we can’t ignore it. People have started to move away from the delta-AIC <2 type argument these days.
Q: For model selection, which is preferred, AIC or delta AIC?
A: Delta-AIC is just AIC less a constant, hence they do the same job.
Q: Can you comment briefly on Markovian changes vs. Random changes in terms of anadromous species, when surveying population abundance and habitat use in-river?
A: For more mobile species, or if a longer time period between seasons (e.g. 5 years) then random changes may be a more reasonable assumption.
Q: Darryl's audio is echoing/cutting out throughout today's talk. Maybe batteries or microphone needs replacement? Thanks for checking for tomorrow.
A: No batteries. More likely to be internet traffic at some point between my microphone and your speakers. If it wasn’t a problem for most then it may be due to your internet connection. Sorry, not much I can do.
Q: Could some of these concepts be applicable to 'creeping environmental phenomena,' i.e. invasive plant spp
A: I guess so.
Q: b&w are true distribution, and the colored grids are what?
A: Same as yesterday. Different colours of blue (dark -> light) indicate lower probabilities of occupancy.
Q: Oregon State Office of BLM has science webinars. I don't know if they are open to all. Contact Karen_Blakney@blm.gov for further details. She is the Oregon Science Coordinator.
Q: Are the occupancy models used in patchy benthic resources, like clams? Usually a fairly high density is needed to run a profitable fishery?
A: Could be, depends on situation, scale and questions of interest.
Q: Is there anywhere to get technical assistance for these methods during an actual project?
A: For detailed assistance there are a few people like me [Darryl] that you can contract to help out. For ‘free’ help there are a few resources, look in the help menu of Program PRESENCE.
Q: Have you used GPS location data for presence data?
A: Not sure what you mean. In what way?
Q: p.s. I mean GPS location data from GPS collars.
A: You possibly could, but there are likely better ways to use that information than within this type of analysis. Depends on what you’re trying to do.
Q: if you have 63 stations and only ONE survey, at single-season model, is possible run the model
A: Not if p<1. Though it depends what your one survey actually is. Sometimes it’s possible to reformat your data to get at the required repeat survey information.
***** Sent 8/25/10 *****
Email Questions
Karl: Do you have a sense for when you anticipate next offering the Modeling Patterns and Dynamics of Species Occurrence course? Issues with vacation this time around. End of summer always a difficult time to take on anything additional.
Paul: We have not discussed it, but I hope we can offer the course again next year. Recordings and handouts are available at ftp://ftpext.usgs.gov/pub/cr/co/fort.collins/Geissler/SpeciesOccurrence/
Gordon: I don't see the recording from Monday on the ftp site yet. Will it be posted soon?
Paul: Please refresh your browser so you can see files that have been added. Mondays WMV file is there but not Tuesdays. I stayed home yesterday to baby sit a sick granddaughter and left the recording at home. She is not very sick.
Karl: Thanks. I can access those recordings at any time? Do I need to register?
Paul: Yes, they are available at any time at ftp://ftpext.usgs.gov/pub/cr/co/fort.collins/Geissler/SpeciesOccurrence/ without registering. However, I will appreciate your registering at https://www1.gotomeeting.com/register/895225144
so that we can get a better count of the participants. I can also provide a certificate of participation if you watch the recordings.
Laura: How do you model co-variates for rare species if there is not enough biology known about the species to determine which covariates are important or would influence detection or presence/absence?
Jim: No magic bullet here. My advice would be to rely on intuition, knowledge of related species, etc., to come up with very small number of potential covariates. This is generally preferable to shopping list approach of recording as many covariates as you can think of. Of course once you collect covariate data you can assess their utility via model selection in the occupancy framework.
Dave: I'm having some problems with GoToMeeting. I was able to join the webinar yesterday, with much difficulty, but I'm not having any luck today. I get the error message "GoToMeeting cannot be started due to a security certificate mismatch". There is a suggestion that downloading/reinstalling the software will fix the problem .... it did yesterday (eventually), but not today. We use GoToMeeting frequently within our institution, and have never had any problems. Any suggestions? Hopefully, you have come across this before. Thanks for your assistance.
Paul: My first thought was that it could be a firewall issue, because some organizations block GoToWebinar. However, this is probably not the case, because you are able to use GoToMeeting.
I cannot help but you could either ask your IT support people or contact GoToWebinar at http://support.gotomeeting.com/ics/support/default.asp?deptID=564
Dave: I think it is a firewall issue also, but our IT guys haven't been able to offer any help. Two of my colleagues in are having the same problem (GoToMeeting worked for lecture 1, but not for lecture 2) Are the dial-in numbers listed in your original email (213-286-1201; access code = 471-019-658) independent of the GoToMeeting software i.e. can we phone in to listen, and just follow along on printed lecture slides?
Paul: Yes, as far as I know. You could also try connecting from home.
Dave: If this is the case, can we then submit any questions directly to your, Jim's or Darryl's emails, to be added to the daily question log? If it helps, we can pool our questions before sending. Thanks again for your help.
Paul: Yes
Christine: Hello, I particpated in the webinar yesterday from my computer at work. Today I am at a different computer but on the same govt. email system/outlook. For some reason, I keep getting a message that due to problem with security certificate the system cannot connect me to the webinar. Not sure why this happens, I used both computers do participate in your webinar back in May and had no problems. Also I did follow instructions and downloaded the webinar program again and that did not make any difference. Today i am listening via phone which is not as effective. Hope to get on the web tomorrow. Thank you for any advice.
Paul: My first thought was that it could be a firewall issue, because some organizations block GoToWebinar. However, this is probably not the case, because you are able to use join the webinar yesterday.
I cannot help but you could either ask your IT support people or contact GoToWebinar at http://support.gotomeeting.com/ics/support/default.asp?deptID=564
Luciana: It was not clear to me why you don't consider a main objective on conservation/management to assess the capability of detecting trends. Suppose an scenario in which we have limited resources to invest on the estimation of population trends of a certain species, which is widely distributed in an area that is difficult to survey entirely. We could run simulation models prior to conduct the surveys, in order to evaluate alternative strategies of samples based on different combinations of survey effort (time, space, intensity). These analyzes could be used to identify the most convenient strategy considering the trade off between time and resources availability.
Jim: My point was that trend detection should not be considered as management. Perhaps it is a component of management, but management/conservation implies some sort of action designed to move the system in a desirable direction. Trend detection, by itself, cannot possible do this.
Husam: I am more interested in knowing about how to handle detection histories of two models first models dealing with interaction between 2 species and second models of multiple seasons and detrmining local extinction and recolonization. I have my own data for the two types. Another thing that I found difficult learning on my own ( I read many tutorials and practicals especially those of John Hynes) is how to read the PRESENCE output if I am dealing with two particular models. Hope that something is covered in the workshop regarding those issues.
Jim: We will cover multi-season and multi-species models explicitly. Can handle both issues with multistate models that are crafted in the right way, and we will try to cover this as well.
Questions Log 2010_08_24
Q: what is the website where I can print out the PowerPoint?
A: ftp://ftpext.usgs.gov/pub/cr/co/fort.collins/Geissler/SpeciesOccurrence/ They are PDF files.
Q: Paul, will everyone attending the webinar get the certificates, or only the ones who ask to
A: I will send a certificate to everyone that attends three or more live webinars. If you watch fewer live sessions, and watch the other recordings, send me an email telling me that and I will send a certificate.
Q: Last webinar the slides would change long before the voice caught up (for Darryl's part). I know you were trying to wait for the people's video to catch up, but it was very confusing for those of us that were synced with your computer
A: Another option is to download the files from ftp://ftpext.usgs.gov/pub/cr/co/fort.collins/Geissler/SpeciesOccurrence/ and follow along.
Q: How are you doing to check our participation on this course to certify us?
A: I get a list of those who login for each session.
Q: I won't be able to watch in real time over the next few days but I will listen afterwards. What should I do so you can verify and certify me?
A: Send me an email after you have watched the recordings saying that you watched them.
Q: Hi Paul...I am trying to download the files at this link "ftp://ftpext.usgs.gov/pub/cr/co/fort.collins/Geissler/SpeciesOccurrence" but I could not...Is the link working well
A: That link should work. If it asks for a user name and password, leave them blank.
Q: There were good questions asked and answered yesterday. Will you be Emailing out all questions and answers to the participants?
A: Yes, but there is a delay.
Q: Is there a file or a site where we could read other people's questions/answers for each session?
A: I will email and post them, but there is a delay.
Q: Where can we find full citations for references given in slides?
A: MacKenzie et al. (2006) book is a first stop. We have other specific references for specific topics.
A: See http://www.fort.usgs.gov/brdscience/SpeciesOccurrence.htm#References
Q: Where can i watch the video recordings?
A: ftp://ftpext.usgs.gov/pub/cr/co/fort.collins/Geissler/SpeciesOccurrence/
Q: So you need an assessment of appropriate habitat for your species as a first order activity?
A: You can assess occupancy whether you have a priori ideas about appropriate habitat or not.
Q: What if you are unable to hit the exact same spot (marine/at-sea surveys).
A: OK, then select some larger sample unit and just insure that each day's sampling device hits somewhere within that unit.
Q: So we can apply this to a stratified random sampling program?
A: Sure, you can estimate occupancy by stratum if you like. Or you could use the habitat factors that define strata as covariates. Approach is very flexible.
Q: P_J is the probability that a species will be detected during the survey? Or within an area in a particular survey?
A: It is the probability that the species will be detected at a single occupied unit in a single survey.
Q: Is a reference list available?
A: We can try to supply this. Mackenzie et al. (2006) should be good start and contain references prior to 2006.
A: See http://www.fort.usgs.gov/brdscience/SpeciesOccurrence.htm#References
Q: Does it change anything if detectability is density-dependent
A: Our approaches permit you to estimate detection probability, so it can vary according to many factors. Density is one of these factors and we describe method for dealing with this explicitly today.
Q: Will the questions/answers available on the net? If so, when?
A: Yes, I will both email them to the group and post them on the web. Unfortunately, there will be a delay.
Q: What about false positives - can they be considered in this modeling process?
A: We will discuss today. Royle-Link (2005, Ecology) contains general model, and new ms in review (Miller et al.) provides some practical approaches to such modeling.
Q: Are there any free software (R?) to perform such models?
A: Free software PRESENCE and MARK are available.
A: See http://www.mbr-pwrc.usgs.gov/software.html
Q: Why not multiply (1-psi) as well for the three occasions?
A: Two possibilities: Present psi and not observer on three occasions or not present (1-psi)
A: Nature flips a single coin to decide whether species is present or not. Only the detection process over the replicate samples requires the survey-specific probabilities (must miss species at all surveys in order to miss it entirely).
Q: How conservative is the estimate of occupancy when the species was present, but not detected (Slide 12 of Single Season Likelihood Approach)?
A: Not sure what you mean by conservative. Our estimates are approximately unbiased when assumptions are met.
Q: Will you post references related to the workshop, moreover than those mentioned on the slides
A: I think so.
Q: Is the model to predict occupancy state based in presence data?
A: Models estimate occupancy using detection-nondetection data.
Q: Could we get an example WinBUGS model fitting a state-space model?
A: The code given in the notes will work.
Q: What are the relevant references for the state-space approach?
A: Royle-Dorazio (2008) book is a good source, and several of the more recent Mackenzie et al. papers show both likelihood and state-space approaches.
Q: Will we hear more about state space models?
A: We will keep bring these up throughout the lectures.
Q: I can't remember what "psi" is
A: probability that a site is occupied
Q: What is the ftp website where all the presentations are archived?
A: ftp://ftpext.usgs.gov/pub/cr/co/fort.collins/Geissler/SpeciesOccurrence/
Q: With the state-space approach, you treat absences as unknown. This sounds similar to how pseudo-abs are treated. Could you discuss?
A: Non-detections are treated as uncertain, and the MCMC machinery tries different values for the true presence/absence of the species at those sites as it proceeds through the iterations/analysis. All of this is based on data that has been collected from the field. It’s quite different to how pseudo-absences are created in species distribution modeling as typically these aren’t based on any field data on detection/non-detections
Q: Can the method be used to estimate total population per area or a collection of sites?
A: What method are you referring to, and by total population are you referring to number of individuals?
Q: Can you choose multi-season models, but have only one-time habitat covariates (and simply use the same values for all the seasons), assuming they wouldn't change? I cannot assume single season model, but took the habitat measures only once.
A: Yes you can, but with multi-season model you still require multiple surveys per season to account for non-detections. I think this can be relaxed sometimes but requires some fairly strict untestable assumptions to be made.
Q: Why we have to use logit link? Why not binomial or another?
A: Theoretically you can use any link function you want. We’ve just stuck with the logit-link here because it’s commonly used. Note the binomial isn’t a link function.
Q: can occupancy and detection probabilities be functions of SURVEY-specific covariates
A: Detection probabilities can, but occupancy probabilities can’t because it’s assumed that occupancy state is actually unchanging for duration of repeat surveys so it doesn’t make since to try and model it with such covariates.
Q: Can the probability of availability be included with these analyses? (Probability that a given species is present, it is doing something that can be detected by the observer)
A: It can, if you have an additional layer/source of information, e.g. surveys where conducted on each of 5 different nights, but each night 2 observers were used and conducted independent surveys.
Q: If temp is continually increasing, say good habitat is 10 C.... Better is 12 C.... but 15 is too warm, how do you keep the logit from returning a 1 for the too high temps?
A: You can use transformations or polynomials of covariates. For example, if you expect the effect to asymptote out to certain value regardless of further increases in the covariate, you might consider a model with log(X) rather than just X. Or, if you think there’s some optimal value for the covariate X, include an X^2 term so you have a quadratic relationship.
Q: What is subscript j in the previous equation again?
A: Survey occasions
Q: How do you deal with modeling covariates that affect detection if for instance you don't know enough about the biology of a species or factors that influence detection? For instance, rare species. [
A: If you can develop a list of potential covariates you can assess how important each appears to be by comparing different models.
Q: Could you repeat again why Habitat D is considered standard?
A: Habitat D is the standard when using the indicator variables HabA – HabC because that category corresponds to the 3 indicator variables all being 0.
Q: Does this advantage hold for spatially missing data (i.e. null data for some independent variables in model inputs...)?
A: I don’t’ quite understand the question.
Q: Is that the same way you indicate a missing observation in program Mark?
A: Yes, you can denote a missing observation with either a ‘-‘ or a ‘.’.
Q: Why is an absence P-1?
A: I don’t understand the question
Q: Hi Darryl or Jim: With respect to missing data, is it reasonable to impute mean covariate values that are associated with missing data?
A: You could, but then are reliant on the mean value being pretty close to the ‘right’ value. Other approach would be to define a model for the covariate values and impute the missing covariate value itself. Easy to do within WinBUGS. But then reliant on covariate model being ok. Much better to not have missing covariate values. ;-)
Q: What happens when we are sampling highly vagile species?
A: All depends on how sampling units, seasons and surveys are defined. More on this in Session 4.
Q: Slide 34: I'm having trouble understanding the difference between "presence" and "use
A: Presence means the species is always there during a season, use mean it's sometimes there during the season. Hence 'occupancy' can depend on how you define a 'season'
Q: Does this mean that sampling continues during seasons when the species should normally not be present
A: No. A 'season' is completely arbitrary, but surveys should be conducted when you have a non- negligible chance of detecting the species. Unless it's a place where the species is absent, of course. More on 'seasons' when we talk about study design.
Q: How do you deal with lack of independence for bird surveys when 1 longer survey period, e.g. 10 min count, is broken down into 2 or 3 "surveys" to estimate detection probability?
A: You could use the covariate approach Jim just described so that you let the probability of redetection be different from the probability of first detection.
Q: Is the assumption that there is closure in each sampling unit, or over the entire survey area (all units combined
A: The former to be sure, and the latter would be preferably to ensure estimated occupancy is biologically meaningful.
Q: Could you please provide the full citation for the Hines et al. 2010 paper?
A: Will make a full reference list available later. Title includes 'tiger on trails' if you want to search for it.
A: http://www.esajournals.org/doi/pdf/10.1890/09-0321.1
Q: I would assume for a species with a large range, sampling units should be very large then to account for closure
A: Will talk more about this in study design session (day 4)
Q: Where can we find the full citations for references that are given in the slides?
A: Will provide a list later.
Q: Would you give some data for exercises later on (and advice regarding software
A: There will be no exercises in this course.
Q: I'm having problem understanding the table of slide 36 about lack of independence.
A: Table is an example for how you might be able to create a covariate (X) based upon the observed data (h) to allow the probability of redetection to be different from the probability of first detection.
Q: Slide 37: In this tiger example, is the pattern of local occupancy being modeled (i.e. dependence on adjacent occupancies) occur because tigers follow trails for > 1km?
A: yes
Q: Can we also use this covariate approach Jim described, in which the probability of first detection influences the probability of latter detection, be used for cases in which we have trap shyness (e.g. animals trapped once will avoid being trapped again)?
A: Yes, same idea as a behavioral response in mark-recapture
Q: In this one, are you assuming that you have ONE tiger? Or is there some underlying thing about clumped versus overdispersed populations in this?
A: not quite sure what you mean, but it's all at the species level. For tiger, their tend to occur either as singletons, or small groups (mums and cubs)
Q: What about stream dwelling amphibians, such that there is spatial autocorrelation?
A: Could be used, depending on question of interest and how you've done sampling.
Q: Wouldn't the probability of occupancy in segment 1 be higher given that segment 2 was occupied? Would we have to account for that in the model as well?
A: We are assuming some form of directionality to the problem.
Q: Will you or Jim talk about spatial autocorrelation in survey units, not in the sense of spatial replicates as with the tiger example, but for actual fixed sampling units that are too close together to be independent spatially?
A: yes, later today
Q: I'm wondering about species that tend to be clumped (not tigers). Where occupancy in one patch might provide information on occupancy in adjacent patches.
A: that would be spatial correlation at a slightly different level. More on this later.
Q: Wondering if these models can be applied to habitat occupancy of invasive plant species.
A: we think so
Q: I´m having some problem with audio, is it only me? Or is it a problem on the connection
A: coming through ok for me
Q: Why we have to use logit link? Why not another?
A: could use any you like. Logit-link is common and equivalent to logistic regression. Only logit link available in PRESENCE
Q: ...Melvin Hooton? (Utah State...)
A: I think the person Jim is referring to is someone Norris
Q: It might be good to lay out which chapter pages in the MacKenzie et al or Royle/Dorazo books are relevant for each day's talks?
A: The sessions pretty much follow a specific chapter from MacKenzie et al, so if you have a copy it should be reasonably obvious
Q: How do you decide how many groups are appropriate if you don't have any idea about covariates
A: You can different models with different numbers of groups and compare results to see which is 'best'. BUT the identified 'best' number of groups is unlikely to be biologically meaningful
Q: Good, thanks. The groups on slide 43 were related to maybe different kinds of habitat that might characterize the sampling units?
A: Groups are unknown for each unit. If it's something you can identify like habitat, you'd measure and use that as a covariate on detection instead.
Q: Can you clarify me the difference of these notations: ∏, Ѳ, λ, ψ? Thanks
A: Some were defined yesterday in session 1 notes, others we try to define before we use them. If you look back and check and let use know which ones we've missed.
Q: what is the r in slide 44?
A: the probability of detecting an individual
Q: on Nonconditional occupancy BRS example, what is the 0.12 in the occupancy estimate?
A: the standard error for the estimate
Q: what is meant by closure? Didn't catch that.
A: will discuss more in next 2 days
Q: I don't understand where the 0.6 figure came from on slide 30
A: 0.6 is the estimate of psi - unconditional occupancy
Q: Does PRESENCE calculate SE's using the delta method or does it have to be done long hand?
A: PRESENCE does it for us
Q: I can´t get the power point slides from the web site, is blocked?
A: only pdfs are available (I think)
Q: I can´t get the power point slides from the web site, is blocked?
A: It should not be blocked. If it asks for a user name and password, leave them blank.
Q: were?
A: ftp://ftpext.usgs.gov/pub/cr/co/fort.collins/Geissler/SpeciesOccurrence/
Q: What would be the maximum proportion of missing data allowed in order to still get trusty results?
A: Depends how much non-missing data you have really rather than degree of missing data
Q: I am not sure you are getting my questions since there is no response.... could you please check this?
A: We have a lot of questions, but we will get to them when we can.
Q: I understand WHAT the 0.6 was on slide 30, I just don't understand how you arrived at that estimate; it just sort of magically appeared.
A: It was estimate from real data in Program PRESENCE
Q: does PRESENCE give you the p-value of the random effect?
A: PRESENCE can not do random effects in maximum likelihood approach
Q: the "complementary log-log link on p" model to account for abundance induced heterogeneity Jim just mentioned - is there some reference or source to check this out?
A: It's referred in Chapter 5 of MacKenzie et al. (2006) Occupancy estimation and modeling (our book)
Q: so with WinBUGS?
A: With WinBUGS you can get average p and is standard deviation. It depends on exactly how you set it up in WinBUGS
Q: That link says that is an error with the FTP server
A: Please try again. It seems to be working now.
Q: The implicit assumption that the number of animals at a unit is constant is most likely false. Can you compensate for this in the model?
A: Depends on the species and what you're defining as a sampling unit, but this abundance approach is one way of doing that.
Q: What does TS stand for?
A: Test Statistic
Q: what does wrt is?
A: with respect to
Q: are we going to receive the scripts of the questions and answers?
A: yes
Q: Can psi and p be calculate for each "site" if there are only 4-5 replicates per "site"? Is there a guideline for how many replicates are needed per site?
A: Even with covariates, you use entire data set for estimation. Number of replicates is an important part of study design (topic Thursday)
Q: Can you access model fit when you include covariates?
A: Yes, example I showed did assume covariates.
Q: The implicit assumption that the number of animals at a unit is constant is most likely false. Can you compensate for this in the model?
A: This was the stuff about abundance-induced heterogeneity that was covered
Q: What is the r in slide 44?
A: Probability that an individual is detected
Q: Can you clarify me the difference of these notations: ∏, Ѳ, λ, ψ? Thanks
A: Upper case pi is product, psi is occupancy, and I am not sure where you saw the lambda.
Q: Good, thanks. The groups on slide 43 were related to maybe different kinds of habitat that might characterize the sampling units?
A: Perhaps, but if we were smart enough to hypothesize that the habitat might be relevant to detection, then we would enter them as covariates rather than use the omnibus mixture model
Q: How do you decide how many groups are appropriate if you don't have any idea about covariates?
A: Yes, a single model is defined by the number of groups it uses. You can test different numbers (seldom need more than 2 or 3) via different models.
Q: What's the full Mackenzie 2006 reference?
A: MacKenzie, D.I., J.D. Nichols, J.A. Royle, K.H. Pollock, L.A. Bailey, and J.E. Hines. 2006. Occupancy modeling and estimation. Academic Press, San Diego, CA. 324pp.
Q: Is there a place where I can get a description of using the delta method for calculating standard errors for conditional occupancy estimates?
A: Seber 1982, Williams et al. 2002 books have this i know. Can't recall whether occupancy book has this.
Q: In reference to slide 51, psi and P appear to be estimated for each site and have different values. How is this done if psi and p are calculated using the entire data set?
A: Covariates differ from site to site, even though the relationship between covariates and parameters is the same over all sites.
Q: Statisticians can use Excel without wretching?
A: you bet
Q: I can't even get the website with the slides to open. It asks me if I want to transfer files with Core FTP. I say yes and the website tries to open but can't connect. I say no, and it doesn't even try to connect. What's up?
A: I don't know. If it continues to be a problem, please contact Paul_Geissler@usgs.gov.
Q: Are some of you available to cooperate for a joined scientific paper or only for consultancy?
A: We both work with many different groups and are available depending on time constraints.
Q: for to use the effect of other species how a covariate...do you use psi or p
A: other species may influence psi and/or p
Q: This is spatial correlation only in the pr (detection)?
A: Tigers on trails was just for detection. Darryl later spoke about occupancy.
A: http://www.esajournals.org/doi/pdf/10.1890/09-0321.1
Q: can you give me a cite of Sargeant et al 2005
A: J. Wildlife Management should be enough for you, right?
Q: when you introduce other species how covariates in PRESENCE, what probability do you introduce? p or psi?
A: other species may influence either parameter
Q: Can you use covariates to deal with different sampling gear, or gear efficiency changing across sampling units?
A: yes, this is an important use of covariates
Q: could you provide a key for the color maps?
A: Darker means higher probabilities of occupancy. Red outline means that places were surveyed.
Q: How would you address the question of how an individual is using its habitat, i.e. you find the animal at a site/sampling area, but you don't know if it is breeding, dispersing, or transitory? Sometimes straight occupancy can be misleading
A: Right, see day 5 on multistate occupancy.
Q: yes, other species may influence either parameter, but when you introduce the probability in PRESENCE, What index do you introduce how a covariate, psi or p?
A: You have a separate design matrix for each kind of parameter in the program.
Q: I'm a bit confused with the used of covariates. If I think that e.g. patch size influences occupancy, should I use it as a covariate to achieve better estimates of occupancy, or as explanatory variable in a e.g. logistic regression? Or maybe both?
A: If detection probability is not 1, then you want to conduct analysis in PRESENCE. You can think of PRESENCE as allowing you to conduct logistic regression in the case where p<1.
Q: Can I have continuous numerical data as covariates?
A: Sure, continuous or categorical is possible.
Q: Will you speak more specifically about number of parameters of a certain model in the future?
A: I don't know how much detail we get into. If I have no covariates and 2 sampling occasions, I have 3 parameters, psi and detection probability for occasion 1 and occasion 2.
Q: Curious to know if you did these calculations on the Weta and detection probability (with observer and day) by hand using the equations or through a stats package (Presence)?
A: PRESENCE, because most of the equations do not yield closed-form solutions (can't find peak of likelihood function analytically).
Q: what about if pr of detection is around 0.95? It would be better logistic regression or PRESENCE?
A: PRESENCE, because it deals with p<1. And if p=1, it gives same answers as logistic regression.
Q: In slide 74, where does -1.07 come from?
A: It’s the effect size associated with the Obs1 covariate from the regression equation given on slide 72
Q: It seems you are more from the "Bayesian Camp", how do you feel about using mixed models, GAMs, etc. to model/predict species abundances/occurrences?
A: Actually I hop between camps. These models are essentially a form of mixed model. Whether a Bayesian or a Frequentist both could use a mixed model, GAM, GLM etc, it’s how they use them to make inference that’s slightly different.
Q: When you add one covariate to psi, for example, than this sums up an additional parameter or does the number of parameters stay the same, since we still have one value of psi, but being dependent on COV? So, the model psi (COV)p(.) has one parameter more than psi(.)p(.) or not?
A: Yes, 1 more parameter as we have to estimate the size of the effect for COV.
Q: Is there a chat value that is too large to continue with occupancy modeling?
A: No more so than with any other form of data analysis. It may be indicating that you need to consider additional covariates as there’s unexplained variation in your data.
Q: On Slide 71, the most complicated model, model fit: p-value=0.20, denotes not a good fit right?
A: p value of 0.2 is not bad at all. There’s insufficient evidence of lack of fit.
Q: I don't regularly consider 0.2 a good fit.....what is the range of "good
A: There’s certainty a degree of subjectivity to when you say there’s enough evidence. First recall that we’re actually testing for lack-of-fit. A p-value of 0.2 says that 1 in 5 of the test-statistics that were generated when the considered model was known to be correct are more extreme than the observed value from the real data. Hence, the observed value isn’t (in my opinion) that unusual. Certainly if the value was more like 0.1 then I’d getting more nervous and considering making adjustments.
Q: what about variability between individuals? Is there any method to account for this kind of heteroheneity? Could we use information about radio tracked individuals for instance?
A: We haven’t given a lot of thought to that. Most of what we’re talking about here is at the species-level
Q: Data augmentation was mentioned under the State Space Approach and I was wondering if you could describe what data augmentation is?
A: Just another name for essentially the same thing.
Q: I'm just wondering if you will help to make the connections between the abstract model theory and how to use the software to run the models. Also, how do you select which software to use?
A: Sorry, these online workshops are primarily on the theory with some examples thrown in. It’s not feasible for us to get into software training with 500+ people online. In the in-person workshops we obviously spend a lot more time with the software.
A: See http://www.proteus.co.nz/workshops.html for workshops.
Q: In this example, what else do you specifically know about the overall sample grid?
A: Like what? All we use is the location of each surveyed grid, and determine which grids are the neighbors of each other.
Q: ...or any stratification?
A: There is no stratification
Q: ...so black ~complete certainty, and white ~no detections, while shades of blue represent ~level of detection truth?
A: Black is very high probability of occupancy (near 1 so near certainty of presence), white is very low probability of occupancy (near 0 so near certainty of absence) and shades of blue are somewhere in between. Everything is in terms of occupancy, not detection.
Q: ...so the second of these two (slide 68) shows a grid following a neighborhood analysis-based adjustment to original "raw" survey data?
A: Kind of, they’re estimated values after accounting for the fact that the probability of particular unit is occupied depends upon how many of its neighbors might be occupied.
**** sent 8/24/10 ****
Emailed Questions:
Satya: It was realy an interesting lecture--really good one. But, I am not very clear about 'Logit function'.
Why do we need to use this link fuction? Can you please explain with simple example?
Paul: See http://en.wikipedia.org/wiki/Logit and http://en.wikipedia.org/wiki/Generalized_linear_model
Bill: I had a hard time getting into Webinar. I think it has to do with our Forest Service System. I got the ppt and listen in on the call it worked fine for me.
Paul: I am sorry you had a problem, but that seems to be a workable solution. If you are on the phone, you can ask questions orally by clicking on the hand icon in the control panel and waiting to be recognized. Watching the recordings is another option.
Jim: Is the webinar this week something different each day or is this one lesson that is given repeatedly for more opportunity to work-it-into schedules?
Paul: There different topics each day. See http://www.fort.usgs.gov/brdscience/SpeciesOccurrence.htm for the schedule.
Questions Log 2010_08_23:
Q: Hi Paul, just to test: how long do the recordings of the presentations will be on the website
A: We will keep them up indefinitely.
Q: Is it necessary open the R program?
A: We will not be having computer demonstrations. Darryl does in his longer in-person workshops with demonstrations http://www.proteus.co.nz/workshops.html
Q: Will we be using the microphone during the webinar?
A: If you have a headset or microphone, click the hand icon in the console, and we will recognize you at an appropriate time.
Q: Hi Paul, the control panel says you are talking but I cannot hear you.
A: Please check your audio and speakers. I think the problem is on your end, because others can hear and your connection looks ok from hear.
Q: If I switch to the telephone mode will I still be able to use the chat to ask question?
A: Yes, everyone can use the questions.
Q: Paul, Can I assume the files posted to 'ftpext.usgs.gov by 'Session', each represent a day's worth of material?
A: Each session is 1 day
Q: The auditory is not always clear (voice gets cut off at times). Do you think this is something to do with my computer/system or is there anything you can do on your end?
A: I do not think there is anything we can do. Jim is clear to me, although we are at different locations. I may be a problem with your internet bandwidth.
A: It may be do to with your internet connection (e.g., current bandwidth) especially if you're on a Wi-Fi connection. Audio is ok for me.
Q: We are not on a Wi-Fi connection, and yet the audio is cutting out periodically for us
A: It may be because of how many other users are on your internet connection and the current internet traffic. If audio via internet is not very good you may want to use the phone-in option.
Q: I believe the quote is, "We need to gather more data before we can definitively conclude that we have failed to meet our objective
A: ;-)
Q: Where can I get the presentations files?
A: ftp://ftpext.usgs.gov/pub/cr/co/fort.collins/Geissler/SpeciesOccurrence/
Q: How do you measure the detectability of a species?
A: Through collecting appropriate data and using appropriate methods. We'll cover this in more detail tomorrow
Q: What is state variable?
A: Some metric to quantify the current state of a population, e.g. population size or proportion of area currently occupied by a species (e.g., current distribution)
Q: and vital rate/s?
A: The things that make the state variable change; survival, recruitment etc.
Q: What would be an adaptive design?
A: There's an initial selection of sample units, if the specs is detected in one of those units, then you survey all of its neighboring units. You then keep expanding until you don't detect the species in any of the recently surveyed units
Q: Can we use data from call counts of pheasants?
A: Could do, depends on the question you're trying to ask
Q: So state variable is like dependent variable and vital rate/s explanatory variable?
A: Depending on the question, they could be either. But more often both would be more like the dependent variables. More on these soon.
Q: If we don`t use/have spatial sampling design during species counting, is it can still be applied?
A: So you have more than 1 survey location, but didn't select those locations according to any suit of sampling design? If so, then strictly speaking, in that case inference is really restricted to just the surveyed locations.
Q: Does your experience with east coasters lead you to believe that good coffee does not exist east of the MS R?
A: Not being a coffee drinker, I wouldn't really know. This is really tough when I supposed to make sense at 6:30 in the morning!
Q: ouch, understood.
Q: Back to slide 41: can Jim clarify what lambda is?
A: Lambda is population growth rate or change. The ratio of the population sizes at 2 different times or places
Q: It will be helpful to say something more about the "strong assumptions" needed to make inferences from presence/speudo absence data
A: Haven't really got time to get into them here, at what assumptions are being made usually depends on the method. But the biggest one that is usually violated is random spatial sampling. The second is the detectability of the species is equal at all points on the landscape
Q: Is there any viable sampling design to estimate the abundance of all species in a community since in this level the majority of species are mainly rare?
A: Without being too flippant about it, anything is viable if your budget is big enough. What you're able to do with a given budget is not an easy question to answer here.
Q: Does it make sense to make inference on a relatively small spatial scale, but with the objective to correlate occupancy with covariates of microhabitat?
A: Depends on the species of interest, and how you're defining 'occupancy'. More on this in study design session
Q: How strong inferences made from RSPF's can be to model species occurrence within a geographic area?
A: Depends on the sampling scheme really. With a good scheme and protocols then inferences will be fairly strong. Note that the RSPF is at the population scale of used/unused unit rather than a RSPF at the scale of individuals animals (which might come from radio-tracking for example).
Q: Habitat selection varies from season to season (especially in birds!!) then how can we generalize the relationship?
A: If you want to account for those changes, then you need to obtain a snapshot of the system within each season, and then model how things change. The multi-season model we talk about on day 3 could be used for this.
Q: So how do you deal with organisms, such as rare mesocarnivores that may only be represented by those alternate collecting methods (e.g., road kills, second-hand observations, etc.?
A: It may be that those types of data aren't actually that useful for that question you're trying to ask. You may have to consider collecting other data like sign surveys, camera trapping etc.
Q: what does "ve" means in slide 62?
A: +ve = positive, and –ve = negative
Q: What`s the relationship between the probability of occupancy and abundance? Do we can get an abundance index?
A: Some folks try to view occupancy as an index to abundance, but we do not think of it in this way. Under certain strict assumptions, you can actually estimate abundance from occupancy data (see MacKenzie et al. 2006, stuff on abundance-induced heterogeneity).
Q: Is there any viable sampling design to estimate the abundance of all species in a community since in this level the majority of species are mainly rare?
A: JDN: Not sure what you mean. If you mean estimate the species abundance distribution (number of animals in each species of community) I don't know of any omnibus method that will work for all species simultaneously. However, you can estimate species richness using species list data across replicates (temporal replicates at different locations or spatial replicates within each of a number of larger areas).
Q: What is a spatial Markov process?
A: http://en.wikipedia.org/wiki/Markov_process
Q: So how do you deal with organisms, such as rare mesocarnivores that may only be represented by those alternate collecting methods (e.g., road kills, second-hand observations, etc.)?
A: JDN: Occupancy modeling assumes no specific type of sampling, so road-kills are fine, recognizing that you can only say things about variation in occupancy among roads within different habitats, perhaps, but not about areas with and without roads (because you have specified no sampling in areas without roads).
Q: is Beals Smoothing = borrowing info across species?
A: Kind of, but not how we would think about it necessarily
Q: which could be a good sampling design to sample the co-occurrence of 2 top predator species, at the same time reducing the rate of false absences?
A: It would depend upon the goals of the study and detectability of each species. Generally you would want to get the probability of detecting a species at least 1 at a unit be be fairly high, 85-95%
Q: It will be helpful to say something more about the "strong assumptions" needed to make inferences from presence/speudo absence data
A: JDN: Detection probabilities should be 1 (no non-detection). Actually, even then I am thinking that our statement is generous. I am not certain how you can say much about where animals are not found if you have only data on where they were found.
Q: Are sample units more analogous to a minimum mapping unit (MMU), or to systematically placed sample points, or are they completely distinct from either of these?
A: They can be whatever makes the most sense for your application. Personally I don’t like the idea of a sample point in the sense that observations inherently have to be tied to some area around that point to be at all meaningful. Therefore, people could use that presumed area to define some form of grid cell.
Q: Back to slide 41: can Jim clarify what lambda is?
A: It is simply a ratio of abundances. Over time, this is a rate of population change. Over space, we would call this relative abundance.
Q: Is it only me that cannot hear Darryl anymore?
A: I can hear him well, although we are on different continents.
Q: So state variable is like dependent variable and vital rate/s explanatory variable?
A: The way I would state this is simply that change in abundance is a function of the rates of survival, reproduction, and movement in and out. So the vital rates are not just correlated with change in abundance - they are the determinants of that change.
Q: What is state variable?
A: A quantity that characterizes current "state" of system (system well-being). It can be used to specify position/health of system at any point in time. Pop size, proportion of patches occupied, number of species present, fraction of habitat patches in each particular habitat type are examples of state variables.
Q: Will the answerers be available following the close of Darryl's portion of today's presentation?
A: We will send emails with questions and answers, but it will take us a while to compile the answers.
Q: and vital rate/s?
A: Vital rates are the rate parameters that bring about change in state variables (e.g., rates of survival and reproduction and movement for population size, state transition probabilities for habitat state models.
Q: Why having `curvature` in the MLE creates problems? Slide 89
A: I probably explained it too quick. Curvature doesn’t cause problems, but the degree of curvature in the likelihood function at the maximum likelihood estimate determines the standard error for that estimate. If you think about the likelihood function as defining a surface, with the MLE being the highest point on that surface (e.g. a hilltop), then the standard error is determined by how ‘sharp’ the hilltop is.
Q: What would be an adaptive design?
A: It is used in cases where organisms are rare and clumped. Stage 1 sampling could be simple random, for example, but then stage 2 sampling is of sample units adjacent to the units at which detections were made in stage 1. It is adaptive in that sense (conditional on what you find during first stage).
Q: Can we use data from call counts of pheasants?
A: Sure, there are no restrictions on method of detection.
Q: Is it possible to create occupancy model based on presence-absence detections (from scat surveys) and, having data on the use-non use of the same units by radiotracked animals, validate the model
A: Provided that both types of data are collected over similar time scales, and that the radio-tracked animals are in the same areas of those leaving the scat, then this could be reasonable. If you’re radio-tracked animals are only a subset of the population and so in some areas where scat surveys were conducted there’s no chance of a radio-tracked animal being there, then the 2 may not agree with neither of them necessarily being wrong.
Q: Can you use inferences from single species vital rates of occupancy (extinction, colonization, etc.) to determine whether habitat patches may be sources or sinks?
A: You could define places with low extinction probabilities as potential sources for example, but really it depends on exactly how you want to define population sources and sinks.
Q: How do you measure the detectability of a species?
A: Many ways. In occupancy, the temporal replication allows us to do this. Imagine going to each location 3 times in consecutive days. If you detect species at a location on only 1 of the 3 days, versus all 3 days, this provides information about detection probability. This should become clear tomorrow.
Q: I believe the quote is, "We need to gather more data before we can definitively conclude that we have failed to meet our objective."
A: It is also very common to claim a need for more data before taking any management action.
Q: Is it fair to say that Bayesian approaches may be more appropriate than maximum likelihood approaches with respect to investigating spatial distributions in a temporal context with the goal of developing predictive management tools
A: Not sure I’d say more appropriate, but at the moment perhaps more practical because it’s theoretically easier to implement them. Whether they actually work properly for a specific analysis can still be touch and go. Sometimes though you could approximate a spatial model using covariates and maximum likelihood.
Q: Sorry, what did the green line means?
A: These were alternative prior distributions that might be considered and resulting posterior distribution.
Q: Will it be possible to provide complete list of literature (references) used in the course
A: Will do.
Q: What does Pr means?
A: probability
Q: Sorry, what does it means "Pr"?? like Pr (extinction)
A: My fault, probability of extinction.
Q: What software is recommended for occupancy modeling?
A: Program PRESENCE has been developed for these models specifically. They are also implemented in Program MARK because of the close similarity with mark-recapture models.
A: See http://www.mbr-pwrc.usgs.gov/software.html
Q: Why the 'link' function is used?
A: Link function is just a transformation so that covariate values that may range between +/- infinity can be used on a dependent variable that has to lie between 0 and 1 (i.e. a probability).
Q: What I meant is if you know of any community study which considers detectability when the state variable is the abundances of the species and not the richness.
A: An Indian colleague is completing a Ph.D. thesis that uses distance sampling to estimate abundance for ungulates in a park area of India. So for the ungulate community, he is indeed including detection probability for all species.
Q: I don't see how you can make sure a very rare mammal species DOESN'T occur in a location, even with long-term camera trapping. How do you usually deal with that uncertainty?
A: If you estimate detection probability for this or a related species, you can make statements about the probability of a species being present, given x visits with y detection probability at each visit.
Q: Did Darryl just reference the PRISM model?
A: On which slide? And which PRISM model?
Q: Can logistic regression be regarded as a RSF?
A: It often forms the basis for one.
Q: Can the presentations be distributed to us before tomorrow's class?
A: Yes, the recordings should be available later today. The PTFs are on the FTP site now. ftp://ftpext.usgs.gov/pub/cr/co/fort.collins/Geissler/SpeciesOccurrence/
Q: For a regional-scale presence/non-presence single species correlation study in the northern Midwest region of the U.S., how important are issues of compounding error from various input data sets with respect to confounding actual causality for an invasive wetland plant species?
A: Question is a bit too general for me to know how to answer. Certainly failure to deal with detection probabilities, in cases where detection is related to habitat, can lead to misleading inferences.
Q: how can I compute de CI of the dependent in a logistic regression? Can I use 1.96 (SE) or there is another way
A: That likely to be a reasonable approximation in many cases, then backtransform the limits from the logistic scale to the real scale
Q: Can the subsamples inside a transect be used as replicates to estimate probabilities rates or the occurrence frequencies used as abundance of a species in place?
A: Not certain what is being asked? We can sometimes use spatial replication (as might involve transect segments) to estimate occupancy.
Q: what does K_L models mean?
A: Kullback - Leibler
Q: It'd be good to get some refs on the criticisms of hypothesis testing Darryl mentioned. Are there specific refs he had in mind?
A: Googling ‘Anderson hypothesis testing’ would give you a pretty good start.
A: http://warnercnr.colostate.edu/~anderson/PDF_files/TESTING.pdf
Q: I did transect to find/count otter's feces after having radiotracked these otters the previous night. My idea would be: build an occupancy model based on faces found (estimating also the detectability), and then, calibrate/evaluate occupancy model based on radiotelemetry (used - non used units by the radiotracked otters, that are territorial and linked to the river). Does this make sense for you?
A: In a general way it makes sense. There are also ways to use both data sources for inference simultaneously.
Q: (I've read something similar in MacKenzie & Nichols, 2004, but this was before the course, hope to catch more of it in these days! :))
Q: Define AIC again?
A: http://en.wikipedia.org/wiki/Akaike_information_criterion
Q: I did not understand why a detection probability <1 could affect colonization estimate both + and -. I could only think of -, as I might not find all newly colonized spots
A: So you estimate colonization from time t to t+1. You may have missed occupancy at lots of places at time t, leading you to think that sites were unoccupied (thus available to colonists) when they were really occupied. So an unoccupied to occupied suit may have really been occupied to occupied.
Q: what mean -ve and + ve effects on graph of slide 41?
A: positi"ve" (+ve) or negati"ve" (-ve)
Q: If color aerial photographs are RGB models of the visible light reflected from Earth's surface, at what point to models with outputs based off this and other models' outputs become less useful than field surveys?
A: I don't understand question, but detection and misclassification probabilities are important.
Q: Please, all this models and assumptions can be applied to marine organisms that live underwater without problems?
A: Sure, conditional on logistical issues. In fact I imagine that detection is often more of an issue for many marine species than for terrestrial species.
Q: Reference to an earlier question, if there are e.g. 300 sampling sites, is it that each site has to be visited at least 3 times to get detection probabilities.
A: If detectability is high, then 3 will be ok, but if lower then it would be too few and you would actually be better to go to fewer sites more frequently. More on this on day 4.
Q: Is there some assumption using spatial replicates to estimates occupancy?
A: Spatial replication in the sense that you’re sampling at multiple points on the landscape is required, at least for the problems that we’ve tended to think about.
Use of spatial replication requires (1) random selection of spatial replicates within the larger sample unit and (2) sampling these replicates with replacement.
Q: How critical is an understanding of the statistical background information to stay up with the course?
A: Not supercritical, but it will help with some of the later material. It’s there for reference if nothing else.
Q: any chance we work on some of these formulas with real data to better understand them
A: We don’t have time to do that in this form of course, but we do for in-person workshops http://www.proteus.co.nz/workshops.html
Q: Do you have to choose if you're interested in Pr(extinction) or Pr(occupancy) or are they both components of the multi-species, multi-season model?
A: Both can be estimated
Q: Which could be a good sampling design to sample the co-occurrence of 2 top predator species, at the same time reducing the rate of false absences?
A: Will depend upon the level of detectability of the species
Q: When you have detectability problems, is it better to repeat the same sample points many times or making more sample points?
A: Will talk more about this on day 4
Q: Where can we find the cited literature that are used in the webinar? Do we get a references list later?
A: Will distribute
Q: For a single species occupancy analysis, how many samples and replicates are needed?
A: Will talk more on these in session 4
Q: Thanks for the answer about the community, and I really meant the species abundance distribution. Do you think it is plausible that all the species have a similar detectability? If so, I still can use the counts to apply the SADs
A: I would think it would be very unlikely for all species in a community to have = detection probabilities.
Q: are we speaking about the possibility to use radiotelemetry data in occupancy models next days? (Maybe the 5th day.)
A: Possibly answered this above, but will very briefly touch on it on day 5
Q: What is a spatial Markov process?
A: Example: Presence of species at location x may depend on presence-absence of the species at nearby locations.
Q: What do you do with very large areas with diverse habitats? How do you define your sampling unit relative to the 'grain' of the region?
A: Depends on the exact question of interest. There no simple answer, sorry.
**** sent 8/23/10 ****
Jeanne: I may not be able to listen to each session live each day, but would like to see the recorded version. Is the session recorded, and how soon would it be available for viewing? Ideally, I'd like to catch up in the evening to any part of a session I missed during the day. Will this be possible?
Paul: I will convert the recordings to windows media format and post them on ftp://ftpext.usgs.gov/pub/cr/co/fort.collins/Geissler/SpeciesOccurrence/ so soon as I can, by the next day at the latest.
Cindy: Is there any way to test if my system will run the webinar? I remember having trouble making the webinar connection, that I had to change my settings from not automatically accepting cookies to auto-accept. I’ve changed the settings but would like to know if things will work BEFORE the webinar starts.
Paul: I will sign on an hour early to the webinar for those who want to test their system and I will accept calls at 970-226-9482 until 10 minutes before the presentation. I will not interrupt the presentation to take calls, but if you have problems you can watch the recordings.
Barbara: I missed to change my time zone when I registered. I put that I'm from Portugal but I didn't change my timezone. How do I do that in order to have the right schedule?
Paul: I suggest that you use http://www.worldtimezone.com/time-us12.html to find the time at your location. It is at 12:30 PM in United States - Denver.
Jana:
I believe I registered some months ago for the upcoming webinar on
occurrence modelling next week. However to my best knowledge I have not
received details on where and how to log in to the sessions. Do I need
anything or will I just find the relevant info on http://www.fort.usgs.gov/brdscience/SpeciesOccurrence.htm come next
week?
Paul: You should have received a reminder sent by GoToWebinar from me with the link to join the webinar. If not, please check your spam folder in case it went there. If your email address is wrong or if you cannot find the message, register again at https://www1.gotomeeting.com/register/895225144 . If you receive duplicate reminders, you can cancel one of them by clicking on a link in the reminder message.
Jim: Have you guys developed any workarounds for folks who have gotomeeting blocked via institutional firewalls, etc.?
When I was with the university of wisconsin this wasn't an issue, but I just found out that it will be in my new position as a government scientist.
Paul: I do not have a workaround, but you can either watch from another location (home) or watch the recordings at ftp://ftpext.usgs.gov/pub/cr/co/fort.collins/Geissler/SpeciesOccurrence/ . People at military bases often have this problem.
U.S. Department of the Interior |
U.S. Geological Survey
Page Contact Information: Paul_Geissler@usgs.gov
