Limits...
Dealing with varying detection probability, unequal sample sizes and clumped distributions in count data.

Kotze DJ, O'Hara RB, Lehvävirta S - PLoS ONE (2012)

Bottom Line: Results did not improve when seasonality and number of trapping days were included in these models as offset terms, but only performed well when the response variable was specified as following a negative binomial distribution.Finally, if seasonal variation of a species is unknown, which is often the case, seasonality can be added as a free factor, resulting in well-performing negative binomial models.Based on these results we recommend (a) add sampling effort (number of trapping days in our example) to the models as an offset term, (b) if precise information is available on seasonal variation in detectability of a study object, add seasonality to the models as an offset term; (c) if information on seasonal variation in detectability is inadequate, add seasonality as a free factor; and (d) specify the response variable of count data as following a negative binomial or over-dispersed Poisson distribution.

View Article: PubMed Central - PubMed

Affiliation: Department of Environmental Sciences, University of Helsinki, Helsinki, Finland. johan.kotze@helsinki.fi

ABSTRACT
Temporal variation in the detectability of a species can bias estimates of relative abundance if not handled correctly. For example, when effort varies in space and/or time it becomes necessary to take variation in detectability into account when data are analyzed. We demonstrate the importance of incorporating seasonality into the analysis of data with unequal sample sizes due to lost traps at a particular density of a species. A case study of count data was simulated using a spring-active carabid beetle. Traps were 'lost' randomly during high beetle activity in high abundance sites and during low beetle activity in low abundance sites. Five different models were fitted to datasets with different levels of loss. If sample sizes were unequal and a seasonality variable was not included in models that assumed the number of individuals was log-normally distributed, the models severely under- or overestimated the true effect size. Results did not improve when seasonality and number of trapping days were included in these models as offset terms, but only performed well when the response variable was specified as following a negative binomial distribution. Finally, if seasonal variation of a species is unknown, which is often the case, seasonality can be added as a free factor, resulting in well-performing negative binomial models. Based on these results we recommend (a) add sampling effort (number of trapping days in our example) to the models as an offset term, (b) if precise information is available on seasonal variation in detectability of a study object, add seasonality to the models as an offset term; (c) if information on seasonal variation in detectability is inadequate, add seasonality as a free factor; and (d) specify the response variable of count data as following a negative binomial or over-dispersed Poisson distribution.

Show MeSH

Related in: MedlinePlus

Predicted catch after trap losses at low activity in low-abundance treatments.Box and whisker plots of the effect sizes (predicted catch) of the analyses performed with five models for original parameters estimated from data on Pterostichus oblongopunctatus abundances, and trap loss at low activity in treatments 1 and 2. Since no trap losses occurred at the high-abundance treatment (treatment 3), losses were zero for the last five box and whisker plots. See Figs. 1 and 3 for more details.
© Copyright Policy
Related In: Results  -  Collection


getmorefigures.php?uid=PMC3401226&req=5

pone-0040923-g002: Predicted catch after trap losses at low activity in low-abundance treatments.Box and whisker plots of the effect sizes (predicted catch) of the analyses performed with five models for original parameters estimated from data on Pterostichus oblongopunctatus abundances, and trap loss at low activity in treatments 1 and 2. Since no trap losses occurred at the high-abundance treatment (treatment 3), losses were zero for the last five box and whisker plots. See Figs. 1 and 3 for more details.

Mentions: Another problem related to the collection of ecological field data is that sample sizes may vary from one time or place to the next. Even with the best-prepared field experiments, ecologists are often faced with unbalanced designs. Designs may be unavoidably unbalanced from the start, for example because an investigator cannot make simultaneous observations at multiple localities and is consequently forced to sample different sites at different times. Samples may also become lost during the observation period, e.g. traps may be lost or broken or observers are unable to carry out all the observations required. If the experimental design is unbalanced, and, equally importantly, if the study organism varies in detectability over time (e.g., seasonal variation in activity), sampling effort may not be comparable between treatments. Simply stated, if samples are lost at different times during the field period (a common feature of studies in urban environments for example, see also [17]), pooling and standardizing the remaining samples over the whole field period may produce gross over- and underestimates of abundance and its variation, at least for species that are abundant or easily detectable only during some part of the season.


Dealing with varying detection probability, unequal sample sizes and clumped distributions in count data.

Kotze DJ, O'Hara RB, Lehvävirta S - PLoS ONE (2012)

Predicted catch after trap losses at low activity in low-abundance treatments.Box and whisker plots of the effect sizes (predicted catch) of the analyses performed with five models for original parameters estimated from data on Pterostichus oblongopunctatus abundances, and trap loss at low activity in treatments 1 and 2. Since no trap losses occurred at the high-abundance treatment (treatment 3), losses were zero for the last five box and whisker plots. See Figs. 1 and 3 for more details.
© Copyright Policy
Related In: Results  -  Collection

Show All Figures
getmorefigures.php?uid=PMC3401226&req=5

pone-0040923-g002: Predicted catch after trap losses at low activity in low-abundance treatments.Box and whisker plots of the effect sizes (predicted catch) of the analyses performed with five models for original parameters estimated from data on Pterostichus oblongopunctatus abundances, and trap loss at low activity in treatments 1 and 2. Since no trap losses occurred at the high-abundance treatment (treatment 3), losses were zero for the last five box and whisker plots. See Figs. 1 and 3 for more details.
Mentions: Another problem related to the collection of ecological field data is that sample sizes may vary from one time or place to the next. Even with the best-prepared field experiments, ecologists are often faced with unbalanced designs. Designs may be unavoidably unbalanced from the start, for example because an investigator cannot make simultaneous observations at multiple localities and is consequently forced to sample different sites at different times. Samples may also become lost during the observation period, e.g. traps may be lost or broken or observers are unable to carry out all the observations required. If the experimental design is unbalanced, and, equally importantly, if the study organism varies in detectability over time (e.g., seasonal variation in activity), sampling effort may not be comparable between treatments. Simply stated, if samples are lost at different times during the field period (a common feature of studies in urban environments for example, see also [17]), pooling and standardizing the remaining samples over the whole field period may produce gross over- and underestimates of abundance and its variation, at least for species that are abundant or easily detectable only during some part of the season.

Bottom Line: Results did not improve when seasonality and number of trapping days were included in these models as offset terms, but only performed well when the response variable was specified as following a negative binomial distribution.Finally, if seasonal variation of a species is unknown, which is often the case, seasonality can be added as a free factor, resulting in well-performing negative binomial models.Based on these results we recommend (a) add sampling effort (number of trapping days in our example) to the models as an offset term, (b) if precise information is available on seasonal variation in detectability of a study object, add seasonality to the models as an offset term; (c) if information on seasonal variation in detectability is inadequate, add seasonality as a free factor; and (d) specify the response variable of count data as following a negative binomial or over-dispersed Poisson distribution.

View Article: PubMed Central - PubMed

Affiliation: Department of Environmental Sciences, University of Helsinki, Helsinki, Finland. johan.kotze@helsinki.fi

ABSTRACT
Temporal variation in the detectability of a species can bias estimates of relative abundance if not handled correctly. For example, when effort varies in space and/or time it becomes necessary to take variation in detectability into account when data are analyzed. We demonstrate the importance of incorporating seasonality into the analysis of data with unequal sample sizes due to lost traps at a particular density of a species. A case study of count data was simulated using a spring-active carabid beetle. Traps were 'lost' randomly during high beetle activity in high abundance sites and during low beetle activity in low abundance sites. Five different models were fitted to datasets with different levels of loss. If sample sizes were unequal and a seasonality variable was not included in models that assumed the number of individuals was log-normally distributed, the models severely under- or overestimated the true effect size. Results did not improve when seasonality and number of trapping days were included in these models as offset terms, but only performed well when the response variable was specified as following a negative binomial distribution. Finally, if seasonal variation of a species is unknown, which is often the case, seasonality can be added as a free factor, resulting in well-performing negative binomial models. Based on these results we recommend (a) add sampling effort (number of trapping days in our example) to the models as an offset term, (b) if precise information is available on seasonal variation in detectability of a study object, add seasonality to the models as an offset term; (c) if information on seasonal variation in detectability is inadequate, add seasonality as a free factor; and (d) specify the response variable of count data as following a negative binomial or over-dispersed Poisson distribution.

Show MeSH
Related in: MedlinePlus