climwin 1.2.3

climwin v1.2.3 includes mostly minor internal changes to make climwin compatible with upcoming R v4.0.0

None

None

Technical changes

• Remove dependency on plyr, which was an artefact of old code.
• Ensure that code is compatible with changes to stringsAsFactors default in v4.0.0

climwin 1.2.2

climwin v1.2.2 includes a number of important bug fixes that may affect ongoing analyses.

Major changes

Nonsensical results with absolute/week

A major bug has been detected when running slidingwin() with cinterval = "week" and type = "absolute". In this particular case, results were nonsensical and would not match similar results generated using cinterval = "day"/"month". This issue was caused by the way we calculated the number of weeks in a year. This bug has now been fixed. Any previous analyses using this approach should be reconsidered.

Minor changes

plotwin now works with a single best window

When using cinterval = "month", it is possible to identify one best window with a large model weight (>0.95). This led to errors in plotwin(), which expects multiple models with small model weights. This issue has now been fixed, and plotwin() can work with results where there is a single best window.

Fixed bug where an error is returned when weights are fixed at 1

slidingwin() allows users to provide weights in the baseline model. However, when these weights are uniform at 1, this led to an error in the way slidingwin() updates the baseline model during the window fitting process. The method for dealing with weights has been changed and this bug has been removed.

climwin 1.2.1

climwin v1.2.1 is now fully compatible with R v3.6.0. To report errors and bugs or ask questions please visit the climwin google group.

R v3.6.0 contains a new default method for the function sample(), which is used inside the slidingwin function in climwin. These changes will affect the reproducibility of previous results run on older R version. Old results can be reproduced by running RNGversion(vstr = "3.5.3"), which will revert back to the old method for conducting sample().

No major changes

Minor changes

K-fold cross validation with a single fold

k-fold cross validation can now be conducted with k = 1.

Plot changes to work with ggplot2 >3.0

With v3.0 of ggplot2 some slight compatibility changes were introduced when creating plots. These have now been fixed.

climwin no longer prints to console

Previous versions of climwin used the print() function to write text in the console. This is now done using message() and warning(), which allows users to suppress information if desired (e.g. with suppressMessages()).

climwin 1.2.0

Our newest version includes a number of important tweaks and bug fixes, as well as an updated help vignette. To report errors and bugs or ask questions please visit the climwin google group.

Major changes

Functionality with ‘nlme’

Older versions of climwin allowed the use of mixed effects models with the package lme4; however, there have been a number of requests for nlme compatability as an alternative. This has been made possible in v1.2.0, with the option to include constant and exponential variance functions for model weighting (varIdent and varExp respectively). Note, however, that k-fold cross validation is currently not available with nlme models.

N.B. Please see minor changes below for other updates relating to mixed effects models.

K-fold cross validation included with weightwin

Along with the use of randwin, k-fold cross validation is an important tool for overcoming potential issues of over-fitting in climwin. K-fold cross validation was previously only functional in slidingwin, but has now been included with weightwin, using the argument k.

Using climate thresholds on weekly/monthly data

Users may run climwin with cinterval == “week”/“month” in cases where only weekly/monthly data is available but also to reduce computational time when using daily data (see news for v1.1.0). This raises potential concerns when using climatic thresholds (e.g. arguments upper, lower, binary). When daily data is available, users may wish to apply thresholds before data is grouped by week/month (e.g. for each week, what is the mean number of days that temperature > x). Alternatively, users may wish to apply thresholds after week/month grouping has occurred (e.g. is mean weekly temperature > x). v1.2.0 now includes a prompt allowing users to select between these two possibilities.

Changes to plots in weightwin

While running weightwin returns a number of plots to help the user track the optimisation process. We have replaced the bottom left plot to show the plotted relationship between the weighted mean of climate and the response variable specified in the baseline model. This allows users to spot any outliers that may be driving patterns in the data.

N.B. This plot does not account for any additional variables included in the baseline model. It should be used as a tool for identifying outliers rather than an accurate reflection of the statistical relationship between the variables.

Minor changes

Use of maximum likelihood for mixed effects models

When using mixed effects models in climwin it is necessary to use a maximum likelihood [ML] approach rather than restricted maximum likelihood [REML]. In v1.2.0 this has now been made mandatory. Any models fitted with REML will be refitted with ML (along with an accompanying message).

N.B. This means that the best model output will also be fitted with ML. Users can manually refit the best model with REML if desired.

Changes to the use of weekly data

The method use to group data at a weekly scale has been changed slightly. This may lead to slight changes in results when using the argument cinterval = "week" compared to previous versions; however, changes should be minimal and the overall interpretation of results should remain the same.

Changes to warnings and messages

Previously, climwin printed any messages and warnings immediately during model fitting; however, this disrupted the progress bar making it difficult for users to follow the progress of their analysis. All warning messages are now displayed at the end of the process. Messages associated with model fitting (such as rank-deficiency when using lmer) are now suppressed. These messages are expected when fitting very small climate windows and so should not be a concern for users.

Bug fixes

Slope statistic

Fixed a bug where the ‘slope’ statistic was calculated with the wrong sign.

climwin 1.1.0

Our newest version adds a number of useful features to climwin as well as a few bug fixes. In addition, we have now created a climwin google group for users to ask questions about the package. Please report any errors or bugs on this forum.

Major changes

Custom model structure

In older versions of climwin, climate data was extracted from the fitted climate window and included as a fixed effect in the specified baseline model. This precluded the use of more complex model structures, such as interactions between climate and other fixed effects or mixed effects models with random climate slopes. We now include a more versatile method for specifying model structure in climwin. This involves the inclusion of a dummy ‘climate’ variable in the baseline model (N.B. the variable must be called ‘climate’). This dummy variable will then be replaced with climate data from each fitted climate window, maintaining the original model structure. See an example of baseline model structure below:

baseline = lm(response ~ climate*sex)

Using this more versatile model construction makes the argument func redundant (i.e. the argument used to specify quadratic, cubic etc. terms). Users can now structure their model manually to test for these effects.

Note: We have maintained functionality for the original baseline model design. If no variable called ‘climate’ is provided the original method will be used.

Dealing with missing data

In previous versions, the argument cmissing could be designated as either TRUE or FALSE. When FALSE, the presence of missing values in any tested climate window would return an error. When TRUE, all climate windows containing missing values would be removed from our analysis. On further consideration, we felt it was unwise to remove data from our analysis. As an alternative, we now provide two methods to estimate the value of NA records. “method1” will replace the value of a missing cell with the mean of the two preceding and following records. “method2” will replace the value of a missing cell with the mean of all other records on the same date. For more detail on dealing with missing data please read our FAQ.

Multiple iterations with weightwin

As weightwin uses an optimisation function, there is the possibility that the function may fail to converge or will converge on a local optima. Because of these issues, a single run of weightwin may not provide an accurate measure of the climate landscape. Users should instead run multiple iterations of weightwin with varying starting parameters to help find the global optima. The argument n in weightwin allows users to specify the number of iterations to run, with starting parameters randomly assigned in each run. weightwin will then return a summary table of results from all iterations.

Minor changes

Non-daily data

The original design of climwin required users to provide their climate data at a daily resolution, even if users had only a single recorded value across each month/week. This caused some confusion with users. climwin can now deal with climate data at a monthly or weekly resolution (i.e. one record for each month/week). Running climate window analysis at a monthly or weekly scale also provides a method for dealing with missing climate data.

climwin 1.0.0

Major changes

climwin version 1.0.0 includes a number of major changes to coincide with the release of our corresponding paper [1]. We have tried to make most of our changes backwards compatible, but any major issues should be reported to the package maintainer (liam.bailey@anu.edu.au).

New function slidingwin:

climwin aims to distinguish between two separate methods for testing climate windows. The commonly used sliding window approach [1] and less common weighted window approach [2]. To ensure the distinction between these two methods is clear, the function climatewin has been made redundant and been replaced with the function slidingwin. Parameters used in slidingwin and climatewin are identical. Users can now conduct a sliding window analysis using slidingwin and a weighted window analysis using weightwin.

randwin for weighted window analysis:

The function randwin can now also be used to conduct randomisations with a weighted window approach (i.e. using the function weightwin). Users must now define whether randomisations are to be conducted using a sliding window (“sliding”) or weighted window (“weighted”) approach with the parameter ‘window’. Note that all parameters from weightwin will be required to run randwin using a weighted window approach (e.g. ‘par’, ‘weightfunc’).

Cohort parameter:

When a group of biological measurements covers two years (e.g. Southern hemisphere species which breed between November - February [2]) use of ‘absolute’ climate windows will cause these measurements to be split across different calendar years. To overcome this issue, we include a ‘cohort’ parameter to our functions.

The cohort variable will determine which biological measurements should be grouped together (e.g. measurements from the same breeding season), and ensure that these measurements share the same reference day. The cohort variable should come from the same dataset as the ‘bdate’ parameter (i.e. variables should have equal lengths).

See our advanced vignette for more details:

vignette("advanced_climwin", package = "climwin")

Spatial parameter:

Climate window analysis often requires large amounts of data to effectively determine periods of climate sensitivity. Often this is achieved through temporal replication (collecting many years of data on the same population) but can also be achieved through spatial replication (collecting data on multiple populations), or, ideally, a combination of the two. The new parameter ‘spatial’ allows users to carry out a climate window analysis with data from multiple populations by linking each set of biological measurements to a corresponding set of climate data.

Users can include data from multiple study sites/populations in their climate dataset (i.e. the dataset used for parameters ‘cdate’ and ‘xvar’), with a new site ID variable included to distinguish between different sites. Similarly, the user can add a new site ID variable to the biological dataset that can be used to link biological measurements to the corresponding climate data. When carrying out analyses the user can then include the parameter ‘spatial’, a list item, containing the biological site ID variable and climate site ID variable respectively. During model fitting, the climate window analysis will extract different climate data for each biological record based on the provided site ID.

N.B. Spatial replication in climate window analysis works on the assumption that all populations share the same period of climate sensitivity. If this is NOT the case, populations should be analysed separately.

See our advanced vignette for more details:

vignette("advanced_climwin", package = "climwin")

Cox proportional hazard models:

Proportional hazard models may often be useful for climate window analyses on phenological data. We have included the ability for users to fit proportional hazard models for the parameter ‘baseline’ using the function coxph. For more detail on understanding the use of proportional hazard models for phenology analysis see [2].

Function pvalue:

In previous versions of climwin climate windows have been compared visually using a number of methods (e.g. deltaAICc distribution, model weights). In this newest version, we have included two metrics that allow for a standard method of distinguishing real periods of climate sensitivity in biological data.

These two metrics, $$P_{C}$$ and $$P_{\Delta AICc}$$, determine the likelihood that a given climate window would occur by chance, given the results of a randwin analysis on the same data. They can be calculated using the new function pvalue. For more information on the effectiveness of the new metrics, please see [1].

‘exclude’ parameter:

Although climwin helps us move away from the selection of arbitrary climate windows the method is inherently exploratory, raising concerns about overfitting. Climate data from short duration time windows are particularly likely to show spurious relationships in climate window analysis [1]. The inclusion of the function pvalue (above) reduces the chance that these short duration windows will be mistaken as ‘real’ climate signals; however, these spurious windows can still cause problems when conducting multi-model inferencing as the short duration windows may be distinctly different from other top models. As a solution, we include the parameter ‘exclude’ in the functions slidingwin and randwin. This allows users to exclude windows of a specific duration and lag to prevent these small windows from interfering with analyses.

Minor changes

Parameter changes:

• ‘furthest’ and ‘closest’ are now combined into a single parameter ‘range’.

• ‘cutoff.day’ and ‘cutoff.month’ are now combined into a single parameter ‘refday’.

• ‘cvk’ parameter is now redundant, replaced with parameter ‘k’.

• ‘thresh’ parameter is now redundant, replaced with parameter ‘binary’.

• ‘type’ now accepts possible arguments ‘absolute’ and ‘relative’ (rather than ‘fixed’ and ‘variable’).

climwin 0.1.2

Fixed bug which caused convergence issues using cross-validation.

climwin 0.0.1

• Initial package release. Forthcoming changes will be noted.