Maddie Minett my opinion is applied to you data extension of OLS, a Multiple Linear Regression (MLR). In Ornithology is very common to regressed your variable (spending Time of bird)to more than two factors (habitats).

First thing I would do is determine how you're going to handle missing data. What proportion of rows have missing values? Are they missing at random? (*If a particular species was less likely to visit a particular feeder than another species, this could be important!). It might be worthwhile to conduct a multinomial regression (before even looking at feeding time) to see if there are differences in species preferences.

Next, take a look at the residual plots. How bad is the hetereoscedasticity? Are the residuals extremely skewed or are they fairly symmetric? Non-normal residuals are not necessarily fatal for conducting parametric tests. Also, time data are rarely normal. I'm no bird expert, but I'd wager your data show a strong positive skew (*a few birds spent a considerably long time at the feeder relative to the rest of the sample). In which case, a model including a gamma distribution (rather than Gaussian) might describe the data better.

Blaine Tomkins Thank-you for taking the time to answer. I really appreciate it.

This test is just one in a long list I've run already. I just didn't want to make the question too long by including all of them. A contingency table and Bonferri adjustment revealed that there is indeed a strong association between species and food preference e.g. Goldfinches and Niger. (I was advised to use a contingency table to analyse the species, food choices and the number of times they were chosen during a university drop-in statistics clinic). Would you say a multinominal regression would be preferable instead, or could it be used alongside this? Is it the multinomial regression that will generate the residual plots? Would this multinomial regression take proportion of time spent being vigilant into account (and answer if different species displayed significantly different proportions of time spent being vigilant at different feeders)?

The main focus of the study is food choice and proportion of time spent being vigilant at the feeder rather than feeding time (feeding time was recorded to be able to calculate the proportion of time spent being vigilant). The gaps in the proportion of time spent being vigilant column are where the bird either went behind the feeder and/or I couldn't determine when the bird first landed on the feeder. There are ~4300 visits across the whole study, and of course the number of each species varies greatly, as does the food choice. Out of all of those visits, the species and food choice were always identifiable, but not time/vigilance at the feeder. From eyeballing the data, I would say the proportion of time spent vigilant column has about 10-20% gaps.

Maddie Minett The study makes a bit more sense now. Thanks for clarifying. Regarding the contingency table v. multinomial regression, either should work for the categorical data. Regression would be preferred if your goal was to create a particular model for making predictions about bird behavior, but it doesn't sound like this is what you want to do. It sounds more like you're just interested in determining whether there are differences based on species (Correct me if I'm wrong). Chi-square/likelihood ratio test on data from the contingency table will give you what you need.

Now, regarding the second research question: proportion of time being vigilant. Using proportions as a DV is tricky. You'll need to consider a lot of things about your data, including: the shape of the distribution (density plot), the relative number of observations at the lower "0" and upper "1" bounds -and how to handle them, determine whether logarithmic transformation of data is appropriate and useful. Transformations often make handling/analyses of proportion data easier (It's a good idea to compare a plot of the transformed data to a plot of the non-transformed data). *If you're using proportions as the DV, forget the residual plot. Proportion data are generally not normally distributed or homoscedastic.

10-20% of missing data is quite a lot. I recommend reading up on MAR, MNAR, & MCAR to help determine the source of the missing data, and how you can handle the missing data based on the type.

Maddie Minett As to which type of analysis to conduct, I'm, admittedly, at a loss given this unique type of proportion data - which is continuous rather than based on count data. My guess is beta regression with a logit link, but I could be wrong. I'll recommend this question with hopes that those more knowledgeable than I can chime in

Maddie Minett my opinion is applied to you data extension of OLS, a Multiple Linear Regression (MLR). In Ornithology is very common to regressed your variable (spending Time of bird)to more than two factors (habitats).

A word of caution before using multiple regression. OLS assumes that residuals are normally distributed. Again, take a look at your residuals. If they are extremely skewed (or something like this), you will get biased estimates. If your residuals come from an exponential distribution, Maximum Likelihood Estimation may be the better choice.

Also, be careful with a DV based on proportions - mainly in the way you interpret them. For example, a difference between 2 ms and 4 ms feeding time is a proportion = .5, whereas a difference between 40 ms and 42 ms feeding time is a proportion = .95 - even though the difference was 2 ms in each case. Since you're only interested in relative comparisons across species, not absolute comparisons, you may wish to transform your DV data using a logarithmic function.