The data series is from changes in bee hive weights per day, so all data points are important. They are not normally distributed (many zeros) hence the need to transform the series. I wanted to use Box-Cox to transform it but then I cannot have negative values.
I am not good with statistics so I was planning on using a box-cox transformation website. http://www.wessa.net/rwasp_boxcoxnorm.wasp
Thank you.
Could you be clearer as to what the model is that you are attempting to test?
Also you likely are able to use some sort of non-parametric test instead of using a transformation.
I am not using a specific model but would like to do segmented (piecewise) regression with the change in hive weight (dependent) and independent weather variables temperature, relative humidity, precipitation, wind and solar radiation to determine threshold values. Only temperature and relative humidity are normally distributed, the rest are skewed due to having many zeros, hence the need for transformation. Only the changes in hive weight contains negative values. I would also like to do path coefficient analysis to determine the effect of the independent variables on each other.
Dear Alexandra,
In this case, you can use a two-parameter Box-Cox transformation or Johnson translation system.
Thank you for your reply. How do you do a two-parameter Box-Cox transformation or Johnson translation system? Could you give examples of software?
29
Why do you want to transform the data? The zeros are real, and the fluctuations are naturally both positive and negative. It seems to me that your zeros may simply be zeros within the limits of your measurement. That is, you may be unable to detect changes in weight that are smaller than a given size.
Could we see a normal quantile plot of your data? This would be really helpful.
Dear Alexandra,
Both parameters of a two-parameter Box-Cox transformation can be estimated In R, for example. In R, the boxcox.fit() function in package geoR will fit the parameters. See https://cran.r-project.org/web/packages/geoR/
Ronán Michael Conroy, here is what a QQ plot looks like with the data.
Sergiy Prykhodko, I am not proficient with R, but I will ask help from someone who is.
188
FIrst, you don't have a huge concentration of zeros. In fact, 25 of your observations are zero to the first six decimal places.
Second, what you posted is not a quantile plot. I've attached one to this post. I have converted your values from lbs (which are silly units, impossible to do maths with) to grammes which at least allow you to express your results without resorting to decimals.
The plot shows weight changes are not normally distributed. The distribution is pretty symmetrical, but with a very large kurtosis.
But the next step requires asking about hives. How many hives have we here? Weight changes may well be clustered within hives. And indeed if these are serial measurements, then we are missing the pattern by graphing them as if they were single independent measurements.
Tell us more about your study design.
- Badges
- Science topic