First of all, thank you for your great work on the panelr package.
I was just running some models as robustness checks. In one of these models, I included an interaction term but only had two waves of data available.
I received this error (I think from lme4): Some predictor variables are on very different scales: consider rescaling. Also the interaction term's coefficient was far off (like a quadrillion times larger than the main effects).
At first, I was looking at my data because I was working with mean-centred variables. But the error was not connected to that. However, since the Giesselmann & Schmidt-Catran paper mentions that the "double-demean" approach only works with T>2 in the abstract, I switched to the conventional "demean". This worked. Also, the interactions (with the default setting) worked in multi-wave panels.
I have now had a short look at the approach and if I am not mistaken, with only two waves of data no within-subject variation remains in the computed interaction term. So differences in my data might only have been there due to floating point arithmetics (i.e., very [!] small differences or as the error message said, indeed a very different scale). I don't know if my explanation is correct because I did not look into the double-demean approach in detail. But I guess, what I saw was due to this approach being applied to only two waves?
Thus, I would suggest to output a meaningful warning in such cases, mention this fact for "interaction.style" in the documentation, or even switch the default in cases of only two waves of data.
Hi Jacob!
First of all, thank you for your great work on the panelr package.
I was just running some models as robustness checks. In one of these models, I included an interaction term but only had two waves of data available.
I received this error (I think from lme4):
Some predictor variables are on very different scales: consider rescaling.Also the interaction term's coefficient was far off (like a quadrillion times larger than the main effects).At first, I was looking at my data because I was working with mean-centred variables. But the error was not connected to that. However, since the Giesselmann & Schmidt-Catran paper mentions that the "double-demean" approach only works with T>2 in the abstract, I switched to the conventional "demean". This worked. Also, the interactions (with the default setting) worked in multi-wave panels.
I have now had a short look at the approach and if I am not mistaken, with only two waves of data no within-subject variation remains in the computed interaction term. So differences in my data might only have been there due to floating point arithmetics (i.e., very [!] small differences or as the error message said, indeed a very different scale). I don't know if my explanation is correct because I did not look into the double-demean approach in detail. But I guess, what I saw was due to this approach being applied to only two waves?
Thus, I would suggest to output a meaningful warning in such cases, mention this fact for "interaction.style" in the documentation, or even switch the default in cases of only two waves of data.
Best regards, Johann