I am currently working on my dissertation project regarding media usage trends and would like to receive some clarification on panel data model that uses time-invariant independent variables and time-variant dependent outcome.
Firstly, summary "mean graph" of 24 months worth of data revealed that there is a general decline among customers in usage of TV. In other words, live TV is becoming less and less popular.
Now, my main interest using PLM regression in R is to find out which variables or variable categories drive that change. For instance, one would think that young adults during 24 months period drastically changed their behaviour because they are more keen in exploring and using new media technologies, such as, video-on-demand, whereas, perhaps, older generations are more in favour of continuing to use traditional media consumption channels, such as, live TV. Therefore, young adults should be strong predictors of changes in TV consumption when compared to elderly.
So far, I've run the following command to produce desired results:
library(plm)
tv_usage_trend <- plm(tv_volume ~ age_group + gender + control_variable + control_variable + control_variable, data = total, index =c("id", "date"), model = "between")
summary(tv_usage_trend)
It can be seen that I used "model = between" function in the PLM package due to my time-invariant predictors, which means that I cannot measure "within" effects and can only look at "between" effects among individuals. The snapshot of my results is as following:
Estimate Std. Error t-value Pr(>|t|)
(Intercept) 384.455 2498.529 0.1539 0.877711
AGE_BANDING65 - 74 479.638 157.961 3.0364 0.002396 **
AGE_BANDING55 - 64 -577.320 191.700 -3.0116 0.002601 **
AGE_BANDING45 - 54 -1487.911 197.152 -7.5470 4.574e-14 ***
AGE_BANDING35 - 44 -2332.972 211.404 -11.0356 < 2.2e-16 ***
AGE_BANDING25 - 34 -2466.860 229.941 -10.7282 < 2.2e-16 ***
AGE_BANDING18 - 24 -2193.473 307.055 -7.1436 9.291e-13 ***
My reference category in here was (Age_BANDING (75+). The results look as expected - young adults significantly differ from the oldest group (at least one * next to the category means significant relationship). However, when investigating the results closer and plotting best fit lines for each category over time I noticed that actually AGE_BANDING (75+) group also changed their behaviour - they are much less likely to watch TV as well.
In other words, this means that there should be no significant difference between the reference category and some of the younger adults categories, but that is not the case. I believe the model does not account for over-time effect on the dependent variable and only compares overall tv_volume means across age groups, which of course then tells the story about which groups are more or less likely to watch TV in general rather than which groups are drivers of TV consumption change.
Any ideas on how I could solve this? I would really really appreciate any help.
Thanks,
Klaudijus
