mixed effect model?

Question

I have 365 days of bike sharing demand data for 15 stations. I am thinking of taking each day data as a data point (n=365*15=5475) and relate daily weather variable as well as land use variable. The land use variable for each station will be the same across the data point. My objective is to develop a Poisson or NB model to see how weather and land use variable affect demand for each station.

Does it violate the independence assumption as I have multiple data point for a station? Do I need mixed effect Poisson model?

Answer 1

Controlling for weather and station-features (what you call land use variables, since the stations never move) will reduce the conditional dependence between stations. Chances are, however, it will not make them conditionally independent. You can use some exploratory analyses to determine whether autoregressive or intraclass effects persist. (lorellograms, variograms, etc.) by performing the independent data analysis and inspecting the residuals.

With residual dependence, it is safest and best to use some methods for dependent data on top of the fixed effects, like random effects by day and by site. It is interesting and curious to note that when the fixed effects do not sufficiently control for the concept they intended to capture, it can attenuate the fixed effect. It is something to note when comparing the two analyses.