Say we observe outcomes for all U.S. states over many years. Some law/policy is introduced and affects some states but not others. The policy may start anytime throughout the year for some states; for example, it could be introduced in January or September of the year. The basic model looks something like this:
$$ \mathrm{ln}(y_{st}) = \gamma_s + \lambda_t + \delta P_{st} + u_{st}, $$
where a logged continuous outcome $y$ is regressed on state fixed effects, year fixed effects, and policy variable $P_{st}$. To be clear, $P_{st}$ is equal to $0$ in any state-year where the policy was not in effect, a value of 1 where the policy was in effect for the whole year, and a value equal to the portion of the year the policy was in effect otherwise. This turns the binary variable into a proportion. For example, in a year before a state's first "full year" under the new law $P_{st}$ could equal .25 or .75 depending upon the month the law was enacted. On the other hand, if the policy is removed later, I suppose the variable could equal a value between $0$ and $1$ in the year it concluded. This approach would involve the researcher inputting fractional values at the extremes (i.e., a state must have a full year of policy exposure to receive a value of 1).
I can't quite wrap my head around how to interpret such a variable. In essence, we turned the binary variable into some sort of "continuous" treatment, where you have a fraction of exposure before your first full year of treatment, but full exposure in the years the policy is in effect. Is my understanding correct?
Once we enter this into a linear model as a continuous predictor, how should we interpret this variable? I suppose I'm looking for insight into how I can compare this interpretation to the binary case.
See this NBER working paper (footnote 33) where they detail the coding of this variable.