Return to Answer

The treatment indicator is allowed to switch ‘on’ and ‘off’ throughout the panel. This is often the case in policy analysis, where some units can have multiple treatment histories. For example, a new law was enacted in a subset of U.S. states at the beginning of 2013, only to be repealed at the conclusion of 2016. Later, legislators in a subset of U.S. states where the law was nullified decide to reintroduce the legislation again in 2018 where it remains in effect. In practice, your treatment dummy should be coded to reflect this reality. However, this could become problematic if policymakers decide to introduce or remove laws/policies based upon past outcomes of the response variable. Review pages 4 through 7 of Lecture 10 for a more in-depth discussion of this.

The treatment variable is allowed to switch ‘on’ and ‘off’ throughout the panel. Put differently, the binary indicator may switch back and forth between 0 and 1 multiple times over time within a unit(s). This is often the case in policy analysis, where some units can have multiple treatment histories. For example, a new law was enacted in a subset of U.S. states at the beginning of 2013, only to be repealed at the conclusion of 2016. Later, legislators in a subset of U.S. states where the law was nullified decide to reintroduce the legislation again in 2018 where it remains in effect. In practice, your treatment dummy should be coded to reflect this reality. However, this could become problematic if policymakers decide to introduce or remove laws/policies based upon past outcomes of the response variable. Review pages 4 through 7 of Lecture 10 for a more in-depth discussion of this.

To be transparent, the mechanics of DD can be reversed. Only a handful of academic studies have executed this well in my estimation. In this context, always treated units would constitute your control group, while the treated group is the subset of units eligible for the treatment in a later time period. It's a comparison of the units exposed to a treatment in all time periods (i.e., the "always 0" group) with the switchers (i.e., changers from 0 to 1). In this sense, DD "in reverse" is simply referring to the fact that the "controls" represent the always treated. Note, under an identification condition involving only treated responses, the identification condition may be tested by future parallel treated paths across the two groups, rather than past parallel untreated paths. Review the work of Kim and Lee 2019 for a neat application of this.

To be transparent, the mechanics of DD can be reversed. Only a handful of academic studies have executed this well in my estimation. In this context, always treated units would constitute your control group, while the treated group is the subset of units that change their treatment status later. Maybe some exposure is removed by the second time period. The "change" is coded as any other treatment in a DD setting. In other words, it's a comparison of the units exposed to a treatment in all time periods (i.e., the "always 0" group) with the switchers (i.e., changers from 0 to 1). In this sense, DD "in reverse" is simply referring to the fact that the "controls" represent the always treated. Note, under an identification condition involving only treated responses, the identification condition may be tested by future parallel treated paths across the two groups, rather than past parallel untreated paths. Review the work of Kim and Lee 2019 for a neat application of this.

To be transparent, the mechanics of DD can be reversed. Only a handful of academic studies have executed this well in my estimation. In this context, always treated units would constitute your control group, while the treated group is the subset of units eligible for the treatment in a later time period. It's a comparison of the units exposed to a treatment in all time periods (i.e., the "always 0" group) with the switchers (i.e., changers from 0 to 1). In this sense, DD "in reverse" is simply referring to the fact that the "controls" represent the always treated. Note, under an identification condition involving only treated responses, the identification condition may be tested by future parallel treated paths across the two groups, rather than past parallel untreated paths. Review the work of Kim and Lee (2019) for a neat application of this.

To be transparent, the mechanics of DD can be reversed. Only a handful of academic studies have executed this well in my estimation. In this context, always treated units would constitute your control group, while the treated group is the subset of units eligible for the treatment in a later time period. It's a comparison of the units exposed to a treatment in all time periods (i.e., the "always 0" group) with the switchers (i.e., changers from 0 to 1). In this sense, DD "in reverse" is simply referring to the fact that the "controls" represent the always treated. Note, under an identification condition involving only treated responses, the identification condition may be tested by future parallel treated paths across the two groups, rather than past parallel untreated paths. Review the work of Kim and Lee 2019 for a neat application of this.