Placebo test for the validity of difference-in-differences analysis results for limited data period I am doing an analysis of a strategy implemented by a four-year-old corporate company and this implementation was given in different cities which all of its employees were received the treatment or not in 2016-2019. The strategy change date is 2017.
I apply the quasi-experimental method (difference-in-differences) and I am concerned about the validity of my results. Therefore, for the validity check of my difference-in-differences analysis, I am planning to do placebo tests with fake treatment dates before and after the actual date of treatment. However, due to the company was established in 2016, there is no data before 2016. Although I can use "2018" as a fake treatment year, I am wondering how can I do a placebo test before the actual date with a single period (just 2016) data and do I need to use any fake treatment year before the actual date? Is it enough to use only 2018 as a fake treatment date for placebo test to check the validity of the actual treatment date (2017) results?
Any alternative suggestions for the difference in differences placebo test or robustness of the difference-in-differences test results would be much appreciated.
 A: It is often difficult to buttress claims of trend equivalence before the onset of treatment with only one pretreatment period. Serial observations are often required to demonstrate parallel trends, which (in my opinion) requires at least three pretreatment periods to support it.

Although I can use "2018" as a fake treatment year, I am wondering how can I do a placebo test before the actual date with a single period (just 2016) data and do I need to use any fake treatment year before the actual date?

Applied researchers often use "event study" estimates to support claims of common group trends. They may, for instance, interact a treatment indicator with a series of time dummies for all years before and after the onset of treatment, leaving out the period immediately before treatment, or some more distant pre-event period, as a reference (baseline). You cannot possibly do this with one pre-event time period. Researchers often do this to test the validity of the difference-in-differences design. Put more simply, you shouldn't be capturing significant treatment effects in the pretreatment epoch. Treatment effects should concentrate around the exposure period. Any strong nonzero effects in the periods leading up to the treatment could be interpreted as selection bias.
You indicated in your question that you wish to treat a post-treatment year (e.g., 2018) as a placebo treatment year. Is the treatment only in effect for one year before it is officially removed? I don't know the intimate details of your study, but I would advise against using 2018 as a fake treatment year. Once treatment is removed, are there any lingering (persistent) effects beyond conclusion of the intervention/treatment? If you're going to use 2018 a placebo treatment year, then your treatment should have no impact on your outcome(s) in that year.

Is it enough to use only 2018 as a fake treatment date for placebo test to check the validity of the actual treatment date (2017) results?

Perhaps. If treatment reverts back to some pretreatment level and there are absolutely no lingering effects that you need to account for, then this could support the validity of your approach. In other words, it shows treatment effects concentrating around your exposure period.
You could also investigate fake outcomes. Suppose you are interested in the effects of your intervention across multiple outcome measures. Try to find outcome variables that would, in theory, not be impacted by your treatment. It is comforting to find insignificant treatment effects with outcomes that do not (should not) vary with treatment exposure. This could be used as a robustness check.
In sum, unless you have some strong theoretical basis for assuming trends do not deviate drastically in the periods immediately before the onset of treatment, it will be difficult to support claims of common trends across treatment and control groups.
The best course of action would be to acquire more pre- or post-event data. You could still do a difference-in-differences design, just be prepared to answer these questions.
