Especificando uma Diferença no Modelo de Diferenças com Vários Períodos

20

Quando eu estimo uma diferença no modelo de diferenças com dois períodos, o modelo de regressão equivalente seria

uma. Yist=α+γsTreatment+λdt+δ(Treatmentdt)+ϵist

  • onde é um manequim que é igual a 1 se a observação é a partir do grupo de tratamentoTreatment
  • e é um manequim que é igual a 1, no período de tempo após o tratamento ocorreud

Assim, a equação assume os seguintes valores.

  • Grupo controle, antes do tratamento: α
  • Grupo controle, após tratamento: α+λ
  • Grupo de tratamento, antes do tratamento: α+γ
  • Grupo de tratamento, após o tratamento: α+γ+λ+δ

Portanto, em um modelo de dois períodos, a diferença na estimativa de diferenças é .δ

Mas o que acontece com relação a se eu tiver mais de um período pré e pós-tratamento? Ainda uso um boneco que indica se um ano é anterior ou posterior ao tratamento?dt

Ou adiciono manequins de ano sem especificar se cada ano pertence ao período pré ou pós-tratamento? Como isso:

b. Yist=α+γsTreatment+yeardummy+δ(Treatmentdt)+ϵist

Ou posso incluir ambos (por exemplo, yeardummy+λdt )?

c. Yist=α+γsTreatment+yeardummy+λdt+δ(Treatmentdt)+ϵist

Em conclusão, como especifico uma diferença no modelo de diferenças com vários períodos de tempo (a, b ou c)?

Tom
fonte
1
Você geralmente usa o modelo b. Observe que no modelo c, será perfeitamente colinear com as dummies de ano, de modo que o modelo não pode ser estimado. dt
standard_error
Seria ótimo se você pudesse explicar por que b é usado geralmente. Talvez dê algumas referências ou apenas explique duas frases.
precisa saber é o seguinte
e no modelo b. você poderia adicionar uma variável contínua por ano em vez de manequins? Como a interpretação dos coeficientes diferiria nesses casos?

Respostas:

19

Yist=α+γs(Treatments)+λ(year dummyt)+δDst+ϵist
where DtTreatmentsdt is a dummy variable which equals one for treatment units s in the post-treatment period (dt=1) and is zero otherwise. Note that this is a more general formulation of the difference in differences regression which allows for different timings of the treatment for different treated units.

As was pointed out correctly in the comments your proposed solution c) does not work out due to collinearity with the time dummies and the dummy for the post-treatment period. However, a slight variant of this turns out to be a robustness check. Let γs0 and γs1 be two sets of dummy variables for each control unit s0 and each treated unit s1, respectively, then interacting the dummies for the treated units with the time variable t and regressing

Yist=γs0+γs1t+λ(year dummyt)+δDst+ϵist
includes a unit specific time trend γs1t. When you include these unit specific time trends and the difference in differences coefficient δ does not change significantly you can be more confident about your results. Otherwise you might wonder whether your treatment effect has absorbed differences between treated units due to an underlying time trend (can happen when policies kick in at different points in time).

An example cited in Angrist and Pischke (2009) Mostly Harmless Econometrics is a labor market policy study by Besley and Burgess (2004). In their paper it happens that the inclusion of state-specific time trends kills the estimated treatment effect. Note though that for this robustness check you need more than 3 time periods.

Andy
fonte
A follow up since I am trying to decide if implementing this with some administrative data is appropriate: Would you say a DD approach is more valid than a CITS design if there are only 4 time points (2 pre and 2 post) in a model? Also, if I have multiple cohorts within waves of data should these be examine separately or in a unified model? Thanks.
bfoste01
@Andy: Can you pls explain, what you mean by s0, s1, and unit-specific time trend? Assuming I have two newspapers (WPT and NYT) and WPT is my treatement group, which of them would be s0 and s1?
user3683131
1
Am I right in thinking that this analysis compares the average pre and post treatment and does not account for secular trends? i.e. if d_t = 0 for all time periods before the switch point, and d_t = 1 for all time periods after, then this analysis is essentially the same as the two time periods one, except the average is taken of all the before/after time periods. Any time trends in the outcome before/after the treatment switch are ignored? I am trying to decide if a DiD model is correct for an analysis I am planning to carry out.
AP30
0

I would like to clarify something (and indirectly address a question in the comments). In particular, it concerns the use of unit-specific linear time trends. As a robustness check, it would appear you are only interacting dummies for treated units (i.e., γ1s) with a continuous time trend. However, it is actually the case that you are interacting a full set of unit/state dummies (unit/state fixed effects) with a linear time trend variable.

Angrist and Pischke (2009) recommend this approach on page 238 in Mostly Harmless Econometrics. Differences in notation can cause confusion. Reproducing specification 5.2.7:

yist=γ0s+γ1st+λt+δDst+Xistβ+εist,

where γ0s is a state-specific intercept, in accordance with the s subscript used in their book. You can view γ1s as the state-specific trend coefficient multiplying the time trend variable, t. Different papers use different notation. For example, Wolfers (2006) replicates a model incorporating state-specific linear time trends. Reproducing model (1):

ys,t=sStates+tYeart+sStatesTimet+δDs,t+εs,t,

where the model includes state and year fixed effects (i.e., dummies for each state and year). The treatment variable Ds,t is when state s adopts a unilateral divorce regime in period t. Notice this specification interacts state dummies with a linear time trend (i.e., Timet). This is yet another representation of state-specific linear time trends in your model specification.

Unit-specific linear time trends is also addressed in another post (see below):

How to account for endogenous program placement?

In sum, you want to interact all unit (group) dummies with a continuous time trend variable.

Paper by Justin Wolfers is below for your reference:

https://users.nber.org/~jwolfers/papers/Divorce(AER).pdf

Tom
fonte