Covariate-Conditional Parallel Trends and Standardization

Why Condition on Pre-Treatment Covariates?

In a two-period difference-in-differences design, the control group’s change in outcome is used to represent how the treated group would have changed without treatment. That comparison is most credible when it is made among units with similar pre-treatment characteristics. For example, treated and control units may have different mixes of age groups, and age groups may have different outcome trends even without treatment. Conditional parallel trends makes the comparison within covariate values; standardization then combines those comparisons using the treated group’s covariate distribution.

The target is the average treatment effect on the treated (ATT): the average effect of treatment for the units that received it. The essential weighting choice follows from that target. Counterfactual changes must be averaged as if the comparison group had the treated units’ covariate distribution, not the control group’s distribution or an arbitrary distribution.

Definition: Let $D=1$ denote units assigned to or receiving the intervention group and $D=0$ denote controls. Let $t=0$ be the pre-intervention period and $t=1$ the post-intervention period. Write $Y_t(d)$ for the outcome at time $t$ under treatment status $d$, and let $X$ denote covariates measured before treatment. The ATT is $$ \operatorname{ATT}=E[Y_1(1)-Y_1(0)\mid D=1]. $$ It is the average difference between treated units’ post-period outcomes under treatment and the post-period outcomes those same units would have had without treatment.

Conditional Parallel Trends and Overlap

The conditional parallel trends assumption says that, at each relevant value of the pre-treatment covariates, the treated and control groups would have had the same mean change in their outcomes in the absence of treatment. It is an assumption about mean changes, not about equal outcome levels at baseline.

Assumption: Conditional parallel trends. For covariate values $x$ represented among treated units, $$ E[Y_1(0)-Y_0(0)\mid D=1,X=x] = E[Y_1(0)-Y_0(0)\mid D=0,X=x]. $$ The covariates $X$ must be measured before treatment. This assumption concerns the untreated potential outcome $Y_t(0)$, including its unobserved post-period value for treated units.

For controls, the observed outcome is their untreated potential outcome. Thus, their observed change within covariate value $x$ provides the conditional mean untreated change on the right-hand side. For treated units, the observed post-period outcome is under treatment, so their untreated post-period outcome remains counterfactual.

Identification also requires overlap: for covariate values occurring among treated units, there must be controls available for comparison. In probability terms, $P(D=0\mid X=x)>0$ for values $x$ in the treated group’s covariate support. Overlap need not hold at covariate values found only among controls, because the target is the ATT. Without overlap for some treated units, their counterfactual change cannot be learned from controls at comparable $X$ values without additional assumptions.

As in a standard potential-outcomes setup, treatment must be well-defined and there must be no interference between units. The pre-period must also be untreated: $Y_0=Y_0(0)$ for both groups, with no anticipation of the intervention affecting pre-period outcomes. Together with consistency, these conditions let observed outcomes stand in for the corresponding potential outcomes. Conditional parallel trends and overlap are additional identifying conditions; they are not guaranteed by the data structure alone.

Deriving the Standardized ATT

Define the observed outcome change as $\Delta Y=Y_1-Y_0$. For controls, consistency and the absence of pre-treatment effects imply $\Delta Y=Y_1(0)-Y_0(0)$. Conditional parallel trends therefore gives, for treated-supported values of $x$, $$ E[\Delta Y\mid D=0,X=x] = E[Y_1(0)-Y_0(0)\mid D=1,X=x]. $$ So the control group’s conditional mean change estimates the mean untreated change the treated units would have experienced at that covariate value.

Within the treated group, the observed change combines its untreated change with the treatment effect. Subtracting the standardized untreated change leaves the ATT:

\begin{aligned} \operatorname{ATT} &=E[Y_1(1)-Y_1(0)\mid D=1]\\ &=E[Y_1(1)-Y_0(0)\mid D=1] - E[Y_1(0)-Y_0(0)\mid D=1]\\ &=E[\Delta Y\mid D=1] - E\!\left[E[\Delta Y\mid D=0,X]\mid D=1\right]\\ &=E\!\left[ E[\Delta Y\mid D=1,X]-E[\Delta Y\mid D=0,X] \mid D=1\right]. \end{aligned}

The second line uses $Y_0(1)=Y_0(0)$, because the pre-period precedes treatment and has no anticipation. The third line replaces the treated group’s mean untreated change with the control-group mean change at the same $X$, using conditional parallel trends. The last line follows from iterated expectations: the treated group’s observed mean change can be written as the average of its covariate-specific mean changes over the treated distribution of $X$.

This is a standardization formula. First calculate a treated-versus-control difference in mean changes at each covariate value; then average those differences over the covariate distribution among treated units. If $X$ has a continuous distribution, the outer average is an integral over the treated distribution. For discrete covariate strata, it is a weighted sum.

Worked Examples

Worked Example: Standardizing Across Two Covariate Strata

Suppose a hypothetical program evaluation divides units into two pre-treatment age strata, $X=A$ and $X=B$. The entries below are hypothetical group sizes and mean outcomes, in outcome units. All four group-by-stratum cells are represented, so there is overlap.

Stratum	Treated $n$	Treated mean $Y_0$	Treated mean $Y_1$	Control $n$	Control mean $Y_0$	Control mean $Y_1$
A	60	10	14	100	8	9
B	40	12	19	100	11	13

In stratum A, the treated mean change is $14-10=4$, and the control mean change is $9-8=1$. The conditional DiD contrast is $4-1=3$ outcome units. In stratum B, the changes are $19-12=7$ for treated units and $13-11=2$ for controls, giving a conditional contrast of $7-2=5$ outcome units.

There are 100 treated units, with 60% in A and 40% in B. Standardizing to that treated distribution gives $$ \widehat{\operatorname{ATT}} =0.60(3)+0.40(5) =1.8+2.0 =3.8 $$ outcome units. Equivalently, the treated mean observed change is $0.60(4)+0.40(7)=5.2$; the standardized control change is $0.60(1)+0.40(2)=1.4$; and $5.2-1.4=3.8$. This interpretation depends on conditional parallel trends being plausible within both strata, not merely on the arithmetic.

Worked Example: Why the Pooled DiD Can Mislead

Consider hypothetical binary covariate strata $X=L$ and $X=H$. Ninety percent of treated units are in L and 10% are in H, while the control group has the reverse composition. In each stratum the untreated outcome change is 0 for L and 10 for H. Suppose treatment adds 2 outcome units in each stratum.

Stratum	Treated share	Control share	Treated mean change	Control mean change
L	0.90	0.10	2	0
H	0.10	0.90	12	10

The conditional trends assumption holds in both strata: the controls’ changes, 0 and 10, equal the treated units’ untreated changes at the same $X$. The treatment effect is 2 in each stratum. The treated group’s pooled observed change is $$ 0.90(2)+0.10(12)=1.8+1.2=3.0. $$ The controls’ pooled change is $$ 0.10(0)+0.90(10)=9.0. $$ The unadjusted pooled DiD is therefore $3.0-9.0=-6.0$ outcome units, despite a treatment effect of 2 in each stratum. Its comparison uses the control group’s covariate mix rather than the treated group’s mix.

Standardizing the control changes to the treated distribution gives $0.90(0)+0.10(10)=1.0$. The standardized ATT is $3.0-1.0=2.0$ outcome units, matching the stipulated treatment effect. This example shows why conditional parallel trends does not, by itself, ensure that the pooled unadjusted DiD identifies the ATT when covariate composition differs.

Worked Example: The Target Determines the Weights

In a second hypothetical evaluation, low- and high-risk strata have different conditional effects. Among treated units, 25% are low-risk and 75% are high-risk. The observed mean changes are 3 and 10 outcome units for treated units in the two strata, and 2 and 4 for controls. Assume overlap and conditional parallel trends in both strata.

The conditional DiD contrasts are $3-2=1$ outcome unit for low-risk units and $10-4=6$ outcome units for high-risk units. Since the target is the ATT, use the treated group’s shares: $$ \widehat{\operatorname{ATT}} =0.25(1)+0.75(6) =0.25+4.50 =4.75 $$ outcome units.

An equal-weighted average would be $(1+6)/2=3.5$ outcome units, but that describes a different weighting target. It is not the ATT for this treated population, which is mostly high-risk. The conditional effects need not be identical for standardization to work; their treated-distribution-weighted average is the ATT.

What the Assumption Does—and Does Not—Say

Conditional parallel trends permits treated and control units to have different baseline outcome levels, and it permits outcome trends to differ across covariate strata. It requires that, within a given $X$ value, the mean untreated change for treated units would match the mean change observed among controls. It does not require equal treatment effects across strata.

The condition is about counterfactual mean changes, so it cannot be verified directly for treated units: their post-period untreated outcomes are unobserved. The observed data can show covariate overlap and describe group-specific changes, but those features do not prove the counterfactual assumption. A standardized estimate is causal only to the extent that the stated design assumptions are credible.

Define the target.
Specify the treated group and the ATT, along with the pre-treatment covariates to use for conditioning.

Check the comparison is available.
Confirm that controls are represented at the covariate values found among treated units. Without this overlap, some treated units lack comparable controls.

Compare changes conditionally.
For each covariate value or stratum, subtract the control mean change from the treated mean change.

Standardize to treated units.
Average the conditional contrasts using the treated group’s covariate distribution to obtain the ATT.

Common Mistakes

Using the control group’s covariate shares. That averages conditional contrasts for a control-like population, not the treated group. For the ATT, use treated shares.
Assuming conditional trends means equal baseline levels. The assumption concerns changes in untreated potential outcomes within $X$; baseline means may differ.
Conditioning on a post-treatment variable. A variable affected by treatment is not a pre-treatment covariate for this standardization and may distort the comparison. Use covariates measured before treatment.
Treating the assumption as a statement about every observed unit. Conditional parallel trends equates conditional mean changes in the relevant populations; it does not require each treated unit to have the same change as a control.
Calling the standardized contrast automatically causal. The calculation identifies the ATT only under the assumptions, including conditional parallel trends, overlap, consistency, and no anticipation.

Key takeaway: Conditional parallel trends identifies the treated group’s untreated change from control-group changes at the same pre-treatment covariate values. Standardizing those changes over the treated group’s covariate distribution yields the counterfactual change needed for the ATT.

Check Your Understanding

Answer each question using the target population and assumptions stated in the tutorial.

State conditional parallel trends in words and explain which potential outcome it concerns.
In a two-stratum setting, conditional DiD contrasts are 2 and 5. If 80% of treated units are in the first stratum, calculate the standardized ATT.
Why can pooled DiD differ from the standardized ATT even when conditional parallel trends holds in every stratum?
What overlap condition is needed to identify the ATT, and why is overlap required only over the treated covariate support?
Explain why equal weighting of stratum-specific contrasts need not estimate the ATT.