A state raises its minimum wage, and researchers want to estimate the policy's effect on employment. Simply comparing employment before and after the change within that state confounds the policy's effect with any other economic trend happening over the same period. Difference-in-differences addresses this by also tracking employment over the same period in a comparable neighboring state that didn't raise its minimum wage, then comparing the change in the treated state against the change in the untreated comparison state — isolating an estimate of the policy's specific effect from broader economic trends that plausibly affected both states similarly.
The core logic behind the method
If both the treated and untreated groups would have followed the same underlying trend in the outcome absent the treatment, then any difference between how much each group's outcome actually changed can be attributed to the treatment itself, since the shared trend cancels out in the comparison — the treated group's actual change minus the untreated group's actual change over the same period isolates the treatment's specific contribution, assuming the shared-trend logic holds.
Why the parallel trends assumption is the load-bearing requirement
This entire logic depends on one central, unprovable assumption: that in the absence of the treatment, the treated and untreated groups would have followed parallel trends in the outcome being studied. If the two groups were already on genuinely different trajectories for reasons unrelated to the treatment — say, the treated state's economy was already growing faster than the comparison state's before the minimum wage change — the difference-in-differences estimate will incorrectly attribute part of that pre-existing divergence to the treatment itself.
Why this assumption can't be directly proven, but can be partially checked
Because parallel trends is fundamentally a claim about what would have happened without the treatment — something that, by definition, can never be directly observed since the treatment did occur — researchers instead check whether the two groups followed similar trends during a period before the treatment was introduced, using this pre-treatment parallel trend as supporting, though not conclusive, evidence that the same parallel trajectory would plausibly have continued absent the treatment.
Why this method has become so widely used in applied policy research
Difference-in-differences provides a genuinely useful way to estimate causal effects from real-world policy changes and natural experiments where randomized assignment isn't feasible, using naturally occurring comparison groups instead, which makes it one of the most commonly applied quasi-experimental methods across economics and policy evaluation specifically because these kinds of natural comparisons are far more often available than genuine randomized trials.
What this means for evaluating research that uses this method
- Check whether a study presents evidence of pre-treatment parallel trends between the treated and comparison groups, since this is the standard, available check on the method's core assumption
- Be specifically skeptical of a difference-in-differences estimate where the comparison group seems structurally different from the treated group in ways plausibly related to the outcome
- Look for robustness checks using alternative comparison groups, since a finding that holds up across multiple reasonable comparison groups is more credible than one relying on a single specific comparison
- Recognize that even strong pre-treatment parallel trends don't guarantee the assumption held afterward — it's supporting evidence, not direct proof
Difference-in-differences is a genuinely elegant solution to a genuinely hard problem — estimating causal effects without a randomized trial — and its entire credibility rests on a single assumption that can only ever be partially, indirectly checked, which is exactly why the choice of comparison group deserves as much scrutiny as the statistical result itself.