Parallel Trends and Pre-Trend Tests

A city imposes a new financial transaction tax and an analyst compares trading volume before and after, in that city versus a neighboring city that didn't change its tax. The whole logic of difference-in-differences rests on one assumption: had the tax never been imposed, trading volume in the taxed city would have continued moving in step with the untaxed city — same ups, same downs, just from a different starting level. If that assumption is wrong — if the two cities were already drifting apart for unrelated reasons — the "effect" the method reports is partly or entirely an artifact of that pre-existing drift, not the tax.

An analogy: two joggers on a track

Picture two joggers who have been running side by side, maintaining a roughly constant gap between them, for the last mile. One jogger then drinks a sports gel; shortly after, the gap between them widens. It's tempting to credit the gel. But this inference only holds up if the two joggers' pace gap was genuinely stable before the gel — if one runner had already been slowly pulling ahead for the whole previous mile, a widening gap after the gel proves nothing new; it's just the same pre-existing trend continuing. Checking that the gap was flat and stable before the gel is the pre-trend test; a stable pre-period gap is the evidence (never proof) that the joggers really were on parallel paths.

The assumption, one symbol at a time

Let $Y_{it}$ be the outcome for unit $i$ (city) at time $t$, and let $Y_{it}(0)$ denote the potential outcome had treatment never happened. Parallel trends states:

$$\mathbb{E}[Y_{it}(0) - Y_{i,t-1}(0) \mid \text{treated group}] = \mathbb{E}[Y_{it}(0) - Y_{i,t-1}(0) \mid \text{control group}]$$

In plain English: absent treatment, the treated group's outcome would have changed from one period to the next by the same average amount as the control group's actually did — not the same level, just the same change. This is unobservable directly (we never see the treated group's untreated counterfactual after treatment starts), so it can't be tested in the post-treatment period. What can be checked is whether the two groups' trends were parallel in the pre-treatment periods, using a pre-trend test: regress the outcome on group, time, and group-by-time interactions restricted to pre-treatment periods, and test whether those interaction coefficients are statistically indistinguishable from zero.

Worked example 1: a clean pre-trend check

Quarterly trading volume (in $m) for the taxed city and the control city, three quarters before the tax:

Q1: taxed 100, control 80 (gap 20). Q2: taxed 103, control 83 (gap 20). Q3: taxed 106, control 86 (gap 20). Q4 (tax imposed): taxed 95, control 89 (gap 6).

The pre-period gap is rock steady at 20 across Q1–Q3, giving reasonable confidence in parallel trends. Post-tax, the gap collapses from 20 to 6, a drop of 14. The diff-in-diff estimate of the tax's effect is roughly a $14m reduction in the taxed city's volume relative to what it would have been, and the flat pre-trend makes this a credible causal read.

Worked example 2: a pre-trend that should raise a flag

Same setup, different pre-period: Q1 gap 20, Q2 gap 24, Q3 gap 28 — the taxed city was already outpacing the control by a growing margin before the tax, at roughly +4 per quarter. Q4 (tax imposed): taxed 95, control 89, gap 6 — a sudden reversal from a growing +28 gap to +6. Naively running diff-in-differences on just the last pre-period and post-period (28 → 6, an apparent −22 effect) would badly overstate the tax's impact, because part of that pre-existing +4/quarter divergence should be projected forward and subtracted out. If the trend had simply continued, the "no-tax" Q4 gap would have been about 32, making the true effect closer to 32 − 6 = −26, or the pre-trend itself might signal the two cities are diverging for unrelated reasons entirely, in which case no clean diff-in-differences estimate is trustworthy at all without further adjustment (e.g., extrapolating the trend, or switching to a synthetic control approach).

What this means in practice

A pre-trend test is a necessary but never sufficient check: passing it (flat, statistically insignificant pre-period interaction terms) is consistent with parallel trends but doesn't prove the assumption holds after treatment, since anything could have changed the moment treatment started. Failing it, however, is strong evidence against the design as-is, and analysts should either find a better control group, extend the model to allow for group-specific trends, or move to an event study design that plots the full dynamic path of estimated effects period by period, making any pre-trend visually obvious rather than buried in a single before/after comparison.