Event Studies on Chinese Case Law: Courts Switch Rules in a Month, Litigants Take a Filing Cycle

A common use of a Chinese judgment corpus is the before-and-after study. A new judicial interpretation takes effect on a known date; you count something in judgments before and after; you report the change. Policy teams do it, empirical legal scholars do it, and legal AI teams do it when they want to know whether a model trained on older text still describes current practice.

The design has a flaw that is easy to miss. A judgment is dated when it is decided, but the parties framed their claims when the case was filed, often months earlier. Indexing events by decision date mixes two populations: cases filed under the old rule and decided under the new one, and cases filed under the new rule. The court's citations switch at the decision date. The litigants' behaviour can only switch at the filing date. The date-indexed series shows a slow slope where there are really two steps.

We measured this on one well-known rule change in our own index. Three findings:

All figures describe our own index of 160,291,678 PRC judgment records on the stated date. They are not national litigation statistics, and a share of judgments is not a share of cases filed: many loan disputes end in mediation, withdrawal or default before any judgment is written. Where a number depends on a choice we made, the choice is stated.

The rule change

For fifteen years Chinese courts handled debts incurred by one spouse during the marriage under Article 24 of the Supreme People's Court's second interpretation of the Marriage Law (2003). In practice it worked as a presumption: a debt incurred during the marriage was treated as a joint debt unless the other spouse could prove one of narrow exceptions. Creditors suing a borrower routinely named the borrower's spouse as a co-defendant.

The Supreme People's Court's interpretation on marital-debt disputes, in force from 18 January 2018, moved the burden. A debt incurred by one spouse in their own name beyond ordinary household needs is no longer joint unless the creditor shows it was used for the couple's shared life or shared business, or that both spouses agreed to it, for example by both signing or one ratifying later. The rule was carried into Article 1064 of the Civil Code from 2021. This summary is informational, not legal advice.

That gives a clean test. If creditors respond to the new burden, fewer of them should plead spousal liability in new loan suits. If courts respond, judgments should start citing the new interpretation. Both are observable in judgment text.

How we measured

• Population. Private-lending cases: cause of action 民间借贷纠纷 or 民间借贷, or 民间借贷 in the title. Judgments with a merits decision: the body contains 判决如下 (“the judgment is as follows”). First instance: the case number contains 民初.
• Behaviour. A judgment raises spousal liability if its body contains 夫妻共同债务, 夫妻共同之债 or 夫妻共同债 (joint marital debt). The phrase appears whenever a plaintiff pleads it, a spouse denies it, or the court decides it.
• Citation. The full title of the 2018 interpretation, 关于审理涉及夫妻债务纠纷案件适用法律有关问题的解释, as a phrase.
• Filing cohort. The year in parentheses at the start of the case number, e.g. （2017）桂0107民初1755号 was filed in 2017. Both full-width and half-width parentheses are matched.
• Exclusions. Three ingest batches whose date field is a publication date rather than a decision date (see our date-field audit), one batch whose bodies are almost entirely empty, and a 7,000-record test set.
• Method. Every count is an exact query over the full index, bucketed by month of decision date. Phrase queries use the same tokenizer as the index; we checked token positions with _analyze first, because a mismatched analyzer silently returns zero (see our note on phrase-query false zeros).
• Reading check. 60 judgments decided March–June 2018, 30 from each filing cohort (Elasticsearch random_score, seed 20261002). In all 60 the phrase appears in a plaintiff's claim, a defence or the court's reasoning about whether the borrower's spouse shares liability for the loan.

Step one: courts switched within a month

Restricting to private-lending judgments that raise spousal liability, the citation of the new interpretation behaves the way an event study expects:

January is partial because the rule took effect on the 18th. From February the share is flat. It also holds inside each filing cohort: among first-instance judgments decided from February to December 2018, the citation share was 60–68% for cases filed in 2017 and 57–61% for cases filed in 2018. Courts applied the new rule to cases that were already pending. For the court's side of the record, decision date is the right clock.

The ceiling of about 60% is not a compliance rate. A judgment can discuss spousal liability without citing anything (for example when the spouse admits the debt), and some cite the rule under a shortened or slightly wrong title. For judgments decided February–December 2018 (97,726 in this population), the full title matches 57,055; the fragment 涉及夫妻债务纠纷案件 matches 58,759; and a miswritten title that inserts 共同 (“joint”), which one of our 60 sampled judgments uses, matches 1,176 that the full-title query misses. A citation series should take the union of the variants, or at least report how much the shortest reliable fragment adds.

Step two: the behavioural series looks gradual

Now the litigants' side. Across all private-lending merits judgments, the share that raise spousal liability:

The obvious reading is a gradual response: creditors took most of a year to absorb the new burden, then kept drifting down as the Civil Code arrived. Some of that may be true. But the first nine months of the slide have a mechanical explanation that has nothing to do with anyone learning anything.

Step three: split by filing year

Every Chinese case number begins with the year the case was filed. Take first-instance private-lending judgments decided in 2018 and split them by that year:

Two things are visible. First, cases filed in 2017 never adjusted. Decided in March or in December, under the new rule and usually citing it, they raised spousal liability in roughly one judgment in five, the same as or slightly above the 2017 level. That is what you would expect: the complaint was drafted, and the spouse named as a defendant, before the rule existed.

Second, cases filed in 2018 started low. Even the fastest ones, decided in February and March, raised spousal liability about 11.5% of the time, against 16.5% for the older cohort decided in the same weeks. The behavioural change is concentrated at filing.

The aggregate is just the mix. The 2017-filed share of first-instance decisions was 78.5% in January, 56.3% in February, 37.9% in March, 24.0% in April, 8.2% in June and 1.1% in December. Weighting the two cohort rates by those shares gives 15.64%, 14.15%, 13.39%, 12.63%, 10.43% and 7.92%; the observed aggregate is 15.75%, 14.16%, 13.46%, 12.68%, 10.47% and 7.95%. The two-cohort model accounts for the series almost exactly.

How much of the year's decline is the cohort switch? If the 2018-filed rate had stayed at its February–March level of 11.5%, the December aggregate would have been about 11.6%. So roughly half of the fall from 15.75% to 7.95% is old cases leaving the pipeline, and the other half is a further drift inside the 2018-filed cohort, from 11.5% to 7.8%.

What the within-cohort drift can and cannot mean

The second half is harder to read, and this is where a careful study has to stop. Within one filing year, cases decided later are not a random sample of cases filed: they are the slower ones. Slower cases differ from fast ones in amount, complexity, number of defendants and whether anyone contests. A drift inside the cohort could be creditors learning through 2018, or it could be selection by case duration. With decision date and filing year alone, the two cannot be separated.

The same selection shows up the other way round. Cases filed in 2017 but decided late in 2018 raise spousal liability slightly more often than those decided early, up to 23.0% in October; judgments with 2018 case numbers decided in late 2019, which include appeals, rise to 13–19%. Long-running cases are different cases. Any analysis that conditions on decision date inside a cohort inherits this.

What the data does support: the step between cohorts is large (about 16.5% against about 11.5% for cases decided in the same months), and it lines up with the filing boundary, not with the decision date of the rule.

The pre-trend that was a denominator

One more trap, which we fell into ourselves. On our Chinese-language data page about marital debt we reported the mention rate among private-lending records with text, and that series was already falling through 2017, from 8.59% in January to 6.65% in December. We suggested treating early 2017 as a candidate break.

That fall is the denominator. The broad population counts every private-lending record with a body, including procedural rulings, withdrawal orders and transfer notices, which rarely discuss the merits. The share of that population that are merits judgments fell from 45.5% in January 2017 to 35.9% in December 2017, with the same year-end dip (36.5%) in November 2016. Restrict to merits judgments and the 2017 series is flat: 18.30% in January, 17.85% in December; restrict further to first instance and the two halves of 2017 are 16.46% and 16.45%. We are correcting the Chinese page.

A denominator that mixes document types will produce trends whenever the mix moves, and the mix moves with court calendars, campaigns to clear backlogs and the way each ingest batch was collected. Fix the document type before reading a pre-trend.

Date integrity checks before you start

An event study is only as good as its clock. Two cheap checks on this example:

• Citations before the rule existed. 1,117 records dated before 17 January 2018 cite the 2018 interpretation, out of 274,455 that cite it at all (0.41%). Those are date errors in the source or our ingest, and they will sit on the wrong side of any cutoff.
• Impossible cohort-date pairs. 730 private-lending merits judgments dated 2017 carry a 2018 case number, out of 920,898 dated 2017 (0.08%). Small, but worth excluding explicitly rather than averaging over.
• Parse coverage. Of 930,920 private-lending merits judgments decided in 2018, 861,001 (92.5%) have a case number whose year we could parse as 2012–2019. Report the rest as a separate bucket, not as zero.

Separately, three of our ingest batches store a publication date in the date field. Left in, they place decisions months or years after they happened and blur exactly the boundary you are trying to measure. We exclude them from every series on this page; a vendor delivery should tell you which records have that property.

A design that holds up

• Pick the clock by actor. What the court does (cite, apply, award) is a decision-date event. What the parties do (sue, plead, name defendants, choose a cause of action) is a filing event. Use the case-number year, or a filing date if you have one, for the second.
• Compare cohorts decided in the same window. Old-filed and new-filed cases decided in the same months share the court, the judges and the rule; the main difference is when the claim was framed. That contrast is cleaner than before-and-after on decision date.
• Expect a pipeline tail and measure it. In this example, 2017-filed cases were 78.5% of first-instance private-lending decisions in January 2018, under 25% by April and about 1% by December. That curve is the smear you will see in any date-indexed series. Different causes of action have very different tails.
• Do not read within-cohort drift as learning without controlling for case duration. Report it, flag the selection, and stop.
• Fix the document type before reading pre-trends; rulings and notices move with calendars.
• Run the impossibility checks above and publish their counts.
• For model evaluation, the same logic applies: a set of “2018 judgments” is mostly 2017 litigation in its first quarter. If you are testing whether a model has absorbed a rule change, slice by filing year as well as decision year.

How to reproduce this on any delivery

# Elasticsearch, body analysed with ik_max_word; date stored as yyyy-mm-dd keyword.
# Population: private-lending merits judgments, first instance
{"bool": {"must": [
  {"bool": {"should": [
     {"terms": {"cause": ["民间借贷纠纷", "民间借贷"]}},
     {"match_phrase": {"title": {"query": "民间借贷", "analyzer": "ik_max_word"}}}
  ], "minimum_should_match": 1}},
  {"match_phrase": {"body": {"query": "判决如下", "analyzer": "ik_max_word"}}},
  {"wildcard": {"caseNo": "*民初*"}}
]}}
# Filing cohort: add {"wildcard": {"caseNo": "（2017）*"}} OR "(2017)*"
# Behaviour: add a should over 夫妻共同债务 / 夫妻共同之债 / 夫妻共同债 (match_phrase)
# Bucket by month with a filters aggregation of date ranges.

Check the case-number prefix forms in your own data first. Full-width and half-width parentheses both occur, and some records carry suffixes such as 之一 after the number; a wildcard on the prefix handles both.

What this page does not claim

• Rates are shares of private-lending merits judgments in our index, not of loan disputes filed in China. Cases that ended without a judgment are not in the denominator.
• “Raises spousal liability” means the phrase appears in the judgment. It does not tell us whether the spouse was sued, held liable or released.
• We do not claim the 2018 interpretation caused the cohort difference. The cohorts were also filed at different times for other reasons, and creditor behaviour may have been moving with commentary before the rule was issued; the flat 2017 series argues against a large anticipatory shift on this measure, nothing more.
• The within-2018 drift is not attributed to learning or to selection; we cannot separate them with these fields.
• The 60-judgment reading check supports what the phrase picks up, not the decimals.
• Monthly record counts reflect our collection windows, not litigation volume. Compare shares, not counts.