Dates in return_expectations don't seem to accept date expressions

[StevenMaude]

This is a repost of a question from @jscuffell in #389.

(I closed that issue in favour of the question here, for visibility. The issue seems more "usage"-oriented than a direct issue with the documentation. It, of course, may turn out that this is feedback to improve the docs.)

Hi there,

I've just started having a go at using Opensafely to generate dummy data to use to plan the data analysis for a separate project. Apologies if I'm misunderstanding what index_date does or it's my study definition code, but I'm getting some inconsistent behaviour with the date in return_expectations. My current study definition is below:

from cohortextractor import (
   StudyDefinition,
   codelist,
   codelist_from_csv,
   combine_codelists,
   filter_codes_by_category,
   patients,
)
from codelists import *

def loop_over_codes(code_list):
   def make_variable(code):
       return {
           f"{code}": (
               patients.with_these_clinical_events(
                   codelist([code],system="snomed"),
                   on_or_after="index_date",
                   returning="date",
                   find_first_match_in_period = True,
                   date_format="YYYY-MM-DD",
                   return_expectations={
                       "date": {"earliest": "2004-01-01", "latest": "today"}, # this date expression
                       "rate" : "uniform",
                   },
               )
           )
       }

   variables = {}
   for code in code_list.keys():
       variables.update(make_variable(code))
   return variables

study = StudyDefinition(
   default_expectations={
       "date": {"earliest": "2004-01-01", "latest": "today"},
       "rate": "uniform",
       "incidence": 0.2,
   },
   population = patients.all(),
   index_date = "2021-09-01",
   
   age = patients.age_as_of("index_date",
                           return_expectations = {"int" : {"distribution": "population_ages"}, "incidence" : 1}),
   
   sex=patients.sex(
       return_expectations={
           "rate": "universal",
           "category": {"ratios": {"M": 0.49, "F": 0.51}},
       }
   ),
   
   **loop_over_codes(codelists),
)

When I do a opensafely run run_all, the cohortextractor does not return any data for the variables which have been created in the make_variable loop (first few lines of input.csv below):

cancer1,cancer3,bornlatam,dementia,cancer2,smi,ckd35,asthma,age,sex,patient_id
,,,,,,,,39,M,911661
,,,,,,,,38,F,571755
,,,,,,,,36,F,681666
,,,,,,,,44,F,369185
,,,,,,,,8,F,524457
,,,,,,,,44,F,983471
,,,,,,,,14,F,876029

Interestingly, if I change the 'earliest' parameter in return_expectations from '2004-01-01' to 'index_date', it works as expected:

cancer1,cancer3,bornlatam,dementia,cancer2,smi,ckd35,asthma,age,sex,patient_id
,2021-09-02,,,,,,,42,F,389551
,,,,,,,,64,F,287564
,,,,,,,,61,F,580975
,,,,,,,2021-09-02,4,F,212551
2021-09-02,,,2021-09-02,,2021-09-02,2021-09-02,,64,M,250708

It's no problem - I can just change the index_date - but wondered if this was the correct behaviour?

Many thanks,
Jamie