Several updates were made to American Community Survey (ACS) write-in response coding operations for 2025 processing due to resource constraints and to take advantage of emerging technical solutions. Subject matter analysts reviewed the resulting files to assess the impact of these changes during their review of the editing process and during their review and acceptance of the weighted data estimates. Based on their thorough review, subject matter analysts identified no special comparison guidance or caution needed specifically due to these changes.  

Ancestry

Normal coding operations were conducted for data collected from January 2025 through May 2025. Beginning with data collected in June 2025, the legacy autocoder was applied following normal procedures; however, to eliminate the clerical coding workload, all uncoded results were sent to a secondary automated process rather than continuing to be clerically coded. The secondary process utilized a fuzzy-matching algorithm to assign the best matching code comparing the write-in string with the master Ancestry code list. Responses that exceeded a string-length quality check were assigned a value of 997 “Not Reported or Classified” consistent with existing clerical coding procedures for uncodable responses for Ancestry. Subject matter analysts determined the resulting coded data were comparable to previous years; however, data users will notice a slight increase in the proportion of “Unclassified or not reported” due to the elimination of clerical coding for this topic.

Field of Degree

Normal coding operations were conducted for data collected from January 2025 through May 2025. Beginning with data collected in June 2025, the legacy autocoder was applied following normal procedures; however, to reduce the clerical coding workload, all uncoded results were then sent to a secondary automated process before being clerically coded. The secondary process utilized a fuzzy-matching algorithm to assign the best matching code comparing the write-in string with the master Field of Degree code list. This improvement reduced the clerical coding workload by approximately 30 percent. Responses that failed a string-length quality check were then sent to clerical coding for completion of all remaining uncoded responses. Subject matter analysts determined the resulting coded data were comparable to previous years.

Industry and Occupation

Normal coding operations were conducted for data collected from January 2025 through July 2025. Beginning with data collected in August 2025, a new autocoder was introduced to support the coding of Industry and Occupation write-ins. The new autocoder uses a large language model (LLM) to assign Industry and Occupation codes, both jointly and individually. It replaced the legacy autocoder which used a logistic regression model first implemented in 2012. The new LLM autocoder produces higher coding rates for Industry and Occupation, both separately and jointly, while coding data faster and more efficiently. This improvement reduced the clerical coding workload by approximately 32 percent. Responses that remained uncoded after the LLM autocoder was applied were sent to clerical coding for completion of the remaining uncoded responses. Subject matter analysts determined the resulting coded data were consistent with previous years.  

For more information on the Industry and Occupation coding, visit Guidance for Industry and Occupation Data Users.

Place of Work

Normal coding operations were conducted for data collected from January 2025 through July 2025. Beginning with data collected in August 2025, a new address matching tool was introduced to support the geocoding of Place of Work write-ins, which increased match rates by approximately 5 percent and reduced the clerical coding workload. Beginning with data collected in September 2025, a new business rule was introduced to strategically remove approximately 40 percent of cases from the autocoding and resulting clerical coding workload to better balance the coding workload and real-time capacity. The cases that were initially removed were later sent to autocoding. Any cases left uncoded after this final step were left uncoded consistent with existing coding procedures, mostly affecting Place of Work below the place level which is not included in standard ACS data products. Subject matter analysts determined the resulting coded data were comparable to previous years.

These changes to ACS write-in response coding operations will remain in place as the ACS Program continues to investigate and implement additional efficiencies and improvements. Likewise, subject matter analysts will continue to assess the coding results to ensure and maintain data quality for all affected topics.