Howard County Community Voices

How we collected it

Everyday Canvassing went door to door across apartment communities in Howard County and talked with residents in person. Canvassers asked open questions about life in the county, such as housing, schools, safety, costs, transportation, and what people would change. People answered in their own words, and a few basic details were noted when residents chose to share them, like age range and whether they rent. Taking part was voluntary, and many conversations happened in languages in addition to English.

How we cleaned it

Before anything went online, every interview was prepared so it would be safe to publish. We removed personal details that could identify someone, such as names, phone numbers, email addresses, and apartment or unit numbers. Where a resident said their own name, we still removed it, because a stranger reading the site should not be able to attach a name to a story. When we took something out, we left a marker in its place, like “[name removed],” so you can see that a detail was protected without seeing the detail itself. Where a passage of a recording was hard to hear, we say so in the text with markers like “[unclear]” rather than guessing. Interviews held in other languages are translated by people, not just software, and labeled. When what you are reading is a canvasser’s written summary rather than a recorded conversation, a badge on the story says so.

We also separated canvassers’ own side notes from what residents actually said, so the published text is the resident’s voice and not staff observations. We corrected the spelling of apartment community names so each place shows up once on the map instead of several times, and we placed each community using a verified address unassociated with a person’s home. Age was recorded two different ways during collection, as ranges and as exact numbers, so we put everything into the same set of age ranges.

In total, we checked and cleaned each interview 4 - 6 times. The first check occurred when each interview was first collected. Our team of staff canvassers flagged sensitive information and corrected mistyped notes. The second check came from our admin team where we redacted flagged sensitive information, corrected more mistyped notes, deleted notes that were not captured well enough to be understood, and deleted all contact info, names, and address info before proceeding to check 3. The third check ran the double-checked interviews without contact info, names, or addresses through a large language model to suggest further removals of information that may look harmless but when pieced together could reveal enough information about any individual that they could be identified by a neighbor, co-worker, boss, or otherwise. When information is pieced together to identify and target a person, this is called “mosaic re-identification.” The third check also standardized formatting and punctuation across all our interviews so they were easier to read, and helped us build short summaries of the long-form interviews where we did not already create them. In the final fourth check, our admin team studied each interview adjusted by a large language model to make sure each interview still maintained the original stories and conversational style of each person we interviewed. When the large language model did not maintain the original story or style well, our admin team revised each interview to make sure they used the original wording of the people we interviewed. We also removed or generalized sentences in interviews in order to reduce the possibility that any person we interviewed could be identified by anyone who is not themselves. Some interviews went through several re-checks of our admin team to ensure the people we interviewed could not be easily identified by the pieces of info they shared in their interview, and that the interview maintained what the people interviewed actually said.

How we categorized it

To make patterns visible, we tagged each interview with the topics it touched on. There are 29 topics, grouped into 8 broader areas. For each topic in an interview, we also marked whether the resident’s assessment of that topic came across as concern, mixed, or positive. Where a resident expressed no view on a topic, we record no sentiment rather than guessing. Our rule was simple: if we could not point to specific words in the interview that supported a tag, we did not apply it.

Note: The topics and the concern, mixed, and positive labels are Everyday Canvassing’s reading of what people said. Another person might sort the same words differently. The residents’ words are theirs; the labels and groupings are our interpretation, which we think makes the bigger picture easier to see.