In the comments on yesterday's post on Somerville building age someone suggested census data for housing units could be helpful. I've now looked into it, but it's not great: Somerville seems to not have been submitting building permit data to the census until recently, and the way condo conversions are tracked overstates growth.

Initially it does look good: we have B25001 "Housing Units" representing estimates for how many housing units there are in various places, including Somerville. Here's are the 1-year estimates from the ACS:

year	number	margin of error
2017	36,788	±1,952
2016	33,565	±1,918
2015	32,226	±1,849
2014	33,085	±2,025
2013	34,375	±2,187
2012	33,079	±1,730
2011	32,794	±2,086
2010	31,604	±1,989
2009	30,198	±2,047
2008	32,621	±2,257
2007	32,903	±2,401
2006	31,781	±2,347
2005	33,627	±2,901

The methodology (pdf) says the estimate comes from:

A surprising thing to me, then, is that the reported margin of error is relatively consistent. If we're starting with a decent estimate from the Census every ten years, and then adjusting it to try to keep it current, I would expect to see a low margin of error in 2010 (and 2000) and then see it rise over the course of each decade.

Anyway, since Somerville doesn't have many (any?) mobile homes we're just talking about New Construction and Housing Loss.

For New Construction they estimate based primarily on building permits:

Building permit numbers come from the Building Permits Survey which claims to have data for 20,000 municipalities. Each one files form C-404 "Report of Building or Zoning Permits Issued and Local Public Construction". The documentation (pdf) has:

Two sets of data are shown for each type of construction:

1. Estimates with Imputation - includes reported data for monthly respondents and imputed data for nonrespondents.
2. Reported Only - includes only reported data for respondents.

Pulling out the data for Somerville (sheet) there are many years where "estimated" differs from "reported", which means Somerville was nor submitting the form every year. Worse, the data doesn't distinguish between "they submitted the data and it was zero" vs "they didn't submit any data". The "reported" data from 2005-2014 is all zeros, so this does sound like failure to report.

In 2015 we suddenly have 604 units reported, after a decade of zeros, and since that's ~3x higher than what we see in 2016-2018 I wonder if some number of previous years of permits got reported all at once in the 2015 data?

Mostly this makes me not trust the ACS housing units numbers here very much: it looks like they stayed flat for years because of missing data and then suddenly jumped up year over year because of perhaps a backlog coming in.

There's also a problem with the way they calculate loss. A lot of new construction in Somerville is replacing existing buildings, some of which are residential. But loss is calculated extremely roughly:

We calculate housing unit loss by applying an annual loss rate to the housing stock. Then, we add to that estimate the number of units lost due to natural disasters. Housing loss rates are derived from the 2009 and 2011 American Housing Survey at the regional level. A unit is counted as lost if a survey was completed in 2009, but it was listed as a non-response (Type C, 30 - Demolished) in the 2011 survey.

This means that if the rate at which units are demolished and replaced changes it will be picked up in the construction numbers but not the demolition numbers and erroneously counted as an increase. Looking at the 2018 building permit survey numbers I see Somerville reporting 32 permits for new 1-unit buildings, which sounds to me like condos. Since those were probably conversions from existing 2-3 unit buildings, we're counting the added units but not the removed ones.

Overall it does not sound like the ACS numbers are very useful. They're missing a lot of data and what data they do have is unclear. I would love to have solid data on this, especially going back a long time, but it still doesn't sound to me like that's been assembled.