Reviewing the 2017 Census Population Estimates
Breaking Down the Key Facts
Yesterday, Census released new population estimates for the nation and the states. These estimates reflect population on July 1, 2017, but also reflect revisions back to the last Census, in 2010. The media covers these figures intensively, with rhetorical awards meted out for the fastest growing and shrinking states.
But my interest is more technical: how is the Census Bureau revising its view of each state’s population?
One way to do this is to take each vintage of estimates from 2011 to 2017, take the annualized growth rate over its period, and extrapolate it out to 2020, giving a “naive forecast” of 2020 population from each vintage. This will give us a consistent estimator of which states Census is continually slashing population outlooks for, and which ones it is raising. I did this graph in my last post laying the groundwork for reading Census updates, so now I’ll duplicate it, with 2017 added in:
From 2011 to 2016 I show in hues of red, lightest for 2011, darkest for 2016. 2017 I highlight in bright blue.
As you can see, some places waffle. But, generally, the blue dots tend to fall further out in the direction revisions were already headed. Alaska and Alabama get more negative, Florida and Idaho get more positive, etc. There are, of course, exceptions: Illinois’ 2020 outlook edged ever-so-slightly upwards after 6 years of worsening outlooks. The same is true of New Mexico. But in most cases, past trends in forecast revisions predict future trends in forecast revisions.
I’ll note that I suggested in my last post that Puerto Rico would probably get a cut: it did, though granted less of a cut than I expected; 2016 population was revised down by nearly 5,000 residents, even as 2017 shows continuing decline. I continue to believe the Census Bureau is somewhat overstating Puerto Rico’s population.
What Revisions Did Census Make?
Now, before I go further, let’s talk about why Census back-year numbers change!
The simplest reason is just that more updated data is available! Data about a given period is only available with a time-delay, so the most recent year estimates are often made with incomplete, provisional, or aggregated data, and may require statistical supplement to get the final estimates. As new and more complete data becomes available for each year, the estimates improve in quality. Census doesn’t have the full-and-complete data used for annual estimations until about 3–5 years after the year being estimated, though they have pretty good data by the 2nd or 3rd year.
Then there’s the more complex reason: methodology changes! For 2017, Census made some changes. In 2016, as I’ve discussed before, Census changed the way they estimate the emigration of foreign-born people in America, substantially increasing their estimate of such movement. In 2017, they’ve added further detail to their estimate, by using 1) more precise year-specific data, 2) more detailed mortality tables for Hispanics. The result is lower emigration estimates, though still above the 2015 estimates. I’ll admit I don’t totally understand why these changes would have this effect, and would appreciate a little bit more detail from Census clarifying what drove this change. I trust they’re doing it all correctly, I’m just professionally curious what in this data drove what ends up being a significant change.
Census also doubled their estimate of native-born emigration, a long overdue change. But this then implies that the changes for foreign-born outflows must have been really significant, enough to offset a worsening native-born component by 40,000 or more outflows per year.
There were also some technical changes to net domestic migration, especially in how the IRS estimates retiree migration. I have some beefs with Census’ use of administrative data for domestic net migration generally, so will hold my fire here, as my critique ends up being a larger meta-critique on Census’ reliance on the incomplete population universe of IRS data.
What Components Changed?
There are two different ways to look at a given component, say, deaths. One is to look at the time trend. The other is to look at revisions.
If we want to look at one state, we can look at revision size and the time trend simultaneously. But if we want to look at all states together, we can’t do both simultaneously. For example, we could look at the time trend and revisions made to domestic net migration in Illinois.
As you can see, in the Vintage 2017 data, net domestic migration got worse from 2016 to 2017. But the Vintage 2017 data makes domestic net migration look somewhat less severe in almost all of the back years! So you can say that Illinois’ net migration fell in 2017… and also that Census raised its estimate of Illinois’ net migration in 2017.
The chart below will show each state and the cumulative revisions made to births, deaths, domestic migration, and international migration in the 2017 vintage estimates, divided by 2010 population. Basically, it will show the broad scale of revisions for each component and state.
In most states, the biggest absolute size of revisions is in the migration categories. That’s to be expected, since those components have the most extensive data problems, the most frequent methodology changes, and are the hardest to pin down to physical events, whereas births and deaths are far easier to geographically allocate.
Now, data visualization nerds are probably howling at me that the graph at left is not a useful presentation of information. And I basically agree. But I want to prevent the full range of information, and what the graph does show is big outliers. Domestic migration was revised up a lot for Connecticut, Delaware, DC, New Mexico, North Carolina, and South Carolina, and substantially down for Washington, South Dakota, Rhode Island, North Dakota, Montana, Iowa, and Arizona. Much of that has to do with changed treatment of retirees to a matched-address-change method rather than a residual-loss method.
For international migration, there were big upward revisions in Texas, Rhode Island, New York, New Jersey, Minnesota, Massachusetts, Maryland, Florida, and California… and big downward revisions for Wyoming, West Virginia, Vermont, Utah, the Dakotas, Montana, Maine, DC, and Alaska. These changes reflect methodology changes which tended to re-allocate migration into a number of hub or gateway states, increase net inflows overall, but increase net outflows of native-born people.
For births and deaths, the range is much smaller. But Alabama, Delaware, South Dakota, Montana, and to some extent Wyoming all saw increases in their birth estimates, while Puerto Rico, Vermont, North Dakota, Massachusetts, and California saw cuts. For deaths, the big upward revisions are in Utah, Vermont, Nebraska, Maine, Kansas, and Arkansas, while downward revisions can be seen in New York, Maryland, Hawaii, and DC.
Throughout the whole country, changes to births and deaths were marginal, mostly what we’re seeing is Census reshuffling the geographic detail of where they assign the births or deaths to, rather than a change in the total number. International migration is the big change.
This does make a second year in a row where big method changes are driving big revisions in international migration. It’s good for Census to improve its method. It would be even better to make all the necessary improvements in one fell swoop rather than jolting the data every year.
Mapping Revisions
That data gives you a lot of detail. But we can also map these changes. The most interesting one to map is net migration. Below, I show a map of the cumulative change in all net migration by state, represented as a percentage of 2010 population.
As you can see, the biggest cuts to net migration are in the upper midwest, mountain states, and major aging-states like WV, VT, and ME. Meanwhile, we can see big improvements in estimated migration throughout the Atlantic coast from Florida to NH, and in many immigrant-receiving states.
We can also look at the revisions made to natural increase, so births and deaths.
So what you can see here is that the birth/death balance was revised more negative throughout basically the whole western part of the country; indeed, almost everwhere west of the Appalachians was breakeven or down, while, again, the Atlantic and Gulf areas do better. Note that the scale for the natural increase revisions is greatly compressed.
When we combine these, we get cumulative revisions, so which states got the biggest changes compared to their 2010 populations?
As you can see, it looks a lot like the migration map. Migration is the biggest component, so this makes sense. But it doesn’t look exactly like the migration map. If you clink the two links and flip between the two windows, you’ll see some differences.
So with revisions covered now… what do we make of the moves for 2017?
How’d My Forecasts Do?
Okay, But I Can Do Better
Not long ago, I published some forecasts of Census’ 2017 population estimates. I’d never done that exercise before, but figured I’d give it a shot. So I did. The results are interesting. The graph below shows the actual 2017 estimate as a percent of 2016 population, my 2017 estimate as a percent of 2016 population, a 2017 forecast based on naive extrapolation from the 2016 vintage estimates, and 2016 population itself.
The gray dot is 2016 population. The black dot is 2017 population indexed to 2016 population from V2017. The blue dot is my forecast of what V2017 would say. The red dot is a prediction of V2017–2017 estimate done by simply extrapolating the V2016 2011–2016 growth rate out to 2017.
So who performs better? Well, my forecast was, on average, wrong by 0.26% for 2017 population. A naive forecast was wrong by 0.32%. So I did marginally better than an extremely rudimentary forecast. My method was the “more correct” method… about the half time. Which is not stellar.
But where my method did well was on avoiding very big misses. The standard deviation of the error margin for my method is 0.19%. For the naive forecast, it’s 0.41%.
The trouble is, coming up with a better method is hard. Census is evidently using a large amount of data sources that aren’t publicly released.
But here’s the trick. Some of the places where I missed up also had revisions. So what happens if we add in revision effects? How does revision scale associate with my forecast error?
Huzzah! I’m not a total doofus! There’s a fairly close association between the scale of Census’ revisions to back year populations and my forecast error. When my forecast error was too high (higher Y axis values), it usually associated with Census having made downward revisions to the baseline population (more leftward X axis values).
All in all, my forecast of 2017 population, then, did reasonably well, given the constraint that Census makes prudential changes to methodology that have large effects on a fairly regular basis.
Growth in 2017
You’ve patiently waited through this long and painful blog post. I shall reward you with maps.
First of all, which states grew the most in 2017?
We can see two distinct gainer-regions: the western states and the southeast. Texas, pieces of the upper midwest, and upper New England (and my always-favorite-Delaware!) also do well. But it’s worth noting, the biggest relative gainer states are not Texas and Florida. They’re South Carolina, Nevada, and Idaho. Maybe that’s crystal-clear evidence for your preferred policy narrative… but it’s not for me.
We can see a swathe of red from Boston to Virginia to Oklahoma to Wisconsin. This region has lots of different stuff going on. New York and Illinois are big losers, with Connecticut, West Virginia, and New Jersey also having meaningful losses. Other states do better… but largely by riding on gains from Illinois and New York. Even gainers like South Dakota and Minnesota are largely benefiting from the fracking-crash in Wyoming and North Dakota.
Aside from those regions, we see other narrative-challenging losses: Alabama and Louisiana are losing people! Both Alaska and Hawaii have major net outflows, while California has meaningful outflows too.
Next up, international migration!
Net international migration is positive in all states. But it’s more positive in some than others. We can see the most significant flows into the northeast urban states, Florida, Hawaii, Texas, and California. There is variation throughout the rest of the country, but nothing particularly distinctive. But it may be worth noting that Montana and West Virginia have the lowest immigration rates.
And what about total population growth?
Several states had population decline in 2017. Wyoming, West Virginia, Illinois, and Alaska had the most significant declines, but there are more modest declines in Hawaii, North Dakota, Louisiana, and Mississippi. Again, maybe that matches your ideological priors about winner/loser places… but it seems fairly broad to me!
The fastest growers are Idaho and Nevada, followed by Utah, Washington, Arizona, Colorado, Texas, Florida, and South Carolina.
A few notes here on some claims I’ve made. I had some dire things to say about Connecticut not so long ago. Specifically, I said:
“See that dip at the end there? That’s a 19,000 person statewide population decline. There’s not really any reason to think that’s going to make a huge turnaround in the near future either. The best they might eke out is population stability, but continued decline is likely. That means that Connecticut’s government is unlikely to face the need or ability to hire lots more people: indeed, if this trend continues, Connecticut could show as much as a 30,000 person population decline from the 2010 to 2020 Census. By 2026, 10 years from the most recent data, population could be 3.51 million, or about 65,000 people less than the 2016 estimate.”
Well, in 2017… population rose by 500 people! This is partly due to a method change in retiree migration, but could also reflect a real stabilization!
Next up: Hawaii! I’ve mentioned several times that there’s an interesting comparison between Hawaii and Puerto Rico. Well… call me a prophet of doom but… Hawaii’s population started falling this year, by about 1100 people. May be a rough time to be an island.
Next up: Illinois. Population kept falling. Guess it’s not a good time to be… a large midwestern state with a booming urban hub? Except, err, Louisiana’s population also started declining this year… so… guess it’s not a good to be… okay this gets confusing. Notably, Mississippi is in its 3rd year of decline. Anybody selling a “warm states have huge advantages” story should perhaps be reconsidering given that Maine’s population growth is making a sharp turn upwards while Louisiana shrinks.
Conclusion
Census revisions matter, especially if, perhaps, your job is to try and forecast population. This was the second year of substantial methodology revisions to Census’ international migration component. Revisions and growth patterns alike defy easy political or ideological explanation. Remember that as you read your local Census population pieces.
Check out my Podcast about the history of American migration.
If you like this post and want to see more research like it, I’d love for you to share it on Twitter or Facebook. Or, just as valuable for me, you can click the recommend button at the bottom of the page. Thanks!
Follow me on Twitter to keep up with what I’m writing and reading. Follow my Medium Collection at In a State of Migration if you want updates when I write new posts. And if you’re writing about migration too, feel free to submit a post to the collection!
I’m a native of Wilmore, Kentucky, a graduate of Transylvania University, and also the George Washington University’s Elliott School. My real job is as an economist at USDA’s Foreign Agricultural Service, where I analyze and forecast cotton market conditions. I’m married to a kickass Kentucky woman named Ruth.
DISCLAIMER: My posts are not endorsed by and do not in any way represent the opinions of the United States government or any branch, department, agency, or division of it. My writing represents exclusively my own opinions. I did not receive any financial support or remuneration from any party for this research.
