A few weeks ago, when I inaugurated Chart of the Week, we looked at U.S. housing starts using annual data from 1890 to 2010. There is usually more than one way to design a chart, especially for such a key data series. I promised then to do a follow-up with more about starts, and here we are. We’ll use this opportunity to highlight three important elements of charting data: (1) re-expression; (2) periodicity; and (3) seasonal adjustment.
First, re-expression. Two friends -- Nino Pedrelli and Ann Danner -- asked how the key chart would look if we made an adjustment for the changing size of the U.S. population. In 1890, US population was about 63 million; and with average household size of about five, we had about 13 million resident households. Today's population stands at about 310 million, or with a household size of 2.6, about 118 million households.
Let's look at the simple transformation of our annual housing data we presented two weeks ago. (Take a look at that chart first, to refresh your memory.)
Today’s first chart simply divides the annual number of total housing starts by the number of households in the U.S., represented by the thick blue line and the first y-axis. In the starts chart from our first posting, prewar housing starts of half a million to a million didn’t look terribly impressive compared to the postwar starts ranging between 1-2 million and above (until recently). But compared to the number of households, in today’s chart, we see that the 20s were a real boom and bust.
Part of the reason for the relatively high rate of starts many decades ago was higher population growth. Today’s Figure 1 also shows the annual rate of population growth, using the red line and the second y-axis. Pre-Depression U.S. population growth usually ran at 1½ to 2% per year, quite a bit higher than today's 1%. (Note also the decline in population growth during the Depression, and the big one-time shifts in resident population associated with movements to and from overseas during the World Wars). Other factors, not addressed directly in the chart, include changes in the rate of depreciation of typical units, rising incomes and concomitant demand for larger and better units, geographic mobility, and the fact that average household size was declining more rapidly a century ago than it’s declining today.
Next let’s look at the issues of periodicity and seasonal adjustment. Many basic real estate and economic indicators come with varying “time signatures,” e.g. annual, quarterly, monthly and so on. We looked at annual starts before partly because annual data are available for the longest time span, back to 1890. Monthly housing starts data for the U.S. are available after 1959. These are available both as seasonally adjusted, and unadjusted.
Figure 2 presents both these monthly series starting in 1987. We start in 1987 instead of 1959 here simply to allow the reader to see patterns within years more clearly.
Seasonal adjustment arises because when we look at one month (or quarter) of data and compare it to the previous period, we often wish to somehow account for regular and fairly predictable changes in certain months or quarters. For example, weather affects construction – more new houses are started in May than in January for obvious reasons (at least in Wisconsin).
Many time series, then come two ways – seasonally adjusted (often expressed at annual rates), or not seasonally adjusted. The seasonal pattern in Figure 2 is obvious, with unadjusted starts spiking in April or May of most years, hitting low points in December and January.
Figure 3 presents the seasonally adjusted data all the way back to 1959; and our final chart combines all three elements by presenting monthly housing starts per 1000 population.
Two key results are evident from Figure 4. First, we see a declining trend in starts per capita over the last 50 years; this trend remains even if we omit the last few years of data. Second, even after accounting for this trend, recent housing starts are the lowest that we've observed per capita since collection of the monthly data began in 1959.
Posts
