“Going With the Flow”—A Comparison of Interstate Elderly Migration During 1970–2000 Using the (I)PUMS Versus Full Census Data

Abstract

Objectives.

We investigate how much state-to-state elderly migration patterns have changed during 1970–2000 and compare the findings from 2 commonly used sources of data, the census flow tabulations and the integrated public use microdata series (IPUMS).

Methods.

We calculate descriptive statistics such as migration rates, the distribution of top destination and origin states, and a new migration Herfindahl–Hirschman Index that measures geographic concentration. Comparisons over time and between data sources are formalized using correlations and regression analyses that permit persistent flow patterns.

Results.

After an increase between 1970 and 1980, elderly migration rates have been stable, with a slight decline. Elderly migration has become less geographically concentrated; the decline of California and Florida and ascension of Nevada and the Carolinas as top destinations are evident. Correlation and regression analyses reveal that migration patterns are overall very persistent over time, especially using census tabulations based on a larger sample.

Discussion.

Elderly migration patterns have been quite stable since 1970. Using the IPUMS, as most migration studies do, exaggerates the changes in elderly migration over time in both descriptive and statistical analyses, a result that is likely due to its smaller sample size and the relative rarity of an interstate move.

A rich literature exists on the patterns and determinants of elderly migration (e.g., Walters, 2002), and several studies analyze how elderly migration patterns have changed over time (e.g., Flynn, Longino, Wiseman, & Biggar, 1985; Lin, 1999; Longino & Bradley, 2003, 2006). None, however, systematically considers changes in interstate migration flows during 1970–2000. This period is the longest in which both origin and destination states are available in census microdata, information necessary to identify out-migration and migration flows. It is also marked with many changes believed to influence interstate migration—for example, general changes such as the widespread use of air conditioning and decreases in long-distance communication costs, as well as changes in the elderly population, specifically such as the growth of two-earner (thus, two-retiree) households and declines in reported elderly disability (e.g., Wilmoth & Longino, 2006).

Moreover, previous studies rely on the Public Use Micro-Samples (PUMS), which is based on, at best, 5% of the U.S. population. By contrast, the full census tabulations are based on approximately one sixth of the U.S. population. The relative infrequency of elderly interstate migration makes sample size important, especially for discerning changes over time. Sample size concerns are likely to grow with the elimination of the 2010 census long form—and the census data it yields. The American Community Survey (ACS), a smaller annual survey replacing it, has already been found to exhibit greater irregularities in age-specific migration flows than the PUMS (e.g., Raymer & Rogers, 2007). Such irregularities are likely exaggerated when investigating changes over time. The extent to which the widely used PUMS exaggerates irregularities compared with census tabulations is unknown.

This report fills both gaps. Using data from the 1970, 1980, 1990, and 2000 full census tabulations and the corresponding (I)PUMS, several measures of interstate elderly migration are compared over time and across data sources. Our analyses confirm some changes reported in past research but also reveal that the smaller (I)PUMS data overstate the degree to which elderly migration has changed.

METHODS

Data

Both sources are census data in which an individual is counted as a migrant if s/he lives in a different state at the time of the census than 5 years prior, which has the limitation of underestimating movement (e.g., Sergeant, Ekerdt, & Chapin, 2008). The state of residence 5 years ago was not coded in the 1960 census microdata, which precludes identifying migration flows and limits our study to 1970–2000. Age is measured at the time of the census; individuals can be up to 5 years younger when the move took place.

The first sources are state-to-state migration flows reported in Census Summary Tape File 3 for 1980, 1990, and 2000 and the Fifth Count File for 1970. We refer to these as “census flows.” These files report the number of individuals who moved between each pair of states and are publicly available in all four census years for ages 5+ but only since 1980 for ages 65+. We therefore requisitioned a comparable tabulation from census for ages 65+ in 1970. Subtracting elderly flows from total flows yields nonelderly flows (ages 5–64 years). We use nonelderly flows to discern whether key findings are unique to the elderly population.

The second source is the Integrated Public Use Microdata Series (IPUMS), created by the Minnesota Population Center in an effort to bring each census’s PUMS into one place and improve the uniformity of coding across years (Ruggles et al., 2004). The PUMS is the census-based data most often utilized in past studies (e.g., Flynn et al., 1985; Lin, 1999; Longino & Bradley, 2003, 2006). Reported at the individual level and containing many characteristics, it provides researchers flexibility in constructing migration flows (e.g., based on different ages or retirement status).

The IPUMS is therefore richer data but is based on smaller samples than the census flows. The census flows are calculated from the census long form, a sample of approximately one in six. Since 1980, the largest single sample for the IPUMS is 5% of the U.S. population. In 1970, the Form 1 State Sample, a 1% sample, is the only one containing migration. IPUMS provides individual weights with which to generate representative statistics. In 1970 and 1980, every individual was given equal weight, whereas in 1990 and 2000, individuals were weighted differently, making the use of weights important (Longino & Bradley, 2003). Using the appropriate weights, we construct the IPUMS flows in the same manner as the census flows—for ages 5–64 and ages 65+—and limit our analyses to individuals moving between the 48 contiguous states.

Migration Measures

Sample size likely becomes more important as the analysis moves from aggregate to disaggregate measures and emphasizes changes over time. Thus, we begin with the interstate migration rate—the percentage of individuals who migrated across states in the 5 years prior to each census.

State-level patterns are compared next. Similar to Flynn and colleagues (1985), we construct measures of the “impact” of migration and each state’s “attractiveness.” For impact, we use the “net” in-migration rate, which equals in-migrants minus out-migrants divided by the state’s elderly population. We list the top 10 “net importers and/or exporters” based on the largest (positive and/or negative) net rate. To measure attractiveness or “volume” of migration, we use the total migrants moving to each destination. This measure disproportionately yields large population states but also reveals how “concentrated” elderly migration is.

Flynn and colleagues (1985) and Longino and Bradley (2003, 2006) calculate the percentage of migrants moving to the top 10 destination states (“receivers”) to measure concentration and find a steady decline over time. We calculate this percentage and propose a richer measure based on the Herfindahl–Hirschman Index (HHI), which is typically used to measure market concentration by summing the squared market share of each firm in a market. It ranges from approximately 0 (a very large number of firms, each with a miniscule share) to 10,000 (one firm with total market share). Our migration HHI squares and sums the percentage of total migrants received by each destination and thus provides information about the entire distribution. The top “senders” and concentration measures among sending states are constructed analogously using state of origin. The minimum value of our migration HHI is limited by the number of possible locations. With 48 contiguous states, an equal distribution of migration flows (i.e., 1/48th or 2.08% of migrants move to each state) yields the minimum value of 48 × (2.08)² = 208. Although the minimum value increases as the number of locations declines, the maximum value of 10,000 is unaffected.

The most disaggregated measure is migration “flows,” and their large number makes the use of summary statistics essential. Correlation coefficients are calculated (a) between census and IPUMS flows in each year and (b) between flows in different years within the same data source. The latter correlations reveal the persistence of flows over time and whether it depends on the data used.

Another measure of persistence is to estimate a flow regression as a function of flow-specific and time period indicators,

(1)

where M_ijt is the number of elderly persons moving from state i to state j in census period t. D_ij is a set of [(48 × 47) − 1] flow-specific dichotomous variables, where Alabama to Arizona is the reference. Three census year dichotomous variables, d_t, are created similarly (1970 is the reference). The R² from this regression measures the proportion of total variation in log-flows that has “not” changed over time (β) other than overall time effects (λ). Using asymptotic results in Stuart and Ord (1991, pp. 1031–1032), we estimate standard errors for each R² and test if the R² across data sources are statistically significantly different. Two variations of equation (1) are estimated as well. Both include the log-origin and log-destination populations to account for changes in population distribution. The second adds the nonelderly migration flow to measure how much of the remaining unexplained variation is explained by changes in nonelderly behavior (e.g., the tendency to move south). All models are estimated for the nonelderly flows for comparison.

RESULTS

Table 1 confirms what others have found that elderly interstate migration is a rare fairly stable event—approximately 4% migrate in a 5-year period (e.g., Longino & Bradley, 2006). Both sources reveal an increase between 1970 and 1980 as in Flynn and colleagues (1985), followed by a slight decline since 1980 (evident in Lin, 1999, among others). The nonelderly population also displays this pattern, so it is not unique to the elderly population. Table 1 also shows that the IPUMS and census yield differences—even at this high level of aggregation and even for the more mobile nonelderly population. The IPUMS rates display more volatility over time than the census, as expected, given its smaller sample size. The IPUMS elderly rate is consistently higher than the census due to both a larger numerator (migrants) and smaller denominator (population). However, this tendency does not extend to the nonelderly population, suggesting it is not a general bias.

Table 1.

Interstate Migration Patterns for Various Age Groups, by Source and Census Year

	Elderly (ages 65+) migration rate (%)		Nonelderly (ages 5–64 years) migration rate (%)
Year	Census	IPUMS	Census	IPUMS
1970	3.66	3.68	9.51	9.20
1980	4.27	4.45	10.80	10.63
1990	4.03	4.35	10.86	10.43
2000	4.07	4.28	9.68	9.67
Average rate	4.01	4.19	10.21	9.98

Figure 1 summarizes the state-level measures described earlier for 1970 and 2000; other years are available on request. Differences between the two sources are emphasized with lines demarking where the rankings diverge and by shading states that appear for only one source. States that exited or entered the top 10 between 1970 and 2000 are in bold italics. The IPUMS and census measures coincide closely, especially in 2000 and for the volume measures. Correlation coefficients between the two sources for each measure and census year are consistently above .97 and are very large (.99+) for volume measures.

Figure 1.

State level elderly migration measures, by data source and year.

Both sources reveal similar patterns over time as well—for example, the ascension of Nevada, Georgia, and the Carolinas and the decline of Florida and California as top destinations, as in Lin (1999). Investigating all four census years reveals that these changes mostly took place between 1970 and 1980 and that most other differences are marginal (e.g., exiting and/or entering states are consistently in the top 20). Both sources also confirm the declining concentration of both destinations and origins, especially during 1980–1990 as in Longino and Bradley (2003, 2006), but our HHI reveals new insights. The concentration ratio suggests similar levels and rates of decline between the concentration of destinations and origins. In contrast, our HHI reveals that destinations are much more concentrated than origins in every year and that the decline in their concentration has been steeper. The IPUMS once again tends to exaggerate the changes over time.

Tables 2–4 reports summary statistics calculated for the state-to-state migration flows. Table 2 shows how the correlation coefficients between the IPUMS and census flows are approximately .99 in each year. Table 3 reports how the correlations over time are also quite large, greater than .9, and as expected decline the greater the time span. The correlation coefficients between the IPUMS and census flows are approximately .99 in each year. The correlations over time are also quite large, greater than .9, and as expected decline the greater the time span.

Table 2.

Summary Statistics for State-to-State Elderly Migration Flows, by Data Source and Year: Correlation coefficients across census years between census and IPUMS

Year	Correlation
1970	.9899
1980	.9983
1990	.9896
2000	.9974

Table 3.

Summary Statistics for State-to-State Elderly Migration Flows, by Data Source and Year: Correlation coefficients between census years, by source

Census					IPUMS
	1970	1980	1990	2000		1970	1980	1990	2000
1980	.9831				1980	.9757
1990	.9245	.9564			1990	.9453	.9718
2000	.9168	.9419	.9853		2000	.9053	.9367	.9777

Table 4.

Summary Statistics for State-to-State Elderly Migration Flows, by Data Source and Year: R² from migration flow regressions, by age group, data source, and model specification.

		Elderly		Nonelderly
		Census	IPUMS	Census	IPUMS
1	Flow and year fixed effects (equation 1)	.9292	.8945	.9691	.9492
2	Add origin and destination populations	.9379	.9025	.9740	.9538
3	Add origin and destination populations plus other age group migration flows	.9408	.9048	.9768	.9683

Note: R² indicates the proportion of variation in migration flows that has not changed over time.

Table 4 reports the R², which further underscore the stability of migration flows. The estimated standard error for each R² is extremely small, ranging from 0.0006 to 0.004, such that the differences in R² across samples are statistically significant from 0, and the 95% confidence intervals are plus or minus less than 0.01.

Census results show that approximately 93% of the total variation in elderly migration flows during 1970–2000 is due to persistent flow patterns. Adding state populations barely increases the R², and our broadest model yields an R² of approximately 0.94, suggesting a positive statistically significant relationship between elderly and nonelderly flows. These results therefore suggest that approximately 7% of the total variation in elderly migration flows represents a real change over time; the remainder is due to persistent flows. Eliminating the year dichotomous variables or the dominant New York–Florida flow barely changes the R² (results not shown).

In contrast, IPUMS results suggest that approximately 10% of the variation in elderly migration can be explained by factors other than time trends and persistent flow patterns. If year dichotomous variables are omitted, that number increases to 17%. The time effects should matter more to the IPUMS, given the different sampling and weighting schemes for each year. Performing these exercises for nonelderly flows yields smaller discrepancies between the census and IPUMS and even greater stability over time.

DISCUSSION

Our analyses confirm what others have found—interstate elderly migration has been remarkably stable since 1970. They lend support to Wolf and Longino (2005) who dispute our “increasingly mobile” society. Our analyses also confirm some subtle but real changes in elderly migration that have been highlighted in previous research that uses only the PUMS (Flynn et al., 1985; Lin, 1999; Longino & Bradley, 2003, 2006), including a declining geographic concentration. Our proposed HHI measure, which accounts for all locations, suggests that the trend is strongest among destinations and in less common locations.

Because our analyses are limited to geographic patterns of migration rather than migrant characteristics and motives, they offer no direct evidence as to why these changes have occurred. These combined subtle changes appear consistent with the increasing average age of the elderly population (defined as age 65+) over time caused by increased longevity and the rapid growth of the population aged 85 years and older. Our analyses, by data necessity, aggregate elderly individuals into those aged 65 years and older. The average elderly person—and potential migrant—is therefore growing older with each new census and more likely to undertake an assistance rather than amenity move, as suggested by Conway and Rork (2009). Interstate migration rates are typically highest among the youngest of the elderly population, and thus, this increased average age of the elderly population suppresses their aggregate mobility, even as near-elderly and young-elderly migration may be growing as in Wolf and Longino (2005) and Conway and Rork. Likewise, a growing motive for assistance likely yields a more diffuse set of origins and—especially—destinations as such migrants pursue something other than a pleasant climate. Congestion in traditional retirement destinations also helps contribute to this trend.

However, using the IPUMS data rather than the full census tends to exaggerate these changes. Although the migration measures we tabulate from the IPUMS are very closely correlated to the census measures at a point in time, the IPUMS has a larger sampling error due to its smaller size that gets magnified when one investigates changes over time. This additional error is especially evident in the flow regressions in which the proportion of total variation not explained by persistent patterns is approximately 50% larger with the IPUMS.

Despite the enormous sample sizes of the IPUMS and the census data, the relative rarity of an interstate move combined with the strong persistence of migration patterns makes sample size an issue, especially for detecting changes over time. The smaller differences found between the nonelderly IPUMS versus census regressions further illuminate this issue, as they are a larger segment of the population and more than twice as likely to move out of state. Our results have implications for future migration research as the ACS becomes the largest sample available for studying interstate migration. The ACS is an even smaller sample, and its migration questions identify only moves in the previous year (instead of the previous 5 years, as in the census) and therefore are measuring a fundamentally different—and rarer still—event. Although 5-year averages of the data can be used that approximate the size of the IPUMS, the migration it is measuring is still rarer and likely more volatile. Our findings therefore caution that discerning credible trends in elderly interstate migration over time will become increasingly difficult in the future.

FUNDING

This work was supported by the National Institutes of Health (NIH; 5R03AG028479-02). Its contents are solely the responsibility of the authors and do not necessarily represent the official views of NIH.