The first test of the migration mortality hypothesis is to examine whether the size of the summer population is correlated with the size of the overwintering population; the hypothesis presumes that it does not. We used the number of monarchs tagged within the summer breeding region as an indicator of late summer population size. The number of monarchs tagged in the Midwest and Northeast portions of the geographic sectors north of 40° latitude north and east of 100° longitude west (Figure 1) are shown in Table 1 for each year from 1998 to 2015. To determine how well the number of tagged individuals explained the variation in overwintering numbers, we subdivided the analyses by Midwest, Northeast and Total area (Midwest plus Northeast). We compared regressions using the Total tagged and Midwest and Northeast tagged alone (Supplementary Table 1). Based on the lowest AIC value, log Total number tagged was the best predictor of overwintering hectares. However, log Midwest tagged was a close second. Because we wished to compare number tagged as a measure of population size to other population size measures that are Midwest-based, we focused on the number tagged in the Midwest. The number tagged in the Midwest alone explained 74% of the variation in overwintering hectares (Supplementary Table 1 and Figure 2). The number tagged in the Northeast was somewhat less strongly correlated with overwintering hectares (Table 2 and Figure 2). The number tagged in the Midwest was correlated with the number tagged in the Northeast (Table 2). The number of individuals tagged in the Midwest and Northeast both declined over the period from 1998 to 2015 as did the overwintering hectares (Table 2).

Two other measures of summer population size, Midwest monarch egg production and Midwest NABA butterfly counts, are shown in Table 1. The number tagged in the Midwest was highly correlated with Midwest late summer egg production and also correlated with midsummer NABA counts for the years from 2006 to 2015 (Table 2). NABA counts and eggs per stem were highly correlated with each other (Table 2). Egg production and NABA counts were also correlated with the overwintering population size (Table 2). Thus, three independent measures of summer population size: number tagged, NABA counts, and eggs per stem, were correlated with each other and correlated with the size of the overwintering population. This result is contrary to the prediction of the migration mortality hypothesis.

Another test of the migration mortality hypothesis is to examine whether migration success, as measured by tag recovery rate, is correlated with overwintering population size and whether it has declined over time. The numbers of tags recovered each year in the Midwest and Northeast and these two regions combined (Total) are shown in Table 3. The recovery rate for the Midwest was used in these analyses because this region has been shown to be the core area for monarch production (Flockhart et al., 2015) and because the number tagged in the Midwest was a good predictor of overwinter population size. For the analysis of recovery rates, the mass mortality years, 2001, 2003, and 2015 were not included for reasons explained in Methods. Midwest recovery rate was not correlated with overwintering population size (Table 2 and Supplementary Table 1).

Although the size of the late summer population explained a large amount of the annual variation in the size of the overwinter population, it is possible that some of the variation in overwinter population size is due to annual variation in migration success. To examine this possibility, we ran a multiple regression using Midwest tagging numbers and Midwest recovery rates (Supplementary Table 1). The equation for the regression of log MW Number Tagged and log MW Recovery Rate on log Overwintering Hectares (OW) was: log OW = −6.3127 + 1.6466 log MW Number Tagged + 0.2461 log MW Recovery Rate (F_{2, 12} = 15.5268, p = 0.0005, R², adjusted = 0.6748). The standardized regression coefficients (src) were 0.8602 for log MW Number Tagged and 0.1772 for log MW Recovery Rate. The square of each standard regression coefficient, indicates the amount of variation explained by each variable. Src² was 0.74 (p = 0.0001) for log MW Number Tagged and 0.03 (p = 0.2739) for log MW Recovery Rate, indicating that log Number Tagged accounted for about 74% of the variation in log OW while log Recovery Rate accounted for about 3%.

There was no decline in the tag recovery rate over the period 1998–2014 for either the Midwest or the Northeast (Table 2 and Figure 3). The recovery rates for the Midwest were correlated with the recovery rates for the Northeast (Table 2).

The recovery rates for the Midwest were correlated with yearly values for NDVI, an index of greenness of the region that provides nectar for migrating butterflies (Table 2). The year 2000, which had the lowest recovery rate, had the lowest NDVI value, and the year 2008, which had the highest recovery rate, had the highest NDVI value (Table 3 and Figure 3).

We also examined whether NDVI itself, could explain any of the annual variation in overwintering hectares (Supplementary Table 1). The standardized regression coefficients (src) were 0.8917 for log MW Number Tagged and 0.1668 for log MW NDVI. Src² was 0.80 (p < 0.0001) for log MW Number Tagged and 0.03 (p = 0.16) for log MW NDVI, indicating that log Number Tagged accounted for about 80% of the variation in log OW while log NDVI accounted for about 3%. While NDVI is correlated with migration success, just like migration success itself, there was no decline in NDVI from 2000 to 2015 (Table 2).

The yearly recovery rates for butterflies from the Northeast were correlated with those from the Midwest (Table 2) but consistently lower (Table 3 and Figure 3). The recovery rate for all the normal mortality years 1998–2014 combined was 0.94% for those tagged in the Midwest and 0.24% for those tagged in the Northeast (Supplementary Table 2). Thus, the average recovery rate for the Midwest was 3.93 times greater than for the Northeast (Supplementary Table 2). For the mass mortality years 2001 and 2003, the ratio of Midwest to Northeast recovery rates were similar to that of all normal mortality years (4.18 and 3.58) but, for the mass mortality year 2015, the recovery rate for the Midwest was not that much greater than for the Northeast (1.19) (Supplementary Table 2). For normal and mass mortality years combined, the recovery rate for the Midwest was 3.65 times that for the Northeast (Supplementary Table 2). Thus, a tagged butterfly from the Midwest was about 4 times more likely to be recovered than a tagged butterfly from the Northeast. Of the butterflies recovered in Mexico from the region above 40° latitude, 89% came from the Midwest with 11% coming from the Northeast. These percentages contrast to those for number tagged with 69% from the Midwest and 31% from the Northeast.

The migration mortality hypothesis presumes that the size of the summer population in the U.S. Midwest is not correlated with the size of the overwintering population in Mexico. We found that the number of monarchs tagged each season in the Midwest was highly correlated with the size of the overwintering population as was the number tagged in the Northeast, although not as highly (Table 2). It could be argued that the number of individuals tagged is more a reflection of the size of the migratory population rather than the summer population. As such, number tagged may already have incorporated some degree of migratory failure and thus would be more likely to be correlated with overwinter hectares. We chose to look at number tagged above 40° latitude and east of 100° longitude because this is the prime summer breeding region for monarchs (Flockhart et al., 2017). Individuals tagged in this region are closer to their natal origin than those tagged below 40° latitude. Butterflies tagged above 40° latitude have traveled southwest for at most two 5 × 5 latitude/longitude sectors (Figure 1) before capture and have 4–6 5 × 5 sectors remaining in their journey. Although it is certainly the case that some percentage of individuals leaving from a particular natal origin point will not make it to a tagging point above 40° latitude, the majority of migration mortality has yet to occur. So, the number of butterflies available to be tagged above 40° latitude is similar to the summer population size.

The underlying premise of the migration mortality hypothesis is that there is a lack of correlation between summer and winter population sizes. Studies showing a lack of correlation used summer population estimates based on data sets including the annual butterfly counts by NABA and butterfly counts made in Ohio and in Illinois (Ries et al., 2015a; Inamine et al., 2016; Saunders et al., 2016). However, there are methodological problems with these summer butterfly counts that make them inaccurate measures of population size (Pleasants et al., 2016, 2017). Sampling is limited in geographic scope and time and focuses on the early rather than the late summer population. More significantly, no counts were made in corn and soybean fields in the late 1990s and early 2000s when monarchs and milkweeds were still present in those fields. Surveys made during that period therefore underestimated the actual size of the monarch population. Survey data from after that period do show a correlation with overwintering population size (Crewe et al., 2019; Saunders et al., 2019) as does egg production for the last two decades that incorporates information on milkweed abundance (Pleasants and Oberhauser, 2013; Pleasants et al., 2017). Because migrating butterflies come from all habitats, including corn and soybean fields, the sampling bias seen in butterfly counts in the late 1990s and early 2000s is not an issue.

In comparing eggs per stem and NABA counts with overwintering population size, we were restricted to the years since 2006 to remove the sampling bias for not surveying in agricultural fields. But in comparing number tagged with overwintering population size we can use data for years before 2006 because butterflies during migration have come from all possible habitats, including agricultural fields. For the years 2006–2015, eggs per stem and NABA counts of summer population size were highly correlated with Midwest number tagged counts (Table 3). All three of these measures of summer population size were correlated with overwintering hectares (Table 2). These results are counter to the foundational assertion of the migration mortality hypothesis that the size of the summer breeding population does not predict the size of the overwintering population (Ries et al., 2015a; Inamine et al., 2016; Agrawal and Inamine, 2018). The relationship between summer population numbers and overwintering hectares suggests that factors affecting monarch breeding population development from March through October, rather than migration success, primarily determine overwintering population size. While physical factors, such as weather and spring and summer temperatures are important determinants of population growth each year (Saunders et al., 2016, 2018, 2019; Crewe et al., 2019), there is also strong evidence for the role of milkweed availability (Pleasants and Oberhauser, 2013; Flockhart et al., 2015; Pleasants, 2017; Pleasants et al., 2017).

Tagging data arguably can provide a better picture of the size of the migrating population than counts made at peninsula points where monarchs stop over during migration. Counts made at Cape May, New Jersey and Peninsula Point, Michigan were not correlated with the size of the overwintering population (Davis, 2012; Badgett and Davis, 2015), but Crewe and McCracken (2015) did find a positive correlation between counts and overwintering numbers for Long Point in Ontario. Possible reasons for the lack of correlation between stopover points and overwintering numbers include the issue of double counting when counts are made more than once a day or when monarchs remain at stopover points for multiple days due to weather. Other reasons are discussed in Pleasants et al. (2016).

While we used numbers of butterflies tagged as an indicator of population size, we recognize that it is not a perfect representation of the summer population size. The tagging effort is influenced by weather events that limit tagging and the times when people are available to tag. In addition, there are few people tagging west of 95° longitude an area known to produce a substantial number of monarchs most years (Supplementary Figure 1). High populations are likely to be underestimated by tagging as well since taggers often run out of tags. It is also the case that many taggers only tag during the beginning of the migration, leaving the tail of the migration underrepresented in the tagging records. There was a 27% decline in the number of tags distributed from 1998 to 2015 (Table 1) (R² = 0.33, p = 0.013). This decline may represent a general loss of interest in tagging, or reduced expectations because of low population size that discouraged tagging. The proportion of tags that are affixed to butterflies reared in captivity has increased (Supplementary Figure 2) possibly reducing the effort to tag wild butterflies. In spite of these limitations, the fact that this index of population size is highly correlated with other measures of summer population size and overwinter population size lends it credibility.

There were 3 years where the number of monarchs tagged in the Midwest was outside the 95% confidence interval in the regression of number tagged vs. overwintering hectares (Figure 2). In the year 2000 Midwest tagging overestimated the size of the overwintering population (Figure 2). The recovery rate for 2000 was the lowest seen (Table 3 and Figure 3) and corresponded to a year when NDVI values were the lowest (Table 3). Thus, in this year, drier than normal conditions in the southern region may have significantly affected migration success and reduced the number of individuals arriving in Mexico relative to what was expected from tagging numbers. The number tagged in the year 2014 also overestimated the size of the overwintering population (Figure 2). Recovery rate and NDVI were not unusual for that year (Table 3). Migration was later than normal in 2014 (Journey North⁵). Late migrations are associated with low recovery rates (Taylor et al., 2019). It is also possible that there was an increased effort in tagging following the report in 2013 of the lowest overwintering population size ever recorded. In 1998, the number tagged underestimated the size of the overwintering population (Figure 2). The recovery rate was not particularly high that year so higher than normal migration success is probably not the explanation. The number of tags applied was the fifth lowest in the record (Table 1), suggesting that tagging simply underrepresented the size of the migratory population.

In the regression of number tagged in the Northeast vs. overwintering hectares, the year 2012 was an outlier (Figure 2). The number of butterflies tagged in the Northeast in 2012 greatly overestimated the actual overwintering population size. There was an early spring in 2012 (Ault et al., 2013) and monarchs colonized the Northeast earlier than normal (Journey North⁵). The resulting population in the Northeast, as represented by the monarchs per hour count at Cape May, New Jersey, was the third highest in the 27 years of that program (Cape May Monarch Monitoring Project⁶). On the other hand, extremely high temperatures in the Midwest, with drought conditions in much of the area (Taylor et al., unpublished data), reduced the 2012 migration in the Midwest and the number tagged in the Midwest to the second lowest total in the record (Table 1).

The number tagged in the Midwest from 1998 to 2015 was correlated with the number tagged in the Northeast with an r value of 0.64 (Table 2). The Midwest and Northeast populations are partially but not completely independent. There is some evidence that the Northeast is colonized by monarchs that were born in the first generation in the Midwest (Miller et al., 2011). But there is also evidence that some of the monarchs colonizing the Northeast have come directly from the Texas generation (Journey North, 2019).

Migration success, as measured by recovery rate, was not correlated with overwintering population size (Table 2). Nor did recovery rates decline from 1998 to 2014 (Table 2, Figure 3) as would be expected if migration success was driving the decline in the overwintering population. That said, migration success may play a small role in determining the annual variation overwintering population size. After most of the variation in overwintering population size was accounted for by summer population size, migration success accounted for an additional 3% of the variation. In addition to migration success, recovery rate is a function of overwinter mortality and detection probability. We have no data on the annual variation in those two factors. High levels of variation in those two factors could swamp out changes in migration success. However, the migration mortality hypothesis predicts a substantial decline in migration success which should be detectable despite variation in the other two factors. The fact that there was a significant relationship between recovery rate and NDVI, a factor suspected of affecting migration success, also indicates that variation in these other two factors are insufficient to undermine the use of recovery rate as a measure of migration success.

Although recovery rates did not decline from 1998 to 2014, they did vary over those years (Figure 3). This variation may be due to a variety of causes, some associated with the conditions during the migration, such as parasite load (Bartel et al., 2011), road kills (Kantola et al., 2019) and nectar availability. However, Saunders et al. (2019) did not find a significant correlation between Ophryocystis elektroscirrha (OE) parasite load and overwintering population size from 2004 to 2015. They also did not find a decline over time in parasite load, suggesting that this factor was not directly involved in the monarch population decline. They did find that the level of greenness along the southern fall migration corridor was correlated with the size of the overwintering population size, after accounting for the role of summer population size. However, they did not find that greenness had decreased over time.

We found that the greenness index, NDVI, was positively correlated with migration success. The lowest NDVI value and the lowest recovery rate was for the year 2000 (Table 3 and Figure 3) and was associated with drought in Texas (drought.gov⁷). The highest recovery rate observed was in 2008 which had the highest NDVI value (Table 3 and Figure 3). The year 2011, when there was also a drought in Texas (Brower et al., 2015), had the second lowest NDVI and the third lowest recovery rate (Table 3 and Figure 3). The fact that recovery rate is correlated with NDVI, a proxy for nectar availability, which we know to be important for monarch migration success (Brower et al., 2015), lends credibility to the use of recovery rate as a measure of migration success.

In addition to normal overwinter mortality caused by predators, broken wings, exposure or weather conditions, death may also result from inadequate lipid stores obtained during migration (Brower et al., 2015) or neogregarine parasite load that butterflies bring with them (Bartel et al., 2011; Altizer et al., 2015). We considered death due to lipid shortage as add-on mortality. We have to consider the possibility that years with poor migration success may also have increased add-on mortality if butterflies arrive in Mexico with reduced lipid supplies. The recovery rate is based on the product of migration success and overwinter mortality, including add-on mortality. What happens if low migration success is always coupled with high add-on mortality? In a poor migration year, lower migration success will have the effect of decreasing the recovery rate, but add-on mortality will have the effect of increasing the recovery rate. Because of these opposing effects on recovery rate, the question is whether we can compare recovery rates for different years and interpret those differences as differences in migration success. Lipid analysis by Brower et al. (2015) can provide some insight into this question. In 2011, which was a drought year in Texas, they found lower lipid levels in butterflies in Texas than in non-drought years. Surprisingly however, they found that November lipid levels of butterflies at the overwintering sites in Mexico were not lower than in non-drought years. The difference in lipid levels in Texas and Mexico could be due to the acquisition of nectar and lipids as the Texas monarchs continued to the overwintering sites through Northeastern Mexico or selective mortality that eliminated monarchs with low lipid levels. While Brower et al. (2015) did find individuals with low lipid levels in Mexico, their data suggest that the frequency of such individuals, and the amount of add-on mortality that might accrue from it, was no different between drought year and normal years. Thus, add-on mortality and overall migration success may not be interrelated. If this is the case, recovery rates are primarily a measure of migration success.

Overall, the recovery rate for monarchs tagged in the Midwest from 1998 to 2014 was about four times greater than for monarchs tagged in the Northeast. Thus, monarchs tagged in the Midwest have a higher probability of making it to Mexico than monarchs tagged in the Northeast. The greater distances from Northeast locations to Mexico may account for lower recovery rates from that region due to longer flight times and greater mortality. In addition, monarchs from the Northeast using the Eastern flyway appear to take two paths, one along the east coast from Maine to at least South Carolina and an interior path along the Appalachians. Monarchs then move west along the Gulf coast from Florida to Brownsville, Texas (Journey North⁵). Brindza et al. (2008) found that butterflies tagged in more inland areas of the Northeast were more likely to be recovered than those tagged in coastal areas. Some east coast migrants may also end up in Florida (Vander Zanden et al., 2018). The Central Flyway, used by the butterflies from the Midwest, provides a clear southwest track to the Texas border with Mexico. Differences in recovery rate between the Central Flyway and the Eastern Flyway may also be the result of differences in wind patterns and temperature. Geographic differences in recovery rate on a finer scale are being investigated.

The significant correlation between the recovery rates for the Midwest and Northeast (Table 2) suggests that, although migration success for the Northeast is generally lower than the Midwest, there is an overarching factor that affects both regions similarly. Both Midwest and Northeast monarchs must traverse similar terrain in southern Texas and northeastern Mexico although they may take different routes to get to that point. The correlation between migration success and NDVI suggests that nectar availability in this region affects migration success of all butterflies, no matter their point of origin. In addition, there may be weather related timing delays in migration from 1 year to the next that can affect migration success for both regions.

The recoveries from the 3 years (2001, 2003, and 2015), with mass mortality due to winter storms, were not included in the analysis because the majority of these deaths were decidedly storm related rather than caused by attrition factors occurring in a normal year. Because recovery rate is based on dead butterflies, to use recovery rate as an index of migration success requires the assumption that normal overwinter mortality is a random variable over time. Mass mortality years violate that assumption and are not be included in the temporal analysis. Yet, recoveries in mass mortality years can tell us; (1) that large numbers of tagged monarchs were still alive at the time of these events and (2) that years with normal mortality and mass mortality both provide a random sample of the population, only differing in sample size, as evidenced by the similarity in the ratio of Midwest to Northeast recovery rates between normal and mass mortality years (Supplementary Table 2).