Demographic analysis of dormancy and survival in the terrestrial orchid Cypripedium reginae



    Corresponding author
    1. Patuxent Wildlife Research Center, US Geological Survey, 11510 American Holly Drive, Laurel, MD 20708, USA, and Swiss Ornithological Institute, 6204 Sempach, Switzerland, and
    • Present address and correspondence: Marc Kéry, CEFE/CNRS, 1919 Route de Mende, 34293 Montpellier Cedex 05, France (tel. +33 04 67 27 76 92; e-mail

    Search for more papers by this author

    1. Patuxent Wildlife Research Center, US Geological Survey, 11510 American Holly Drive, Laurel, MD 20708, USA, and Swiss Ornithological Institute, 6204 Sempach, Switzerland, and
    2. West Virginia Wesleyan College, 59 College Avenue, Buckhannon, WV 26201, USA
    Search for more papers by this author


  • 1We use capture-recapture models to estimate the fraction of dormant ramets, survival and state transition rates, and to identify factors affecting these rates, for the terrestrial orchid Cypripedium reginae. We studied two populations in West Virginia, USA, for 11 years and investigated relationships between grazing and demography. Abe Run's population was small, with moderate herbivory by deer and relatively constant population size. The population at Big Draft was of medium size, with heavy deer grazing, and a sharply declining number of flowering plants up to the spring before our study started, when the population was fenced.
  • 2We observed dormant episodes lasting from 1 to 4 years. At Abe Run and Big Draft, 32.5% and 7.4% of ramets, respectively, were dormant at least once during the study period for an average of 1.6 and 1.3 years, respectively. We estimated the annual fraction of ramets in the dormant state at 12.3% (95% CI 9.5–15.8%) at Abe Run and at 1.8% (95% CI 1.2–2.6%) at Big Draft. Transition rates between the dormant, vegetative and flowering life-states did not vary between years in either population. Most surviving ramets remained in the same state from one year to the next. Survival rates were constant at Abe Run (0.96, 95% CI 0.93–0.97), but varied between years at Big Draft (0.89–0.99, mean 0.95).
  • 3At Big Draft, we found neither a temporal trend in survival after cessation of grazing, nor relationships between survival and the number of spring frost days or cumulative precipitation during the current or the previous 12 months. However, analysis of precipitation on a 3-month basis revealed a positive relationship between survival and precipitation during the spring (March–May) of the previous year.
  • 4Relationship between climate and the population dynamics of orchids may have to be studied with a fine temporal resolution, and considering possible time lags. Capture-recapture modelling provides a comprehensive and flexible framework for demographic analysis of plants with dormancy.


Demographic analysis is essential for understanding variation in plant numbers in space and time. The sessile nature of adult plants makes demographic analysis of their behaviour easier than for most animals, which often escape from observation (Harper 1977). However, many perennial plants employ dormancy as a form of vertical escape, and can stay below-ground for one or several growing periods. Seed dormancy and winter bud dormancy are familiar, well-studied phenomena that enable many plants to germinate and sprout in the season most conducive to growth and development. In a similar fashion, perennials employ dormancy when they overwinter as underground corms, rhizomes or tubers, which typically send up shoots in the spring after dormancy is broken by various environmental cues. Little is known about why these perennating organs sometimes do not sprout but continue in the dormant state for another year or more. This type of extended dormancy is widespread among plants (Lesica & Steele 1994) and has been identified as a major problem for plant population models (Menges 2000). It is especially common in terrestrial orchids (Wells & Willems 1991) where dormant episodes lasting up to 12 years have been reported (Tamm 1972); however, in most cases they only seem to last for one to a few years. The biological significance of this extended dormancy is unclear, although it has been linked to avoidance of environmental stress (Lesica & Steele 1994). Frequency of dormancy has been related to climatic conditions in some studies (Shefferson et al. 2001) but not in others (Wells 1967; Hutchings 1987b). In these studies, drawing sound conclusions is challenged by the fact that without excavation, it is impossible to determine whether a plant not recorded above-ground is dormant or in fact dead.

Most previous demographic studies on orchids have made strong assumptions about the duration of dormant episodes and the timing of the eventual death of plants not seen during later years of a study (e.g. Hutchings 1987a,b; Gregg 1991; Wells et al. 1998; Primack & Stacy 1998). As suggested elsewhere, these methods may produce unsatisfactory and biased estimates of demographic parameters (M. Kéry, K.B. Gregg, M. Schaub, unpublished report). In contrast, capture-recapture models do not rely on such strong assumptions (Nichols 1992). These models have been developed by animal population ecologists for the frequent case when not every individual studied is always detectable. They were first used to estimate the size of populations and later also for estimation of survival rates and other demographic quantities. (See Williams et al. 2002 for a thorough overview of many of these models that may also be of interest to plant ecologists.) The situation in orchids is not really that different from that in animals such as deer, in that a marked individual that is not seen during a survey may be dead or may just have been missed (because of dormancy in the case of the orchid). Capture-recapture models have been used increasingly by plant ecologists (Naylor 1972; Alexander et al. 1997; Slade et al. 2003; Kéry 2004). They are particularly well suited for the study of plants with dormancy, and enable us to estimate survival and state transition rates, to test factors that affect these rates, and to compute the fraction of the population in the dormant state (Shefferson et al. 2001, 2003).

We apply capture-recapture methods in an analysis of ramet demography in two contrasting populations of the showy lady's slipper Cypripedium reginae, in order to estimate rates of survival and transitions between life-states. One population was small and relatively stable, whereas the other was medium sized and recovering from a sharp decline caused by herbivory by white-tailed deer Odocoileus virginianus (Gregg 2004). We investigate whether this difference is reflected in rates of survival and of state transitions. We were also interested in determining whether cessation of deer grazing caused any directed change in demographic rates, and what other factors might affect these rates. Previous studies on orchids have indicated an influence of climate on numbers flowering and on the fraction in the dormant state (e.g. Shefferson et al. 2001). In an analysis of another orchid, Cleistes bifaria, we found effects of climate on demographic rates that were delayed by 1 year (M. Kéry et al., unpublished report). In our analysis of the larger of the two populations of C. reginae, we therefore focus on climatic factors during both the current and the previous year.


study species

The perennial orchid Cypripedium reginae Walter is widely distributed throughout the north-eastern United States but may be experiencing a decline in numbers, especially in its southernmost locations (Rooney 1985). Obligate outcrossing (Stoutamire 1967) is carried out by syrphid flies and flower beetles (Vogt 1990) and by two species of megachilid leaf cutter bees (Guignard 1887). Typically, one to two flowers are borne on ramets (= stems) that may reach almost 1 m in height. The ramets arise from underground rhizomes that produce one primary shoot bud each year for continued growth in the subsequent spring, as well as one or two auxiliary buds that may replace the main shoot bud, should it be damaged, or, in the case of mature, robust plants, may produce additional aerial shoots (Waterman 1949). Over time, asexual reproduction may produce clones with multiple shoots. Kennedy (2003) found that stem clusters in C. reginae are composed of more than one genotype in 26% of cases but that, with one exception, all stems sampled from the two populations used in this study had identical genotypes for all 10 of the allozyme loci resolved. Thus, there was very little genetic variation in either population.

study sites and populations

We studied two forest populations in West Virginia, USA. The Abe Run population (994 m a.s.l.) is located over Greenbrier limestone in Tucker County in an open forest and shrub wetland with speckled alder/willow thickets (Tom DeMeo, US Forest Service, personal communication). The area is fed by groundwater seepage from surrounding limestone slopes. The soil is usually water saturated, with pH ranging from 6.2 to 6.8. The orchids are in clumps scattered throughout a triangular area of 66 × 54 × 78 m. Irregular monitoring between site discovery in 1979 and the early 1990s reported from 0 to 15 flowers and occasional browsing by deer (West Virginia Heritage Program, unpublished data, 1992). Moderate herbivory occurred from 1992 to 1994, when 9–46% of ramets were grazed. Ramet numbers were fairly stable from 1997 to 2001, with a mean of about 50 ramets present each year (Gregg 2004). Deer exclusion wire mesh fencing 1.22-m tall was placed over two orchid clumps in 1993 and over another in 1995.

The Big Draft population (785 m a.s.l.) is located in Greenbrier County, in an opening approximately 43 m long by 20 m at its widest point. The site is a white oak–white pine forest (Tom DeMeo, personal communication) over thin-bedded limestones (Linda Tracy, US Forest Service, personal communication). Soil pH ranges from 5.8 to 6.4. The orchids grow in several clusters at the side of a creek. In the early 1980s, 500–950 flowers were counted on over 600 ramets (Eye 1975; US Forest Service, unpublished correspondence, 1985, 1987). From 1986 to 1988, white-tailed deer (Odocoileus virginianus) grazed large portions of 65–95% of the ramets (Gregg 2004). Recovery of pre-herbivory stem height, flower production level and leaf area took from 9 to 12 years; however, even after 12 years the number of stems was only 29% of the pre-herbivory population size. From 1992 to 1999, on average about 170 ramets were counted annually above ground. Although the entire opening was enclosed with 2.44-m tall, open mesh wire fencing in 1989, the population produced no flowers in either 1989 or 1990 (Gregg 2004).

field methods

We studied Abe Run from 1993 to 2003 and Big Draft from 1989 to 1999. Scoring the life-state of a genet with ramets in more than one state is arbitrary; hence, we chose the ramet as demographic unit in our study. Each year about 3–4 weeks after flowering, we censused the entire population, marked new ramets, checked previously marked ramets, and recorded their height and life-state (vegetative, flowering, or not present above-ground). Seedlings (ramets = less than or equal to 7 cm tall with narrow leaves) were excluded from this study because they were observed at Abe Run only in 2002. Ramets were marked using plastic tags and by mapping their position.

Detectability of marked above-ground ramets is the product of the probability to detect a tag and the probability to detect an above-ground ramet, given that its tag has been detected. As we believe that every marked above-ground ramet was indeed detected when its tag had been detected, we estimated detectability as 1 minus the proportion of undetected tags. At Big Draft, only 5 of 2053 (0.2%) previously marked ramets or their tags were not detected. Hence, average detectability was 0.998. When Abe Run was censused by KBG from 1998 to 2003, none of 451 previous tags went undetected. For the years before that, we do not have the data to estimate detectability, but believe that it must be near-perfect as well: censuses were done very carefully, the ground is open around the orchids, orchids are large and grow in a carefully mapped, small area that is easy to census. Hence, we assume near-perfect detectability of above-ground ramets at both sites, and use methods that make the assumption that detectability equals 1. Note that our estimation methods do not make any assumptions about unmarked plants, the detectability of which may well be < 1. All one needs to assume here is that virtually all marked plants are detected.

As is typical of many lady's slippers, ramets often grew out of interwoven rhizome mats at both study sites, so it was not possible to distinguish different genets. Ascertaining genet identity for each ramet would require a combination of excavation and genetic analysis of every stem, neither of which was possible. Instead, we determined visually whether ramets were isolated or belonged to a ramet cluster capable of sharing resources by clonal or mycorrhizal connections. Cluster boundaries were usually obvious. When in doubt, a ramet was scored as isolated if > 24 cm away from any other ramets. We could not determine whether clusters contained ramets of more than one genet but it was deemed possible (Kennedy 2003).

For each marked ramet, census results can be written concisely as a detection history, i.e. a string of zeroes and letters, that indicates in what year and state a marked ramet was recorded (letters) and in what year it was not recorded above-ground (zeroes). Consider the following hypothetical detection history from an 11-year study, 00VVVF0V000. This ramet was first observed and marked as a vegetative (V) ramet in year 3 of the study, came up again as vegetative the following two years, after which it flowered (F) in year six. It was not seen the following year, reappeared as vegetative in year eight but was not seen in year nine or thereafter. The presence in a detection history of internal zeroes (year 7) and terminal zeroes (years 9–11) is typical for demographic studies of plants with dormant states. Their interpretation is a challenge for demographic estimation. In general, it is not clear whether a zero represents a dead or a dormant ramet. In our study, however, detectability of a ramet was c. 1 when it was present above ground; hence, an internal zero in the detection history must represent a ramet in the dormant (D) state. For terminal zeroes, however, a ramet could be either dead or dormant. Dormant ramets are in an unobservable state unless they are excavated, which we did not do for conservation reasons.

probabilistic models for detection histories

Capture-recapture models estimate demographic parameters, such as survival and state-transition rates, when not all ramets are detected with certainty. The models thus have parameters for detectability to yield unbiased estimates of demographic rates. We used single- and multistate capture-recapture models to properly deal with uncertainty about the fate of a ramet with terminal zeroes in its detection history.

In single-state models, all observable, above-ground life-states are collapsed so the detection history reduces to a string of 1 (ramet recorded alive above ground) and 0 (ramet not recorded above ground) values. Using the Cormack-Jolly-Seber (CJS) model (Lebreton et al. 1992), the detection history can be written as a function of parameters for survival, φ, and detection p, given survival. If, as in our study, the detectability of marked above-ground plants equals one, the complement of the estimated detection rate, 1 − p, is the fraction of the population that is dormant (Shefferson et al. 2001). Reduced-parameter versions of the model assume that parameters may be the same across years for either survival or detectability. Important assumptions of this model family are that for each year, probability of survival and of detection is the same for all ramets. Ramets are further assumed to be independent from each other.

We tested these assumptions by bootstrapping the deviance of the most general CJS model {φyear, pyear}, with year-specific parameters for survival and detection, 1000 times in each population. For both populations, the model did not fit very well (P ≤ 0.001). We assumed that this lack of fit was due to overdispersion, which is frequent in count data. Overdispersion means that there is a larger variance in the data than expected from a multinomial distribution. This may be due to non-independence of ramets that belong to the same genet. We calculated an overdispersion factor c by dividing the deviance of the model for the data at hand by the mean deviance from the 1000 bootstrap replicates. This resulted in estimates of c of 1.37 at Abe Run and 1.76 at Big Draft. Both indicate fairly small amounts of overdispersion (Burnham & Anderson 1998). By correcting for overdispersion in this way we state that there is more uncertainty about the model parameters than in the multinomial model. This makes SEs wider and tends to make the best model chosen by AICc (see below) less complex.

Multistate (MS) models are an extension of the CJS model to two or more life-states between which transitions may occur (Arnason 1973; Brownie et al. 1993). In addition to parameters for survival (φ) and detection (p), they have parameters (ψmn) for rates of transition from state m to state n. Transition rates are defined conditional on survival, so for each state m they sum to 1 over all possible states n.

Theoretically, for a plant with three life-states D, V and F, the conditional probability of the detection history in our example above can be written as:

image(eqn 1)

where inline image is the survival rate of a ramet in state m at period i (from year i to year i + 1), inline image is the transition rate of a ramet in state m in (given its survival from i to i + 1), and inline image is detection probability for a ramet in state m and year i. D, V and F denote the dormant, vegetative and flowering states. inline image is an algebraic shorthand for the probability that a ramet is not detected after year 8 in any state (as denoted by the dot). It can be expressed as a recursive function of time- and state-specific probabilities of survival, transition and detection. See Lebreton et al. (1992) for the χ term in the CJS model.

The term in the first square brackets in equation 1 reflects uncertainty associated with an internal zero in the detection history in year 4. In our study, pV = pF= 1 and pD = 0, hence, there is no uncertainty about the state of the ramet for internal zeroes. Hence, the term simplifies to become inline image. The term in the second square brackets in equation 1 reflects uncertainty associated with the terminal zeroes during years 8–10. A ramet may die between year 7 and 8 with probability inline image or it may survive, move to another state, remain undetected and not be seen again after year 8. However, as above, this term again simplifies to become inline image. By a similar reasoning, inline image can be written recursively as a function of the other parameters. This expression can then again be simplified to yield the following for the entire expression in the last square bracket in equation 1:

inline image

For both the CJS and multistate models, a product-multinomial likelihood can be written and, based on the number of ramets with a certain detection history, maximum likelihood parameter estimates found numerically by programs such as MARK (White & Burnham 1999) or M-SURGE (Choquet et al. 2003). When fitting MS models, we defined three states (D, V and F) and set detectability at 0 for the unobservable, dormant state and at 1 for each observable state (vegetative and flowering). In contrast to the CJS model, no goodness-of-fit test for MS models is available in program MARK. However, for both data sets lack of fit of the CJS model to the data was small. The less restrictive MS model is likely to reduce part of the lack of fit by allowing for state-dependent transition rates. We therefore assumed that our most general MS model fitted the data sufficiently well to allow meaningful interpretation of model selection statistics and standard errors.

The most general MS model possible for our situation, {φstate×year, ψstate×year}, would have fully state- and time-specific parameters for both survival and state-transitions. Note that the model contains no parameters to be estimated for the known and fixed detectability. This model is overparameterized and not identifiable without adding constraints on some parameters (Kendall & Nichols 2002; M. Kéry et al., unpublished report). This means that some or all of its parameters may not be estimable (Catchpole & Morgan 1997). In particular, it is impossible to know whether survival of the dormant state varies over time, and thus in the model, survival of the dormant state needs to be set equal to that of another state (such as inline image), if time-specific estimates of survival are desired. However, even when all parameters in a model may be identifiable in principle, they need not be so in practice due to sparse data. The most general model for C. reginae for which we were able to obtain numerical solutions in MARK was φyear, ψstate×year. This model had the same year-specific survival rates for all life-states, and state- and year-specific transition rates.

modelling of survival and transition rates

To test for variation in vital rates between years, we set parameters constant across years and then compared this model with another model where year-dependent vital rates were assumed. We used the small-sample corrected version of Akaike's information criterion (AICc) to rank models. AICc is a model selection criterion that expresses the bias/variance trade-off for simple vs. complex models. The model with the lowest value of AICc represents a best compromise between lack of fit and complexity, both of which are undesirable. Convention is that fit of two models is approximately equal when the difference in AICc between them is less than 2 units. We also computed model weights w that quantify the support for each model relative to all others in the set of models that were thought a priori to be useful and compared by AICc. For the CJS model, where a goodness-of-fit test is available in program MARK, we based all model selection on the overdispersion-adjusted criterion QAICc.

To test for effects of climate on survival rates of C. reginae, we constrained time-specific parameters to be a logit-linear function of these covariates. (We did not conduct a covariate analysis for state transitions, as there was no evidence of time-variation in them; see Table 1.) Based on earlier work (Shefferson et al. 2001), we chose to examine possible effects on survival rates between two flowering periods in the interval {i, i + 1} of cumulative precipitation in the current year (June–May during {i, i + 1}) and in the previous year (June–May during {i − 1, i}), and mean spring temperature and number of freezing days in the current spring (both from March to May during {i, i + 1}). In addition, we also split up the current and previous years into 3-month intervals and checked for effects of cumulative precipitation during these shorter time-windows. Climatic data were obtained from the National Oceanic and Atmospheric Administration (NOAA), Asheville, North Carolina, USA, for White Sulphur Springs, WV, approximately 8 km from the Big Draft site. These analyses could only be conducted for Big Draft because our sample size for Abe Run was too small. As data for Big Draft comprised only 11 years and climatic covariates were correlated with each other, it was not sensible to fit complex models and we fitted models with single covariates only.

Table 1. Analysis of time-variation in survival and state-transition rates in two populations of Cypripedium reginae in West Virginia, USA. Model selection is shown for a multistate model with detection probability fixed at 0 and 1 for dormant and above-ground plants, respectively. The most general model is φyear, ψstate×year, with a set of year-specific survival rates (φ) and a set of separate, year-specific transition rates (ψ) for each state. Reduced-parameter models are obtained by setting one or both parameter sets equal across time. Notation: No. parameters = number of estimable parameters; Deviance = model deviance, AICc = small-sample AIC, ΔAIC = difference of a model's AICc and the AICc of the model with the lowest AICc, w = Akaike weights
ModelNo. parametersDevianceAICcΔAIC w
(a) Abe Run
φconst, ψstate 7684.7 995.5 0.000.90
φconst, ψstate×year59567.7 999.9 4.390.10
φyear, ψstate16678.91008.813.210.00
φyear, ψstate×year67561.31015.119.520.00
(b) Big Draft
φyear, ψstate16837.92157.2 0.000.95
φconst, ψstate 7862.02163.0 5.850.05
φyear, ψstate×year66766.42191.334.100.00
φconst, ψstate×year57789.92195.338.110.00

To express the amount of variation in survival (φ) explained by a climatic covariate, we expressed the reduction in deviance D achieved by fitting a covariate X as a percentage of the maximally achievable reduction in deviance by fitting one parameter for every year for survival rates: (Dconst} − DX})/(Dconst} −Dyear}). We conducted capture-recapture analyses with program MARK (White & Burnham 1999) and all other analyses with GenStat (Payne et al. 1993).


occurrence of ramet dormancy

At Abe Run, 98 ramets of C. reginae were marked from 1993 to 2003. Twenty-one appeared for the first time in the last 2 years of the study, so dormancy could not be observed for them. Of the remaining 77 ramets, 25 (32.5%) went dormant (i.e. had internal zeroes in their capture history) at least once during the study period. Four ramets had more than one episode of dormancy. Dormant episodes (n = 30) lasted 1 year (60%), 2 years (23%), 3 years (10%) and 4 years (7%). At Big Draft, 258 ramets were marked from 1989 to 1999. Thirty appeared for the first time in the last 2 years of the study. Of the remaining 228 ramets, 19 (7.4%) exhibited one dormancy episode each during the study period. Dormant episodes (n = 19) lasted 1 year (79%), 2 years (16%) and 3 years (5%). The proportion of ramets that went dormant at some time during our study period was higher at Abe Run than at Big Draft (inline image = 27.16, P < 0.0010). However, mean duration of dormant episodes was not significantly different (Abe Run, 1.6 years; Big Draft, 1.3 years; Mann–Whitney U-test, z = 1.21, P = 0.23), nor was the distribution of the duration of dormant episodes different (log-linear model, test of duration–population interaction, inline image = 3.25, P = 0.36).

We used the Cormack-Jolly-Seber (CJS) model to estimate the proportion of ramets in the dormant state each year. For both populations, the most parsimonious model was {φconst, pconst}. It yielded estimates for the average annual fraction dormant of 12.3% (95% CI 9.5–15.8%) at Abe Run and of 1.8% (95% CI 1.2–2.6%) at Big Draft.

relationships between dormancy, ramet clustering and ramet height

We tested whether being an isolated ramet or a member of a cluster of ramets might affect the tendency to undergo dormancy. When combining the data from both populations, we found no association between dormancy and isolation (loglinear model; test of the dormancy × isolation interaction, inline image = 0.03, P = 0.85). We also tested whether undergoing extended dormancy might affect clustered and isolated ramets differently. Ramet heights before and after a dormancy episode were known for seven isolated ramets and for 33 ramets belonging to clusters. For isolated ramets, mean heights were significantly shorter after the dormancy episode (20.6 ± SE 5.2 cm) than before (32.8 ± SE 7.0 cm; paired t-test, P = 0.04). By contrast, mean heights of clustered ramets were not significantly different before (43.1 ± SE 3.1 cm) and after a period of dormancy (42.3 ± SE 3.2 cm; paired t-test, P = 0.70). For the period when both populations were studied (1993–99), ramets were taller at Abe Run (51.8 cm, SE 1.13) than at Big Draft (41.9 cm, SE 0.62; F1,1309 = 81.16, P < 0.001).

estimates of survival and transition rates

The most parsimonious multistate models had constant, state-specific transition rates for both populations with constant survival at Abe Run (Table 1a) and year-specific survival at Big Draft (Table 1b). Other models had little support (w ≤ 0.10). Based on the two most parsimonious models, annual survival rate of C. reginae was estimated at 0.96 (95% CI 0.93–0.97) at Abe Run and varied from 0.89 to 0.99 (mean 0.95) at Big Draft.

Ramets most frequently remained in the same state from one year to the next except for dormant ramets at Big Draft, which most often moved to the vegetative state (Table 2). There were marked differences between Abe Run and Big Draft. A higher percentage of dormant ramets remained dormant at Abe Run while a larger proportion of flowering ramets remained flowering at Big Draft. There was a tendency for more exchanges between the dormant and vegetative state, and between the vegetative and flowering state, than between the dormant and flowering state.

Table 2. State-transition rates in two populations of Cypripedium reginae in West Virginia, USA. The table shows estimates from the most parsimonious multistate model in Table 1const, ψstate). Transition rates from state m to nmn) sum to 1 over all possible states n for each state m. Hence, estimates of ψmn for m = n were obtained by subtraction and SEs by the delta method (Mood et al. 1974)
Transition m to nmn)Abe RunBig Draft
Dormant to dormant0.50 (0.092)0.25 (0.106)
Dormant to vegetative0.41 (0.083)0.59 (0.106)
Dormant to flowering0.08 (0.037)0.16 (0.073)
Vegetative to dormant0.13 (0.024)0.02 (0.004)
Vegetative to vegetative0.68 (0.030)0.83 (0.012)
Vegetative to flowering0.20 (0.025)0.15 (0.011)
Flowering to dormant0.06 (0.018)0.01 (0.006)
Flowering to vegetative0.30 (0.034)0.18 (0.018)
Flowering to flowering0.64 (0.036)0.81 (0.019)

factors affecting survival at big draft

At Big Draft, grazing by deer was stopped in early 1989, just before our demographic study began. We expected that survival rates might increase during our study. However, we found no such trend. A model representing this hypothesis had no support from the data and only explained 3% of the total temporal variation in survival (cf. linear time-trend model in Table 3). Analysis of precipitation on a 3-month basis suggested a positive relationship between survival and precipitation during the spring of the previous year (March–May i − 1). More rainfall during the previous spring increased the chances of surviving to the following year and explained 43% of the variance in survival rates of C. reginae at Big Draft (Table 3, Fig. 1). Thus, we found a climatic effect with a 1-year lag on a component of the population dynamics of this orchid.

Table 3. Covariate analysis of survival rates of Cypripedium reginae at Big Draft with annual survival constrained to be a logit-linear function of each covariate X in the model φX, ψstate. Numbers of parameters are 16, 8 and 7, for the model with unstructured time-variation, the covariate models, and the constant model, respectively. Notation is as in Table 1, with the addition of Slope of estimate, which shows the direction of the effect. NA = not applicable
ModelSlope of estimateDevianceAICcΔAIC w Deviance explained (%)
Unstructured time variation (φyear, ψstate)NA837.92157.2 2.580.12100
Time (linear trend)+861.32164.3 9.710.00  3
Temperature (Mar–May i)855.42158.5 3.860.06 27
Frost (Mar–May i)+859.92162.9 8.310.01  9
Precipitation previous 12 months861.82164.910.280.00  1
Precipitation current 12 months859.92162.9 8.320.01  9
Precipitation (Jun–Aug i−1)854.22157.2 2.600.12 33
Precipitation (Sep–Nov i−1)856.52159.6 4.980.04 23
Precipitation (Dec–Feb i−1)+858.42161.4 6.790.01 15
Precipitation (Mar–May i−1)+851.62154.6 0.000.45 43
Precipitation (Jun–Aug i)+857.52160.6 5.960.02 19
Precipitation (Sep–Nov i)862.02165.010.450.00  0
Precipitation (Dec–Feb i)855.32158.3 3.720.07 28
Precipitation (Mar–May i)855.22158.3 3.670.07 28
Constant survival (φconst, ψstate)NA862.02163.0 8.430.01  0
Figure 1.

Relationship between precipitation during March to May in the previous year and survival of Cypripedium reginae at Big Draft, WV. The y-axis shows annual survival rate in the interval (i, i + 1) estimated from a multistate capture-recapture model (φyear, ψstate). The x-axis shows precipitation (in mm) during Marchi−1 to May i−1. Precipitation pertains to part of the interval between the previous (i − 1) and the current year (i) and survival rate to the interval between the current (i) and the following year (i + 1).


fraction dormant and rates of survival and state-transition

Using capture-recapture methods for demographic analysis of the showy lady's slipper, Cypripedium reginae, we estimated that 2% and 12% of the ramets in two contrasting populations were dormant every year and that ramets had high survival rates (0.95–0.96) at both sites, respectively. A diversity of methods and possible problems with previous approaches to demographic estimation in plants with dormancy (M. Kéry et al., unpublished report) make comparisons with many other studied species difficult (e.g. Hutchings 1987a,b; Hutchings 1989; Gregg 1991; Wells & Cox 1991; Waite & Farrell 1998; Willems & Melser 1998).

Compared with the two previous studies that have used capture-recapture in species of smaller stature, C. reginae had a lower fraction dormant, and higher survival. In genets of Cypripedium calceolus ssp. parviflorum, average survival was 0.88 and the average fraction dormant was 32% (Shefferson et al. 2001). In ramets of Cleistes bifaria, average survival rate was 0.86 and average fraction dormant was 29% (M. Kéry et al., unpublished report). Note that in our studies of Cleistes and Cypripedium, genet survival rates would be higher and genet dormancy lower than corresponding values for ramets, because when a genet consists of several ramets, it is less likely that all of them will die or become dormant at the same time.

Above-ground states in C. reginae most often remained the same from one year to the next, while in the closely related C. calceolus, vegetative genets went dormant most frequently, and only flowering genets most often remained in that same state from one year to the next (Shefferson et al. 2003). It seems typical of many orchid species that once plants have reached the flowering state, they are highly likely to continue flowering and less likely to become dormant in future years (Wells 1967; Mehrhoff 1989; Gregg 1991; Falb & Leopold 1993; Shefferson et al. 2003). We assume that high proportions of ramets remaining in above-ground states are evidence of healthy populations, as in less vigorous or declining populations a higher proportion of above-ground plants entered prolonged dormancy (Mehrhoff 1989; Gregg 1991). Even though the fraction dormant was about six times higher in C. reginae at Abe Run than at Big Draft, the relatively low values at both sites are probably indicative of favourable conditions. Other capture-recapture studies have also found variation over space and time in the fraction of dormant genets or ramets (C. calceolus ssp. parviflorum, Shefferson et al. 2001; Cleistes bifaria, M. Kéry et al., unpublished report) as have studies using conventional methods (e.g. Brzosko 2002).

possible relationships between dormancy, size, clustering and survival

In Cypripedium calceolus, Brzosko (2002) reported much more dormancy of smaller genets (i.e. those with a mean of 1.0–2.1 ramets genet−1), than of larger genets (mean of 2.5–12.4 ramets genet−1). However, Brzosko (2002) reported on genet dormancy, while we investigated dormancy of individual ramets. Isolated ramets of C. reginae, which could be likened to small plants, did not undergo dormancy more often than ramets in clusters. Plants of Cypripedium reginae are relatively large compared with those of Cleistes bifaria or Cypripedium calceolus ssp. parviflorum. Whether there is a general tendency for species with larger size to have less dormancy and higher survival should be tested by studying additional species. There are several ways to measure plant size, e.g. ramet height, number of ramets genet−1, and size differences associated with the various life-states. Different methods may be combined to help our understanding of the relationship between size and dormancy.

For isolated ramets, reappearance following dormancy was associated with lower stem height than before dormancy. This may reflect a cost due to lack of photosynthesis and entail a reduction of stored reserves and of subsequent flowering probability. A similar effect was found in Solidago when intact genets exhibited more growth than those of which rhizome connections had been severed (Schmid et al. 1988). Without connections to other ramets, isolated ramets may have access to too few resources to survive a year with no photosynthesis without displaying a clear cost. This may make a dormancy episode more risky for isolated than for connected ramets, leading to the prediction that isolated ramets might undergo dormancy less often than those belonging to clusters. However, this hypothesis was not supported, perhaps because of the small number of isolated ramets in the study. We had also hypothesized steadily increasing ramet survival rates following elimination of grazing at Big Draft in early 1989, but no such increase occurred. One possible explanation is that individuals weakened the most by the severe herbivory had already died in the year or two prior to the flowering season when our study began.

factors affecting population dynamics in orchids

Many factors may combine to determine numerical changes in plant populations. Climate is an important driver of plant population dynamics and climatic effects have often been suggested for orchids (e.g. Hutchings 1987a; Gregg 1991; Wells et al. 1998; Brzosko 2002). Two previous studies using capture-recapture methods to analyse orchid population dynamics (Shefferson et al. 2001; M. Kéry et al., unpublished report) found correlations between climate and both survival and state transition rates. Our analysis of Cypripedium reginae supports these findings, although, in contrast to Cleistes bifaria, there was no evidence of temporal variation in transition rates. However, we also note that the testing for climatic effects is a complex issue, as there are so many different ways to quantify climate and even more ways to present climatic variables, e.g. by taking averages or totals over different time windows.

Density dependence might be another important factor for the population dynamics of plants. In clonal plants, there is an additional level of complexity, as neighbours may belong to the same genet. Competition between related ramets may differ from competition between unrelated neighbours, especially when ramets are physiologically integrated (Hartnett & Bazzaz 1983). It would be interesting to study density and kinship effects on demographic rates of orchids, especially with respect to dormancy.

methodological considerations

All models make assumptions. Even though ad hoc, model-free methods may, in fact, make even more, hidden, assumptions, it is still necessary that model assumptions should be met as fully as possible. Our capture-recapture approach makes three assumptions that may not be perfectly met: independence of fates, equality of rates, and perfect detectability of the above-ground states.

A typical challenge in plant demography is the question of what constitutes an individual. In many clonal plants, genets are much harder to identify than are ramets, so many studies take ramets as their units. Sharing of genet identity between ramets will introduce some non-independence of their fates, which violates the model assumption, as well as any other statistical analysis of demographic data. Such non-independence may be an important source of overdispersion. It will not bias estimates, but will make estimated standard errors too small and support models that contain too many parameters that appear significant. Overdispersion can be estimated and model selection adjusted for, as we have done. Therefore, when applying capture-recapture models to plant demography where the ramet is the unit, we should perhaps not be too concerned about a small degree of lack of fit. This might reflect overdispersion rather than a structural inadequacy of the model, at least as long as ĉ ≤ 3 (Lebreton et al. 1992). For both populations in our study, the estimated overdispersion factor was well below that.

The assumption of identity of rates is likely to be violated somewhat. At least at Big Draft, survival rates were different for the three life-states and this may introduce bias into the estimation of the fraction dormant in the CJS model. It may be possible to express the fraction dormant as a function of the parameters of a multistate model (Fujiwara & Caswell 2002) and this might yield better parameter estimates.

We also assume that all ramets are always detected when above ground. In our study, detection probability of above-ground states was indeed very close to 1 but in most studies this may not be the case. Shefferson et al. (2001) show that the fraction dormant will be overestimated if this assumption is untrue. The assumption can be tested, e.g. by a double-observer or similar kind of repeated survey (Kéry & Gregg 2003), and demographic estimates can be adjusted accordingly.

Capture-recapture models in general, and multistate models in particular, are data hungry. Sample size may therefore limit modelling possibilities, as can be seen in our Abe Run population; no covariate modelling was possible for the multistate model. In such cases, some insight might be gained by covariate modelling in the CJS model, where both survival (φ) and 1 minus the fraction dormant (p) can be made a function of covariates. Similarly, the most appropriate way to test for any relationship between clustering of ramets and the transition rates into the dormant state would be to explicitly model it for two groups of plants, those that belong to a cluster and those that do not. However, we were restricted to a more indirect approach for want of a larger sample size.

Compared with other orchid populations, dormancy was fairly uncommon in our Big Draft population. This means that terminal zeroes in a ramet history at Big Draft commonly mean the ramet had in fact died. (Note that capture-recapture models do not permit the determination of which ramet dies and when.) However, for two reasons we still prefer a capture-recapture approach to estimation in this situation. The first is that one does not know the prevalence of dormancy in a population without adopting proper estimation methods. Using capture-recapture methods enables estimates of demographic parameters under less stringent assumptions than do previous methods (M. Kéry et al., unpublished report). The second is that capture-recapture models provide a unified framework for estimation and testing even in the absence of dormancy, and useful programs such as MARK or lately M-SURGE (Choquet et al. 2003) are readily available.

It is not possible to estimate independent, time-dependent survival rates for the dormant state. Dormant plants are in an unobservable state, in the same way that a marked sea turtle is absent and therefore unobservable at its breeding beach during non-breeding years (Fujiwara & Caswell 2002; Kendall & Nichols 2002). With plants though, one has the advantage that unobservable individuals are not entirely out of reach. Great gains in insight could be achieved if part of the plants could be excavated in some of the years when they are not recorded above ground. This would provide the information necessary to make a multistate model fully identifiable for the case of dormancy. This should be exploited in future demographic studies of plants with dormant states.


We thank Michael Hutchings, Lindsay Haddon, Michael Schaub, Richard Shefferson and two anonymous referees for comments, Marge Enders and Wallace Kornack for censusing Abe Run during 1993–2000, and Carl Colson, Jeff Gregg, Jim Miller and others for field assistance. We thank G. C. White for providing researchers with the MARK program. MK wants to thank Bernhard Schmid for continuing support. MK was funded by grant 81ZH-64044 from the Swiss National Science Foundation. KBG acknowledges financial support from the United States Forest Service, USDA, and protective management of the Big Draft site by the Forest Service and of the Abe Run site by the West Virginia Division of Natural Resources, Canaan Valley State Park.