Appendix S1 How I used the result from published studies When capture-recapture and capture-recovery models were fit to data using the frequentist (likelihood maximization) approach, authors reported year-specific estimates of the model parameters. In the following I describe how I extracted the correlation between n and h from these year-specific estimates. The appendix is based on populations in which a birth pulse is followed by a period of anthropogenic mortality. It is formulated using sport hunting as an example, but results are valid for other forms of anthropogenic mortality (fishing, collision, etc.). This appendix does not provide details about the building of likelihood functions. 1) Reparameterization of Seber’s capture-recovery model Seber’s capture-recovery model {Brownie, 1985 #40; Williams, 2002 #73}, denoted CR in Table 1 of the main text, applies to datasets in which recoveries occur only from individuals dead from the hunting. The parameters of the Seber model are the annual survival probability 𝑆 = 1 − 𝑛 − ℎ and Seber’s recovery probability 𝑟 = ℎ ℎ+𝑛 𝜆, where 𝜆 is the probability that a dead marked individual is reported as such, given that it died from anthropogenic cause. An alternative formulation is Brownie’s recovery probability 𝑓 = ℎ𝜆. Brownie’s parameterization renders the computation of 𝑐𝑜𝑟𝑟(𝑛, ℎ) easier, but Seber’s parameterization makes the initial capture-recovery model design and fitting more straightforward, especially since it provides a better framework for examining the effects of covariates on S independently of the recovery probability REF (Gauthier and Lebreton 2008). To study compensation, one wants to extract the correlation between n the annual natural mortality and h the annual hunting mortality. To do that, one needs to know 𝜆 the band reporting rate. The expected value and temporal variance of this parameter were obtained from external studies (reward bands, radio-transmitters; see Appendix S3). The reparameterization is then performed via the delta-method which provides the following system of five equations, where E, cov and Var stand for temporal expected value, temporal covariance, and temporal variance: Eq. S1 𝑐𝑜𝑣(𝑆, 𝑟) ≃ − 𝐸(𝜆) 2 (𝑐𝑜𝑣(𝑛, ℎ)(𝐸(𝑛) − 𝐸(ℎ)) + 𝑉𝑎𝑟(ℎ)𝐸(𝑛) − 𝑉𝑎𝑟(𝑛)𝐸(ℎ)) (𝐸(ℎ) + 𝐸(𝑛)) 𝑉𝑎𝑟(𝑆) = 𝑉𝑎𝑟(𝑛) + 𝑉𝑎𝑟(ℎ) + 2𝑐𝑜𝑣(𝑛, ℎ) 2 𝐸(ℎ) 𝐸(𝜆)2 2 2 𝑉𝑎𝑟(𝑟) ≃ ( ) 𝑉𝑎𝑟(𝜆) + 4 (𝐸(𝑛) 𝑉𝑎𝑟(ℎ) − 2𝑐𝑜𝑣(𝑛, ℎ)𝐸(ℎ)𝐸(𝑛) + 𝐸(ℎ) 𝑉𝑎𝑟(𝑛)) 𝐸(ℎ) + 𝐸(𝑛) (𝐸(ℎ) + 𝐸(𝑛)) 𝐸(𝑆) = 1 − 𝐸(𝑛) − 𝐸(ℎ) 𝐸(ℎ) 𝐸(𝑟) ≃ 𝐸(𝜆) { 𝐸(ℎ) + 𝐸(𝑛) From this system of equation it follows that (MAPLE output not simplified or factorized): Eq. S2 Var(λ)Var(r)E(r)E(λ) − E(λ)3 cov(S, r) + Var(λ)E(λ)cov(S, r)E(S) − 2E(λ)2 E(r)cov(S, r)E(S) − 2E(r)Var(λ)cov(S, r)E(S) 2 𝑐𝑜𝑣(𝑛, ℎ) = −Var(S)E(λ) + Var(λ)E(r)2 + Var(λ)E(S)2 E(r)2 − 2Var(λ)E(S)E(r)2 − Var(λ)Var(r)E(r)2 − Var(r)E(λ)2E(r) + Var(r)E(λ)3 E(r) + E(λ)3 cov(S, r)E(S)⁄E(λ)3 (E(λ)3 + V −Var(λ)E(λ)cov(S, r) + 2E(λ)2 E(r)cov(S, r) + 2E(r)Var(λ)cov(S, r) + 2Var(S)E(S)E(λ)2 − Var(S)E(S)2 E(λ)2 2 2E(λ)2 E(r)cov(S, r)E(S) + 2E(r)Var(λ)cov(S, r)E(S) + Var(S)E(λ)2 − Var(λ)E(r)2 − Var(λ)E(S)2 E(r)2 + 2Var(λ)E(S)E(r)2 𝑉𝑎𝑟(ℎ) = +Var(λ)Var(r)E(r)2 + Var(r)E(λ)2 E(r)2 − 2E(λ)2 E(r)cov(S, r) ⁄E(λ)2 (E(λ)2 + Var(λ)) 2 2 2 −2E(r)Var(λ)cov(S, r) − 2Var(S)E(S)E(λ) + Var(S)E(S) E(λ) ) −2Var(λ)Var(r)E(r)E(λ) + 2E(λ)3 cov(S, r) + Var(r)E(λ)4 − 2Var(λ)E(λ)cov(S, r)E(S) + 2E(λ)2 E(r)cov(S, r)E(S) + 2E(r)Var(λ)cov(S, r)E(S) 𝑉𝑎𝑟(𝑛) = ⁄E(λ)2 (E(λ)2 + V +2Var(λ)E(S)E(r)2 + Var(S)E(λ)2 − Var(λ)E(r)2 − Var(λ)E(S)2 E(r)2 − 2Var(r)E(λ)3 E(r) + Var(λ)Var(r)E(r)2 + Var(λ)Var(r)E(λ)2 2 2 3 2 2 2 2 +Var(r)E(λ) E(r) − 2E(r)Var(λ)cov(S, r) − 2E(λ) cov(S, r)E(S) + 2Var(λ)E(λ)cov(S, r) − 2E(λ) E(r)cov(S, r) − 2Var(S)E(S)E(λ) + Var(S)E(S) E(λ) 𝐸(ℎ) = −E(r)(−1 + E(S))/E(λ) { 𝐸(𝑛) = 1 − 𝐸(𝑆) + E(r)(−1 + E(S))/E(λ) 2) Capture-recovery model with independent measure of anthropogenic mortality rate This model (denoted CR-H in Table 1 of the main text) applies to datasets similar to those in part 1, but for which in addition a measure of hunting pressure is available each year. It is a variant of the CR model. The measure of hunting pressure, denoted H, can be the ratio between total number of individuals harvested and total population size {Gauthier, 2001 #231}, or an independent assessment of the rate at which individuals are killed per unit of hunting effort multiplied by a measure of that hunting effort {Baird, 2000 #241}. This method also applies to known-fate data (radio tracked individuals), in which case H is the proportion of individuals known to have died from hunting (e.g., {Creel, 2010 #233}. Then, the relationship between hunting pressure and overall survival probability S can be implemented as a built-in regression, which is generally assumed linear {Gauthier, 2001 #231; Barker, 1991 #1019; Creel, 2010 #233}: Eq. S1 𝑆 = (1 − 𝑛0 )(1 − 𝑏 ∙ 𝐻 ) where n0 (one minus the intercept, mortality in the absence of hunting) and b (slope of the regression) are parameters to be estimated. b is a direct measure of the level of compensation. Complete additivity corresponds to b = 1 and complete compensation corresponds to b = 0. The compensation-additivity rate can be shown to be 𝐶 = 1 − (1 − 𝑛0 ) ∙ 𝑏. The associated correlation coefficient 𝑐𝑜𝑟𝑟(𝑆, 𝐻) can then be used to compute 𝑐𝑜𝑟𝑟(𝑛, ℎ) using: Eq. S2 𝑐𝑜𝑟𝑟(𝑛, ℎ) ≃ −𝑐𝑜𝑟𝑟(𝑆, 𝐻)√𝑉𝑎𝑟(𝑆)𝑉𝑎𝑟(𝐻) − 𝑉𝑎𝑟(𝐻) √𝑉𝑎𝑟(𝑆) + 𝑉𝑎𝑟(𝐻) + 2 ∙ 𝑐𝑜𝑟𝑟(𝑆, 𝐻)√𝑉𝑎𝑟(𝑆)𝑉𝑎𝑟(𝐻)√𝑉𝑎𝑟(𝐻) As highlighted by {Lebreton, 2005 #74}, if H is an imprecise measure, this can lead to a downwards bias in the absolute value of b (i.e., bias towards compensation). If H is assumed to be very imprecise, one may want to use hunting effort only instead of H, e.g., the number and activity level of hunters {Rolland, 2010 #272}, or even a proxy for hunting effort, e.g., the maximum number of individuals that hunters are legally allowed to harvest (bag limit; Péron et al. submitted). Then, the result of the regression still informs the presence of compensation or additivity, but generally cannot be used to obtain 𝑐𝑜𝑟𝑟(𝑛, ℎ). 3) Cause-specific capture-recovery model This multistate model {Schaub, 2004 #364; Servanty, 2010 #104} applies to datasets where the recoveries from hunting are supplemented by either live reencounters (observations of live individuals) or recoveries from other mortality causes (e.g., road-kills, power-line electrocutions, sick and injured individuals). The latter data often stem from reports by the general public. For clarity they are called retrieval data hereafter. In the first case (reencounter data), the model is noted CRP in Table 1 of the main text. In the second case (retrieval data) the model is noted CRR in Table 1 of the main text. Parameters n, h, and λ are as in previous cases the annual natural mortality, hunting mortality, and reporting rate of hunter-killed individuals. The difference is that n and h are directly estimable (not confounded with λ), and thus the correlation of n and h is directly implemented. New sets of parameters are p the probability of detecting live individuals, and 𝜆𝑛 , which is the probability for a marked individual that died from causes other than hunting to be reported as such. Note that identifiability or estimation issues may arise when using this type of model {Catchpole, 2001 #253}. The version of this model in which all parameters are time-constant is not identifiable. {Schaub, 2004 #364} report that models where at least h and 𝜆𝑛 (or conversely n and λ) are time-dependent are identifiable, but warn against the risk of flawed results since the non-identifiable time-independent model is nested within the identifiable time-dependent model. 4) Capture-recapture model with independent measure of hunting pressure This model (denoted CP-H in Table 1 of the main text) applies to datasets in which only live reencounters exist, but each year an independent measure of hunting pressure is available. This type of dataset is more typical of by-catch exploitation {Rolland, 2010 #263; Veran, 2007 #247} than sport hunting, since it implies that harvested individuals are never reported as such (λ = 0). The modeling framework is the standard capture-recapture model {Lebreton, 1992 #26}, in which the estimated survival corresponds to “apparent” survival, i.e., the product of true survival and site-fidelity (individuals may be alive but have exited the monitored area). The usual notation for apparent survival is φ. The other parameter set is reencounter probability p. If one assume no interaction between anthropogenic mortality and site-fidelity, the regression is the same as in part 2 above (replacing notation S by φ).