A flexible approach to age dependence in organizational mortality. Comparing the life duration for cooperative and non-cooperative enterprises using a Bayesian Generalized Additive Discrete Time Survival Model

This paper proposes a new estimation model to capture the complex effect of age on organization survival. Testing various theoretical propositions on organizational mortality, we study the survival of French agricultural cooperatives in comparison with other firms with which they compete. The relationship between age and mortality in organizations is analyzed using a Bayesian Generalized discrete-time semi-parametric hazard model with correlated random effects, incorporating unobserved heterogeneity and isolating the various effects of time. This analysis emphasizes the specificity of the temporal dynamics of cooperatives in relation to their special role in agriculture.


Introduction
The survival of firms is a major subject of study in economics, industrial organization or organizational ecology, with a large number of works, using various approaches, addressing this issue (Geroski, 1995;Geroski et al., 2010;Hannan, 1998Hannan, , 2005Jovanovic, 1982Jovanovic, , 2001Martimort, 1999;Mata and Portugal, 1994;Simons and Ingram, 2004). The literature on organizational mortality has identified four different effects of age on firm mortality (Hannan, 1998;Le Mens et al., 2011). The first effect, the liability of newness (P1), is a negative effect of age on mortality. According to standard industrial organization approach, the market acts as a selection for initially non-performing organizations that could not later adapt themselves (Geroski, 1995). For organizational ecology, the new roles and functions of these new organizations come at some cost; the firm faces constraints on capital which may limit its development capacity; it is more fragile due to the non-stability of inner social interactions (the problem of organizational learning); and, finally, it has not gained full legitimacy, lacking stable relationships with customers. The second effect of age, the liability of adolescence (P2), emphasizes the existence of an increased mortality rate after the first years (Brüderl and Schusseler, 1990). Organizations have some "immunity" endowments (Hannan, 1998), that enable them to live for a certain period after their creation. Rational actors will put an end to an organization only if they have had enough information on its negative performance. 1 During the first years, the founders are more determined to make an effort to save the organization, the probability of death taking the form of an inverted U. P2 suggests an increase in mortality and a negative effect of age on the mortality rate. The third effect of age, the liability of obsolescence (P3), is a positive effect of age on the mortality rate due to environmental change. The notion of structural inertia is used to explain this effect: founders leave their mark on the organization. This makes its transformation costly and increases mortality (Hannan et al., 2006). Since the distance between the initial and the current environment increases with the age of the organization (Barron et al., 1994), the mortality rate should increase with each successive time period. Age is not a causal factor, but an indicator of the gap between the current environment and the environment at the time of foundation. Hannan (1998) points out, however, that initial endowments and resources of companies protect them from bankrupt in a particular environmental context, but can also potentially protect them when this environment changes. They may lead to the reduction or even the elimination of this effect in relation to others. The final effect of age, the liability of senescence (P4), emphasizes a positive effect of age on mortality due to the effects of bureaucratization. This effect is typical in industrial organization (Martimort, 1999). An older organization is characterized by greater 1 See the equation developed by Schary (1990) to explain firm exit in a sector. difficulty in controlling the activity of its employees or members, who may relax their innovation efforts. In organizational ecology, bureaucratization leads to greater organizational inertia. The old organizations function by tending to rely on traditional solutions which, in time, limit effective collective action (Barron et al., 1994).
The result of the combination of these four effects is relatively unknown (Coad, 2017). Hannan et al. (2007) propose an integration of these different effects to emphasize a positive impact of age on mortality in organizational ecology (domination of P3 and P4 over P1 and P2), and a negative effect, followed by a positive effect, in industrial organization (domination of P1 and P4 over P2 and P3). Different structures within the sectors may appear (Hannan, 2005): if there is a dominance of P1 and P2, the first entrants dominate the sector for long periods; in the case of dominance of P3 and P4, we see waves of Schumpeterian creative destruction. As pointed out by Le Mens et al. (2011), mixed empirical results highlight, for some, the increased mortality of young organizations, for some, the fact that this mortality occurs after a certain time, and, for others, that older organizations have the greatest chance of dying. The study of agricultural cooperatives is particularly relevant to test these effects because the agri-food industry and wholesale trading are characterized by a multiplicity of organizational structures: cooperatives and non-cooperatives, small and large companies, etc. (Boone and Ozcan, 2014;Maietta and Sena, 2008;Ingram, 2003, 2004;Sykuta and Cook, 2001).
Cooperatives are of particular interest to economists because of their unique ownership structure and the incentives associated with this structure (Fulton and Giannakas, 2013). The presence of cooperatives in agriculture is also a the subject of intense debate. Are these structures relics of the past, surviving due to the ideologies that generated them in the 19th century (social Catholicism, Republicanism, etc.), 2 but destined to disappear and to be replaced by more effective organizations, more suited to deal with market pressure (Cross et al., 2009)? Or, conversely, are they a modern structure, consistent with the new demands of a globalized agriculture, capable of organizational innovation ? The distinctive features of cooperatives as members-owned businesses (Birchall, 2013) can give them special survival comparative advantages. With regards to the literature Ozcan, 2014, 2016;Burdin, 2014;Frenken, 2014;Kitts, 2009;Monteiro and Stewart, 2015;Risch et al., 2014;Rousselière and Joly, 2011), an important empirical contribution is the identification of the various effects of time, a distinctive feature of the cooperative vs. the non-cooperative enterprise.
Our main contributions to the overall work involving the study of the various effects of time on the mortality of organizations are twofold. First, at the empirical level, we rely on confidential and nearly-exhaustive data of the Annual Business Surveys 1984-2006, provided by the French Ministry of Agriculture. 3 To our knowledge, this is one of the first analyses of the French agri-cultural cooperative sector, one of the most developed and innovative in the world (Bijman and Iliopoulos, 2014;Charriere and Aumond, 2016). Contrary to previous studies using the same database (Carrere et al., 2011;Rousselière and Joly, 2011), thanks to our time-varying model, we underline the lower importance of unobserved heterogeneity on the survival of cooperatives, suggesting that unobserved variables may have a smaller impact. Second, we propose an original contribution to panel data econometrics and survival models in particular (Singer and Willett, 2003), which allows us to identify the factors determining the survival of cooperatives, and to distinguish the latter from other organizational structures. To isolate these effects, we estimate an original Bayesian generalized additive discrete-time semi-parametric hazard model with correlated random effects, incorporating unobserved heterogeneity. This is the first attempt in the literature on organizational mortality to estimate a smooth effect of time on survival. This model belongs to the more general family of GAMM (Generalized Additive Mixed Models). As discussed in the conclusion, this framework can be easily extended to take into account other survival features.
The remainder of this paper is structured as follows. Section 2. provides a review of the existing literature, emphasizing the specificity of cooperatives with respect to four potentially different effects of time on mortality. In section 3., we present our database and justify the focus on the cohort of enterprises created after 1984. We develop eventually our Bayesian econometric framework for estimating the form of the survival function in section 4.. In section 5., we present and discuss the results which underline the specificity of the temporal dynamics of cooperatives. Our main conclusions are summarized in section 6..

Earlier works on the impact of time on the survival of cooperatives
The dynamics of survival of cooperatives is a combination of the four effects described above. According to P1, their ability to mobilize non-market resources (e.g. free labor from members in case of difficulty, or indivisible reserves), to rely on their members' commitment (Cechin et al., 2013;James and Sykuta, 2006;Nunez-Nickel and Moyano-Fuentes, 2004), or the fact that their founders follow other goals than strictly economic ones, suggest a low mortality rate in the early years (Pérotin, 2004), especially because of the lower entry rate. Indeed, it may be harder for a cooperative to enter the market because founders may have more difficulties to access resources or bank loans and, therefore, face harsher credit constraints (Simons and Ingram, 2004;Chaddad et al., 2005). Because the liability of newness may be related to the small size at the entry, Boone and Ozcan (2016) show that cooperatives outlive corporations if investments size at founding is large.
From a different perspective, agricultural cooperatives appear to be victims of the liability of include data from more recent surveys in our longitudinal analysis.
adolescence (Hansmann, 1999). This corresponds to the end of the original ideology of "end of honeymoon", described by Pérotin (2004). Cooperatives can, indeed, run the first few years at a lower cost, by incorporating adjustment mechanisms in the payroll in the case of workers' cooperatives, in exchange for contributions of raw materials in the case of agricultural cooperatives). Similar to non-profit organizations, cooperatives benefit from the strong involvement and enthusiasm of their founders. Pérotin (2004) highlights that the effect of adolescence is much more pronounced for cooperatives than for other companies: the initial selection is more stringent and it is easier to temporarily adjust costs in order to increase their chances of survival. Due to their enthusiasm, cooperative's members do not close the cooperative upon the first observation of poor performance. For example, in the simple case of a supply agricultural cooperative, Fulton (2001) underlines that in case of low commitment and high price, members may exit the cooperative or force the board to a merger with another cooperative or an Investor-Owned Firm (IOF) (Banerjee et al., 2001). Cazzufi and Moradi (2012) show that due to the positive effect of membership size on survival, exit can be related to a decrease in commitment.
Regarding the effect P3, studies show that cooperatives can be victims of changing institutional contexts. According to Nunez-Nickel and Moyano-Fuentes (2004) and Ingram (2003, 2004), cooperatives are more sensitive to (un)favorable changes in the regulatory environment, but have a greater ability to adapt to macroeconomic fluctuations. Staber (1992) pointed out that agricultural marketing cooperatives are highly resistant to recessions. In France, cooperatives may be sensitive to changes in public policies that affect them: the 1992 reform of the cooperative status, the 2003 reform of the common agricultural policy favoring producer organizations, the reform of milk quotas (Hovelaque et al., 2009), the impact of international trade treaties, etc. Cooperatives have a special role in agricultural industries, compared to other organizational forms, and have been supported by public policies (Valentinov, 2007;Iliopoulos, 2013). The favorable policy treatment, that cooperatives have enjoyed, is partially due to the belief that cooperatives have a pro-competitive effect (Fulton and Giannakas, 2013). Iliopoulos (2013) extends this argument and summarizes four main reasons for this: 1. countervailing market power: agricultural cooperatives provide individual farmers with an institutional mechanism that increases their bargaining power against upstream and downstream partners in food supply chains, and corrects for excess supply induced prices; 2. market failure: agricultural cooperatives address various forms of market failures, acting as a competitive yardstick that improves market performance; 3. linking supply and offer: agricultural cooperatives improve the coordination of supply with demand for farm commodities, leading to prices more consistent with production costs; 4. community development: agricultural cooperatives are instrumental for achieving community development goals and for facilitating the integration of low-income producers into community life.
Previous researches show, for example, a positive impact of operational funds on the performance of French cooperative in the fruits and vegetable sector (Camanzi et al., 2011). Therefore, we may underscore a relative low long term mortality for the whole population of cooperatives, in comparison to non cooperative enterprises.
Finally, according to Holström (1999), on the long run, agricultural cooperatives appear to be victims of P4. The accumulation of procedures or the lack of leadership could lead to pronounced effects of organizational inertia, especially in large cooperatives Fulton (2001). The results of Kitts (2009) show the same trend for community-based organizations. However, the control methods used in cooperatives in relation to the needs of their members appear a priori to be more flexible forms of organization. They have, as shown in large cooperative groups, the ability to find original solutions to the risks of bureaucratization, e.g. Italian cooperative consortia, the variable-geometry unions in agricultural cooperatives, the Mondragon industrial complex, etc. Hansmann, 1996).
The methodology that we use here aims to identify not only the general relationship between age and mortality, but also to identify the presence or absence of each of the four potential effects listed above. This is an important difference with respect to previous studies, that use less flexible methods, or evaluate only the effect of legal status on mortality (Burdin, 2014;Monteiro and Stewart, 2015;Frenken, 2014). Particular attention is given to the specificity of cooperatives in seeking a unique temporal dynamic, and to the specific characteristics and contextual elements affecting them.

Presentation of the data
Provided by the French Ministry of Agriculture for the years 184-2006, the Annual Business Survey provides information on the economic and financial activity of French enterprises. This survey constitutes a unique longitudinal database on firms. Its pertinence to our study lies in the fact that it is a mandatory and comprehensive survey, which, unlike non-mandatory surveys, does not suffer from non-random non-response and general errors related to sampling. From the data at our disposal, we can reconstruct for each unit a time of presence in the database. Depending on the circumstances, these times of presence in the database can be associated with life cycle units. All units are subject to right-censored survival time, in the sense that we have a partial observation of their survival for a time during the observation period. Their survival time is at least equal to their presence time in the population at risk. This problem is addressed routinely by survival models. Enterprises for which the date of birth is unknown raise another problem. They are also subject to left truncation, because they are not identified prior to the first available date in the investigation. Here, information on these units cannot be associated with age. As a result, and following Singer and Willett (2003) and Rabe-Hesketh and Skrondal (2012), these units are excluded from the analysis.
The legal status of companies are grouped into four broad categories: (i) cooperatives, (ii) private limited liability corporations (type A), (iii) public limited liability corporations (Type B), (iv) and other types (other natural or legal persons and legal persons under public law). The descriptive statistics of French enterprises used in the present analysis are reported in Table 5 of Appendix A.1. There are 1,631 cooperatives in the data set, each observed on average 10.21 times, and 7,644 non cooperative enterprises, with an average of 8.91 observations each. 54% of the cooperatives in the dataset were still functioning in 2006. This rate is 46% for private limited liability companies, and only 19% for public limited liability companies.
Different studies have highlighted the need for control variables in order to isolate pure age effects. Regional affiliation have been found as having a significant impact on the survival of businesses, which may benefit from local conditions favorable to their development (Basile et al., 2017;Fritsch et al., 2006;Gagliardi, 2009;Kalmi, 2013;Simons and Ingram, 2004). For our study, we used the geographic division of agricultural cooperatives regional federations. This information is used only as a control variable, as this division becomes part of a particular history and specific to agricultural cooperation.
An industry effect has also been noted: if the activity takes place in a growing sector, then the chance of survival is higher (Mata and Portugal, 1994;Tsvetkova et al., 2014). The industries have, therefore, been recoded into five sets: the Wholesale and Distribution, the Milk and Dairy Products, Meat and Poultry, the Beverage industries, and Others. These last sector is, by definition, much more heterogeneous and difficult to interpret. Cooperatives are more present in the Wholesale industry, and relatively less present in the Meat and Poultry industry.
Size has also been found as playing an important role (Esteve-Pérez et al., 2004;Geroski et al., 2010;Boone and Ozcan, 2016): large companies can indeed benefit from economies of scale, and from an opportunity to diversify their activities. Diversification reduces the risk of closure because poor conditions in a market can be offset by better conditions in another. For other authors (Jovanovic, 1982;Ericson, 1995), size is an indicator of differences in efficiency between firms. These differences stem from experience, managerial skills, technology, and organization. Large companies may have an advantage in fund-raising, can benefit from better taxation rules, and can more easily recruit the most qualified workers. New entrants to industries with high capital intensity have more difficulty surviving because of the significant level of resources needed to achieve the minimum threshold size (Fritsch et al., 2006). Size can be expressed in different ways: turnover, number of employees, investments, etc. For ease of interpretation, annual turnover, tangible and intangible investments are expressed in thousands of euros. The number of employees is expressed in employees at the end of the fiscal year (December 31). A large size is the sign of a strong capacity to mobilize resources and also of a favorable place a priori in the context of competition between organizations for the same resources (Carroll and Hannan, 2000). It increases, other things being equal, the chances of survival of the organization, the liability of smallness, (Freeman et al., 1983). We also control for the intensity of exports, express as a share of total turnover. Esteve-Pérez et al. (2008) show that exporting SMEs (Small and Medium Enterprises) face a significantly lower probability of failure than non-exporters.
Finally, we expect mixed effects of sectoral diversification strategies on the survival of cooperatives (Trechter, 1996). On one hand, diversification may have a positive impact on performance and, therefore, on survival; on the other hand, it also expresses an increasing heterogeneity among members, that may threaten the efficiency of governance (Hansmann, 1996). The sectoral diversification strategies are measured by the Evenness index (Shannon, 1948): where H is the Shannon-Wiener index, S is the number of categories, and p i is the share of the category i among all the categories. This diversity index ranges from 0 (for an enterprise present in only one sector) and 1 (present in all sectors). The value of the Evenness diversification index is relatively close for cooperatives and non-cooperative corporations with a mean between 0.05 and 0.07.
To our knowledge, most of previous studies did not control for the possible correlation between explanatory variables and unobserved heterogeneity. To address the issue of time-varying variables, one can include them in the model with their value at creation. As emphasized by various authors (Carroll and Hannan, 2000;Cefis and Marsili, 2005;Le Mens et al., 2011;Rousselière and Joly, 2011), the characteristics of the organization at creation strongly influence its fate. Another option is that time-varying variables are included with a lag of one year. Note, that this method leads to drop the establishments that exited in the first year of existence, omitting thereby a lot of valuable information (Fackler et al., 2013). We choose an alternative specification of correlated random effects, following Mundlak (1978), Wooldridge (2010), and Blanchard et al. (2014). The correlated random effects model for non-linear estimations relaxes the assumption of strict exogeneity of covariates. The individual effect depends on the mean observation per individual for each covariate, and permits to partially correct for the unobserved heterogeneity (Bache et al., 2013). Our model becomes a discrete time survival analysis with correlated random effects, as proposed by Blanchard et al. (2014).
An important difference with respect to previous works on the shadow of death (Blanchard et al., 2014;Griliches and Regev, 1995;Carreira and Teixeira, 2011) is that we don't include efficiency measurement into the econometric model. The shadow of death is defined as a pattern of pre-exit decreased productivity. However, there is no consensus in the literature on what the cooperative maximizes and, therefore, on any efficiency index. It can be, for example, the utility or welfare of its members (as in Fulton and Giannakas (2001) or Giannakas and Fulton (2005)), or its profit with a patronage refunded to its members (as in Agbo et al. (2015)). For example, Cazzufi and Moradi (2012) show a negative impact of profits on survival for cooperatives. 4

A Generalized Additive Discrete Time Survival Analysis Model with unobserved heterogeneity
The survival analysis (Singer and Willett, 2003) is based on the estimation of the determinants of the hazard (or failure or mortality) for a given unit. There is an ongoing debate in the literature on the advantages and disadvantages of the use of continuous-time and discrete-time models for survival analysis. For example, Carroll and Hannan (2000, p.110) advocate the use of a continuous-time model in order to accumulate empirical results. Authors in organizational ecology largely favor this approach and have a preference for the piece-wise continuous model. 5 Unfortunately, continuous-time survival methods assume that all survival times are unique, and that there are no pairs of individuals with identical or tied survival times (Rabe-Hesketh and Skrondal, 2012). On the contrary, discrete time survival methods allow many subjects to share the same survival time, due to recording reasons. A firm disappears only because the firm is not recorded in the survey the next year. In this case, the discrete time analysis framework has been shown to be more efficient Allison (2010).
Our proposition to use a Generalized Additive Discrete Time Survival Model can be viewed as a compromise between discrete time and continuous time modeling, taking into account the uncertainty about the exact moment of exit. It also goes beyond the traditional approaches to duration analysis, split between non-parametric approaches (such as the canonical Kaplan-Meier model) and (semi-)parametric approaches (based on complementary log-log) (Jenkins, 1995). In this section, we expose our model, which is the Bayesian version of the Additive Discrete Time Survival Model with frailty described in Tutz and Schmid (2016).
The standard cloglog (complementary loglog)model (Jenkins, 1995) can be conceived as a proportional hazard model with discrete-time, i.e. as the discrete-time analogue of the Cox model. For an observation i of an enterprise j, the hazard model, with a proportional hazard model structure, leads to the factorization (Rabe-Hesketh and Skrondal, 2012): where h(z|X ij ) is the hazard (or failure or mortality) of observation i of enterprise j at time z, 6 h 0 (z) is the baseline hazard, X is the vector of P explanatory variables x p , and β p the estimated coefficient of x p .
When we observe the survival time only by intervals, we observe that an integer value T ij = t if z t−1 < Z ij < z t , with Z ij the continuous survival time. The hazard in discrete time is given by: with S 0 the baseline survival function.
The model can be adapted to account for unobserved heterogeneity with an additional term u j : h(t|X ij ) = h 0 (t) · exp(β 1 · x 1ij + · · · + β P · x P ij + u j ) with u j ∼ N (0, ψ). u j can be interpreted as the impact of potential unobservable variables (such as management skills of managers) or omitted variables (such as imitation strategies of enterprises established nearby). The unobserved heterogeneity term, specific to each observed enterprise j, is usually assumed to follow a normal distribution. 7 In the case of the correlated random effects model, we relax the hypothesis of strict exogeneity of u with the distribution D(u j |X ij ) = D(u i |X ij ): 8 ln [− ln(1 − h(t|X ij ))] = β 1 · x 1ij + · · · + β P · x P ij + β a1 ·x 1ij + · · · + β aP ·x P ij + α t + u j (7) 6 Note, that in our case t is equivalent to the age of the enterprise. 7 Various ways to model the unobserved heterogeneity are developed in Nicoletti and Rondinelli (2010). 8 In the standard random effects model, we have D(u j |X ij ) = D(u j ).
where bars indicate the mean of each time-varying variable.
The cloglog of the Generalized Additive Mixed Model (GAMM) is a straightforward application of the cloglog with mixed effects to the Generalized Additive Model (GAM) framework developed by Wood (2006): with f (.) a smooth function of the time variable t. Time is introduced as a non parametric term using thin plate regression splines. The smooth function f (.) is modeled as follows: with µ w ∼ N (0, σ 2 ) and w k (.) the thin plate regression spline function. According to Wood (2003), w k (.) is the default approach to smooth terms because they are the optimal smoother of any given basis dimension/rank. The interest of the smooth term is that it addresses the issue of unstable estimations when, as t increases, only few observations are at risk and estimates become unstable. In this case the result are jumps in the estimated function that occur when the hazards are plotted against time (Tutz and Schmid, 2016). 9 Finally, we can extend our model using smooth terms for various covariates: 10 ln [− ln(1 − h(t|X ij , Z i,m ))] = β 1 · x 1ij + · · · + β P · x P ij + β a1 ·x 1ij + · · · + β aP ·x P ij where f 1 (.) is a smooth function of time variable t, and f m (.) are the smooth functions of covariates B.
We can calculate two different indicators in order to assess the importance of unobserved heterogeneity. The first one is the estimated residual intra-class correlation among the latent responses for two observations of the same enterprise: with ψ the variance of u j . The second indicator is the median hazard ratio (Rabe-Hesketh and 9 Note, that in conjunction to GAMM modeling, the use of Bayesian estimation is an additional way to address this issue, where median and robust parameters are used to measure central tendency (Buerkner, 2017). 10 Note, that our strategy can be easily extended to capture risk factors that may change with time, as proposed by Kauermann et al. (2005) using smooth interaction terms (Wood, 2003). Skrondal, 2012): with Φ −1 the normal inverse function. HR median is the median relative change in the hazard when comparing two observations with the same covariates, but corresponding to two different cooperatives (Austin et al., 2017).
Estimations are conducted with package Brms for R (Buerkner, 2017), based on Stan, a C++ program performing Bayesian inference and optimization (Gelman et al., 2015). Bayesian modelling accounts for uncertainty and sparse data (Gelman et al., 2014a). The Bayesian estimator does not generally allow analytical solutions. Recourse to draws from the posterior parameter distribution are required. Stan is a highly efficient program for high-dimensional and multilevel models with correlated parameters using the No-U-Turn Sampler (NUTS) algorithm (Buerkner, 2017). The NUTS algorithm (Hoffman and Gelman, 2014) removes autocorrelation quicker than frequentist and Bayesian alternative algorithms. 11 In order to increase the effective sample size (ESS), different chains based on different draws may be estimated.
We used weakly informative priors (Independent Cauchy prior distribution with center 0 and scale 2.5 on population parameters and a half Student-t prior with 3 degrees of freedom on the group level parameter), following Gelman (2006) and Buerkner (2017). These weakly informative priors are a good compromise between a fully informative prior and a non-informative prior (that corresponds to traditional frequentist analysis and often leads to unstable estimates), and can be used as routine in real world applications (Gelman et al., 2008).
There are different strategies to include time effects in a survival analysis: year dummies (Bontemps et al., 2013), smooth polynomial representations of time (Monteiro and Stewart, 2015), i.e. linear, quadratic and cubic effects, log of age, or incorporating period effects (Kitts, 2009;Varum and Rocha, 2010). In his survey on firm age effects, Coad (2017) notes that as individual firms change over time, adding a cohort effect distorts the representation of firm-level ageing viewed at the population level. Period effects can be included in order to control for potential macroeconomic (crisis) or institutional effects (change of economic regulations) on the survival dynamics of the enterprise (Varum and Rocha, 2010). In our case, the periods correspond to various changes in the cooperative regulation. 12 For Singer and Willett (2003), the completely general specification of time with dummy variables lacks parsimony and yields fitted hazard functions that can fluctuate erratically across consecutive time periods, due to nothing more than simple sampling variation (Fahrmeir and Wagenpfeil, 1996).
A three step approach is advocated by Cefis and Marsili (2005) or Bontemps et al. (2013) in order to select the appropriate specification: identification of a suitable parametric distribution using a non-parametric bivariate model, identification of temporal variables and control variables to isolate the effect of age on survival, and, lastly, estimation of a semi-parametric multivariate model complementary log-log. Following Singer and Willett (2003), and because of the inconsistency of stepwise model selection (Kass and Raftery, 1995), we used instead a one-stage approach based on Bayesian model selection (BMS) (Ando, 2010;Raftery, 1995). The Widely Applicable Information Criterion (WAIC) implemented in Vehtari et al. (2017) is a fully Bayesian method for estimating point-wise out-of-sample prediction accuracy from a fitted Bayesian model using the log-likelihood evaluated at the posterior simulations of the parameter values (Gelman et al., 2014b). Our benchmark estimation is compared to alternative specifications, a lower WAIC highlighting a better fit. We can also calculate the Akaike weights that are analogous to posterior probabilities of models, conditional on expected future data (McElreath, 2016): with ∆WAIC i the WAIC difference between the model i and the model with the lowest WAIC.

Results
Based on weakly informative priors, our benchmark model is based on 4 chains and 2,000 iterations of which the first 1,000 are a warm-up to calibrate the sampler, leading to a total of 4,000 posterior samples. 13 The Stan algorithm is highly efficient as the autocorrelation of the Markov chains disappear quickly (refer to Figure 1 for the estimations of the age parameter). We obtain therefore a large ESS.
The potential scale reduction factor on split chains,R, and ESS are reported for the parameters of interest in Tables 6 and 7 of Appendix A.2. At convergence,R = 1 (Gelman et al., 2014a). 14 The WAIC is minimal for the GAMM cloglog model and generates a posterior probability of 1 (see Table 1). As expected, a time dummies approach to time effect lacks parsimony, but surprisingly less than other approaches, such as the use of log, linear, cubic, or quadratic effects of age. This suggests that the impact of age on mortality can not be summarize in our case by polynomial representations, but requires a more flexible approach.
The results for the GAMM cloglog model are reported in Table 2 for cooperatives and in Table  3 for non-cooperative enterprises. For cooperatives, results underline a negative impact of the  number of employees, turnover, and tangible investment on mortality. The number of past legal statuses has opposite effects on cooperatives (increases the hazard probability) and on noncooperative enterprises (decreasing the hazard probability, although not significantly). These results are in line with the empirical findings of Boone and Ozcan (2016).
In Table 4, we report the various measures of unobserved heterogeneity. ρ is equal to 0.229 for cooperatives and 0.264 for non-cooperative enterprises, underlining the importance of unobserved variables for the survival of non-cooperatives. As reported in Table 8 of Appendix A.3, the difference is significant according the robust Bayesian estimation test developed by Kruschke (2013). Concerning the other measure, HR median permits a comparison of the magnitude of this general contextual effect with that of model covariates (Austin et al., 2017). As shown by the values of HR median , the hazard ratio which compares the observation with the larger hazard to the observation with the smaller hazard from two different cooperatives, exceeds 1.594 in 50% of the samples. For non-cooperative enterprises, this median hazard ratio is 1.756. As reported in Tables 6 and 7 of Appendix A.2, these magnitudes are larger than for the independent variables.
As shown by Allison (1999) and Williams (2009), direct comparison of the coefficients of non- linear models across groups can be invalid and misleading. Differences in the degree of residual variation across groups can produce apparent differences in slope coefficients that are not indicative of true differences. A valid alternative is to compare predicted probabilities (Long, 2009). 15 In Figures 2 and 3, we report the median of the predicted probabilities as a measure of central tendency. The age effect on mortality has been estimated as a smooth effect. Figure  4 represents this effect for the 4 categories of enterprises. The age dependence in mortality is different for the various organizational forms. After the first three years, mortality increases with age for non-cooperative enterprises (with a higher hazard for public than private corporations with limited liability, and the highest hazard for enterprises with other legal status), with a highest point between years 19 and 21. For cooperatives, the probability of mortality is the lowest at the age of 3 and 13, and the highest for years 7 and 19, suggesting a combination of the various liabilities.
We report the marginal effects of age on mortality. We see that the marginal effect is positive for the first two years, between 5 and 8 years, and finally between 15 and 20 years, underlining some specific effects of liability of newness, adolescence and obsolescence.
Finally, we compare the marginal effects of period on mortality. There is a small increase in mortality for the second period only for the cooperatives (after the change in the cooperative legislation), while the other period effects are slightly similar. On the opposite, there is a huge decrease for other forms (private limited, public limited and others) for the last two periods. The 1992-1995 period was characterized by important changes in cooperation regulation which create incentives to merger between cooperatives (Filippi et al., 2012).

Discussion and conclusion
Our work aimed to propose an original estimation method, that is more flexible than conventional survival regressions proposed in the literature on firm age, and able to disentangle the various effects of time (age and period), while being in line with the principles of parsimony. We tested the relationship between age and mortality in firms in the food and wholesale industries characterized by a diversity of organizational structures. To do this, in the absence of adequate distribution of a survival function a priori, we estimated a Bayesian semi-parametric discretetime model with smooth terms, taking into account unobserved heterogeneity and time-varying covariates.
The Bayesian GAMM presented in this paper is a very flexible framework and can be straightforwardly extended. It can also accommodate variable selection using Lasso prior on the population-level effects (Park and Casella, 2008) or spatial analysis, being therefore an alternative to classical models based on the inclusion of contextual variables (e.g. (Basile et al., 2017)). Indeed, one can consider the possibility of spatial autocorrelation, which is only imperfectly captured by our region variable. Federations may play a role in the differences in survival (Herbel et al., 2015). A process of imitation between comparable firms, or, conversely, of competition for the same resources (Simons and Ingram, 2004) may also be a source of isomorphism at local level (Nilsson et al., 2012). Also, as noted in the theoretical models of interaction between cooperatives and non-cooperatives in agricultural and food markets (Drivas and Gian-  nakas, 2010), a strong interaction between these different forms exists with respect to strategic choices (e.g. differentiation in terms of quality of products offered by agricultural cooperatives from their business competitors). Similarly, on the empirical level, the development of cooperatives happened to compensate for the exit of non-cooperative corporations (Simons and Ingram, 2004). In other contexts, "cooperative beehiving" has been observed. Hakelius et al. (2013) characterized this process as the phenomenon of members who de-associate themselves from large cooperatives and form smaller entities. These new dynamics of spatial autocorrelation can be addressed using smooth terms for latitude and longitude (Fahrmeir and Kneib, 2011).
Our GAMM model captures a smooth effect of age on mortality, that is different for cooperative and non cooperative enterprises, while periods seem to have the same effects on the mortality of both firm types. We find evidence of the liabilities of newness, adolescence, and obsolescence. The marginal effects of age being positive for the first two years, between years 5 and 8, and years 15 and 20. The effect for years 5-7 is also highlighted by Coad (2017), after which firm performance tends to stabilize, at least in relative terms. In our case, we observe one additional peak, suggesting the necessity for cooperatives to adapt themselves to a changing context. This effect is corroborated by a small increase in the mortality during the 1992-1995 period, which included important changes in the French legislation for cooperatives. A striking result is the higher importance of unobserved heterogeneity for non cooperative enterprises than for their cooperative counterparts. Unobserved variables (such as managerial ability or specific human capital of the direction) seem to have a greater impact on the survival of non cooperatives enterprises, than on that of cooperative enterprises characterized by a democratic and more collective governance (Diaz-Foncea and Marcuello, 2013;Pozzobon and Zylbersztajn, 2013).
Cooperatives appear to be highly resistant to the change in the social and economic environment, leading Nunez-Nickel and Moyano-Fuentes (2004) to consider the cooperative ownership structure as a buffer for structural and conjectural shocks or economic depressions. Cooperatives may be less profitable than investor-owned firms, but operate more efficiently, present a stronger financial position (Soboh et al., 2012), and have a stabilizing effect on employment (Delboni and Reggiani, 2013) and price (Muller et al., 2017) with respect to shocks. In line with our empirical findings, this "stability argument" (Chevallier, 2011) for public policy goes beyond the traditional pro-competitive argument for the promotion of cooperatives in developed economies.

A Appendix
A.1 Descriptive Statistics       Figure 5: Rho difference between cooperatives and non cooperatives using BEST test (Kruschke, 2013)