Maximum likelihood estimation in the logistic regression model with a cure fraction

. Logistic regression is widely used in medical studies to investigate the relationship between a binary response variable Y and a set of potential predictors X . The binary response may represent, for example, the occurrence of some outcome of interest ( Y = 1 if the outcome occurred and Y = 0 otherwise). In this paper, we consider the problem of estimating the logistic regression model with a cure fraction. A sample of observations is said to contain a cure fraction when a proportion of the study subjects (the so-called cured individuals, as opposed to the susceptibles) cannot experience the outcome of interest. One problem arising then is that it is usually unknown who are the cured and the susceptible subjects, unless the outcome of interest has been observed. In this setting, a logistic regression analysis of the relationship between X and Y among the susceptibles is no more straightforward. We develop a maximum likelihood estimation procedure for this problem, based on the joint modeling of the binary response of interest and the cure status. We investigate the identiﬁability of the resulting model. Then, we establish the consistency and asymptotic normality of the proposed estimator, and we conduct a simulation study to investigate its ﬁnite-sample behavior.


Introduction
Logistic regression is widely used to model binary response data in medical studies.An example of a binary response variable is the infection status (infected vs uninfected) with respect to some disease.A logistic regression model can be used to investigate the relationship between the infection status and various potential predictors.If Y i denotes the infection status for the i-th individual in a sample of size n (Y i = 1 if the individual is †Corresponding author infected, and Y i = 0 otherwise), and X i denotes the corresponding (p-dimensional, say) predictor, the logistic regression model expresses the relationship between Y i and X i in term of the conditional probability P(Y i = 1|X i ) of infection, as: where β ∈ R p is an unknown parameter to be estimated.An extensive literature has been devoted so far to statistical inference in logistic regression models.Estimation and testing procedures for this class of models are now well established and are available in standard statistical softwares.In particular, the maximum likelihood estimator of β is obtained by solving the following score equation: Asymptotic results (consistency and asymptotic normality) for this estimator were given by Gouriéroux and Monfort (1981) and Fahrmeir and Kaufmann (1985), among others.We refer the reader to Hosmer and Lemeshow (2000) and Hilbe (2009) for detailed treatments and numerous examples.
In this paper, we consider the problem of estimation in the logistic regression model with a cure fraction.In medical studies, it often arises that a proportion of the study subjects cannot experience the outcome of interest.Such individuals are said to be cured, or immune.The population under study can then be considered as a mixture of cured and susceptible subjects, where a subject is said to be susceptible if he would eventually experience the outcome of interest.One problem arising in this setting is that it is usually unknown who are the susceptible, and the cured subjects (unless the outcome of interest has been observed).Consider, for example, the occurrence of infection from some disease to be the outcome of interest.Then, if a subject is uninfected, the investigator usually does not know whether this subject is immune to the infection, or susceptible albeit still uninfected.
Estimating a regression model with a cure fraction can be viewed as a zero-inflated regression problem.Zero-inflation occurs in the analysis of count data when the observations contain more zeros than expected.Failure to account for these extra zeros is known to result in biased parameter estimates and inferences.The regression analysis of count data with excess zeros has attracted much attention so far.For example, Lambert (1992) proposed the zero-inflated Poisson (ZIP) regression model for count data with many zeros.This was further extended to a semiparametric ZIP regression model by Lam et al. (2006).We refer to Dietz and Böhning (2000) and Xiang et al. (2007) for a review of various other extensions of the ZIP model.Other popular models are the zero-inflated binomial (ZIB) regression model (see, for example, Hall (2000)), and the zero-inflated negative binomial (ZINB) regression model (see, for example, Ridout et al. (2001)).Recently, Kelley and Anderson (2008) proposed a zero-inflated proportional odds model (ZIPO) for ordinal outcomes, when some individuals are not susceptible to the phenomenon being measured.Various other models and numerous references can be found in Famoye and Singh (2006) and Lee et al. (2006).
In our paper, we consider the problem of estimating a logistic regression model from binary response data with a cure fraction, when the cure probability is modeled by a logistic regression.This can be viewed as a zero-inflated Bernoulli regression problem, where logistic link functions are used for both the binary response of interest (the probability of infection, say) and the zero-inflation probability (the probability of being cured).The literature on zero-inflated models is extensive but to the best of our knowledge, the theoretical and numerical issues related to the statistical inference in this model have not been yet investigated.In this paper, we intend to fill this gap.We first investigate the identifiability question in this model.Then, we turn to the problem of estimation.The estimator we propose is obtained by maximizing the joint likelihood for the binary response of interest and the cure indicator.We prove the almost sure asymptotic existence, the consistency, and the asymptotic normality of this estimator.Then, we investigate its finite-sample properties via simulations.
The rest of this paper is organized as follows.In Section 2, we describe the problem of logistic regression with a cure fraction, and we propose an estimation method adapted to this setting.The proposed procedure is based on a joint regression model for the binary response of interest and the cure indicator.In Section 3, we investigate the identifiability of this model, and we state some regularity conditions.In Section 4, we derive the asymptotic properties of the resulting estimator.Section 5 describes a simulation study, where we numerically investigate the small to large sample properties of this estimator.A real data example illustrates the methodology.A discussion and some perspectives are given in Section 6.

Notations and the model set-up
Let (Y 1 , S 1 , X 1 , Z 1 ), . . ., (Y n , S n , X n , Z n ) be independent and identically distributed copies of the random vector (Y, S, X, Z) defined on the probability space (Ω, A, P).For every individual i = 1, . . ., n, Y i is a binary response variable indicating say, the infection status with respect to some disease (that is, Y i = 1 if the i-th individual is infected, and Y i = 0 otherwise), and S i is a binary variable indicating whether individual i is susceptible to the infection (S i = 1) or immune (S i = 0).If Y i = 0, then the value of S i is unknown.Let X i = (1, X i2 , . . ., X ip ) ′ and Z i = (1, Z i2 , . . ., Z iq ) ′ be random vectors of predictors or covariates (both categorical and continuous predictors are allowed).We shall assume in the following that the X i 's are related to the infection status, while the Z i 's are related to immunity.X i and Z i are allowed to share some components.
The logistic regression model for the infection status assumes that the conditional probability P(Y = 1|X i , S i ) of infection is given by log if {S i = 1}, and by , where β = (β 1 , . . ., β p ) ′ ∈ R p is an unknown regression parameter measuring the association between potential predictors and the risk of infection (for a susceptible individual).
The statistical analysis of infection data with model (1) includes estimation and testing for β.Without immunity (that is, if S i = 1 for every i = 1, . . ., n), inference on β from the sample (Y 1 , X 1 , Z 1 ), . . ., (Y n , X n , Z n ) can be based on the maximum likelihood principle.When immunity is present, deriving the maximum likelihood estimator of β is no longer straightforward: if Y i = 0, we do not know whether {S i = 1}, so that (1) applies, or whether {S i = 0}, so that (2) applies.
One solution is to consider every individual i such that {Y i = 0} as being susceptible that is, to ignore a possible immunity of this individual.We may however expect this method to produce biased estimates of the association of interest (such a method will be evaluated in the simulation study described in section 5).Therefore in this paper, we aim at providing an alternative estimation procedure for β.This can be achieved if a model for immunity is available, as is explained in the next section.

The proposed estimation procedure
A model for the immunity status is defined through the conditional probability P(S = 1|Z i ) of being susceptible to the infection.A common choice for this is the logistic model (see, for example, Fang et al. (2005) and Lu (2008Lu ( , 2010) ) who considered estimation in various survival regression models with a cure fraction): where θ = (θ 1 , . . ., θ q ) ′ ∈ R q is an unknown regression parameter.
Remark 1.We note that the model defined by ( 1)-( 2)-( 3) can be viewed as a zero-inflated Bernoulli regression model, with logit links for both the binary response of interest and the zero-inflation component.As far as we know, no theoretical investigation of this model has been undertaken yet.Such a work is carried out in the following.
Let ψ := (β ′ , θ ′ ) ′ denote the unknown k-dimensional (k = p+q) parameter in the conditional distribution of Y given X i and Z i .ψ includes both β (considered as the parameter of interest) and θ (considered as a nuisance parameter).Now, the likelihood for ψ from the independent sample (Y i , S i , X i , Z i ) (i = 1, . . ., n) (where S i is unknown when Y i = 0) is as follows: We define the maximum likelihood estimator where l n (ψ) := log L n (ψ) is the log-likelihood function.In the following, we shall be interested in the asymptotic properties of the maximum likelihood estimator β n of β, considered as a sub-component of ψ n .We will however obtain consistency and asymptotic normality results for the whole ψ n .Before proceeding, we need to set some further notations.

Some further notations
Define first the (p × n) and (q × n) matrices and let W be the (k × 2n) block-matrix defined as where 0 ab denotes the (a×b) matrix whose components are all equal to zero (for any positive integer values a, b).Let also C(ψ) be the 2n-dimensional column vector defined as , and B θ (ψ) = (B θ i (ψ)) 1≤i≤n are n-dimensional column vectors with respective elements .
Then, simple algebra shows that the score equation can be rewritten as ln (ψ) = WC(ψ) = 0.
Note that the size of C(ψ), W, and D(ψ) depends on n.However, in order to simplify notations, n will not be used as a lower index for these vector and matrices.In the next section, we investigate the question of parameter identifiability in model ( 1)-( 2)-(3).
C2 Let ψ 0 = (β ′ 0 , θ ′ 0 ) ′ denote the true parameter value.β 0 and θ 0 lie in the interior of known compact sets B ⊂ R p and G ⊂ R q respectively.C3 The Hessian matrix ln (ψ) is negative definite and of full rank, for every n = 1, 2, . . .Let λ n and Λ n be respectively the smallest and largest eigenvalues of WD(ψ 0 )W ′ .There exists a finite positive constant c 2 such that Λ n /λ n < c 2 for every n = 1, 2, . . .

C4
There exists a continuous covariate V which is in X but not in Z that is, if β V and θ V denote the coefficients of V in the linear predictors (1) and (3) respectively, then β V = 0 and θ V = 0.At a model-building stage, it is known that V is in X.
The conditions C1, C2, C3 are classical conditions for identifiability and asymptotic results in standard logistic regression (see, for example, Gouriéroux and Monfort (1981) and Guyon (2001)).The condition C4, which imposes some restrictions on the covariates, is required for identifiability of ψ in the joint model ( 1)-( 2)-(3) (we may alternatively assume that the continuous covariate V is in Z but not in X).In the following, we will assume that V is in X but not in Z, with β V := β l for some l ∈ {2, . . ., p}, and for the i-th individual, we will denote V i by X il .The condition C4 is discussed in greater details in the following two remarks.
Remark 2. We may relate the identifiability issue in model ( 1)-( 2)-( 3) to the problem of identifiability of mixtures of logistic regression models, which was investigated by Follmann and Lambert (1991).Follmann and Lambert (1991) considered the case where there is a finite number c of components in the mixture (we consider here the case where c = 2, with one degenerate component) and the mixing probabilities are constant (here, the mixing probabilities given by ( 3) are allowed to depend on covariates).The authors have shown that finite mixtures of logistic regressions are identifiable provided that the number of unique covariate combinations values is sufficiently large.C4 can be viewed as a sufficient condition for achieving the same kind of requirement.A similar condition appears in Kelley and Anderson (2008).
To understand C4, note that if X i = Z i , then exchanging the parameters β and θ in (1) and (3) yields the same likelihood value L n (ψ), which is a cause of model non-identifiability.A similar remark holds if we invert the linear predictors β ′ X i and θ ′ Z i .The condition C4 evacuates these problems.
First, by asking one of the covariates to be significant in one and only one linear predictor, C4 prevents β ′ X and θ ′ Z from being of the same form, and the parameters are thus not exchangeable.Secondly, by assuming that we know, prior to model fitting, that there exists a covariate V which is in X but not in Z, C4 will force each linear predictor to be attached to the correct corresponding model ( 1) or (3).These facts are illustrated in a web-based supplementary document available at the following address: http://perso.univ-lr.fr/jfdupuy/supplementary.pdf.There, we provide the results of a simulation study which investigates numerically the identifiability of model ( 1)-( 2)-( 3).For each of the models considered in this study, we assume that C4 is satisfied: the linear predictors β ′ X i and θ ′ Z i share three covariates (one is continuous, two are discrete), and an additional continuous covariate is included in X i .Using the procedure described in Section 2, maximum likelihood estimates are obtained for β and θ, and are averaged over N = 1000 samples (we considered several combinations of sample size, proportion of immunes, proportion of infected among the susceptibles).Both parameters β and θ appear to be identifiable (the averaged estimates appear to be close to the true parameters, including those corresponding to the three shared covariates).
Remark 3. The condition C4 does not appear to be too restrictive in practice.Consider the example of the transmission of some disease by breastfeeding.If every child in the sample is breastfeeded, it can be expected that the length (in days, say) of the breastfeeding period (a continuous covariate) will influence the probability of infection, while the susceptibility probability will rather depend on risk factors such as say, the mother's infection status.It is also worth noting that the consequences of C4, in terms of model-building, are rather mild.At a model-building stage, we may be tempted to incorporate all available covariates in both linear predictors (1) and (3), and to remove irrelevant factors by using backward elimination.The condition C4 slightly restricts this fitting strategy, by imposing that one relevant continuous covariate is incorporated in one (and only one) linear predictor.This should often be doable in practice, since the statistician often gets some prior knowledge (from the clinicians, epidemiologists,. . . ) about the dataset to be analyzed.
We are now in position to prove the following result: Theorem 1 (Identifiability).Under the conditions C1-C4, the model ( 1)-( 2 Proof of Theorem 1. Suppose that L 1 (ψ) = L 1 (ψ * ) almost surely.Under C1 and C2, there exists a positive constant c 1 such that for every x ∈ F , z ∈ G, and This can be rewritten as Now, under the condition C4, taking the partial derivative of both sides of (5) with respect to the l-th component of x (X il is a continuous covariate) yields since the right-hand-side of (5) does not depend on x.Thus, it follows that Differentiating both sides of this equality with respect to the l-th component of x further yields (β − β * ) ′ x = 0, which implies that β = β * under C1.It remains to show that θ = θ * , which reduces to the identifiability problem in the standard logistic regression model.We have that θ = θ * under C1 (see Guyon (2001) for example), which concludes the proof.

2
We now turn to the asymptotic theory for the proposed estimator.

Asymptotic theory
In this section, we establish rigorously the existence, consistency and asymptotic normality of the maximum likelihood estimator β n of β in model ( 1), obtained from a sample of binary response data with a cure fraction.In the sequel, the space R k of k-dimensional (column) vectors will be provided with the Euclidean norm, and the space R k×k of (k × k) real matrices will be provided with the spectral norm (we will use the same notation • for both).We first prove the following result: Theorem 2 (Existence and consistency).Under the conditions C1-C3, the maximum likelihood estimator ψ n exists almost surely as n → ∞, and converges almost surely to ψ 0 , if and only if λ n tends to infinity as n → ∞.
Proof of Theorem 2. The principle of the proof is similar to Gouriéroux and Monfort (1981) but the technical details are different.Three lemmas are needed.The first lemma essentially provides an intermediate technical result.Its proof is postponed to the appendix.
Lemma 1.Let φ n : R k −→ R k be defined as Then there exists an open ball B(ψ 0 , r) (with r > 0) such that φ n satisfies the Lipschitz condition on B(ψ 0 , r) that is, and 0 < c < 1.
Assume that as n → ∞, ψ n exists and converges almost surely to ψ 0 , but η n (ψ 0 ) does not converge almost surely to 0. Then there exists a set Ω ⊂ Ω with P( Ω) > 0, such that if ω ∈ Ω, there exists ǫ > 0 such that for every m ∈ N, there exists n ≥ m with η n (ψ 0 ) > ǫ.Now, let t = ǫ d(1+c) , with d > 1 sufficiently large so that t ≤ r, where r is such as in Lemma 1. Then for every ψ ∈ B(ψ 0 , t), the following holds: where the second to third line follows by Lemma 1.Therefore, for every ψ ∈ B(ψ 0 , t), and we conclude that for every ψ ∈ B(ψ 0 , t), η n (ψ) > ǫ(1 − 1 d ) > 0. Since η n ( ψ n ) = 0, ψ n cannot belong to B(ψ 0 , t) for large n, which implies that ψ n does not converge almost surely to ψ 0 .This is the desired contradiction.
ln (ψ 0 ) converges almost surely to 0 if and only if λ n tends to infinity as n → ∞.
Proof of Lemma 3. We first prove that the condition is sufficient that is, we assume that λ n tends to infinity as n → ∞.Define the (2n × k) matrix V = (D(ψ 0 )) where I 2n denotes the identity matrix of order 2n.To see this, note that For every i = 1, . . ., n, ] by independence between the individuals, and By Gouriéroux and Monfort (1981) (proof of Lemma 4), if (7) holds, Λ n /λ n < c 2 for every n = 1, 2, . .., and λ n tends to infinity as n → ∞, then ln (ψ 0 ) converges almost surely to 0. We now prove that the condition is necessary.Assume that λ n does not tend to infinity as n → ∞.By Gouriéroux and Monfort (1981) (proof of Lemma 4), (V ′ V) −1 V ′ U (and therefore (WD(ψ 0 )W ′ ) −1 ln (ψ 0 )) cannot converge to 0, which concludes the proof.

2
We now turn to the convergence in distribution of the proposed estimator, which is stated by the following theorem: Theorem 3 (Asymptotic normality).Assume that the conditions C1-C3 hold and that ψ n converges almost surely to ψ 0 .Let Σ n = WD( ψ n )W ′ and I k denote the identity matrix of order k.Then Σ 1 2 n ( ψ n − ψ 0 ) converges in distribution to the Gaussian vector N (0, I k ).
Proof of Theorem 3. A Taylor expansion of the score function is as where ψ n lies between ψ n and ψ 0 , and thus ln ( The two terms in brackets in (8) converge almost surely to I k .To see this, we show for example that Σ and Note also that ψ n converges almost surely to ψ 0 (that is, for every ω ∈ Ω, where Ω ⊂ Ω and P( Ω) = 1).Let ω ∈ Ω.By the same arguments as in the proof of Lemma 1, for every ǫ > 0, there exists a positive n(ǫ, ω) ∈ N such that if n ≥ n(ǫ, ω), then  Eicker (1966), this convergence holds if we can check the following conditions:

Study design
In this section, we investigate the numerical properties of the maximum likelihood estimator β n , under various conditions.The simulation setting is as follows.We consider the following models for the infection status: and the immunity status: where X i2 is normally distributed with mean 0 and variance 1, and Z i2 is normally distributed with mean 1 and variance 1.An i.i.d.sample of size n of the vector (Y, S, X, Z) is generated from this model, and for each individual i, we get a realization (y i , s i , x i , z i ), where s i is considered as unknown if y i = 0.A maximum likelihood estimator β n of β = (β 1 , β 2 ) is obtained from this incomplete dataset by solving the score equation ( 4), using the optim function of the software R.An estimate is also obtained for θ = (θ 1 , θ 2 ), but θ is not the primary parameter of interest hence we only focus on the simulation results for β n .
The finite-sample behavior of the maximum likelihood estimator β n was assessed for several sample sizes (n = 100, 500, 1000, 1500) and various values for the percentage of immunes in the sample, namely 25%, 50%, and 75%.The case where it is known that there are no immunes in the sample was also considered.In this case, there is no missing information about the infection status and therefore, this case provides a benchmark for evaluating the performance of the proposed estimation method.We also considered different values for the proportion of infected individuals among the susceptibles.The desired proportions of immunes and infected were obtained by choosing appropriate values for the parameters β (the parameter of interest) and θ (the nuisance parameter).The following values were considered for β: i) model M 1 : β = (−.8, 1) (using these values, approximately 30% of the susceptibles are infected), ii) model M 2 : β = (1, .7)(approximately 70% of the susceptibles are infected), iii) model M 3 : β = (−.8,0) (approximately 30% of the susceptibles are infected), iv) model M 4 : β = (1, 0) (approximately 70% of the susceptibles are infected).

Results
For each configuration (sample size, percentage of immunes, percentage of infected among susceptibles) of the design parameters, N = 1500 samples were obtained.Based on these 1500 repetitions, we obtain averaged values for the estimates of β 1 and β 2 , which are calculated as 2,n , where β 2,n ) is the estimate obtained from the j-th simulated sample.For each of the parameters β 1 and β 2 , we also obtain the empirical root mean square and mean absolute errors, based on the N samples.When β 2 = 0 (respectively β 2 = 0), we obtain the empirical power (respectively the empirical size) of the Wald test at the 5% level for testing H 0 : β 2 = 0 (models M 1 and M 2 , see Tables 1 and 2) (respectively models M 3 and M 4 , see Tables 1 and 2).The null hypothesis H 0 : β 2 = 0 is the hypothesis that the predictor X 2 does not influence the risk of infection of susceptible individuals.The results are summarized in Tables 1 and 2.
Tables 1 to 2 about here From these tables, it appears that the proposed maximum likelihood estimator β n provides a reasonable approximation of the true parameter value, even when the percentage of immunes is high.While the bias of β n stays limited, its variability increases with the immune fraction, sometimes drastically when the sample size is small.Consequently, when the sample size is small (n = 100) and/or the immune proportion is very high (75%), the power of the Wald test for nullity of the regression coefficient β 2 can be low, compared to the case where there are no immunes.But we note that for moderately large to large sample sizes (n ≥ 500), the dispersion indicators and the power of the Wald test indicate good performance of the maximum likelihood estimate, even when the immune proportion is up to 50%.The level of the Wald test for nullity of β 2 is globally respected except, for every immune proportion, when the sample size is small (n = 100).
We compare these results to the ones obtained from a "naive" method where: i) we consider every individual i such that {Y i = 0} as being susceptible but uninfected, that is we ignore the eventual immunity of this individual, ii) we apply a usual logistic regression analysis to the resulting dataset.The results of such "naive" analysis for model M 1 are given in Table 3 (the results for models M 2 , M 3 , M 4 yield similar observations and thus, they are not given here.However, the complete simulation study is available from the web-based supplementary document mentioned above).

Table 3 about here
From this table, it appears that ignoring the immunity present in the sample results in strongly biased estimates of β.The bias of the intercept estimate increases with the immune proportion.At the same time, the estimate of the regression coefficient β 2 is biased towards 0 for all values of the immune percentage and sample size.This results in a very low power for the Wald test of nullity of β 2 , and in a wrong interpretation of the relationship between the covariate X 2 and the binary response Y .
The quality of the Gaussian approximation to the large-sample distribution of β 2,n was also investigated.For each configuration of the design parameters, histograms of the β (j) 2,n (j = 1, . . ., N ) are obtained, along with the corresponding QQ-plots.The plots for the model M 1 are pictured on Figures 1 to 4 (the plots for the models M 2 , M 3 , M 4 are given in the web-based file).

Figures 1 to 4 about here
From these figures, it appears that the normal approximation stated in Theorem 3 is reasonably satisfied when the proportion of immunes is moderate (25%), provided that the sample size is sufficiently large (n ≥ 500, say).Consider the case when β 2 = 0.When the immune fraction is large (50%), the normal approximation still appears reasonable, provided that the sample size is at least 1000, or eventually 1500.When the immune proportion is very large (75%), the distribution of β 2,n can be highly skewed, in particular when the sample size is small.Consider the case when β 2 = 0. Then the finite-sample distribution of β 2,n appears to be symmetric, with heavy tails however, especially when the sample size is small.When the immune fraction is about 50% and the sample size is greater than or equal to 500, the normal distribution appears to fit reasonably well the distribution of β 2,n .Overall, these results indicate that a reliable statistical inference on the regression effect in the model ( 1) with a cure fraction should be based on a sample having, at least, a moderately large size (n ≥ 500, say) when the immune fraction is moderate (25%), or a large size (n ≥ 1000, say) when the immune proportion is large (50%).

A real data example
In this application, we consider a study of dengue fever, which is a mosquito-borne viral human disease.A dengue infection confers a partial and transient immunity against a subsequent infection (see Dussart et al. (2011)).We consider here a database of size n = 528, which was constituted with individuals recruited in Cambodia, Vietnam, French Guiana, and Brazil (Dussart et al. (2011)).Each individual i was diagnosed for dengue infection and coded as Y i = 1 if infection was present and 0 otherwise.Note that if Y i = 0, then the i-th individual may either be immune at the time of analysis (due to a temporary immunity acquired following a previous infection) or susceptible to dengue infection, albeit not infected.We aim at estimating the risk of infection for those individuals, based on this data set which also includes the following covariates: age (a continuous bounded covariate) and weight (coded 0 in case of underweight and 1 otherwise).We first ran a standard logistic regression analysis of the model logit P(Y = 1|age, weight) = β 1 + β 2 age + β 3 weight.The results are displayed in Table 4.Then, we estimated the parameters β 1 , β 2 and β 3 using the methodology described in Section 2. Note first that the eventual immunity imparted by a past infection is only transient, thus there is no reason why an older individual (who has therefore been exposed longer to the risk of dengue fever) would have a greater probability of being immune than a younger one.In fact, individual susceptibility to the dengue infection may rather depend on whether the individual benefits or not from some preventive and control measures (such as the application of insecticides to larval habitats in his area, or appropriate water storage and waste disposal practices).Such informations are not available in our dataset.Age was therefore taken as the variable V in condition C4, and we fitted to the data the model ( 1)-( 2)-(3) with logit P(Y = 1|age, weight, S = 1) = β 1 + β 2 age + β 3 weight and logit P(S = 1| weight) = θ 1 + θ 2 weight.Since the Wald-type test of "θ 2 = 0" was not significant, we removed the weight from the model for susceptibility, resulting in a constant proportion of immunes.The final results of this fitting procedure are given in Table 4.

Table 4 about here
The fitted model produced the following estimate: 1−exp(0.497)/(1+exp(0.497))≈ 0.38 for the probability of being immune.Then, as expected, the estimated probabilities of infection obtained from our approach are larger than the ones derived from a standard analysis that does not take account of the possible immunity.For example, the probabilities of infection for individuals respectively aged 30 and 10 years old, with "normal" weight, are estimated by 0.29 and 0.55 (standard logistic regression) and 0.31 and 0.86 (from our approach).It is expected that underweighted subjects (those considered to be under a healthy weight) will have higher risks of infection.The probabilities of infection for underweighted individuals respectively aged 30 and 10 years old are estimated by 0.48 and 0.73 (standard logistic regression) and 0.97 and 0.99 (our approach).While both approaches provide the same qualitative conclusions: the probability of dengue infection is higher for younger individuals and in case of underweight (caused by malnutrition for example), they differ on their estimations of the risk of infection.Our approach takes account of the eventual immunity imparted by a past infection and therefore, it is reasonable to think that the resulting estimations of the infection probabilities provide a more realistic picture of the infection risk for this data set.In particular, the estimates provided by our approach suggest that underweight constitutes a major risk factor for dengue infection, irrespectively of age.

Discussion and perspectives
In this paper, we have considered the problem of estimating the logistic regression model from a sample of binary response data with a cure fraction.The estimator we propose is obtained by maximizing a likelihood function, which is derived from a joint regression model for the binary response of interest and the cure indicator, considered as a random variable whose distribution is modeled by a logistic regression (the proposed joint model can thus be viewed as a zero-inflated Bernoulli regression model, with logit links for both the binary response of interest and the zero-inflation component).We have established the existence, consistency, and asymptotic normality of this estimator, and we have investigated its finite-sample properties via simulations.Several open questions now deserve attention.The estimation approach proposed here relies on our ability to correctly specify the model for the binary immunity status.It is therefore of interest to investigate the effect of a misspecification of this model (and in particular, of the link function).The techniques and results by Czado and Santner (1992) may be useful for that purpose.Another issue of interest deals with the inference in the logistic regression model with a cure fraction, in a high-dimensional setting.We have established the theoretical properties of our estimator in a low-dimensional setting that is, when a small number of potential predictors are involved.Several recent contributions (see for example Huang et al. (2008) and Meier et al. (2008)) have considered the problem of estimation in the logistic model (without cure fraction) when the predictor dimension is much larger than the sample size (this problem arises, for example, in genetic studies where high-dimensional data are generated using microarray technologies).Extending our methodology to this setting constitutes another topic for further research.
surely to 0. Moreover, for n sufficiently large, there exists a positive constant c 4 < ∞ such that almost surely, Σ − I k converges almost surely to 0. The almost sure convergence to k follows by similar arguments.

Note:Figure 1 .Figure 2 .Figure 3 .
Figure 1.Histograms and Q-Q plots for β2,n in model M1, with no immunes in the sample (the percentage of immunes is given in brackets).n is the sample size.All results are based on 1500 simulated datasets.

Table 4 .
Dengue fever data analysis